AI papers — 2026-07-31

Problem: Chemistry literature search tools return ranked lists of documents rather than the specific, cross-paper findings scientists and AI agents actually need to assemble answers.

Method: AskChem converts each paper into atomic, typed, provenance-carrying "claims" (grounded by a source DOI and a verbatim quote or evidence locator) using two LLM extraction pipelines — a high-throughput abstract extractor (GPT-5-mini) and a deeper full-text PDF extractor (Gemini 3.1 Pro via Vertex AI batch, with an earlier legacy slice using GPT-4o/GPT-4o-mini). Over this shared claim store it builds three complementary structures: a stabilized faceted taxonomy (reaction type, substance class, application, technique, mechanism, claim type, data, time) for hierarchical browsing/retrieval via reciprocal rank fusion over FTS5 and dense-vector search; an evidence graph with typed relations (supports, contradicts, extends, derives_from, cites_as_evidence); and an exploratory "Living Taxonomy" that places papers under governing scientific principles. The system is exposed via web UI, REST API, SDK, and MCP server, and evaluated on a new benchmark, AskChem-Bench (30 cross-paper chemistry questions), with a GPT-5.5 reader grounded in retrieved claims.

Key results:

• Index scale: 2.4M claims from 147K papers (1925–2026), 307K taxonomy nodes, 171,342 evidence-graph edges, 4,931-node Living Taxonomy covering 1.1M claims / 361K paper placements.
• 100% of claims are source-grounded (claim type, source DOI, verbatim quote); evidence-graph edge-type precision estimated at 97.9% (143/146 decidable edges correct) via expert audit of 148 sampled edges.
• On AskChem-Bench, GPT-5.5 + AskChem achieves 100% resolvable DOIs (vs. 88.3% for the ungrounded reader) and the highest citation density among 5 systems at 18.1 verified DOIs/answer (vs. 9.6 ungrounded, 7.5 Paperclip, 10.7 Edison Scientific, 7.9 NotebookLM), plus best mean relevance score (2.15/3) and highest coverage of recent high-impact papers (18.5%).
• In a worked example, the ungrounded GPT-5.5 fabricated 6 of 14 cited DOIs, while the AskChem-grounded version had all 22 citations resolve correctly.

Why it matters / caveats: By making the "claim" rather than the paper the retrieval unit, AskChem substantially reduces citation fabrication for AI-assisted literature synthesis and provides open (MIT code, CC-BY data) infrastructure usable by both humans and agents. Limitations acknowledged by the authors: the corpus covers only a fraction of chemistry, abstract-only extraction is shallower than full-text, LLM-generated claims/relations/taxonomy placements can be wrong, the benchmark only covers 30 questions (measuring groundedness, not full factual accuracy or utility), taxonomy retrieval gain is not isolated, and Edison Scientific's closed agentic system still produced more citation-linked quantitative detail and a slightly higher on-topic rate.

Problem: GUI agents are typically optimized for simulated benchmarks, leaving a large gap between benchmark performance and reliable, long-horizon, cross-platform real-world execution on mobile, desktop, web, and information-seeking tasks.

Method: Qwen-UI-Agent (a continuation of MAI-UI) is a foundation GUI agent built on four components: (1) an environment infrastructure spanning scalable sandboxes (MobileWorld-based mobile sandbox rebuilt on redroid containers instead of KVM emulators, OSWorld-based computer-use sandbox, a Playwright/Chromium/FastAPI browser environment, and a DeepSearch environment using Serper and Jina Reader) plus a real-device mobile runtime with over 100 physical devices and 150+ apps governed by a health-aware scheduler and a VLM-based judge distinguishing model vs. environment failures; (2) a unified action space interleaving GUI actions (click, type, drag, etc.), CLI actions (cli_command), API calls, and user-interaction actions (ask_user, terminate), supporting batched multi-action decision steps; (3) an AutoResearch-style agent-driven data flywheel with domain capability bootstrapping (rejection sampling into an SFT corpus, following MAI-UI) and an iterative refinement loop where agents synthesize tasks/verifiers, evaluate trajectories, and diagnose failures; and (4) training combining SFT, Action RL for correcting recurring action errors, and verifier-guided Online RL scaled to trajectories exceeding 100 turns using roughly 10,000 concurrent rollout environments with a model-adaptive curriculum, plus a lightweight harness layer enabling proactive mobile-notification-triggered services and cross-platform (mobile+computer) workflows.

Key results:

• MobileWorld-Real (400+ tasks, 100+ apps): 92.2% success, beating Gemini 3.1 Pro, Claude Opus 4.8, GPT-5.6 Sol, and Seed 2.1 Pro by 6.0, 7.5, 6.8, and 3.5 points respectively; 97.5% on AndroidDaily; 82.1% on MobileWorld (14.6 points over Opus 4.8).
• OSWorld-Verified: 79.5% (second overall); OSWorld-v2: 13.9% binary success, 40.0% partial-progress score, exceeding MiniMax M3 and Qwen 3.7 Plus by 17.7 and 18.5 points while using 58.4% and 21.7% fewer steps per task.
• WebArena: 73.6% (beating Claude Opus 4.8, GPT-5.5, Gemini 3.1 Pro by 1.7–8.3 points); DeepSearch: 64.1% on BrowseComp, 75.0% on BrowseComp-ZH; GUI grounding: 81.5% on ScreenSpot-Pro, 97.5% on ScreenSpot-V2, 92.6% on MMBench-GUI L2.
• Over 40% of computer-use action outputs are batched, substantially shortening trajectories; the model outperforms the Qwen base model on agentic benchmarks (e.g., Terminal-Bench 2.0, Claw-Eval) while staying comparable on general reasoning (MMMU-Pro, MMLU-Pro).

Why it matters / caveats: Demonstrates that closing the sim-to-real gap with a large real-device mobile runtime, hybrid GUI+CLI batched actions, and an automated agent-driven data/training flywheel can beat frontier closed-source models (Opus 4.8, Gemini 3.1 Pro, GPT-5.6 Sol) on real-world mobile and competitive computer/browser tasks; caveat is that much of this evaluation relies on the authors' own new benchmarks (MobileWorld-Real, AndroidDaily) and a self-built real-device fleet, which may limit independent reproducibility.

Problem: Agent memory today is bolted on as an external module (e.g., RAG-style retrieval and context concatenation), which decouples memory from the backbone, blocks end-to-end gradient optimization, and adds inference latency.

Method: The authors formalize "memory foundation models" with a native, dynamically-evolving memory state and native memory procedures baked into forward computation, then build Metis, the first prototype, on top of Qwen3.5 (4B/9B/27B) backbones. Each Transformer layer gets a Metis block — a local memory block (a dense key-value memory matrix M and a query-key normalization vector S, updated via a Gated Delta Network-based update rule) and a hyper memory block (learnable importance-weighted token aggregation, plus memory key/value/query projections) — whose output is fused with standard causal attention via a memory attention mechanism. Metis is mid-trained (backbone frozen, only memory parameters trained) on 357,137 synthesized primary-data samples (~406M tokens across remember/forget/update/reflect operations, built from 27 public benchmarks like LoCoMo, LongMemEval, ZsRE, TOFU) plus 609,443 auxiliary samples (multi-entity binding, selective forgetting, post-memory dialogue) using three objectives: memory reconstruction, memory operation, and a regularization objective for interference/pollution robustness.

Key results:

• On the no-context MemOps (Gold) benchmark, Metis-27B reaches 24.76 avg. vs. 10.91 for the Qwen3.5-27B backbone and 2.50 for Temp-LoRA-27B (no-context settings).
• On LoCoMo (Gold) memory-QA, Metis-27B scores 43.42 avg. vs. 16.73 for Qwen3.5-27B (no context) and 37.68 for Temp-LoRA-27B.
• On NextMem contextual generation, Metis-27B reaches 50.82 avg., roughly double the 17.75 of the Qwen3.5-27B no-context baseline.
• Ablations show removing adaptive aggregation (using last-token hidden state instead) causes the largest degradation (e.g., overall avg. drops ~61%), and removing query-key normalization causes a ~28% drop, confirming these architectural components are load-bearing.
• Metis shows OOD transfer on ATM-Bench (e.g., Metis-27B: 66.06 on open-ended vs. 40.24 for Temp-LoRA-27B) though results are more mixed on MemDaily.

Why it matters / caveats: Native memory updates require only a gradient-free forward pass (frozen weights at inference) and can run largely in parallel with standard attention, offering an efficiency and end-to-end-optimization advantage over RAG/external memory. However, the authors report clear limitations: performance degrades on long-term/high-capacity scenarios because information is compressed into fixed-size latent parameters, forgetting remains the hardest operation, and semantically similar facts can be confused in the latent space — so native memory is presented as complementary to, not a replacement for, external memory.

Problem: Progress toward recursive self-improvement (RSI) requires AI systems that improve the process of building AI (AI4AI), but no existing public system jointly spans scalable executable task construction, execution-grounded agent post-training, and an evolutionary search harness that deploys the trained agent in long-horizon machine learning engineering (MLE).

Method: The authors introduce OpenMLE, a full-stack system with three components: OpenMLE-Gym, which builds 5,758 quality-gated executable MLE tasks from Curated Anchors, Kaggle Datasets, and Kaggle Competitions; OpenMLE-ERL, which trains four shared atomic operators (Draft, Improve, Debug, Crossover) via execution-grounded supervised fine-tuning (26,259-example corpus) followed by reinforcement learning with adaptive reward bounds, entropic upper-tail advantage weighting, and fitness-proportional parent selection (score/variance/visit-cooling); and OpenMLE-Evo, an AIRA-Evo-style population search loop that composes these trained operators using structured experience cards, a three-factor (quality/progress/novelty) parent-selection utility, and operation-triggered, operator-conditioned memory synthesis. This pipeline is used to post-train Frontis-MA1-35B (from Qwen3.6-35B-A3B) and a companion Frontis-MA1-30B (from Qwen3-30B-A3B-Thinking-2507), evaluated on MLE-Bench Lite and NatureBench Lite.

Key results:

• On MLE-Bench Lite (12-hour/task budget, single RTX 4090 capped at 12GB VRAM), Frontis-MA1-35B with OpenMLE-Evo improves Medal Average from 39.39% (base Qwen3.6-35B-A3B) to 60.61%, and reaches 71.21% with OpenMLE-Evo-Max, exceeding GPT-5.5+Codex (68.18%) and approaching GPT-5.6 Sol and Kimi K3 (72.73% each).
• The 30B companion model reproduces the effect: 34.85% to 53.03% Medal Average under OpenMLE-Evo (66.67% with Evo-Max), confirming the gain is not backbone-specific.
• Matched-harness comparisons show OpenMLE-Evo beats general-purpose Claude Code/Codex scaffolds and original AIRA-Evo (e.g., 53.03% to 60.61% Medal Average for Frontis-MA1-35B against original AIRA-Evo).
• On NatureBench Lite, holding the harness fixed the trained model raises Match-SOTA from 50% to 70% (vs. base); holding the model fixed, OpenMLE-Evo raises Match-SOTA from 20% to 50% (vs. original AIRA-Evo).
• OpenMLE-Evo cuts total model-token usage by 41.7% (129.3M to 75.3M) and prompt tokens by 50.3% versus original AIRA-Evo on matched trajectories, while increasing new-best validation updates per million tokens by 84.3% (1.77 to 3.27).

Why it matters / caveats: The results demonstrate that training and search gains compose (post-training plus evolutionary search each contribute independently and stack), and that both transfer beyond competition-style Kaggle MLE to scientific AutoResearch tasks — a concrete, measurable step toward AI4AI/RSI. The authors explicitly caveat that OpenMLE does not yet realize full RSI: it optimizes primarily for measured execution outcomes rather than judging which research directions are worth pursuing, and its evolutionary search remains an external harness bolted onto trained operators rather than a general coding agent that can initiate its own actions.

Problem: Video world models that predict future frames directly in pixel space leave the underlying physical dynamics implicit, often producing visually plausible but physically inconsistent outcomes.

Method: PhiZero learns "physical language," a compact discrete representation of state transitions, via a reason-then-render pipeline with two components: a Physical Language Tokenizer (a spatiotemporal encoder plus a transition-level Q-Former that compresses adjacent latent-frame pairs, discretized with finite scalar quantization into a 25K-symbol vocabulary) that conditions a pretrained Wan2.2-5B diffusion decoder (fine-tuned via LoRA rank 32) to reconstruct video from the first frame and physical-language tokens; and a Physical Language Reasoner, initialized from Qwen3-VL-4B with its vocabulary extended by FSQ-index symbols, trained autoregressively to predict physical-language sequences from a first frame and text action intent. Training uses a hierarchical data curation pipeline (50K-hour raw video pool filtered to 10K hours for tokenizer pretraining, 5M four-second clips for tokenizer SFT and reasoner pretraining, 1M motion-rich clips for reasoner SFT), a joint temporal/spatial curriculum, and a pure-noise warm-up stage to prevent the decoder from ignoring the physical-language conditioning.

Key results:

• Best IQ-Score (41.2) on Physics-IQ Verified, best Physics Score (3.01) and Overall (2.97) on PhyGround, and best Physics Adherence (4.88) and Total (8.19) on WorldModelBench, outperforming Wan2.2-14B, Sora 2, Cosmos3-Super, Grok-Video, and Veo3.1 baselines.
• Tokenizer reconstructs video using only 256 discrete physical-language symbols (vs. 44,800 continuous tokens for the Wan2.2-5B VAE) while achieving the best PSNR (28.9)/SSIM (0.903) among compact tokenizers.
• Ablations show removing the diffusion decoder, transition-level Q-Former, or pure-noise warm-up each degrades reconstruction (e.g., PSNR drops from 28.9 to 26.6 without the diffusion decoder); prompt-enhanced Wan2.2-5B baseline reaches only 26.6 IQ-Score vs. PhiZero's 41.2.
• Demonstrates zero-shot transfer: encoding a human-motion video's physical language and decoding it onto an edited first frame transfers motion to a Unitree G1 humanoid or a Sharpa dexterous hand, and enables sim-to-real transfer on LIBERO videos, without target-specific training.

Why it matters / caveats: By disentangling state-transition dynamics from visual appearance into an explicit discrete "language," PhiZero enables fine-grained action-conditioned simulation, interactive rollouts, and cross-embodiment/sim-to-real motion transfer that pixel-space world models struggle to support; the paper notes further limitations and future work are discussed in its appendix (not detailed in the excerpted text).

Problem: Text-to-video models struggle to generate physically consistent dynamics because compressed text prompts leave the temporal/physical evolution of a scene implicit, forcing the model to infer both process and appearance simultaneously ("Causal Opacity").

Method: VideoCoCo is an agentic dual-engine framework that uses executable Blender Python code as a process-level chain of thought: a coding agent synthesizes a self-contained Blender program from the prompt, which is executed in a sandboxed simulation engine to render a deterministic, low-fidelity "white-clay" spatiotemporal draft; an instruction agent then composes an appearance-focused editing instruction from the prompt and draft, and a draft-conditioned generative video editor (initialized from the OmniWeaving base generator) turns the draft into a photorealistic video via draft-conditioned editing. To adapt the editor to simulated drafts, the authors build VideoCoCo-3K, a dataset of 3,000 draft-instruction-target triplets generated by running the pipeline at scale and using Seedance 2.0 as a teacher editor to produce photorealistic targets; the editor is then fine-tuned (LoRA found best) with a standard conditional denoising diffusion objective.

Key results:

• Raises OmniWeaving's average PhyGenBench score from 0.475 to 0.558, the best score among all systems compared (including Wan2.2-TI2V-5B, Cosmos-Predict2.5, LTX-Video, Kling, Sora, CogVideoX).
• Raises OmniWeaving's average VBench-2.0 physical-plausibility score from 52.18% to 77.88% (+25.70 points), with best mechanics (92.31%, +29.52 pts) and thermotics (72.92%, +20.84 pts) scores.
• Ablation shows the tuning-free variant (draft conditioning alone, no editor training) already improves PhyGenBench average from 0.475 to 0.506, isolating the contribution of the executable draft itself; LoRA-tuned editor reaches the best 0.558, outperforming full fine-tuning (0.535) despite updating far fewer parameters.
• Largest gains concentrated in material (+0.133) and thermal (+0.078) dynamics on PhyGenBench, categories where appearance-driven priors are weakest.

Why it matters / caveats: Demonstrates that executable code (rather than text plans or sparse keyframes) can serve as a dense, inspectable, and controllable intermediate representation for physically grounded video generation. The authors note limitations: added inference latency and constraints from the Blender simulator's expressiveness, making complex phenomena like turbulent fluids hard to synthesize zero-shot; future work proposes integrating specialized physics engines (e.g., Taichi) and distilling the executable priors into end-to-end models.

Problem: Decoder-only language models entangle long-term memory and reasoning in a single parameter set, so long-term memory cannot be pretrained, scaled, or swapped independently of the base model.

Method: The paper scales up Memory Decoder (Cao et al., 2026), a standalone transformer module trained via KL-divergence distribution alignment plus a language-modeling loss to imitate a kNN retriever's output distribution, then interpolated with a frozen base model's next-token distribution at inference. To make this feasible at pretraining scale (207B deduplicated Pile tokens, 300B token training budget, memories up to 6.9B parameters), the authors build a distributed Faiss pipeline combining OPQ256 embedding compression, an IVF index with an HNSW quantizer sharded into IndexIVFPQ shards, and parallel GPU search, plus sparse kNN distribution storage and distributed streaming for training. The same recipe is applied to domain memories (1.7B parameters) for biology, law, and finance using Biology-Instructions, DISC-Law-SFT, and FinTrain corpora, tested on Pythia, Qwen3, and OLMo-2/3 backbones.

Key results:

• Pairing a 6.9B general memory with frozen Pythia-410M raises average score (17 benchmarks) from 29.86 to 37.34, surpassing Pythia-12B (37.24) with 39% fewer total parameters.
• 1.7B domain memories improve the average across BioInst, LawBench, and FinEval by more than 9 points at every Qwen3 scale (0.6B–14B), e.g. +17.96/+8.97/+3.04 points on the three domains for Qwen3-14B-Base.
• Cross-vocabulary transfer to OLMo-2-7B and OLMo-3-7B (using only 20% of standard training budget) improves domain averages by 4.26 and 7.77 points respectively.
• Ablations show memory training (vs. matched CPT attached module) adds 8.71–10.21 points on BioInst and 1.68–3.93 on LawBench; memory also raises verbatim-extraction EM@8,16 from 42.4% to 49.7% and domain-anchor completion accuracy from 22.6% to 56.5% versus CPT.

Why it matters / caveats: Demonstrates that independently scaling a swappable, pretrained parametric memory is a more parameter-efficient path to improving LLM performance than scaling the backbone, with domain memories reusable/swappable at inference without retraining the base model. The authors note the offline kNN indexing/retrieval cost during memory pretraining remains a growing preprocessing burden, and the interpolation coefficient is currently fixed rather than adaptive.

Problem: Existing agentic visual reasoning MLLMs invoke external tools indiscriminately, failing to recognize when tools are actually needed, and the accuracy gains from tool use on hard problems are largely canceled out by errors tools introduce on problems the model could already solve without them.

Method: The authors define two metrics, Mode Adaptiveness (MA) and Tool Effect (Tool-Gain/Tool-Harm), to diagnose this behavior across models like Thyme, DeepEyesV2, CodeV, and Metis, then build Beacon, an 8B agentic visual reasoning model based on Qwen3-VL-8B-Instruct, trained via an SFT-then-RL pipeline. SFT uses Gemini 3.1 Pro-generated and refined code-assisted reasoning trajectories on hard examples; RL uses GRPO augmented with two novel mechanisms: the Necessity-Aware Adaptive Reward (NAAR), which gives full reward to correct text-only answers, partial (0.25) reward to correct code-based answers when text-only suffices, and full reward to code-based answers otherwise, and Hint-Guided Capability Expansion (HCE), which injects Gemini 3.1 Pro-generated answer-free hints into rollouts for problems the policy fails entirely, then strips the hints during policy update to transfer the discovered tool-use behavior back to the hint-free model.

Key results:

• Beacon-8B achieves the best average performance among open-source models across 13 benchmarks (V*, HRBench, VisualProbe, RealWorldQA, BLINK, BabyVision, ChartQAPro, MathVista, MathVision, VisualPuzzles, GameQA, TIRBench), ranking first on 11 of 13 and beating its Qwen3-VL-8B-Instruct base by an average of 6.07 points.
• Beacon shows the largest tool-induced accuracy gain over tool-free reasoning (ΔAcc = +1.96%) versus under +1% for Thyme, DeepEyesV2, CodeV, and Metis.
• Beacon attains the highest Mode Adaptiveness (MAmean = 58.83%) and the largest gap between tool-induced gain and harm (ΔTE = +3.14%), compared to near-zero ΔTE for prior models.
• Ablations show NAAR alone yields the best Mode Adaptiveness, HCE improves ΔTE over plain GRPO, and combining both gives the best overall accuracy (58.98% average) with ~40% of previously "all-wrong" rollout groups recovered into useful learning signal via hint-guided re-rollouts.

Why it matters / caveats: The work reframes evaluation of agentic visual reasoning around whether tool use is genuinely necessary and beneficial rather than raw accuracy, showing that most prior tool-augmented MLLMs provide little real advantage over text-only reasoning; Beacon demonstrates that explicit adaptive-reward and hint-guided RL design can produce authentic capability gains, though it is a preprint ("work in progress") and results are reported only on the paper's own 13-benchmark suite.

Problem: RAG paradigms (lexical, dense, graph-based, agentic) are typically evaluated on different benchmarks at a single fixed corpus size, so how their accuracy and cost actually scale as corpora grow to enterprise scale has remained unclear.

Method: The authors build a 28-tier, strictly nested corpus ladder (1,144 to 511,959 documents, 1.7M to 601M tokens, growing 1.25x per rung) on EnterpriseRAG-Bench, holding 500 questions and a fixed "bedrock" of 1,144 relevant/adversarial documents constant while background corpus size grows. They evaluate seven native pipelines across four paradigms — BM25, DenseRAG (Qwen3-Embedding-0.6B), HippoRAG 2, MS-GraphRAG, LightRAG, LinearRAG, and a File-System Agent (ReAct-style tool loop with an 80-LLM-call budget) — all sharing one reader model (Qwen3.6-27B via vLLM) and one judging protocol, with token-level metering of build and query costs. Matched controls (Agent+BM25 retrieval swap, a typed graph-access-layer for agentic querying, paraphrase and top-10 depth controls) isolate agentic policy from retrieval substrate.

Key results:

• Scale-dependent crossover: File-System Agent leads at the bedrock (77.4 vs BM25's 74.7), but BM25 overtakes it around 10M corpus tokens and leads every larger tier, reaching 50.5 vs 30.7 (File-System Agent) and 29.9 (DenseRAG) at the full 601M-token corpus — a nearly 20-point gap.
• File-System Agent's query cost balloons from 226K to 343K tokens/question (39-60x BM25's ~5.8K), with budget exhaustion rising from <7% to 31% at full scale, while BM25's cost stays nearly scale-invariant.
• Graph methods hit construction walls: HippoRAG 2 extrapolates to 2.9B build tokens (~3 instance-days) at full scale, LightRAG is super-linear (b=1.36) projecting to ~102B tokens (~4 instance-years) and fails to complete beyond 2,826 documents.
• Mechanism control (Agent+BM25): swapping the File-System Agent's raw-tree search for BM25 candidate discovery raises its full-scale score from 36.9 to 69.4 while cutting tokens/question from 895K to 101K (~9x reduction), showing the failure is candidate discovery, not reasoning.
• Robustness holds across protocols: binary re-scoring preserves rankings at all 9 shared scales, and an independent judge agrees on 96.2% of pooled alignment verdicts.

Why it matters / caveats: For enterprise-scale corpora (10^5-10^6 documents), the results argue BM25 is the strongest scalable default and sits on the Pareto frontier without any LLM-based index construction, while agentic search is best used after ranked retrieval narrows candidates rather than as a substitute for global ranking. Caveats: findings are on one synthetic enterprise benchmark with one reader/judge model and a fixed 80-call agent budget, and graph-based methods' large-scale accuracy is unmeasured (construction-infeasible) rather than proven worse.

Problem: In open-ended domains without verifiable rewards, using contexts (extracted before training) as distillation supervision stops providing useful signal once the student absorbs them, while naively updating contexts during training destabilizes optimization via loss surges and conflicting target distributions.

Method: The authors decompose the reverse-KL objective of On-Policy Context Distillation (OPCD) into a distillation term (KL to the geometric mean of context-conditioned teacher distributions, qgeo) and a conflict term (−log Z, measuring disagreement among context-conditioned teachers), proving the conflict term contributes no direct gradient. Based on this, they propose Flux-OPD, an iterative training paradigm (K iterations of context extraction + context distillation) that computes a "contextual difference signal" Δk = log qgeo,k − log q0 (context-conditioned vs. context-free teacher) and injects it as a correction onto the stable context-free teacher anchor q0, rather than distilling directly toward context-conditioned teachers as OPCD/OEL do. A contextual weighting strategy scales the correction strength λk using the conflict term δk=−log Zk (via scaling and clipping, controlled by threshold τ and bounds λmin/λmax) to downweight conflicting contexts. Evaluated on prompt optimization for video generation (VPO 10K SFT prompts, Qwen3-VL-Instruct 4B student/8B teacher and Qwen2.5-VL-Instruct 7B student/32B teacher) and medical QA (RaR-Medicine ~18K questions, Qwen3 1.7B student/8B teacher), against OPD, OPCD, and OEL baselines.

Key results:

• On HealthBench (medical QA, Qwen3 8B→1.7B), Flux-OPD scores 20.61 total vs. 19.06 (OPD), 19.63 (OPCD), 19.66 (OEL) — best among all student variants, though still below the 8B teacher's 37.38.
• On VBench with Qwen3-VL-Instruct 8B→4B, Flux-OPD reaches 81.48 (Wan2.1-VACE-1.3B) and 79.69 (CogVideoX-2B) total, beating OPD/OPCD/OEL and even slightly exceeding the teacher's 81.48 average on one generator; OPD/OPCD/OEL degrade below the initial student on prompt optimization.
• Across 3 independent runs, Flux-OPD consistently beats OPD (VBench mean 78.06 vs 77.39; HealthBench mean 20.30 vs 19.51).
• Ablations show each component matters: removing evolving contexts, contextual correction, or contextual weighting all reduce HealthBench score (e.g., full method 20.61 vs. 19.66 without correction/weighting, vs. 19.73–20.03 with static correction strengths instead of adaptive weighting).
• Flux-OPD shows stable training loss/gradient norm (unlike OEL's loss surges) and better OOD generalization on IF-Eval (34.38 vs. OPD's 31.79 prompt-level strict accuracy).

Why it matters / caveats: The reverse-KL decomposition gives a principled explanation for why naively distilling toward context-conditioned teachers underperforms, and offers a general recipe (anchor + correction + conflict-based weighting) for using dynamic, student-adaptive supervision signals in reward-free domains. Gains are demonstrated only on two tasks (video-prompt optimization and medical QA) with relatively small student models (1.7B–7B), and several hyperparameters (τ, scaling vs. clipping, λ bounds) are hand-tuned per setting rather than learned.

Problem: Multi-reference image editors that place several named people into shared contact actions (embrace, carry, grapple) routinely produce fused limbs, invented extremities, and interpenetrating bodies, but existing benchmarks and VLM-as-judge checklists saturate near ceiling and fail to detect these anatomical/geometric errors.

Method: The authors mine 2,500 reference-target editing triplets from video (Pexels stock, Harmony4D, CHI3D) by computing a contact-density curve φ(t) from person-box IoU, pose proximity, optical flow, and occlusion cues, cropping low-contact "valley" frames as identity references and high-contact "peak" frames as held-out targets, then reverse-captioning edit instructions with a VLM (this is MPIE-Bench, spanning 405 scenes, 14 interaction categories, and four contact-density levels C0–C3). They pair this with MPIE-Eval, a six-axis protocol that keeps standard Count/Identity/Instruction/Quality axes but adds two new axes, Anatomy and Interaction, computed from meshes reconstructed by a frozen, public multi-person mesh recovery model (Multi-HMR): Anatomy scores unexplained human-like mass via a leftover-mask/blob analysis (weighted mix of attach/orphan/struct/residual sub-scores), and Interaction scores convex-hull penetration and surface-gap proxies mixed according to whether the prompt requires, forbids, or leaves unspecified contact.

Key results:

• Across ten editors (GPT-Image-2, Gemini-3-Pro-Image, Seedream-5-Pro, FLUX.1-Kontext, DreamO, OmniGen2, UNO, ACE++, BAGEL, FireRed-Image-Edit), mesh Anatomy tops out at 0.65 (Gemini) and mesh Interaction at 0.72 (Seedream) — no single editor is strong on both axes.
• The same images score above 0.95 (up to 0.98–0.99) on a VLM checklist judge, exposing severe ceiling-saturation versus the mesh-based scores.
• A five-rater human study (N=170 hard C2–C3 units) shows the mesh mapping beats the VLM judge on 9/10 checklist items by Spearman correlation with human gold.
• Track A vs. Track B (letterbox→1024²) resolution-normalized rescoring keeps open-model Anatomy rank Spearman at 0.89 (0.83 across all ten), and Anatomy top-3 unchanged; rankings hold under ablation of every weight/gate/threshold (rank Spearman ≥0.98).
• Identity shows the steepest closed/open gap (0.49–0.58 vs. 0.02–0.21).

Why it matters / caveats: The work shows VLM-as-judge contact evaluation is fundamentally unreliable for anatomical plausibility, motivating physically-grounded 3D mesh metrics as a more human-aligned alternative for benchmarking multi-person editors. Caveats acknowledged by the authors: absolute human agreement remains only moderate (treated as a ranking check, not a per-image oracle), the hull-penetration term is a ranking proxy rather than a true physical volume, and Interaction currently uses whole-body proximity without enforcing a specific contact locus (e.g., hand-to-hand).

Problem: Existing embodied-AI datasets fragment the perception-action loop across viewpoints, modalities, and spatial scales (egocentric-only, lab-only mocap, or short atomic clips), so no resource jointly captures synchronized first-person/third-person video, whole-body and hand motion, object 6-DoF pose, audio, and tactile signals during long, naturalistic home activities.

Method: The authors build the Ambient Capture Engine (ACE), a two-configuration capture rig (table-scale and room-scale) deployed in two real furnished homes, using OptiTrack PrimeX 22 motion capture (28 cameras total), ZED One/GoPro exocentric RGB cameras, the custom ACE-Ego-Head-V02 Lite 4-fisheye headset, Manus motion-capture gloves, and ACE-Sense-Glove Lite tactile gloves, all hardware/software-synchronized to an OptiTrack clock via QR-code "clock glance" alignment and hand-eye/ArUco-board calibration. This produces ACE-Data-0 (150 hours, 17M frames, 75,000 episodes, 200 task categories, 50 participants) with goal-level (not scripted) instructions across atomic HOI, chained HOI, and human-scene interaction (HSI) tasks; annotations (object 6-DoF pose/mesh/bbox/motion trail, SMPL-X body and MANO hand poses, tactile pressure maps, Gemini-3.1-pro-preview-generated language captions) are mostly derived by direct measurement/projection rather than model estimation. A three-level benchmark then evaluates tactile-from-vision, human motion estimation, and hand-motion estimation (ego- and exo-view) using over 30 existing SOTA methods (e.g., PressureVision, TouchAnything, GVHMR, WHAM, HaMeR, WiLoR, HaWoR).

Key results:

• Best tactile-from-vision model (TouchAnything) reaches only 0.71 temporal accuracy and 0.16 Contact-IoU, versus near-zero for PressureVision, showing pressure localization from egocentric video remains largely unsolved.
• Across 22 human motion estimators, Procrustes-aligned pose errors (PA-MPJPE) are competitive with lab benchmarks (best ~55.7mm, SMPLest-X), but world-frame trajectory error (WA-MPJPE) stays far higher (best ~217mm, Human3R), revealing a large local-pose-vs-global-trajectory gap.
• For egocentric hand-motion recovery, world-frame trajectory error is 98–102mm (Dyn-HaMR, HaWoR) versus only 63mm for the best exocentric method (HaPTIC) at comparable PA-MPJPE (~10mm), indicating egomotion estimation, not finger articulation, is the dominant error source.
• Dataset scale: 150 hours, 17M frames, 75,000 episodes, 50 participants, 200 task categories, >50 tracked object instances, captured from 8+ synchronized viewpoints per moment.

Why it matters / caveats: By making nearly all annotations directly measured (via calibrated multi-sensor tracking) rather than model-estimated, ACE-Data-0 offers a rare "ground-truth" testbed exposing where current perception methods fail under home-scene occlusion, long horizons, and egomotion — useful both as a diagnostic benchmark and as training data for imitation learning, world models, and VLA systems; the benchmark itself shows current human/hand motion and tactile-inference methods still generalize poorly to real, cluttered, long-horizon domestic settings.

Problem: Existing role-playing agent (RPA) benchmarks make candidates continue a fixed, externally authored dialogue history under a user-independent rubric, which conflates the RPA's own ability with the inherited history's quality and ignores that user satisfaction varies across individuals.

Method: The authors introduce PALATE (Person-Aligned LLM-Simulated-User Assessment with Tailored Evaluation), built on a pool of 300 bilingual character cards (240 original + 60 IP anchors) and a frozen 10-card evaluation panel. For five real annotator users with 5,133 turn-level satisfaction-labeled turns, they train per-user LoRA simulators (rank 16, α=32) on top of Qwen3.5-35B-A3B Instruct via next-user-action SFT (free-text continuation or a [QUIT] token), letting each simulator freely co-construct multi-turn trajectories with candidate RPAs rather than continuing a borrowed history. A meta-prompt automatically induces a frozen personalized experience rubric per user from their annotated training history, evaluated alongside shared Generic (12-dimension, turn-level) and Session (11-dimension, whole-session) rubrics; scoring uses GPT-5.5 as the main judge with Claude Opus 4.8 and DeepSeek V4 Pro as cross-family rejudges.

Key results:

• A controlled fixed-history experiment shows high-quality borrowed histories inflate mean overall scores by ~0.21 (five-point scale) and degraded histories deflate them by ~0.13, confirming fixed-history evaluation measures Q(c|H) rather than intrinsic RPA ability.
• Per-user LoRA simulators reach a 2AFC fooling rate of 0.561 (macro, vs. 0.500 chance) and identity-consistency score of 77.6/100, far above prompted-LLM baselines (fooling rate 0.242, consistency 61.7).
• Personalized rubric scoring (with next-user-reaction signal) achieves 0.613 macro pairwise agreement with human satisfaction on held-out turns, versus 0.507 for the Generic rubric and 0.467 for a reconstructed MiniMax baseline.
• Across 16 candidates (GPT-5.4/5.1, Claude Opus 4.8/Sonnet 4.6, DeepSeek V4 Pro/Flash/V3.2, Gemini 2.5 Pro/3.5 Flash, Qwen3-Max, Qwen3.5-35B-A3B, Seed 2 Mini, Grok 4.3, GLM-5.1, MiniMax-M2-her, CoSER-Llama-3.1-70B), no single model wins on all tracks or all users: GPT-5.4 leads Generic quality, Claude Sonnet 4.6 leads Session, and different users (U1–U5) each prefer different top candidates (Qwen3-Max, DeepSeek V4 Pro, Claude Opus 4.8, Claude Sonnet 4.6).
• Ranking stability across repeated rollouts is high (Spearman ρ ≥ 0.959), and cross-judge rejudging with Claude Opus 4.8/DeepSeek V4 Pro preserves wide-margin orderings on Personalized and Generic tracks.

Why it matters / caveats: The benchmark reframes RPA evaluation from a single scalar leaderboard to interpretable user–RPA performance profiles, showing that generic quality, long-horizon session capability, and person-specific satisfaction are not proportional to one underlying capability. A key limitation is scale: only five extensively annotated users are used, and there is no end-to-end human-ranking reference spanning all 16 candidates, so cross-candidate calibration against broader human judgment remains future work.

Problem: Existing video captioning models cannot explicitly ground local visual elements (subjects, objects, attributes) in a video caption to specific images from a set of candidate reference images, including rejecting distractor images and consistently grouping multiple references of the same entity.

Method: The authors introduce the task of multi-reference image-grounded video captioning and propose RefCaptioner, a two-stage post-training framework built on Qwen3-VL-8B-Instruct. Stage one is mixed-data supervised fine-tuning (SFT) with LoRA on an equal mix of a manually verified multi-reference caption set and a general detailed video-caption set, teaching reference selection and phrase-level <Image_i> tag placement while preserving descriptive coverage. Stage two applies Hierarchical Coverage-Discounted GRPO (HCD-GRPO), which optimizes the full LLM (vision encoder and merger frozen) using a dual-branch reward: a Factual Caption Reward (keypoint coverage plus video-QA-based error discounting) and a Multi-Reference Grounding Reward combining correctly-bound reference coverage, Reference Binding Accuracy, Distractor-Aware Evidence Suppression (DAES), and Cross-Reference Semantic Coherence (CRSC). Training data comes from a constructed corpus of 20,000 videos and 171,354 reference images, and the paper also introduces MRVBench, a 462-video (185 AIGC + 277 real-world), 3,831-reference-image, 2,172-QA-pair benchmark for evaluating factuality and grounding.

Key results:

• On MRVBench, RefCaptioner achieves the best MRVScore (0.888) among open-source models, close to Gemini-3.1-Pro (0.897) and above GPT-5.4 (0.870).
• It attains the best or tied-best score among open-source models on every grounding metric (e.g., Ref-Bind 0.967, Eff-Bind 0.912, Subj-R 0.817, Subj-F1 0.869), surpassing both Gemini-3.1-Pro and GPT-5.4 on Ref-Tag-R, Subj-R, and Subj-F1.
• On the hardest robustness setting (13+ reference images), RefCaptioner scores 0.769 (recall×binding×distractor-rejection product) vs. GPT-5.4's 0.703 and Qwen3.6-35B-A3B's 0.398.
• On general captioning benchmarks it improves over its Qwen3-VL-8B base by 5.66–6.08 points on VDC background/main-object/detailed-description accuracy, and by 1.30 (AR) and 1.70 (CR) points on VCapsBench while maintaining comparable IR.
• Human GSB evaluation on caption-conditioned video reconstruction shows RefCaptioner preferred over every baseline (e.g., 76% Good vs. 10% Bad against Qwen3.6-35B-A3B).
• Ablations show HCD-GRPO's factual-caption reward is most important for VQA, DAES for distractor rejection (Dist-Rej drops from 0.985 to 0.914 without it), and CRSC for subject recall (Subj-R drops from 0.817 to 0.778 without it).

Why it matters / caveats: The work shows phrase-level reference grounding can be learned jointly with caption generation rather than added post-hoc — a two-stage "generate then refine with tags" baseline clearly underperforms RefCaptioner's single-pass approach on reference recall and binding, suggesting the joint training recipe (SFT + HCD-GRPO) is key. The training corpus itself will not be publicly released due to licensing restrictions, only the MRVBench test set and annotations.

Problem: General vision-language models (VLMs) can reason about task-level goals but lack reliable fine-grained spatial perception (localization, segmentation, depth, pose), while specialist vision models provide precise perception but cannot compose their outputs into task-level decisions.

Method: SpatialCLI is a three-stage Call–Learn–Internalize framework: (1) Call exposes four specialist vision models — Locate (Locate Anything + Grounding DINO), Segment (SAM 3), Depth (Depth Anything 3), and Pose (Orient Anything V2 + VGGT) — as ReAct-style callable tools; (2) Learn applies Cold-Start SFT (using Qwen3.5-397B-A17B as teacher to generate filtered tool-use trajectories) followed by agentic RL with GRPO (DAPO-style clip-higher, token-level reduction) using outcome rewards; (3) Internalize converts successful SpatialCLI-RL trajectories via Progressive Evidence-Grounded Trajectory Verbalization (turn-wise evidence consolidation + global trajectory verbalization) into Dual-View Capability Internalization training data (Capability-Internalization View and Tool-Use View, jointly optimized as L_CI = L_internal + λL_agentic). The method is instantiated on Qwen3-VL-8B-Instruct, Qwen3.6-35B-A3B, and Qwen3.6-27B, trained on a shared pool of 37,000 tasks from Vlaser, MindCube-Train, BOPASK-Trajectory, BOPASK-Object-Rearrangement, and RefSpatial. The paper also introduces SpatialCLI-Bench, a 516-example six-choice VQA benchmark for compositional perception spanning localization, segmentation, depth, and pose, built via Gemini 3.1 Pro annotation plus specialist-model evidence filtering and human verification.

Key results:

• On MindCube, SpatialCLI raises Qwen3-VL-8B-Instruct from 29.3% to 84.6% with tools, surpassing GPT-5.6 Sol with tools (72.1%), while retaining 73.8% without tools after internalization.
• On SpatialCLI-Bench, SpatialCLI-8B reaches 91.3% with tools and 72.7% without tools, up from 35.3% for the untrained base model without tools.
• Frontier model GPT-5.6 Sol scores only 48.8% on SpatialCLI-Bench, exposing a clear gap that SpatialCLI Tools close to 72.9% (+24.1 points).
• Ablations show structured tool returns (coordinates/polygons) beat visual-only red-border returns by 20.1 points on SpatialCLI-Bench, and Full Dual-View internalization outperforms One-Pass Dual-View and single-view baselines (72.7 w/o tools vs. 64.5 for one-pass, 45.0 for final-answer-only supervision).
• Tool-use gains shrink with model scale (at most 1.0 point variation across 8B/35B-A3B/27B), but internalization (Capability Internalization Index) keeps scaling with both training data exposure and model capacity.

Why it matters / caveats: The results show tool use and internalized (tool-free) reasoning can coexist and reinforce rather than trade off against each other, pointing toward VLMs that absorb specialist perceptual skills rather than depending on external tools at inference time. The authors note the approach is limited by specialist-tool coverage/reliability, is restricted to structured perceptual outputs and perception-centric tasks, and has not yet been extended to multimodal generation outputs or integrated with VLA action tools.

Problem: It is unclear whether multimodal large language models that generate sketches, annotations, or intermediate images during reasoning actually rely on those visual states, since existing benchmarks emphasize final-answer accuracy without diagnosing whether self-generated visual actions are relevant, faithfully rendered, and actually used.

Method: The authors introduce See2Think, comprising See2ThinkBench (1,200 visually dependent problems across 12 task categories spanning 2D structured, 3D scene, and real-world reasoning, built from sources like RBench-V, EMMA, M3CoT, PRISM-Bench, CLEVR, SuperCLEVR, VLABench, DROID, and IntPhys2, with caption-only solvability filtering to remove text-dominant shortcuts) and Visual Action-of-Thought (VAoT), an interleaved protocol recording textual thoughts, structured visual actions, and externally rendered states. Four inference settings are compared: CoT (text-only), VAoT-NoRender (action planning without rendering), VAoT (closed-loop with rendering), and VAoT-WrongRender (task-relevant corrupted feedback to test behavioral dependence). Four models — GPT-5.5, GPT-o3, Gemini 3.5 Flash, and Qwen3-VL-32B-Instruct — are evaluated, with GPT-5.4 used as an external judge scoring Action Relevance, Render Faithfulness, and Feedback Uptake on each trajectory.

Key results:

• No single setting dominates: GPT-5.5 and Qwen3-VL-32B-Instruct perform best with CoT, GPT-o3 best with VAoT, and Gemini 3.5 Flash best with VAoT-NoRender; overall accuracy ranges roughly 34.8–43.8% across settings/models.
• Action Relevance is near-saturated (0.958–0.994 average) across models, but Render Faithfulness is much lower (0.550–0.705), identifying faithful rendering as the main bottleneck after action selection.
• Under VAoT-WrongRender, 3D scene reasoning accuracy drops by over 10 percentage points (up to 15.5 points at high Feedback Uptake), showing strong behavioral dependence on visual feedback in that domain, while corrupted feedback changes the semantic answer in 32.7–54.2% of samples across models even when net accuracy effects are smaller.
• Qwen3-VL-32B-Instruct has the highest overall Feedback Uptake (0.871) despite lower final-answer accuracy, showing visual-state utility and visual-state dependence are distinct; a human audit of 480 pooled ratings finds 92.9–96.9% of process-level judgments at least partially reasonable.

Why it matters / caveats: The work shows "thinking with images" is a conditional, model- and environment-dependent capability rather than a universal accuracy booster, and disentangles whether visual actions are merely present from whether they are faithfully executed and genuinely used — useful for diagnosing when visual chain-of-thought is genuine versus decorative. Limitations include coverage of only four models, reliance on an external renderer and automatic judge (despite human validation), and sensitivity of the WrongRender intervention's measured effect to intervention quality.

Problem: On-policy self-distillation (OPSD) for reasoning LLMs is brittle in practice because it always forces the student to imitate a fixed privileged teacher, with no explicit control over how aggressively to move away from the reference policy.

Method: The authors show vanilla OPSD is exactly the β=1 case of a KL-regularized policy-optimization family, where the reward is the teacher-to-reference log-ratio and β weights a KL penalty anchoring the student to a reference policy πref; they derive the family's closed-form optimal policy as a geometric interpolation between πref and the privileged teacher pT, then realize it tractably via token-level logit interpolation (softmax over a weighted sum of reference/teacher logits) scheduled with a bounded linear curriculum (teacher weight wk from wstart to wend over K steps) rather than optimizing the RL objective directly. Training uses β-OPSD loss with return-to-go credit assignment (discounted cumulative future token-level log-ratio mismatch, discount γ) instead of vanilla OPSD's myopic local-token KL gradient. Evaluated on Qwen3-1.7B/4B/8B fine-tuned with LoRA on the OpenThoughts math reasoning subset, compared against base model, SFT, GRPO, and vanilla OPSD baselines.

Key results:

• On Qwen3-1.7B, β-OPSD improves avg@12 over vanilla OPSD by 9.16 points (AIME 2024), 5.27 points (AIME 2025), 2.78 points (HMMT 2025), and 5.74 points on average.
• Gains persist at scale: +1.76 average points on Qwen3-4B and +1.66 on Qwen3-8B over vanilla OPSD, outperforming SFT and GRPO baselines across all three model sizes.
• Ablations isolate contributions: the logit-interpolant target alone (vs. direct teacher target) improves avg@12 by up to 6.03 points (AIME 2024); return-to-go credit assignment alone improves avg@12 by up to 5.55 points (AIME 2025) over local token-gradient weighting.
• Best interpolant construction uses a dynamic (stop-gradient current) student endpoint with a fixed privileged teacher endpoint; best schedule tested is linear wstart=0.5→wend=0.8.

Why it matters / caveats: The work reframes on-policy self-distillation as a principled, controllable instance of KL-regularized policy optimization, offering a cheap distillation-based way to approximate an RL solution without costly rollouts or advantage estimation. Caveats: gains shrink at larger model scale and occasionally underperform vanilla OPSD on individual benchmarks (e.g., AIME 2024 for Qwen3-4B, HMMT 2025 for Qwen3-8B); evaluation is limited to math reasoning benchmarks and LoRA fine-tuning, and the optimal interpolation schedule is left as an open question.

Problem: Full self-attention in Diffusion Transformers scales quadratically with sequence length, making high-resolution image and long-video generation increasingly costly, and visual diffusion lacks a systematic hyperparameter-transfer and scaling-law framework analogous to what language models have.

Method: Chimera is a single-stream visual diffusion backbone that packs text, image, and video tokens into one sequence and processes them with Kimi Delta Attention (KDA, linear-complexity O(N) recurrent state tracking) interleaved 3:1 with Multi-head Latent Attention (MLA, compressed-KV global attention), plus modality-aware short convolutions (a "Canon Layer") before each KDA update that replace positional embeddings entirely (NoPE) via a single temporal-major raster scan. Capacity is added with sparse Mixture-of-Experts FFNs (56 routed experts, top-8, auxiliary-loss-free balancing), identity hyper-connections (iHC, a simplified doubly-stochastic-free variant of mHC), and sandwich RMSNorm. To scale this heterogeneous architecture, the authors introduce HeteroP, a module-wise hyperparameter-transfer scheme (extending µP/CompleteP) that derives separate transfer ratios per tensor from its functional fan-in and network depth, enabling a consistently tuned model family (59M–19.3B total parameters) for fitting Chinchilla-style compute-optimal scaling laws over activated parameters, training tokens, and image:video data ratio.

Key results:

• Under matched training compute, dense Chimera reaches the same training loss as a matched full-attention Wan 2.1 (2B) baseline with 1.7x fewer FLOPs; the complete Chimera (MoE + iHC + HeteroP) widens this to 7.3x fewer FLOPs (6.27×10^19 vs. Wan's 4.29×10^20 FLOPs at loss 0.149).
• Final model: 11B total / 2B activated parameters, trained with only ~600 H100-days versus Z-Image-Turbo's ~12.4K H100-days (20x less compute).
• Achieves GenEval 0.82 and DPG-Bench 85.12, matching FLUX.1-dev and Z-Image-Turbo on GenEval and beating both on DPG-Bench.
• Zero-shot extrapolation from 5-second training clips to 30-second video: FID degrades only 6.5% (77.1→82.1) and FVD only 20.9%, versus 50.5%/33.9% for Wan2.1-T2V-1.3B and 53.6%/33.7% for HunyuanVideo-1.5.
• KDA/MLA backbone supports 1.68x longer sequences on one 80GB A100 than a matched MHA/MLA backbone and runs 2.14x faster at 255K tokens; fused Triton convolution kernel gives 2.2–2.3x forward speedup and up to 4x lower peak activation memory.
• Fitted scaling laws: compute-optimal image pretraining balances model size and tokens nearly evenly (N_opt ∝ C^0.48–0.52), while video pretraining tilts modestly toward model size (N_opt ∝ C^0.53–0.56); optimal image:video sample ratio drifts from ~4:1 to ~3:1 with compute.

Why it matters / caveats: Demonstrates that language-model-style hybrid linear/global attention and Chinchilla scaling laws can be adapted to visual diffusion with substantial compute savings and strong long-video length generalization, but the model is trained on a comparatively modest budget (~600 H100-days) and still falls short of the strongest large-scale image generators like Seedream 3.0.

Problem: Interactive video world models lack a single control interface that is both frame-level precise and general across arbitrary dynamics families (e.g., human motion, camera, robot manipulation, gameplay), since symbolic/text interfaces are imprecise and structured motion signals or self-reconstructing latent-action models (LAMs) entangle action with appearance.

Method: ShadowDancer introduces "shadow pairs" — two independent renders of the same underlying dynamics d with independently resampled appearance c, x=R(d,c) and x̃=R(d,c̃) — constructed at scale via the Shadow Library (paired-rendering scripts across Blender/SMPL-X animation, ManiSkill robotic simulation, GTA-/Cyberpunk-style and Unreal Engine open-world games, and DL3DV camera trajectories). It trains a latent action model (LAM) with "cross-shadow prediction": an inverse-dynamics encoder reads latent actions z from the source transition (xt, xt+1) while a forward decoder predicts the other shadow's next frame from (x̃t, zt), using a β-VAE objective (β=0.01, latent dim dz=32), so appearance cannot help prediction and only shared dynamics survives in z. The frozen LAM and a frozen 3D-VAE then condition a fine-tuned SkyReels-V2-1.3B I2V DiT backbone (flow matching, AdaLN modulation plus zero-initialized cross-attention), which is further converted into a block-causal autoregressive world model for streaming rollout, with demonstrations stored as reusable "action assets" (z, s).

Key results:

• Average blinded win rate of 86% over baselines in long-action rollout comparisons (judged via 2AFC by VLM "Fable 5").
• Against Olaf-World baseline on action transfer: PSNR improves from 18.2→22.4 (human motion), 13.0→17.0 (first-person combat), 12.6→17.4 (third-person action), 14.0→22.6 (robot manipulation); camera control ATE drops from 0.021→0.003 and RPE from 0.072→0.005.
• Long-horizon rollout favor rates over baselines: 94% vs Olaf-World, 88% vs Yume-1.5, 78% vs LingBot-World 2.0 on action control (similarly high on action fidelity and long-horizon consistency).
• Ablations show pairing (not assets alone) drives the gain: paired-latent + assets reaches PSNR 16.35 vs 14.92 for unpaired-latent + assets, isolating cross-shadow pairing as the source of transfer improvement.
• ShadowDancer roughly halves Fréchet Video Distance relative to Olaf-World on three reconstruction families, and reproduces unseen post-training actions (a modded character's gait, a two-handed sword swing) on held-out maps.

Why it matters / caveats: The approach lets any demonstrated clip become a reusable, label-free action asset that transfers across environments without motion estimators or fine-tuning, unifying disparate action families under one latent interface. Limitations: it requires pre-authored action assets at deployment, and true shadow pairs are hard to obtain for real-world video (which enters training only as degenerate self-pairs for appearance realism, not identifying supervision), so the method currently relies heavily on synthetic/simulated data sources.

Problem: Parent-order execution (splitting a large trade into smaller child orders to minimize execution costs) traditionally relies on static market assumptions (e.g., TWAP, Almgren-Chriss) or task-specific trained policies, neither of which adapts well to changing market conditions without redesign.

Method: The authors propose PACE (Plan-Ahead Controlled Execution), a hierarchical LLM framework with a Planner module that decomposes a parent order into N long-horizon sub-plans by scoring 5-minute time slots and blending the resulting allocation with a TWAP baseline via a weight λ, and an Executor module that adjusts quantity within each sub-plan relative to the TWAP baseline using an LLM-generated adjustment score scaled by γ. They evaluate on Shenzhen Stock Exchange Level-1 snapshot data (1,680 randomly generated parent orders from April 2026) using ChatGPT-5.4 and DeepSeek-v4-flash as the underlying LLMs, against TWAP, Almgren-Chriss (AC), XGBoost, and LSTM baselines, with a custom Matcher/Evaluator backtesting environment under aggressive and passive order-submission settings.

Key results:

• PACE (DeepSeek-v4-flash variant) beats the strongest baseline by 0.65 bps and TWAP by 1.02 bps (aggressive setting) / 0.71 bps and 1.07 bps (passive setting), significant via bootstrap (p=0.002 and p=0.014).
• Estimated economic impact: ~USD 6.5–10 million in annual execution-cost savings for a fund trading USD 100 billion/year, at an LLM API cost of only ~USD 30 for the full experiment (1,680 orders, ~USD 35.6M traded value).
• Ablations show both Planner and Executor modules contribute (removing either reduces gains from +1.02 to +0.66/+0.40 bps), and gains persist across low/high volatility groups and depth-limited matching rules.
• Behavioral analysis: higher Planner confidence correlates with better performance (regression coefficients 4.78–7.34, p<0.05/0.01), opposite of human overconfidence effects; Executor trades earlier under low time pressure rather than procrastinating (time-pressure coefficient significantly negative, e.g., -0.16, p<0.01).

Why it matters / caveats: This is presented as the first systematic study applying LLMs to trade execution ("how to trade") rather than trade selection ("what to trade"), requiring no market-model assumptions or task-specific training, with gains that generalize across two different LLMs. Caveats: results are backtested on a single market (Shenzhen A-shares) and one month of data, gains are modest in absolute bps, and the authors note live-trading validation and richer inputs (news, microstructure, cross-asset signals) are left to future work.

Problem: Memory-augmented LLM agents typically replay retrieved experiences verbatim into the context regardless of whether they fit the current state, which causes negative transfer when the retrieved memory was formed under different conditions.

Method: MemHarness decomposes memory-guided decision-making into retrieval, contextual memory reconstruction, and action generation, all handled by a single unified policy (Qwen2.5-7B-Instruct backbone) built on the verl-agent framework. At each step the policy retrieves top-3 experiences (via BGE-M3 embeddings over a Milvus vector memory bank) and compares each retrieved experience's source observation against the current history window to critique and rewrite it into state-specific guidance (or output <EMPTY> to fall back to self-reasoning) before generating an action. The full retrieve-reconstruct-act pipeline is trained end-to-end with GRPO using sparse task-outcome rewards (10/0) plus a small format-adherence bonus, after a brief cold-start SFT stage (200 memory-augmented trajectories + 200 trajectory-to-memory summarization pairs per benchmark, distilled via GPT-5.1 from AgentGym seed trajectories).

Key results:

• Achieves 85.2% average success rate on ALFWorld and 75.6% on WebShop, beating GRPO by +8.8% and +9.5% respectively, and beating Gemini-2.5-Pro by +23.1% (ALFWorld) and +39.7% (WebShop) despite using only a 7B model.
• On ALFWorld out-of-distribution layouts, MemHarness reaches 85.9% success rate vs. 76.3% for "RL + Raw Memory" (verbatim replay).
• Ablations show naively combining memory with GRPO (e.g., Mem0+GRPO, SimpleMem+GRPO) degrades performance, while removing reconstruction ("w/o reconstruction") drops ALFWorld from 85.2% to 79.6%, and even disabling memory at test time ("w/o memory," 83.0%) still beats plain GRPO (76.4%), indicating the reconstruction objective intrinsically improves reasoning.
• Swapping the learned reconstruction module for a generic zero-shot Qwen2.5-7B-Instruct rewriter drops ALFWorld performance from 85.2% to 77.7%, showing end-to-end RL-trained reconstruction outperforms prompted rewriting.

Why it matters / caveats: The work reframes agent memory around cognitive-science-style reconstructive recall rather than static retrieval, offering a concrete recipe (retrieve → critique/reconstruct → act, trained jointly via GRPO) that measurably reduces negative transfer and improves OOD robustness; evaluation is limited to ALFWorld and WebShop with a single 7B backbone, and the authors note future work is needed to scale to larger models and more open-ended environments.

Problem: Latent world models (JEPAs) predict how actions change a scene, but converting a desired latent change into an actual action still requires expensive test-time search (e.g., CEM/MPPI rollouts), leaving representation learning and control decoupled.

Method: INTACT (INtent-To-ACTion) augments a LeWM-style forward JEPA (with SIGReg regularization) with one shared conditional action operator, the INTACT Predictor Gη, called twice per transition on an identical four-slot input grammar [z, m, z⊙m, embedded previous action]: once with an attached "local" physical intent (zt+1 − zt) for inverse-dynamics grounding, and once with a stop-gradient "goal" intent (sg(zg) − zt) that serves as a deployable, search-free control interface. At deployment (Algorithm 2), the model recurrently composes the INTACT Predictor mean with the unchanged Forward Predictor to emit a "Direct" action-chunk plan requiring zero candidate sampling; an optional "Guarded A" verifier runs a small local CEM (K=128, I=3, σ0=0.25) centered on this plan. The method is trained and evaluated on the four official LeWM tasks (PushT, OGBench Cube, DMC Reacher, TwoRoom) both single-task and with one shared multi-task encoder across all four domains.

Key results:

• One-epoch, zero-search single-task models reach 85.78% (PushT), 100.00% (Cube), 97.67% (Reacher), and 97.89% (TwoRoom) success; optional local CEM verification (384 vs. 9,000 candidates, 23.44× fewer) raises macro SR to 96.86%, beating pure CEM 300×30 by 16.00 points.
• Paired E5 Cube audit (n=300 matched starts): INTACT Direct 98.7% vs. LeWM CEM 300×30 67.0% (+31.7 points), with 97 favorable vs. only 2 adverse success flips.
• Inference latency drops from 1.48 s (CEM 300×30) to 2.9–5.5 ms for Direct control (~300× planner-latency reduction).
• Shared four-task encoder reaches 89.39±0.77% macro Direct SR, improving all four tasks over matched shared-encoder LeWM by 5.66/32.23/12.56/42.44 points, and exceeding published task-specific LeWM.
• Predicted–expert action-family kNN alignment correlates with success rate at r=0.954 (CKA r=0.897) across 45 checkpoints, versus only r=0.815 for pointwise action R²; the corresponding LeWM–CEM kNN–SR relationship is negative (r=−0.873).

Why it matters / caveats: By coupling forward prediction with an action-aligned intent representation, INTACT removes a major model-based search bottleneck for high-frequency control without an extra policy-training stage, though the authors note it is not an end-to-end VLA latency benchmark and the approach is validated only on four simulated LeWM benchmark domains.

Problem: Existing LLM agent memory systems store interaction content but do not model which peer agents can be trusted and under what conditions, so a central model in a multi-agent system cannot reliably aggregate peer answers when it cannot directly verify them (e.g., in RAG tasks where supporting evidence is hidden).

Method: Σ-Mem maintains two real symmetric matrix states per event: per-peer historical competence matrices {M_p} and a peer relationship matrix G, both updated via decayed, bounded rank-one updates from post-decision correctness feedback, with stability guaranteed by Weyl's inequality bounding eigenvalue change per update. The memory is read via peer-specific residual steering injected into the upper decoder blocks of a frozen central model (producing a Yes/No utility score combined with a posterior over latent peer correctness states for peer selection), or read directly without response content via M-routing (pick peer with highest memory readout) or M-weighted voting. Training uses a mixed supervised dataset of 2,963 events across math reasoning, RAG, and code generation, with peer answer generators Gemma-3-4B-it, Phi-4-mini-instruct, and Qwen2.5-Coder-7B-Instruct, evaluated on five Qwen-family central models (Qwen3-0.6B/4B/8B, Qwen3.5-4B/9B).

Key results:

• On the CF@90 counterfactual reliability-shift benchmark, Σ w/ G raises Qwen3-0.6B accuracy from 46.22% to 71.10%.
• Generalizes to unseen peers (Llama-3.2-3B-Instruct, BitCPM-CANN-3B) in enlarged 4- and 5-peer pools, improving accuracy in most CF settings despite training with only three peers.
• Improves base models in 27 of 30 cases on six OOD benchmarks (PIQA, MMLU, OpenBookQA, SciQ, BBH, SuperGLUE); on BBH, Qwen3-4B improves from 20.38% to 28.66% and Qwen3-8B from 19.29% to 28.17%.
• Response-free direct memory readouts (M-Route, M-Vote) outperform both majority voting and the best fixed peer baseline across the full OOD evaluation set; accuracy on OOD data rises consistently as feedback availability increases from 5% to 100% (ablated on Qwen3.5-4B/9B).

Why it matters / caveats: Establishes reliability memory as a reusable, model-agnostic mechanism for trust-aware coordination that requires no retraining of underlying LLMs. A noted limitation is CF@50, an intentionally ambiguous data stream where half the history supports each peer preference, causing Σ-Mem to faithfully accumulate misleading evidence and sometimes reduce accuracy — the authors frame this as a faithfulness property rather than a defect, and flag adaptive balancing of historical vs. current-response evidence as future work.

Problem: Multimodal agentic VQA systems are evaluated almost solely on final-answer accuracy, an aggregate signal that cannot reveal whether a correct answer came from grounded evidence, language priors, or accidental error cancellation, and that hides failures like citation-backed entity hallucination and repair-time error amplification.

Method: LedgerMind treats an agent trajectory as a provenance-constrained state machine centered on a Structured Evidence Ledger, where every tool output is normalized (via a deterministic mapping M) into typed entries (PERCEPTION/RETRIEVAL/DERIVATION) with confidence, lifecycle status (ACTIVE/STALE/CONFLICTED/DROPPED), and dependency links, and downstream claims may only cite active entries. A Three-Layer Grounding Protocol checks claims via Support Coverage, an Entity Consistency Check (ECC), and a Numeric Coherence Check (NCC); an Adaptive Dual-Path Dispatcher routes queries to a DIRECT path or a FULLPIPE (planning, OCR extraction, crop-zoom, Dual-Read Verification, retrieval, grounded reasoning) path based on a complexity classifier; and an Event-Triggered Verification-and-Repair engine applies only seven typed operators (DROP, REFRESH, RETRY, SWITCH, ACQUIRE, STOPANDANSWER, ABSTAIN), yielding a formal Provenance Non-Amplification guarantee (Proposition 1). The framework is training-free and applied to six frozen backbone MLLMs (GPT-4o/GPT-5.5, Gemini-3-Flash/3.1-Pro, Claude-Sonnet-4.6/Opus-4.7, Kimi-K2.6) across VTC-Bench, MMStar, MMMU, MMMU-Pro, EMMA, MC-Search, and an in-house Hard-200 stress set, with faithfulness audited via a Symmetric Reasoning Faithfulness Audit (S-RFA) using Gemini-3.1-Pro as judge.

Key results:

• On VTC-Bench, LedgerMind with Gemini-3-Flash reaches 58.9% (new SOTA), lifting Gemini-3.1-Pro by +11.8, Gemini-3-Flash by +12.4, and GPT-4o by +23.3 points over native backbones.
• On EMMA, LedgerMind attains 58.29% overall (+9.58 over the best thinking-mode baseline), with +16.15 pp on Math and +16.02 pp on Physics.
• On Hard-200, every backbone improves by +11.2 to +19.7 points overall (no negative cells in the delta heatmap); Kimi-K2.6 jumps from 19.5% to 46.0% (+26.5) on BrowseComp-VL.
• On MC-Search, LedgerMind (Claude-Opus-4.7) raises F1 from the best official baseline's 41.78% to 61.28%, HPS from 31.35% to 57.82%, and nearly halves RD from 0.89 to 0.54.
• Ablations on MMMU-Pro (Gemini-3-Flash) show removing the ledger is the most damaging change (Overall drops from 68.89 to 53.50, −15.39 pts), followed by removing typed repair (−8.49), confirming each component is individually necessary.

Why it matters / caveats: The work reframes agent evaluation around trajectory-level faithfulness rather than answer accuracy alone, and its provenance guarantee only ensures repair cannot fabricate unsupported content, not that tool outputs themselves are correct. The Coding subset of EMMA saw a small negative change (−0.18 pp), and the added grounding/dispatch machinery relies on frozen backbones plus an external LLM judge (Gemini-3.1-Pro) for auditing, which introduces its own dependency and cost.

Problem: Training computer-use agents requires stateful, resettable, login-gated application environments, and once such synthetic environments can be generated in bulk, the bottleneck shifts from environment count to the behavioural depth, targeting, and quality of what's inside each one.

Method: Echoverse is a two-phase factory of GitHub Copilot SDK agents (builders, verifiers, triagers, fixers using Playwright and MCP database tools) that compiles hand-written seeds into FastAPI/SQLite/React applications verified against machine-checkable claims (Phase 1), then grows a task corpus grounded in live database state with database-derived verifiers (Phase 2). A co-evolution loop reads every graded rollout twice — as environment/task/verifier repairs and as SFT training signal — before training a Qwen3.5-9B model via verifier-filtered, rejection-sampled distillation (STaR-style) on trajectories generated by GPT-5.4, followed by an RL stage using group-relative policy gradient (GRPO-style, unnormalized advantage per Liu et al. 2025) with a composite reward combining the grounded database verifier and a dense per-step GPT-4.1-vision judge.

Key results:

• Across 12 training worlds and 14 evaluation splits, the 9B model improves from 36.5% to 67.1% average task success, within 14 points of teacher GPT-5.4 (80.7%), matching/beating it on ECHOMAIL, ECHOBANK, and both nested-filter splits.
• Shallow vs. deep environments: shallow-world training drops live-site accuracy below base (80.0% → 75.0% on Allrecipes) while deep-world training raises it (80.0% → 85.0%; 48.0% → 65.0% on Hugging Face).
• Repairing a single environment (ECHOSTAY) across two co-evolution rounds lifts the trained model from 16.2% to 38.5% solve rate; fixing a broken guest-count control alone recovered 15/24 blocked booking tasks (48% → 78%).
• RL on 5 worlds raises held-out judged score from 58.8% to 68.0% after ~2 epochs; capability-targeted training on datepicker/nested-filter worlds lifts held-out (never-trained) widget families from 34.0%→57.3% and 62.8%→84.8%, and transfers to Online-Mind2Web (29.5%→34.3%).

Why it matters / caveats: Demonstrates that environment depth and a repair-before-training discipline matter more than raw environment count, and that grounded (database-based) verifiers make the same synthetic worlds usable for both SFT and RL at scale, unlike live-web benchmarks which lack reset, throughput, and trustworthy reward. The released benchmark covers only 4 of the 12 worlds (ECHOSTAY, ECHOFORGE, and two capability worlds), and results rely on a single 9B base model (Qwen3.5-9B) and a single teacher (GPT-5.4), limiting generality claims.

Problem: Existing scalable generative models (autoregressive, diffusion, flow) avoid mode-blurring by factoring the generation procedure into many small steps, which makes them fundamentally not end-to-end and caps their "generative expressivity" regardless of how far parameters and data are scaled.

Method: The paper introduces Explorative Modeling, which instead factors the training loop: at each step the model draws K candidate generations (Forward XM) or K data targets (Reverse XM) and trains only on the closest match, implemented as a simple best-of-K for-loop (Algorithm 1/2). Exploration is added on top of existing recipes — Diffusion/Flow models, Jumpy generative models, and Masked Diffusion Language Models (MDLM) — to form hybrids like XDiffusion, XJumpy, and XMDLM, and is also used standalone to build end-to-end Explorative Policy (behavior cloning) and an Explorative World Model (goal-conditioned planning), evaluated against Diffusion Policy and Diffuser respectively; image experiments build on the Representation Autoencoder (RAE) and SiT recipes on ImageNet 256x256, video experiments use Something-Something V2.

Key results:

• Adding exploration to RAE reaches baseline performance with 6.2x less data and 4.1x fewer FLOPs; on SiT it gives 2.5x less data and up to 52% better FLOP efficiency, with gains more than doubling at 3x the compute.
• A Large model exploring 5 modes (XM-5) outperforms an XLarge model with 47% more parameters and no exploration.
• XRAE (XM-2) reaches near-state-of-the-art unguided FID of 1.43 and FDr6 of 3.91 on ImageNet 256x256, converging ~300x faster than the standard SiT recipe.
• Gains from exploration grow with scale: 7%→36% as data scales and 13%→23% as model size scales; video generation FVD improves from 37.5 to 30.0 (less overfitting) with more exploration.
• Explorative Policy matches/beats Diffusion Policy on Robomimic tasks (e.g., 96% vs 94% on Square) using 1 inference step versus 100; Explorative World Model matches or beats Diffuser on Maze2D with 16-256x fewer inference steps (e.g., 80x less compute on average).

Why it matters / caveats: The results position "exploration" as a third pretraining axis alongside parameters and data, one whose benefits appear to compound with scale rather than saturate, and show a concrete substitutability between factoring generation and factoring training that lets models become more end-to-end without sacrificing quality; however, Forward XM's per-candidate compute cost grows with K, and the largest experiments in the paper still use roughly four orders of magnitude less compute than foundation-model training runs, so extrapolation to that scale remains unverified.

Problem: Attention residuals let each Transformer sublayer retrieve a softmax-weighted mixture of all prior layer outputs, but the single shared query forces every feature subspace to read the depth history through one distribution, and this forced compromise worsens as model width grows.

Method: The authors introduce Multi-Head Attention Residuals (MHAR), which reshapes the single (d,) routing query of attention residuals (Kimi, 2025) into H per-subspace heads that each run an independent softmax over the depth history (source list of embeddings plus every attention/MLP output), making the read block-diagonal; the reshape adds zero parameters and negligible FLOPs, and H=1 recovers attention residuals exactly. They train decoder-only, Qwen3-style Transformers from scratch on FineWeb-Edu at 100M/350M/1B scales, compare against a standard Transformer, hyper-connections, and single-head attention residuals, set H equal to the number of KV heads as a hyperparameter-free default, and also build fused Triton forward/backward routing kernels and an identity-preserving delta-attention-residual conversion for grafting MHAR onto a pretrained Marin-8B model during mid-training on a ~1.9T-token anneal corpus (anneal_pt_v3).

Key results:

• MHAR improves validation loss over the standard Transformer at every scale: −0.049 (100M), −0.080 (350M), −0.063 (1B), versus single-head routing which degrades from −0.039 (100M) to +0.140 worse-than-baseline (1B), widening MHAR's advantage over single-head from 0.010 to 0.168.
• MHAR beats hyper-connections at all three scales (by 0.035, 0.050, 0.061), and yields relative perplexity gains of −10%/−15%/−19% on WikiText-2 at 100M/350M/1B with improved LAMBADA accuracy at every scale.
• Fused Triton kernels raise attention-residual training throughput from 0.2–0.5× to 0.55–0.88× of baseline throughput at near-baseline peak memory, and speed up the routing operation alone by 2.0–5.3× over torch.compile.
• On 8B mid-training via delta attention residuals, MHAR adds +3.2 GSM8K (paired McNemar p=0.004) and +3.1 GPQA (p=0.038) over a schedule-matched plain continual-pretraining control, with MMLU/MATH/code statistically unchanged.

Why it matters / caveats: Setting H = number of KV heads gives a near-free, tuning-free upgrade to attention-residual architectures, and a direct probe of trained queries confirms learned subspace disagreement (not source collinearity) as the mechanism driving the benefit at scale. Caveats: gains at 8B mid-training are modest and reasoning-sided only, the compute-equivalent gain estimate (1.3–1.5×) may be inflated by the flat-exponent fixed-step training ladder used, and broader downstream benchmark coverage is left to future work.

Problem: Deployed LLM agents increasingly persist long-term memory as agent-organized directories of markdown files (e.g., Anthropic's memory tool, Claude Code's memory folder), but nobody has tested whether an agent can actually keep such a filesystem store organized as it grows, or whether that organization ever pays off.

Method: The authors formalize filesystem memory around three roles sharing one store: a management agent that integrates incoming chunks and reorganizes the store (create/rewrite/merge/split/move/delete), a search agent that answers queries with citations, and (in the skill setting) a fixed execution agent whose trajectories become chunks to distill into skills. They instantiate this on conversational memory (LoCoMo, PersonaMem 32k/128k, REALTALK) and procedural memory (ALFWorld, 140 tasks), comparing six memory variants (closed-book, BM25 chunk retrieval, verbatim session dump, LLM-foldered sessions, LLM-reorganized store in "condense" vs. "preserve" versions, and a fully agent-curated store) across tool harnesses (six-op file tools mirroring Anthropic's memory tool, sandboxed shell, BM25-augmented search) and a model strength ladder (gpt-5.4-nano, gpt-5.4-mini, gpt-5.4 as management/search agents; gpt-4.1 vs. gpt-4.1-mini as execution agents), with gpt-5.4-mini (low reasoning effort) as the fixed LLM judge.

Key results:

• Organization's clearest unambiguous payoff is search cost: on the largest stores (PersonaMem), reorganized/curated stores cut per-query search cost roughly in half versus verbatim dumps (e.g., 128k tier: 1.6¢ vs. 3.9¢ per query).
• No memory shape wins correctness everywhere — the cheap "Foldered sessions" variant (LLM only sorts files into folders, zero content edits) ties or tops LoCoMo (86.1%), REALTALK (77.6%), and PersonaMem 128k (76.2%), while the fully agent-curated store scores worst on PersonaMem 32k (37.5% vs. 78.1% for the verbatim dump).
• A default reorganization pass silently condenses content unless a "keep every fact" rule is added; on REALTALK this condensation nearly halves correctness (41.2% vs. 77.6% for the preserving version).
• In the skill setting the best memory variant flips with execution-agent strength: a verbatim episode log wins under a strong executor (gpt-4.1: 87.1% task success) but a distilled "curated skills+mem with guidance synthesis" store wins under a weaker one (gpt-4.1-mini: 76.4% vs. 66.4% for the episode log).
• Store shape is mainly a signature of the management model, not of scale: on the same PersonaMem 128k conversation, file count under gpt-5.4-nano/mini/gpt-5.4 is 122/2/105 files (non-monotonic), while taxonomy adherence erodes as stores grow for all but the strongest management agent.

Why it matters / caveats: The paper turns "filesystem memory works because agents keep it organized" from an assumption into an open design space, showing organization reliably buys search economy but not better answers, and that curation quality is capability-gated (management-agent strength) and consumer-dependent (weak vs. strong execution agents want different memory forms) rather than a free win. Results are bounded to the horizons tested (up to ~271 chunks / 140 tasks) and rely on a single LLM judge for conversational correctness, with the authors noting judge noise of about ±1.3 points on LoCoMo.

Problem: Lossy verification schemes for speculative decoding promise extra speedup by relaxing strict distributional matching between draft and target models, but prior work evaluates them under curated settings that obscure how much they silently distort the output distribution and degrade generation quality.

Method: The authors give a unifying theoretical analysis showing lossy verification methods reduce to two paradigms: truncation-based verification (e.g., typical acceptance in Medusa, and SpecCascade, which accept draft tokens if they lie in an allowed set defined by min-p or η-sampling) and collaborative verification (e.g., Collaborative Decoding via Speculation (CoS) and lenience-based relaxation, which interpolate draft and target distributions). They derive closed-form decompositions of the induced output distributions, prove a Lemma characterizing truncation sampling's effect on Block Efficiency (BE), and a Proposition comparing per-token KL divergence under standard SD versus EAGLE-3 tree verification. They evaluate empirically on Qwen2.5-72B/0.5B (standard SD) and LLaMA-3.1-8B with its official EAGLE-3 draft, across MATH, MBPP+, INCLUDE, and BFCL benchmarks, measuring Block Efficiency, Decoding Speed, and task accuracy/Pass@1.

Key results:

• Truncation-based verification's apparent gains largely stem from the underlying truncation sampling itself, not the verification mechanism: under standard SD, matched-pair performance gaps stay small (at most 1.4 points), but under EAGLE-3 the gap widens sharply — SpecCascade's gap grows from -0.38 to -1.68 points and typical acceptance's from -0.32 to -6.32 points, reaching -8.8 points on INCLUDE.
• Typical acceptance falls below even the plain EAGLE-3 baseline on all four benchmarks, and SpecCascade falls below it on two, while true truncation sampling stays at or above baseline throughout.
• On the accuracy-gap case study (Figure 1), the gap between the true min-p baseline and SpecCascade widens from +0.38 percentage points on GSM8K to +6.67 points on AIME as task difficulty increases.
• Ablation on MBPP+ isolates the "overshoot ceiling" (capping draft probability at p(x)/ℓ) as the key mechanism behind lenience-based relaxation's success: applying only the ceiling matches lossless verification's Pass@1 (~75%) while still improving Block Efficiency, whereas the adaptive-interpolation component alone inherits CoS's severe quality/speed tradeoff (Pass@1 drops from 66% to 50% as λ decreases from 0.8 to 0.2).

Why it matters / caveats: The paper shows lossy speculative decoding methods should be benchmarked against distribution-matched (not default) baselines, and identifies "overshoot suppression" of draft over-confidence as the actionable design principle for safe acceleration. Limitations: experiments are restricted to Qwen2.5/Llama-3.1 model families with fixed draft-target pairings, reasoning/code/multilingual/tool-use benchmarks only (no open-ended dialogue), and reported speedups use hardware-agnostic Block Efficiency rather than measured wall-clock time.

Problem: RL search agents like Search-R1 that generate free-form natural-language queries suffer from "retrieval-equivalence collapse," where surface-diverse queries converge on nearly identical retrieved evidence, eroding the within-group return contrast that group-relative RL optimizers (e.g., GRPO) depend on.

Method: The authors propose Harness-G, which replaces free-form query generation with finite action selection over a programmatically induced paragraph–sentence–entity tripartite graph (built without a generative LLM, via sentence splitting, NER, and dense encoding): at each step the policy chooses Select(sentence), Lookup(entity), or Answer from a bounded, deduplicated, deterministic action menu. On top of this interface they introduce Structured Non-myopic Credit (SNC), combining (1) frontier-relative advantage — a frozen "answerer" scorer computes teacher-forced gold-answer probability gains for the chosen action versus previewed same-state alternatives — and (2) structured enablement credit, which backpropagates downstream gains along provenance edges to earlier "bridge" actions; the resulting token-level SNC advantage is combined additively with the GRPO outcome advantage. Training uses Qwen2.5-1.5B/3B-Instruct with GRPO (group size 8, batch size 128, 120 steps, ≤6 turns), evaluated on six QA benchmarks (2WikiMultiHopQA, HotpotQA, MuSiQue, NQ, PopQA, TriviaQA) against baselines including Search-R1, IGPO, R1-Searcher, and Graph-R1.

Key results:

• Highest average F1 at both model scales: outperforms the strongest baseline Graph-R1 by 10.74 F1 points at 1.5B (50.83 vs. 40.09) and 3.98 points at 3B (55.24 vs. 51.26).
• On multi-hop datasets at 3B, F1 gains over Graph-R1 reach 7.97 (2Wiki), 9.12 (HotpotQA), and 5.95 (MuSiQue) points.
• Ablations: switching from free-query to menu interface alone adds >17 F1 points (>35 points on MuSiQue under outcome-only reward); removing both SNC terms (λ=0) drops F1 by 3.08–4.55 points across multi-hop datasets.
• SNC adds only 9–11% wall-clock overhead per step with zero extra rollouts (~170 GPU·h for a full 3B run); graph construction costs $0 in API calls vs. $2.81–$4.14 for LLM-extracted graph baselines.
• Under train-one/eval-all O.O.D. cross-validation, Harness-G wins 21 of 30 dataset pairs and raises mean O.O.D. F1 from 44.10 to 47.38 versus Graph-R1.

Why it matters / caveats: The paper argues action-space/interface design is a complementary axis to reward design for stabilizing and improving RL search-agent training, showing gains are especially pronounced for smaller models (1.5B) with limited capacity. Harness-G remains text-only, and the authors note extending structured actions and SNC to multimodal evidence as future work; it also trails Graph-R1 by 1.84 F1 on the single-hop NQ dataset.

Problem: Token compression methods for omnimodal LLMs (OmniLLMs) typically let one modality (e.g., audio) guide pruning of the other (e.g., video), but audio and video relevance for the same query often peak at different moments, causing this unidirectional guidance to discard answer-critical tokens under aggressive compression.

Method: The authors propose OmniScope, a training-free, modality-decoupled compression framework built on Qwen2.5-Omni that treats the query as a shared semantic anchor but estimates visual and audio salience independently — visual relevance via CLIP ViT-L/14@336px cosine similarity between frames and the query, and audio relevance via a gated cross-modal similarity between the model's own audio encoder embeddings and query token embeddings (using embedding-norm gating plus decaying score propagation across seconds). Visual tokens are pruned with Anchor-Delta Spatio-Temporal Compression (AD-STC), which alternates "anchor" frames (DPC-KNN density-peak selection for spatially distinctive tokens) and "delta" frames (adaptive joint spatio-temporal irreplaceability score or pure temporal-difference score depending on compression ratio), while audio tokens are compressed via per-second bipartite soft matching (token merging) rather than dropping. Evaluated on four audio-video benchmarks (WorldSense, DailyOmni, OmniVideoBench, Video-MME) at Qwen2.5-Omni 3B and 7B scales, compared against Random pruning, FastV (A&V), and OmniZip.

Key results:

• At 25% token retention on the 7B model, OmniScope loses only 0.35 points average accuracy vs. the Full Tokens baseline (51.65 vs 51.35), compared to a 1.55-point drop for OmniZip and 1.57 for Random.
• Delivers up to 3.53x prefill speedup (6299ms → 1784ms on 7B at 25% retention) and over 15% GPU memory reduction (28.31GB → 24.00GB on 7B), plus ~2.1x end-to-end speedup (10.965s → 5.284s).
• At 45% retention, accuracy is nearly lossless and even slightly exceeds the uncompressed baseline in some settings (+0.30 on 7B, +0.03 on 3B).
• Ablations show query-aware per-modality budget allocation beats uniform/cross-guided allocation by up to 1.65 points, and audio token merging outperforms uniform sampling, average pooling, random drop, and energy-based filtering (best variant 52.70 vs. 49.95 avg for energy-based).

Why it matters / caveats: The paper quantifies that ~78.3% of query-video pairs show only weak audio-visual salience correlation, undermining a common design assumption in prior omnimodal compression (e.g., OmniZip) and suggesting decoupled, query-anchored salience estimation is a more robust default; the approach is training-free and reuses existing encoder outputs (no extra memory overhead like OmniZip's full attention-matrix extraction), though it relies on an external CLIP scorer for vision, adding a fixed latency cost that is only amortized in longer-generation scenarios.

Problem: It is unclear whether Deep Research agents can resist apparently credible but factually false documents that enter their long-horizon search-and-synthesis workflow, ultimately adopting false conclusions in their final reports.

Method: The authors build MisKnow-Agent, a framework that constructs a per-task misleading-knowledge blueprint (canonical false conclusion, an institution pool stratified by authority level High/Medium/Low, and an adoption-evaluation instruction), then generates controlled documents across four presentation styles (Paper, News, Blog, Post) using DeepSeek-V4 Pro as the generator; candidates go through search-guided refinement and cross-model verification by five verifier LLMs (GLM-5, Kimi 2.6, DeepSeek-V4 Pro, Qwen3.5-397B, Intern-S1-Pro) plus manual screening. Applied to 100 tasks from DeepResearch Bench, this yields 5,933 quality-controlled misleading documents. They evaluate false-conclusion adoption rate (FCAR) on DeerFlow and WebThinker (each paired with DeepSeek-V4 Pro, Qwen3.5-397B, Intern-S1-Pro) and on closed-source Gemini Deep Research, and test two mitigations: a pre-research verification-enhanced prompt and a post-research claim-by-claim refinement agent.

Key results:

• Injecting a single misleading document raises mean FCAR from 0% (no-injection control) to 54.7% across six framework–LLM configurations.
• Exposure timing matters far more than search rank: FCAR is 40.5% (cold start) and 44.2% (mid-research) but jumps to 85.5% when misleading knowledge is introduced right before final synthesis, while search-result rank changes FCAR by at most ~1.7 percentage points.
• Source cues matter: high-authority sources yield 61.0% FCAR vs. 46.2% for low-authority (14.8-point gap), and Paper-style presentation yields 61.0% vs. 37.5% for Post-style (23.5-point gap) — the largest single-factor effect found.
• Framework/model choice matters: DeerFlow averages 67.0% FCAR vs. 55.0% for WebThinker; Gemini Deep Research shows similar patterns (54% high-authority, 27%→54% FCAR from 1 to 3 injected documents).
• Defenses reduce but do not eliminate adoption: pre-research, post-research, and combined defenses yield FCAR ranges of 37–57%, 20–58%, and 15–62% respectively (vs. 60–76% baseline), and combining defenses is not consistently better than either alone.

Why it matters / caveats: The findings show a previously under-characterized reliability failure mode specific to long-horizon Deep Research workflows (as opposed to short-form RAG), where even a single well-presented misleading document — verified as false by every automated checker — can still be adopted as a conclusion, especially if introduced late in the pipeline; the study is limited to English/Chinese web-style injected documents on DeepResearch Bench tasks and a small set of open- and closed-source systems, so generalization to other agents/domains is untested.

Problem: There is no existing framework for simulating and studying how multiple persona-based LLM agents negotiate a shared itinerary in group travel planning.

Method: AI Tour Meeting instantiates multiple LLM participants, each configured with a persona (name, background, personality, preferences, goal, role, tone, explanation style) that role-plays via a system prompt, and runs a turn-based conversation-phase/voting-phase loop (Algorithm 1) with actions such as search, ask, reflect, propose, satisfied/pass, and accept/reject/score. The framework offers five turn rules (round robin, inviting, facilitating, random, parallel) with balancing/volunteer options and five voting rules (majority, unanimous, single decider, most pleasure, least misery), plus a Python API, GUI, and monitoring dashboard. Local models are served with vLLM; validation uses Qwen3.5-2B/4B/9B, gpt-oss-20b, and gpt-5.4-mini, with gpt-5.4-mini also used to synthesize meeting scenarios.

Key results:

• All tested models except Qwen3.5-2B achieved 100% meeting completion rate; Qwen3.5-2B only reached 82% completion and 58% consensus with a 14.2% constraint-error rate, versus ≤3.1% for larger models.
• Under increasing preference conflict (aligned → mixed → conflicting, 50 meetings each, Qwen3.5-9B), consensus rate dropped from 100% to 94%, average satisfaction fell from 8.93 to 7.13 (1–10 scale), and the "victim rate" (satisfaction ≤4) rose from 0.7% to 11.3%, with turns growing from 10.3 to 25.9.
• Speaking-order analysis (Latin-square rotation) showed no significant acceptance-rate difference by position under aligned preferences, but under conflicting preferences the first speaker's proposals were accepted only ~32% of the time versus ~46–48% for later speakers.
• Scaling to 5 and 10 participants maintained 100% completion (96% consensus at M=10), with turns growing faster than participant count (12.8 → 25.4 → 68.6 turns for M=3,5,10).

Why it matters / caveats: The framework reproduces human-like deliberation dynamics (more conflict → longer discussions, lower satisfaction, order effects favoring later speakers) and is proposed as a research tool for studying LLM agent behavior and evaluating group-travel recommender systems; the code is restricted to research use under NTT's proprietary license, and results depend on a synthetic-scenario setup generated by GPT-5.4-mini rather than real user data.

Problem: Sparse mixture-of-experts (MoE) models are often assumed to benefit from "geometric complementarity" (co-selected experts covering disjoint representation directions), but prior evidence conflates route coherence, candidate quality, and candidate-by-context interaction, so it is unclear whether this assumption actually holds or explains multi-expert value.

Method: The authors introduce an Expert Subspace Separation Index (ESSI, normalizing between-expert subspace distance by local within-expert tangent dispersion via Grassmann chordal distance) and a prefix-controlled 2×2 factorial design that crosses the selected candidate expert vs. the strongest unselected rival with the actual vs. a matched alternative routing context, yielding a difference-in-differences interaction term D. They apply this alongside frozen-route NLL interventions and a controlled Top-1-vs-Top-2 training study to six open MoE models: OLMoE-1B-7B, Mixtral-8x7B-v0.1, DeepSeek-MoE-16B, Qwen3-30B-A3B-Base, Gemma4-26B-A4B, and Qwen3.6-35B-A3B, with the factorial/functional protocols run on OLMoE, Mixtral, and DeepSeek.

Key results:

• ESSI stays near 1 across 18 model-layer cells (range 0.776–1.060, median 0.969), showing expert subspaces overlap substantially rather than forming a hard directional partition.
• Actual routes still fit tokens better than matched alternatives: the alternative-to-actual residual ratio exceeds 1 in all 18 cells (range 1.012–1.901, median 1.283).
• Across all 39 factorial cells (OLMoE, Mixtral, DeepSeek), the selected candidate explains more residual variance than the rival in every cell (Aactual > 0), but the interaction D is negative in all 39 cells with every 95% CI below zero (e.g., OLMoE/Mixtral macro D = −0.05548 [−0.05780, −0.05340]; DeepSeek macro D = −0.05770 [−0.06165, −0.05405]) — the actual context narrows rather than amplifies the candidate's advantage.
• Despite this geometric narrowing, later experts remain functionally useful: adding a later expert reduces next-token NLL in 24 of 39 frozen-route comparisons (the other 15 are statistically inconclusive), and a controlled Top-2 vs. Top-1 training study (matched active compute, parameters differing by <0.04%) favors Top-2 in all 3 seeds (Δ validation loss = 0.1016 ± 0.0025).

Why it matters / caveats: The findings challenge the common design/pruning heuristic that geometric subspace overlap implies redundancy — "coherent overlap" shows routing selects strong, useful experts from a shared geometric neighborhood without needing disjoint linear coverage, so output-level (functional) interventions are necessary before using input-subspace similarity as a proxy for redundancy or pruning value. Caveats: conclusions rest on a linear, rank-128 router-input metric, cover only three architectures for the factorial/functional analyses, and the training study is a single small-scale matched-compute setting.

Problem: Vision-language models (VLMs) struggle to answer questions over long visual contexts (large image collections or hour-long videos) because attention-based retrieval fails to distinguish query-relevant frames from distractors, and processing all visual tokens at once is infeasible under GPU memory limits.

Method: The authors first show that scoring retrieval via cosine similarity between a query and mean value vectors ¯v (rather than query-key attention) gives a stronger relevance signal, then introduce ReToken, a single learnable embedding appended to the question and trained with a class-balanced binary cross-entropy retrieval loss against ground-truth frame relevance labels, using a lightweight final-layer projection matrix (Wr) while keeping the VLM (Qwen3VL-8B, InternVL3.5-8B) frozen by default; an optional partial-tuning variant additionally fine-tunes the first 1-3 VLM layers with a combined retrieval + next-token generation loss. Training uses only the MIRAGE multi-image QA dataset (combining RetVQA, SlideVQA, WebQA, and synthetic LLaVA-150K data); inference uses a two-pass retrieve-then-answer pipeline over a persistent visual KV cache, following ReKV's chunked sliding-window encoding for long videos.

Key results:

• On Visual Haystacks (C=50 distractors), ReToken improves Qwen3VL-8B by 13.4 points and InternVL3.5-8B by 12.4 points accuracy (>20% relative gain).
• Zero-shot transfer to long video: +8.0 points on LVBench with Qwen3VL-8B (average video length over an hour), and +6.8 recall points at K=1 on QAEgo4DTest-MC vs. uniform sampling.
• Value-space retrieval beats query-key attention: recall@1 of 78.0% (Qwen3VL) and 83.8% (InternVL3.5) vs. 65.7%/78.8% for query-key scoring in a controlled two-image setting; attention-based retrieval averages only 5.1% recall@1 across layers on QAEgo4DTest-MC vs. ReToken's 8.5%.
• Adds only ~0.4 seconds per-question latency overhead on a single H100 while both training and long-video inference fit on one H100 GPU.

Why it matters / caveats: ReToken adds just one token and one projection matrix yet substantially closes the gap to an oracle "ground-truth KV cache" upper bound, offering a cheap, model-agnostic retrieval mechanism that generalizes from image QA training to video without video-specific supervision. Limitations include reliance on a two-pass forward pass with expanded early-layer attention (added memory/time cost), training data limited to multi-image QA (no video training data), frame-independent scoring that underperforms on tasks requiring cross-frame/temporal reasoning (e.g., it hurts on LVBench summarization questions, -5.2 points), and the InternVL3.5 partial-training budget (1 epoch) leaving some results ceiling-limited.

Problem: Existing bias-mitigation methods for LLMs (data filtering, fine-tuning/alignment, embedding-space post-processing, or SAE-based interpretability) either act on all weights indiscriminately, risking degradation of unrelated capabilities, or require costly additional training, so there is no lightweight, training-free way to locate exactly which weights encode demographic bias.

Method: The authors propose Fairness Pruning, implemented in the open-source OptiPFair library, which builds minimally contrastive prompt pairs (identical except for one demographic attribute, tokenization-matched) covering five categories — Age, Gender, PhysicalAppearance, RaceEthnicity, Religion — via two new HuggingFace datasets (fairness-pruning-pairs-en: 14 attribute pairs/70 prompt pairs; fairness-pruning-pairs-es: 20 attribute pairs/100 prompt pairs, aligned with BBQ and EsBBQ). It registers a forward hook at the input of down_proj in SwiGLU MLP blocks to compute a per-neuron BiasScore (mean absolute difference in activations between paired prompts), optionally combined with a PPM (Peak-to-Peak Magnitude) structural-importance term into a FairnessPruningScore, then causally validates neurons by zeroing rows/columns of gate_proj, up_proj, and down_proj. Evaluation uses lm-evaluation-harness (WikiText, MMLU, ARC-Challenge, HellaSwag EN/ES) and BBQ/EsBBQ bias scores on Llama-3.2-1B, Llama-3.2-3B, and Salamandra-2B.

Key results:

• Bias signal at down_proj_input consistently peaks in the model's final layer across all 30 model/language/category combinations tested (L15 in Llama-3.2-1B, L27 in Llama-3.2-3B, L23 in Salamandra-2B), with 54-94% of the top-200 biased neurons per category concentrated there.
• Zeroing at most 40 neurons in Llama-3.2-1B (0.031% of the 131,072 total MLP width) yields 99.49% mean retention (97.93%-101.34% range) across five general-capability benchmarks.
• Category circuits are largely distinct: pairwise Jaccard overlap at Top-1% is only 0.204-0.324 in Llama models, meaning 66-88% of candidate neurons are unique to their category; Gender shows the lowest overlap with other categories.
• Effects on aggregate bias are non-monotonic and often sign-inverting rather than a clean reduction: e.g., Religion·Top-20 on Llama-3.2-3B shifts bias_ambig from 6.00% to 1.50% (-4.50pp) while bias_disambig flips sign from +3.84% to -1.67%; in 6 of 8 Llama-3.2-1B experiments the ambiguous and disambiguated BBQ scores move in opposite directions.

Why it matters / caveats: The results demonstrate demographic-bias processing and general model capabilities sit on dissociable, surgically prunable circuits, but because BiasScore is unsigned it mixes bias-amplifying and bias-suppressing neurons, so zeroing produces bidirectional destabilization rather than reliable mitigation — the authors frame this as groundwork for a future signed/directional score using asymmetric corpora. The study is limited to explicit-attribute benchmarks (BBQ/EsBBQ), models under 4B parameters, and MLP-only analysis (attention excluded).

Problem: Large self-supervised speech models (e.g., WavLM, data2vec) achieve strong speech emotion recognition (SER) but are too large (~94M parameters) for edge deployment, and existing multi-teacher knowledge distillation methods use fixed teacher weights and logit-only transfer that ignore per-batch teacher reliability shifts and inter-sample relational structure.

Method: AMRD distills two frozen SSL teachers (data2vec Large and WavLM Base+) into lightweight CNN students (LightSERNet+, MobileNetV3, ResNet18, EfficientNet-B0) using two components: (1) SVM-Based Dynamic Teacher Weighting, which fits a one-class SVM on each teacher's per-batch pairwise cosine similarity matrix of logits to produce a quality score, softmax-aggregated into per-batch teacher weights for soft-target KL distillation; (2) Relational Similarity Matrix Distillation (RSMD), which z-score- and l2-normalizes teacher/student features into Pearson-correlation-based relational similarity matrices (RSMs) and minimizes MSE between student and teacher RSMs (uniformly averaged across teachers). The total loss combines cross-entropy, temperature-scaled KD-KL divergence with SVM-weighted soft targets, and RSMD. Evaluated on IEMOCAP (leave-one-session-out 5-fold) and CREMA-D (5-fold CV) against baselines KD, DKD, CKD, and RKD.

Key results:

• On IEMOCAP, AMRD achieves the highest WA and UA across all four student architectures, with the largest gain on MobileNetV3 (47.21% UA, +2.9 points over the best single-teacher baseline, WavLM+DKD at 44.27%).
• On CREMA-D, AMRD leads on 3 of 4 students; LightSERNet+ reaches 56.37% UA, +1.7 points over the best single-teacher result (D2V+KD, 54.65%).
• AMRD matches or exceeds the best single-teacher baseline in 7 of 8 student-dataset settings.
• The smallest student, LightSERNet+ (0.78M parameters, >120x smaller than WavLM Base+'s ~94M), reaches 52.30% UA on IEMOCAP and 56.37% UA on CREMA-D, improving over no-distillation baselines by 2.8 and 3.4 points respectively.
• Ablations show SVM weighting and RSMD each contribute independently (e.g., SVM weighting alone boosts EfficientNet-B0 UA on CREMA-D by 1.8 points, 39.75%→41.51%), and combining both gives the best result in 3 of 4 tested settings.

Why it matters / caveats: Demonstrates that adaptive, training-only distillation mechanisms (no inference-time cost) can enable on-device SER under extreme model compression. Limitations acknowledged by the authors include using only two teachers (scaling to more teachers is untested), audio-only modality (no facial/linguistic fusion), and in-corpus-only evaluation (no cross-corpus/cross-language generalization tested); one student-dataset pair (MobileNetV3 on CREMA-D) showed no benefit from distillation at all.

Problem: Autonomous vehicle (AV) safety testing needs richer behavioral models of dangerous pedestrians than isolated action labels like "crossing" or "retreating" can provide, and the authors' original 12-archetype taxonomy still left several recurring, distinct behavior patterns unmodeled.

Method: Building on prior work (PedAnalyze ontology and the original Pedestrian Archetypes paper's 12 types), the authors continue annotating YouTube dash-cam footage of risky, near-miss, and collision pedestrian-vehicle interactions using a four-step process: annotate behaviors with the PedAnalyze ontology (crossing, retreating, running, pausing, ignoring traffic, hesitating, etc.), review annotated scenarios to find recurring behavior combinations not covered by the existing 12 archetypes, select representative video evidence and extract key frames, and derive essential vs. optional behavior tags per archetype from a data-driven frequency threshold (essential if present in ≥40% of examples, optional if 10-39%).

Key results:

• Introduces 7 new pedestrian archetypes: Con Artist, Foreigner, Influencer, Protester, Confronted, Pseudo Pedestrian, and Street Vendor, extending the taxonomy from 12 to 19 total archetypes.
• Each new archetype is defined with essential/optional behavior tags (e.g., Con Artist: essential = Collision, Looking, Fall, Not-Cross, Run into traffic, Cross without crosswalk) and illustrated with real dash-cam video frame sequences from cited YouTube sources.
• Essential/optional behavior classification derived from thresholds of ≥40% (essential) and 10-39% (optional) occurrence frequency across annotated video examples.

Why it matters / caveats: The expanded taxonomy gives AV developers a broader, more communicable set of pedestrian behavior models for annotation, simulation, and safety-test scenario generation, targeting rare but high-risk pedestrian behaviors. The work is a preprint based on qualitative annotation of a small set of YouTube videos rather than a large, statistically validated dataset, so archetype prevalence and generalizability are not rigorously established.