AI papers — 2026-08-12

Problem: Existing Agentic AI paradigms (Digital Agents organized around software-state transformations, Embodied Agents around physical-state transformations) do not make the evolving state and agency of an individual person the primary object of modeling, intervention, and evaluation, leaving personal assistants, health agents, and companions fragmented across memory, personalization, sensing, and domain-specific support.

Method: The paper introduces "Combodied Agents," a human-centric paradigm formalized as a closed loop of event-based multimodal perception (producing governed "event-evidence records" from language, speech, vision, physiology, motion, social, environmental, and clinical/institutional data), longitudinal and correctable memory (episodic, semantic, trajectory, goal, relationship, intervention-response, and user-control memory components), a Personal World Model (PWM) that predicts calibrated distributions over future human states, events, and outcomes under alternative decisions/interventions via a formalized latent-state and action-conditioned dynamics framework (Eqs. 1-9), and an admissible Intervention Policy (Pareto-optimal action selection under consent, uncertainty, safety, reversibility, and escalation constraints) with an action taxonomy (Inform, Remind, Recommend, Coach, Nudge, Reflect, Coordinate, Protect, Escalate, Execute). It also proposes a three-stage cloud-to-edge deployment architecture and a scenario-centered evaluation matrix with agency-preservation metrics, referencing existing resources such as LaMP, LongLaMP, LongMemEval, MemoryBank, iOSWorld, PHIA, PH-LLM, and Health-LLM as partial but insufficient benchmarks.

Key results: Not stated — this is a conceptual/framework paper with no reported experiments, benchmark scores, accuracy figures, or other quantitative results. It instead identifies missing benchmarks (Table 8) across health/wellness, emotional support, learning, eldercare, personal life management, workplace wellbeing, and personal guardian/safety scenarios, and proposes a taxonomy/evaluation matrix (Table 7) rather than reporting numeric outcomes.

Why it matters / caveats: The paper argues that optimizing purely for task automation and engagement risks eroding human understanding, judgment, and autonomy, and proposes redirecting Agentic AI design and evaluation toward sustained human benefit and preserved agency rather than maximal automation; a key caveat it states itself is that no unified public benchmark or empirical validation currently exists for the full Combodied Agent loop, and that a complete, high-fidelity "Human Digital Twin" of a person remains unattainable.

Problem: The paper addresses the lack of a unifying survey and framework for co-evolution in agentic systems — cases where multiple agents and their environment jointly and continually reshape each other's evolution, as opposed to single-entity self-evolution bounded by a static learning context.

Method: The paper is a survey that proposes a progressive three-stage taxonomy of co-evolution — Stage 1 Agent–Agent Co-Evolution (adversarial, collaborative, and organizational adaptation among agents in a fixed environment), Stage 2 Agent–Environment Co-Evolution (extending adaptation to tasks, feedback, and interaction spaces), and Stage 3 Meta Co-Evolution (making the evolution mechanism Ω itself evolvable via a self-generated revision process Γ_t) — and formalizes agentic systems as S=(A,E) with agents a_i=(m_i,h_i) (model backbone plus harness), defining co-evolution via coupled state-transition equations. It reviews and categorizes existing methods (e.g., GAN, RARL, AlphaStar, SPAG, ACE-Safety, CORY, POET, PromptBreeder, Gödel Agent, RQGM, among many others) under this taxonomy rather than training a new model.

Key results: Not stated (this is a literature survey with a conceptual taxonomy and qualitative cross-paper synthesis rather than original experiments; Figure 4 aggregates "cross-paper evidence for the effect, consistency, and convergence of co-evolution" but no specific accuracy/benchmark numbers are given).

Why it matters / caveats: The taxonomy offers a unified framework for understanding how agentic systems can progressively shed human-engineered constraints toward open-ended improvement, and identifies open challenges in dynamic evaluation, scaling co-evolution across components, and safety/governance of increasingly autonomous evolutionary processes; the authors note Meta Co-Evolution (Stage 3) is still nascent (only RQGM meets their full definition) and that safety/governance issues are identified as desiderata without concrete safeguards or protocols.

Problem: Existing 4D generation methods either reconstruct 4D scenes from decoded RGB video (introducing distribution mismatch and error propagation) or bind 4D prediction directly to a specific video generator (requiring retraining when the generator or conditioning regime changes), leaving no reusable interface that bypasses RGB while working across multiple video generators.

Method: The paper proposes Latent-to-4D, which treats the final denoised VAE latent of a video diffusion transformer (DiT) as a reusable interface to explicit 4D prediction, bypassing RGB decoding entirely. Its core component, the Latent-to-4D Alignment and Refinement (L4AR) network, uses a fixed trilinear resampling operator plus a learned 3D convolution to align the video VAE latent to the token grid of a pretrained 4D reconstructor (4RC), then refines it via alternating frame-wise and global spatiotemporal self-attention (31-block hierarchy, rank-16 LoRA adaptation) before a 4D decoder predicts per-frame cameras and world-space point maps. Training uses a frozen Wan VAE, freezes the video DiTs, and supervises only the alignment module, LoRA updates, and prediction heads on 1,143 clips from six reconstruction datasets; evaluation uses two new benchmarks, Text4D-200 and I4D-200.

Key results:

• On Text4D-200, beats matched same-latent Wan+4RC cascades by 2.88–3.45 points in DINO-F1 (e.g., Ours (Wan2.1-14B) 57.01 vs. Wan2.1-14B+4RC 53.56).
• On I4D-200, scores highest on all metrics, with a 5.81-point DINO-F1 gain over Wan2.2-I2V-A14B+4RC (61.60 vs. 55.79), outperforming π3, Any4D, and native 4DNeX (28.33).
• A single checkpoint (trained on ~1,143 clips) transfers unchanged across two text-to-video DiTs and one image-to-video DiT sharing the same VAE.
• Human evaluation (50 raters): preference for Ours exceeds 50% on every criterion, e.g. Text-to-4D overall quality 65.7%, Image-to-4D overall quality 70.6%.
• DiT-residual sensitivity test: at ρ=0.6, point-map drift is 0.0053/0.0047 (Ours) vs. 0.3827/0.3160 (RGB-decode+4RC baseline).
• Component ablation shows removing 3D Conv, frame-wise attention, or global attention all substantially degrade results.

Why it matters / caveats: Supports shared video-model VAE latents as an effective, reusable interface for 4D generation that generalizes across compatible DiTs without retraining; evidence is limited to models sharing a common VAE convention, and the DINO-based evaluation is an appearance-dependent proxy rather than a metric-accuracy measurement.

Problem: Existing articulated object reconstruction methods require explicitly observed motion across multiple articulation states, so the paper addresses reconstructing part-level geometry and joint parameters from a single rest-state (closed) observation alone.

Method: The paper introduces Rest2Art, a pipeline that uses an explicit mesh as an intermediate representation to iteratively co-refine part hierarchies (from a VLM, GPT-5.2/Qwen3-VL) and segmentation masks (SAM3), fusing multi-view evidence onto the mesh via confidence-weighted accumulation and graph-based label propagation; joint parameters are then estimated by generating articulation videos with Wan2.2 plus a VBVR LoRA, extracting 2D trajectories with CoTracker3, and fitting rigid-body (revolute/prismatic) joint models via reprojection-error minimization constrained by mesh geometry. Evaluation uses the Articulated Containers Dataset (ACD, from HSSD and ABO) and the MultiScan dataset, with meshes reconstructed via 2DGS.

Key results:

• On ACD-HSSD: axis angular error 11.35°, axis position error 0.85 dm, joint type accuracy 74.49%, vs. Articulate AnyMesh (11.35°/1.31 dm/67.12%) and ArtGS (53.12°/1.37 dm/57.71%).
• On ACD-ABO: 4.78° angular error, 0.20 dm position error, 73.38% type accuracy.
• On MultiScan (real-world): 48.94% coverage, 17.23° angular error, 0.26 dm position error, 65.78% type accuracy, outperforming REArtGS and Articulate AnyMesh on angular error, position error, and type accuracy.
• Co-refinement ablation (GPT-5.2+SAM3): hierarchy tree accuracy improves from 62.3% to 72.5%.
• Video-generation design ablation reduces hallucination rate from 23.1% to 11.5%.
• Mesh-backend ablation: 2DGS achieves segmentation IoU 0.97/mAP 0.99, best among 2DGS/SVRaster/3DGS.

Why it matters / caveats: Shows cross-validating noisy pretrained-model outputs against explicit mesh geometry can yield accurate, physically plausible articulation without observed motion, enabling digital-twin creation from single closed-state images/scans; the authors note future work is needed for more complex kinematic structures and improved physical plausibility of the underlying video generation models.

Problem: Directly optimizing static pretrained-feature-space Fréchet distance (FD) objectives for generator post-training causes "Fréchet hacking," where the trained metric keeps improving while visual quality and Fréchet alignment in other, unused feature spaces stagnate or worsen.

Method: The paper proposes Adversarial Fréchet Distance (AdvFD), which augments a static FD-Loss objective (over frozen SigLIP, MAE, Inception encoders) with a learnable adversarial representation ψω trained alternately via a G-step (generator minimizes static + adversarial discrepancy) and D-step (representation maximizes it). Real-feature whitening (centering/covariance-normalizing features relative to real statistics, provably affine-reparameterization invariant) prevents trivial feature-norm inflation from destabilizing the min-max game. Evaluated for one-step class-conditional ImageNet-1K 256×256 generation on JiT and pixel MeanFlow (pMF) backbones at B/L/H scales.

Key results:

• JiT-B: AdvFD improves FID 1.00→0.79, FD-r6 5.53→3.92, FD-r3 8.45→6.03 (21–29% relative gains over FD-loss).
• JiT-L: FD-r6 3.24→2.01, FD-r3 5.46→3.20 (38–41% gains).
• pMF-H: FID/FD-r6/FD-r3 0.77/1.89/2.69 → 0.74/1.74/2.50.
• A universal adversarial perturbation optimized only for Inception FID reduced FID from 3.31 to 2.56 while introducing visible artifacts, illustrating static-metric exploitability.
• Whitening-location ablation: no whitening → FID 10.69 (severe degradation) vs. AdvFD's 0.79.
• JiT-H AdvFD (953M params) reaches FID 0.72 vs. real-image reference FID 1.68.

Why it matters / caveats: Shows a dynamically learned, whitened comparison space can mitigate reward/objective hacking in distribution-level generative training without added sampling cost across multiple backbones and scales; the paper does not report the added training-time compute/memory overhead beyond noting some encoders required LoRA due to memory constraints.

Problem: Existing self-improving coding agents (e.g., DGM, HGM) condition each self-modification step on only a single agent's single failure trajectory, ignoring richer comparative signals available in the growing archive of past agent variants and their evaluation trajectories.

Method: The paper introduces Mendel Gödel Machine (MGM), which extends the HGM tree-search self-improvement framework by splitting self-modification into three sub-operators: clonal mutation (single-trajectory edit), reaction-norm mutation (edits using an agent's trajectories across multiple tasks), and cross-lineage hybridization (edits using a reference agent's trajectory on the same task from a different lineage). MGM retains HGM's Thompson-sampling policy and adds a global failed-task pool; validated theoretically (additive fitness landscape model) and empirically on Qwen3.6-35B-A3B and DeepSeek-V4 backbones using SWE-bench Verified/Pro/Multilingual and Polyglot.

Key results:

• On Polyglot-60, MGM raises accuracy 50.8%→93.2%, vs. HGM's 77.9%.
• On SWE-bench Verified-60, MGM raises accuracy 68.3%→78.3%, vs. HGM's 73.3%.
• On the full 225-task Polyglot benchmark, MGM (Qwen3.6-35B-A3B) reaches 93.3%, surpassing GPT-5 with ~117× fewer parameters.
• Cross-benchmark generalization: MGM reaches 26.7% on SWE-bench Pro (+10.0 pts) and 55.0% on Multilingual (+13.3 pts); HGM instead drops on Pro (−3.4 pts).
• Ablation on Polyglot-60: full MGM 93.2%; without reaction-norm mutation drops to 79.7%; without cross-lineage hybridization drops to 74.6%.

Why it matters / caveats: Shows richer comparative conditioning on archived trajectories yields agent scaffolds that generalize better across benchmarks and backbone LLMs without extra evaluation budget or token cost; a theoretical caveat is that the diagnostic-advantage proof is under an idealized additive fitness landscape, validated via simulation rather than direct measurement in the real experiments.

Problem: Existing agent benchmarks evaluate short, single-turn, self-contained requests in static environments, so no benchmark measures whether an LLM life agent can be proactive, operate in a dynamic living world that changes on its own, and remain coherent across a long, multi-week horizon.

Method: VibeLifeBench comprises 200 scripted long-horizon tasks (20 per domain across 10 everyday-life domains) run in a simulated world of 22 mock service backends exposing 288 tool interfaces, driven by 7,453 scripted events. Each task is graded by 12,261 fine-grained, weighted, deterministic checks reading only observable artifacts. Evaluation runs on the TERRARIUM multi-turn agent infrastructure with the openclaw harness, testing seven frontier models (Claude Opus 5, GPT-5.5, Gemini 3.5 Flash, Claude Opus 4.8, GLM-5.2, Kimi-K2.6, DeepSeek-V4-Pro) three times per task.

Key results:

• Best model (Claude Opus 5): avg@3 = 32.5; worst (DeepSeek-V4-Pro): avg@3 = 21.1. All seven models fall in a narrow 21–33 band.
• Per-tier pass rates are lowest on highest-weight cross-stage (18.2–31.0%) and final (23.7–32.8%) checks vs. per-stage (34.3–44.9%).
• Proactivity (16.0–33.6%) and Persistence/bookkeeping (18.9–28.0%) are the weakest capability axes.
• Per-stage pass rate declines 10–15 points from the first third to the last third of the timeline for every model.
• Per-domain avg@3 for the top model ranges from 21.8 (team building) to 51.1 (shopping).

Why it matters / caveats: Shows a large, consistent gap between frontier LLM agents' passive tool-use competence and the proactive, persistent behavior needed for real-life long-horizon assistance, with no model competent across all ten domains; check-to-capability-axis assignment is done by keyword matching (indicative, not exact).

Problem: Existing omni-modal dialogue systems can understand multimodal input and produce spoken replies, but their responses remain visually disembodied, lacking a coordinated, dialogue-derived speaking video avatar.

Method: Ex-Omni-2D uses a Qwen3-8B dialogue backbone (with Qwen3-VL-2B vision tower) that outputs a structured Visual Thought Plan (VTP) plus response text, then a Qwen3-TTS-0.6B-initialized Speech Generator produces native 16-codebook acoustic units as a shared interface for waveform decoding and video conditioning. A full-sequence Teacher Video Generator (Wan2.1-T2V-1.3B + OmniAvatar-1.3B LoRA) is distilled into a few-step block-causal Streaming Student using AnyFlow's flow-map plus a novel Prefix Streaming mechanism. Training proceeds in four stages and is evaluated on VoiceBench, OmniCharacter, and VBench/AudioBox/SeedTTS/SyncNet metrics.

Key results:

• Streaming Student: SC/IQ/DD/Sync-C rise with steps — 4-step (93.65/57.40/32.00/3.90, 26.5 FPS); four-GPU end-to-end RTF at 4 steps = 1.293, vs. Teacher RTF = 26.917.
• First audible speech at 2.308s and first playable video chunk at 3.142s (4-step pipeline).
• OmniCharacter: Ex-Omni-2D averages 3.938 across Fluency/Coherency/Consistency vs. 3.537 for the Qwen3-8B backbone.
• VoiceBench: AlpacaEval 4.28, CommonEval 3.71, BBH 58.70 (second-highest after Qwen2.5-Omni).
• Prefix Streaming vs. no-prefix: Sync-C 3.69→3.90; last-minus-first consistency error reduced 21.4%.
• VTP first-frame field accuracy only 43.00% (LLM-judge grounding check); other fields average 85.38%.

Why it matters / caveats: Offers a data-efficient way to give omni-modal dialogue agents synchronized visual presence without paired query–text–speech–video datasets; the Streaming Student trades quality for speed, single-request end-to-end RTF stays above 1 (not real-time), and reference-speaker similarity leaves room for improvement.

Problem: Standard LLM benchmarks measure performance only along a model's nominal, most-confident decoding path, obscuring whether its reasoning is robust once real-world constraints force it off that path.

Method: The paper introduces Decoding-Level Taboo, a zero-prompt, zero-training diagnostic implemented as a Hugging Face transformers logits-processor that masks the top-i candidate logits only at word-initial decoding steps, forcing the model to reconstruct semantically valid alternatives; severity is quantified via "Injected Surprisal" and outcomes via a "Conditional Retention Ratio" R. Applied to base and instruct checkpoints of Qwen2.5 (0.5B–72B), Gemma-3, Llama-3 (8B/70B), and OLMo-2, evaluated with 2-shot CoT on GSM8K, MMLU, TriviaQA, and HumanEval.

Key results:

• On GSM8K at taboo rank i=1, instruction tuning raises retention 8%→48% (Qwen2.5-7B), 25%→90% (Gemma-3-12B); Llama-3.1-8B shows no gap (14% vs 16%).
• Scale effect: Qwen2.5 base retention at i=1 rises from 6% (0.5B) to 59% (32B); aligned versions reach 93% at 32B.
• Llama-3.1-70B recovers the alignment effect absent at 8B: retention rises 36%→75% at i=1.
• HumanEval collapses to near-0% Pass@1 for all models/ranks, serving as a negative control.
• Mid-word (vs. word-initial) masking ablation collapses accuracy to 0.01–0.05 across all families at matched mask width.
• Precision robustness: median |acc_4bit − acc_bf16| = 0.03 over 50 (model, dose) cells.

Why it matters / caveats: Offers a cheap, decoupled probe of latent reasoning resilience with negligible runtime overhead and no retraining; limitations include English-only evaluation, open-weight dense architectures only (no MoE/proprietary/multimodal), a single masking rule, and no multiple-comparison correction across the full evaluation grid.

Problem: Skills accumulated by self-evolving agents through repeated append-only edits grow much longer than their genuinely new procedural content, and existing compression methods either ignore skill structure or require costly task rollouts.

Method: SkillZip represents a skill as a typed contract C(S) = ⟨I, G, T, C, O, E⟩ and formulates compression as a minimum-description-length objective selecting the shortest library-plus-residual representation subject to a hard coverage constraint. It runs in a One-Shot mode (scanner → LLM contract extraction → candidate reuse proposal → dynamic-programming cover → template rendering) and a continual Zip-on-Write mode. Evaluated on three agent backbones (Qwen3.7-Max, Qwen3.6-Plus, Kimi K2.6) over BFCL-v4 Web Search, LiveMathematicianBench, and SpreadsheetBench.

Key results:

• Skill length grows ~5.2× (average) by round 5 of self-evolution across benchmarks.
• SkillZip achieves 27.1%–36.9% compression (avg 31.2%) while macro-average task score (0.577) slightly exceeds the uncompressed evolved skill (0.570).
• Vs. SkillReducer baseline: higher compression (31.2% vs 9.2%) and higher task performance (0.577 vs 0.544).
• Compression cost: SkillZip averages 286 seconds and 4–8 LLM calls with zero task rollouts, vs. SkillReducer's 587–1331 seconds/34–80 LLM calls plus 40–80 validation rollouts — a 3.5× average speedup.
• In a 16-round continual setting, activating Zip-on-Write from round 1 caps skill growth at ~1.6×–1.9× vs. 2.5×–3.7× uncompressed.

Why it matters / caveats: Enables lossless-by-construction consolidation of self-evolved agent skills without consulting downstream tasks or rollouts during compression; its preservation guarantee is limited to requirements the structural parser actually extracts, so coverage depends on parser fidelity.

Problem: Existing benchmarks for LLM-based mobile assistants fail to evaluate the ability to locate and integrate personal information scattered across multiple apps, since they typically provide required parameters explicitly or restrict retrieval to a single document.

Method: SPIEval is a human-curated benchmark of 250 tasks (50 per capability) grounded in 4,335 fictional personal records across 10 simulated apps. It evaluates five cognitive capabilities — reasoning, disambiguation, integration, preference inference, multi-intent decomposition — using 21 tools in multi-turn interaction. Tested on nine LLMs (GPT-5.5, Gemini 3.1 Pro, Claude Opus 4.8, DeepSeek-V4-Pro, GLM-5.2, Hy3, Kimi K2.6, Qwen3.7-Plus, Seed-2.1-Pro) at highest and lowest reasoning effort.

Key results:

• Best model GPT-5.5 (xhigh) achieves only 57.3% overall accuracy; weakest (Kimi K2.6, none) achieves 16.4%.
• 79% of failures stem from committing to wrong parameter values.
• Removing retrieval entirely raises average accuracy from 35.5% to 66.8%; returning all matching records at once (vs. paginated) improves accuracy only marginally (35.5%→36.0%).
• Increasing reasoning effort yields an average gain of 13.8 points.
• Across 126,279 retrieval calls, substring queries account for 98.5%; only 9.5% of retrievals are field-specific.

Why it matters / caveats: Shows information localization — not reasoning or execution — is the primary bottleneck for current LLM-based mobile assistants, exposing a reliability risk when acting on scattered personal data; presented as a benchmark/diagnostic contribution rather than a new model or training method.

Problem: In long-horizon deep research agents, accumulated context grows rapidly while the marginal value of additional retrieved evidence declines, driving up token cost, latency, and noise without a clear understanding of where pruning should be applied.

Method: The paper formalizes marginal value estimation as a stage-aware pruning problem with three intervention points — Pre-Retrieval, Post-Retrieval, Pre-Synthesis — in a tree-structured pipeline built on GPT-Researcher, comparing heuristic scoring rules (MMR, Geometric Residual Novelty, Centroid Drift, DPP, Submodular Coverage), an LLM-as-judge strategy, and a learned pre-retrieval controller. Evaluated on 100 queries from DeepResearchGym (Researchy Questions dataset), with checks on DeepResearch Bench.

Key results:

• Unpruned baseline: 29.0 nodes, 375.4k tokens, 3422.6s runtime, 57.83 overall quality.
• Post-Retrieval MMR (one-stage): reduces tokens to 114.6k (~69.5% lower), runtime to 1379.8s (~59.7% lower), retaining 56.62 quality (~97.9% of baseline).
• Two-stage CD+SC: highest two-stage quality at 59.47 (+1.64 over baseline), 63.4% lower token cost.
• Three-stage MMR: strongest compression overall (73.3% token reduction) but quality drops to 55.90.
• No pruned one-stage method surpasses baseline on the KPR+KPC relevance metric.

Why it matters / caveats: Pruning stage placement matters more than the specific scoring rule — early pruning yields the largest efficiency gains, late pruning mainly refines quality; results are tied to one fixed pipeline/benchmark and may not generalize, and the learned controller shows only limited, inconsistent gains over heuristics.

Problem: Long-document visual question answering over many high-resolution pages is computationally costly and suffers from "context rot," degrading accuracy as context length grows.

Method: InSight-doc is an agentic, retriever-free framework built on Qwen3-VL-8B-Instruct that starts from a low-resolution document view and iteratively emits zoom-in tool calls within a multimodal chain-of-thought. Trained via SFT on a 17.9K-example corpus of zoom-in trajectories (distilled from a two-agent InSight-o3 pipeline), followed by RL using GRPO with a binary accuracy reward on 19.2K hard examples.

Key results:

• On document VQA (DUDE, MP-DocVQA, MMLongBench-Doc, LongDocURL), InSight-doc-8B improves average accuracy over Qwen3-VL-8B by 16.4 points (50.5%→66.9%) at low resolution.
• RL adds substantially beyond SFT alone: average accuracy rises from 56.6% (SFT) to 66.9% (SFT+RL) at r=0.25.
• Efficiency: at 50 DPI, matches near-100-DPI baseline accuracy using 58% fewer tokens and 16% lower latency; on longest documents, 69% token reduction and 71% latency reduction.
• Hallucination: unanswerable-question F1 improves by 24.6 points (DUDE) and 25.9 points (MMLongBench-Doc) at r=0.25.
• RL raises LongDocURL evidence-box coverage to 82.3% and nearly eliminates "stuck" trajectories (down to 0.1%).

Why it matters / caveats: Pushes the accuracy-efficiency Pareto frontier for long-document understanding by treating resolution as an adaptively agentically-controlled resource; evaluation is limited to a single backbone (Qwen3-VL-8B-Instruct) with no cross-family experiments, and cross-paper comparisons are not fully controlled for backbone/data/resolution.

Problem: Supervised finetuning of open multilingual translation LLMs is limited by scarce, unevenly distributed parallel (reference) data, especially for low-resource languages.

Method: Starting from SFT MiLMMT-46-v0.1 models (1B/4B/12B, 46 languages), the authors apply GRPO with a reward averaging XCOMET-XXL and COMETKiwi, gated by language-ID via OpenLID-v3, on a filtered subset (31,572 of 263,982 candidate instances) with references discarded. SFT and RL checkpoints are then linearly interpolated (α=0.5) to produce MiLMMT-46-v1.0; on-policy distillation (GKD/PG-OPD) from the 12B teacher into 1B/4B students is also studied. Evaluated on FLORES+ and WMT24++.

Key results:

• Averaged across scales, RL+interpolation improves WMT24++ reference-free XCOMET by 2.75 points and COMETKiwi by 2.44 points over SFT.
• MiLMMT-46-12B-v1.0 achieves the highest reference-free XCOMET/COMETKiwi among evaluated systems, including Google Translate, Gemini 2.5/3 Pro, and GPT-5; surpasses GPT-5's average spBLEU on FLORES+ (35.14 vs. 34.94).
• MiLMMT-46-1B-v1.0 outperforms TranslateGemma-4B on every reported metric despite 1/4 the parameters.
• Interpolation at α=0.5 recovers 2.79–4.21 spBLEU points relative to pure RL endpoints at a cost of only ~0.53–0.56 XCOMET points.
• On-policy distillation approaches but does not exceed the RL+interpolation quality frontier.

Why it matters / caveats: Shows reference-free RL with QE rewards and checkpoint interpolation can push open multilingual MT models to match or beat proprietary systems on reference-free metrics; the SFT lexical-overlap metric (spBLEU) declines after RL, and distillation to smaller models doesn't exceed direct per-scale RL.

Problem: After a sparse mixture-of-experts (MoE) recommendation model has been trained, there is no method to convert its full expert bank into a smaller standard MoE checkpoint under an explicit serving/expert budget without adding a compression-specific online module.

Method: UniMoMo is a post-training compression framework that formulates expert-bank reduction as calibration-conditioned constrained graph coarsening: it profiles each expert's response to shared unlabeled calibration tokens to build a functional-affinity graph, applies layer-adaptive protection to shield high-traffic experts, then greedily coarsens clusters to a target expert count. Merged experts are reconstructed via traffic-weighted parameter averaging plus a least-squares correction, followed by brief fine-tuning. Evaluated on RankMixer (and four other backbones) across Amazon Beauty, KuaiRec, TenRec, MicroVideo, KuaiVideo, and Ebnerd.

Key results:

• Four-expert (8→4) checkpoints achieve 99.92%–102.30% of Origin MoE NDCG@10 with A100 speedups of 1.28×–1.63×.
• Aggressive two-expert, top-1 checkpoints achieve 98.36%–104.24% of Origin MoE NDCG@10 with speedups of 1.47×–2.21×.
• Beats matched baseline MergeMoE on six-block Amazon Beauty/KuaiRec/TenRec by NDCG@10 gains of 0.0004–0.0051.
• Expert-count sensitivity: 4 vs. 8 experts reduces latency 40.36% while retaining 99.37% of NDCG@10.
• Across four additional backbones, UniMoMo ranks first in all 24 AUC/GAUC metrics, gaining 0.51–1.37 points over the strongest alternative.

Why it matters / caveats: Lets a trained recommendation MoE checkpoint be exported at multiple serving budgets with no added inference-time module and often matched or improved ranking quality; evidence is limited to RankMixer checkpoints with 8 source experts and single-A100 inference, and partition quality depends on calibration traffic being representative.

Problem: State-of-the-art visual document retrieval (VDR) relies on multi-billion-parameter vision-language models that are slow and memory-expensive to index at corpus scale.

Method: DistilVDR is a 524M-parameter end-to-end single-vector VDR system built via "dual-student distillation": two independently trained students (a 454M document encoder combining InternViT-300M-448, ModernBERT-base, and a projection head; and a 70M DistilBERT-base query encoder) regress via a pointwise cosine alignment loss onto cached embeddings from a frozen 8B Qwen3-VL-Embedding-8B teacher, with no contrastive term or negative sampling. Two variants (HiRes, Fast) differ only in visual-tile budget. Evaluated on the full 22-dataset ViDoRe suite.

Key results:

• DistilVDR-HiRes: 61.74 average NDCG@5 (86.9% of the 8B teacher's 71.05); DistilVDR-Fast: 59.98 (84.4% of teacher) at 3× smaller visual-token budget.
• HiRes beats the strongest reproduced sub-1B baseline (colSmol-500M, 53.01 avg) by 8.73 points.
• HiRes exceeds DSE-Qwen2 (2.2B) and ColPali v1.3 (2.9B) despite being 4–6× smaller.
• Efficiency: DistilVDR-Fast hits 99.04 docs/sec at 2.10 GB peak VRAM (~18× faster than the 8B teacher); index footprint is 15.6× smaller than the strongest sub-1B multi-vector baseline.
• Adding contrastive supervision on top of cosine distillation yielded no NDCG@5 improvement at any weight tried.

Why it matters / caveats: Shows bilateral cosine distillation from a single frozen teacher can produce a compact retriever dominating from-scratch small multi-vector approaches on quality and every deployment axis; no baseline was retrained under a matched budget, only one teacher was used, and both variants remain 7–10 points behind the strongest 4–8B multi-vector models.

Problem: Existing outdoor embodied-AI benchmarks lack the photorealism, structural complexity, and continuous large-scale navigability needed to evaluate embodied agents in real-world urban districts.

Method: 360CityArena is built on a "Realistic Virtual World" reconstruction of Akihabara, Tokyo, from 602 interconnected 360° video segments across 85 streets, rendered as a navigable pose graph in Unity. 175 manually authored tasks span three categories/seven subcategories, evaluated with four protocols. Baseline agents wrap proprietary and open-source LMMs (GPT-5, Claude Sonnet 4.5, Gemini 2.5 Flash, Qwen2.5-VL-32B-Instruct, InternVL3.5-8B/38B) in a ReACT-style loop, benchmarked against a 5-participant human baseline.

Key results:

• Human baseline achieves 77.3% average accuracy vs. the best model, Gemini 2.5 Flash, at 17.1%.
• Image-based landmark search outperforms language-based search for most models (e.g., GPT-5: 48.0% vs. 16.0%).
• GPT-5's Environment Understanding accuracy drops 28.0%→18.5% from Easy to Hard.
• Fuzzy-match GPT-5 judge attained 97.9% agreement with human majority vote (κ=0.937).
• Human–best-LMM gap exceeds 36 percentage points even at the most lenient distance threshold tested.

Why it matters / caveats: The large human–model gap demonstrates current LMM-based agents are far from competent at city-scale embodied navigation, particularly exploration strategy and perception under visual clutter; exploration is confined to pre-recorded video trajectories (no free 3D movement), and coverage is limited to a single urban district.

Problem: In query-based plain mask transformers, the standard max-probability-mask inference rule is not explicitly optimized during training, so the query selected at inference often does not produce the most accurate mask, and superior predictions from intermediate decoder layers are discarded.

Method: iFAN combines Adjusted Probability-Mask Ranking (APMR) — a linear+sigmoid mask-quality head trained with soft-IoU/BCE supervision plus a pairwise ranking loss — with Cross-Layer Self-Distillation (CLSD), which selects the best intermediate decoder layer as a "teacher" and distills its score map into the final layer. Applied to EoMT (DINOv2, ViT-S/B/L/G) and PMT (DINOv3, ViT-L) baselines on COCO, ADE20K, and Cityscapes.

Key results:

• Overall average gains: +1.20 PQ, +1.30 AP, +0.63 mIoU across benchmarks with negligible added parameters, FLOPs, or latency.
• Best EoMT-iFAN: 59.6 PQ (COCO), 53.9 PQ (ADE20K) — within 0.3 PQ of best reported results.
• Instance segmentation (COCO): best EoMT result rises from 48.8 to 50.6 AP.
• Ablation (EoMT ViT-L, COCO): baseline 56.00 PQ → full iFAN 57.00 PQ.
• Diagnostics: disagreement between max-prob-mask and max-IoU queries drops from 64.45% to 27.49%.
• Gains diminish as backbone scales: +1.6 PQ on ViT-S vs. +0.8 PQ on ViT-G.

Why it matters / caveats: Closes much of the gap to heavier segmentation architectures while adding virtually no inference-time cost, since both components are training-only; gains diminish with backbone size, and the paper notes a transcription error in main-text results corrected only in the appendix.

Problem: Existing implementations for benchmarking time-series dataset similarity methods are fragmented and hard to extend, leaving similarity metrics evaluated under inconsistent experimental settings.

Method: TSDS-Toolbox is a configuration-driven (YAML) pipeline with Data, Similarity (Wasserstein Distance, MMD, Optimal Transport, Match-and-Deform, DTW, Euclidean, LCSS), Evaluation (OOD/transfer-learning for classification and forecasting via fine-tuned Lag-Llama and Time-MoE), and Analysis layers. Experiments use all supported methods on 25 GluonTS datasets sampled into 100 windows each.

Key results:

• Forecasting OOD with Time-MoE: Match-and-Deform achieves the strongest correlation (0.619), followed by Wasserstein Distance (0.604).
• DBA reducer outperforms PCA in most settings, e.g. classification with DTW (0.237 vs −0.033).
• No single similarity metric achieves the highest correlation consistently across all four evaluation tasks.

Why it matters / caveats: Enables reproducible, standardized benchmarking of time-series dataset similarity for source-dataset selection; the central finding — that no metric consistently correlates best across tasks — means similarity choice is task-dependent with no default recommendation.

Problem: Existing VLM jigsaw-puzzle benchmarks use rectangular cuts that create ambiguous ground truth and are limited to low grid densities, making it hard to rigorously measure joint visual-content and geometric-constraint reasoning.

Method: JigShape is a benchmark of 95,468 jigsaw-puzzle instances built with tab-and-blank interlocking piece edges across four grid densities (4×4 to 16×16), plus a paired no-shape ablation set. Five frontier VLMs are evaluated zero-shot, and Qwen3-VL-8B/Gemma3-12B are supervised fine-tuned on a mixed-grid training set.

Key results:

• Zero-shot: only GPT-5.5 exceeds random baseline on 4×4 (69.65% Piece Accuracy, 26.40% Exact Match vs. 6.25% random); all other models score near chance.
• Scaling cliff: GPT-5.5 drops from 69.65% (4×4) to 4.37% (8×8) to 0.67% (12×12).
• SFT models reach >97% Piece Accuracy on 4×4 but fall to 0.35–0.40% (near-random) on 16×16.
• Removing tab-blank shape constraints drops SFT models from 97%+ to 10–14% — barely above random.

Why it matters / caveats: Shows current VLMs (zero-shot or fine-tuned) cannot maintain constraint satisfaction as puzzle complexity grows, and fine-tuned models mostly exploit geometric shape cues rather than integrating visual content, exposing scalable geometric reasoning as an unsolved challenge.

Problem: Existing benchmarks evaluate data-science code generation or generic GUI/computer-use skills in isolation, so it is unknown whether AI agents can autonomously execute complete, long-horizon, multi-tool data-science workflows inside real operating-system environments.

Method: DSAgentBench extends OSWorld with data-science tooling (Jupyter, VS Code, Chrome, Kaggle API, OpenML, SQLite) on an Ubuntu VM, comprising 275 human-authored, dual-annotator-verified tasks spanning six lifecycle stages, each with a deterministic Python evaluator. 15 closed- and open-source vision-language agents are benchmarked under a fixed 15-action step budget.

Key results:

• Claude-4.6-Sonnet is the strongest agent: 56.70% overall task success (A11y setting), vs. human performance of 85.09%.
• GPT-5 is second-best at 29.81%; all open-source agents score below 1% success.
• Performance drops monotonically with difficulty (Claude-4.6-Sonnet: 86.68% easy → 49.41% hard) and for multi-stage vs. single-stage tasks.
• Increasing the step budget from 15 to 50 yields only marginal gains (24.54%→25.81% for GPT-4o).
• Open-source agents fail almost entirely from grounding errors (97–98%).

Why it matters / caveats: Exposes a substantial capability gap between current agentic systems and human performance on realistic cross-tool data-science workflows; open-source agents were evaluated only in screenshot-only mode, and manual error analysis covers only a subset of trajectories.

Problem: The paper addresses the lack of a rigorous, epistemically-labeled mathematical account of Power Law Graph Attention (PLGA) and the PLDR-LLM architecture — specifically whether PLGA truly contains scaled dot-product attention (SDPA) as an exact special case, and whether the previously reported "collapse" of its learned attention operator to an input-invariant constant at inference is a proven theorem or only a measured empirical phenomenon.

Method: The paper formally defines PLGA — which replaces SDPA's fixed bilinear score form QK⊤ with a learned bilinear operator G_LM built from a query Gram matrix via a deep residual "metric learner," an elementwise power law, and a NOTEARS-style DAG regularizer — inside the decoder-only PLDR-LLM architecture (RoPE, SwiGLU FFN, 32,000-token SentencePiece vocabulary, context length 1024). Claims are labeled theorem/conditional theorem/measurement/conjecture; algebraic proof cores are machine-checked in Lean 4 over mathlib; empirical claims are re-measured via a numerical audit on a released checkpoint.

Key results:

• SDPA is proved (and Lean-verified) to be exactly the G_LM = I point of the PLGA family.
• Measured operator invariance: ~10⁻⁶ to 10⁻¹¹ for near-critical models, and 0 at float resolution for the model trained on 41B tokens.
• Caching (KV-cache + G-cache) gives an empirically measured ~3× inference speedup.
• Cached vs. sequential/block scoring agree on the published TruthfulQA probability-mass metric within 5×10⁻⁵ per item.
• On the audited checkpoint, historical-row logits were bitwise invariant across 2,048,000 randomized suffix-perturbation comparisons.
• Under a nonresonance assumption, a frozen learned operator at head width 64 spans 4032 score-function dimensions unreachable by SDPA with RoPE.

Why it matters / caveats: Reframes PLGA/PLDR-LLM's headline claims at their true epistemic strength, separating what is algebraically proven from what is only measured, and backs core proofs with Lean 4 formalization. Key caveats: the inference-collapse theorem holds only conditionally on empirically observed (not proven) exact operator invariance, derived cache-fidelity bounds are explicitly stated as "far too loose to certify bit-identical decoding," and the self-organized-criticality framing is an explicit phenomenological hypothesis rather than established physics.