AI papers — 2026-09-02

Problem: Existing student simulators either mimic a real student's behavior (state-tracking models) or follow tutor guidance fluently (LLM role-play), but no method reliably does both, limiting their use as proxy feedback for training adaptive AI tutors.

Method: StudentSim fine-tunes Qwen3-4B-Instruct with LoRA in two stages: Stage 1 pools single-turn and multi-turn (problem, wrong response, tutor guidance, corrected response) records across all students in a domain (multi-turn ratio 0.2) to learn shared behavior patterns; Stage 2 specializes the adapter per-student on that student's own records. The authors also build StudentSimEval, a 60-student benchmark (30 chess/Lichess, 15 L2-English/EFCAMDAT, 15 math) scoring behavioral fidelity (F) and guidance responsiveness (R). A trained chess StudentSim is also used as a GRPO reward model to train an AI tutor.

Key results:

• Chess: StudentSim F=0.5150/R=0.9067 vs. GPT-5.4 F=0.2316/R=0.7186 and Maia2 F=0.4535/R=0.2721.
• L2: F=0.5624/R=0.6417 vs. GPT-5.4 F=0.5141/R=0.5950.
• Math: F=0.6384/R=0.9181 vs. GPT-5.4 F=0.6121/R=0.7099.
• A chess tutor RL-trained with StudentSim as reward beats both a no-RL baseline and a GPT-5.4-reward tutor on expert human ratings of accuracy, guidance quality, and personalization.

Why it matters / caveats: Demonstrates that faithful, responsive per-student simulators can serve as scalable, practical reward signals for tutor optimization; the tutor-RL result is explicitly presented only as a proof of concept in chess, not a best-tutor claim, and generalizing the reward function to free-form domains is left open.

Problem: Adapting general VLMs to driving via text-only VQA supervision fails to constrain 3D scene structure and risks catastrophic forgetting of general visual/world knowledge needed for OOD reasoning and shared cockpit-driving deployment.

Method: Qwen-Drive-1.0 keeps the pretrained Qwen3.5-4B VLM architecture unchanged and attaches two external modules: a BEV perception head (depth-based view transform + query-based BEV transformer) jointly trained for 3D detection, semantic occupancy, and BEV map segmentation, and a Planning Expert that generates future ego trajectories via flow matching conditioned on cached VLM keys/values. Training uses a staged recipe mixing driving supervision (perception, driving VQA, causal reasoning) with general-purpose VLM data, plus an RL stage (Qwen-Drive-1.0-RL) tuned on WOD-E2E RFS annotations.

Key results:

• 3D perception: 43.95 mAP / 60.99 map mIoU on nuScenes; 43.45 mAP / 71.27 map mIoU on OpenScene.
• Driving VQA average 69.43 vs. best baseline (Qwen3.5-4B) 63.52; causal reasoning average 58.30 vs. 22.05 for second-best Gemma4-12B; Ego3D RMSE drops 40.9% to 7.78.
• General VLM benchmarks: stays within ~1 point of base Qwen3.5-4B on knowledge/reasoning (66.41 vs 67.40 avg) and surpasses it on spatial understanding (53.96 vs 52.99 avg).
• Motion planning: PDMS 90.7 on NAVSIM; RFS 7.91 on WOD-E2E test split (8.45 on validation split, exceeding the 8.13 human-driver reference).

Why it matters / caveats: First reported unified VLM for driving combining 3D perception, VQA, and planning without changing the base VLM architecture, aimed at shared cockpit-driving compute platforms; results are self-reported against a mix of open baselines and the model is only at 4B scale.

Problem: Prior looped-Transformer studies compare against baselines matched only on parameter count, conflating any architectural benefit of looping with the extra FLOPs the loop spends, leaving unclear whether looping helps under truly matched compute.

Method: SMELT (Sparse MoE Transformer, middle layers Loop Twice) loops the middle 50% of an MoE Transformer's layers twice while jointly matching per-token FLOPs, total non-embedding parameters, and KV cache versus an unlooped Baseline — achieved by narrowing hidden dimension, raising expert count, scaling looped residuals by 1/2, and adjusting attention head size/GQA ratio. The recipe was found via ablations over loop span, effective depth-to-width ratio, and loop count (2 vs 3 vs 4 visits), then scaled to four sizes up to 54B non-embedding parameters and four sparsity levels (~85–97%), fitting separate Chinchilla-style scaling laws for SMELT and the Baseline.

Key results:

• SMELT reaches lower validation loss than the Baseline at every scale/sparsity tested; Compute Efficiency (CE) Gain rises with compute: 6.8–10.0% FLOPs saved at 10^20 FLOPs, 14.7–18.0% saved at 10^21 FLOPs (compute-optimal frontier).
• SMELT outperforms Baseline on DCLM and MMLU by more than validation loss alone predicts; per-domain CE Gain is highest on Code (20.4%) and Finance/Math-STEM (16.8%).
• Advantage grows with sample length and number of in-context examples.
• Mechanistic analysis: the second visit substantially amplifies (rather than overwrites) first-visit residual updates and reduces attention sink, redirecting mass to content-relevant tokens.

Why it matters / caveats: Provides the first multi-scale, fully compute/parameter/KV-matched evidence that looping is a genuine architectural win rather than an artifact of extra FLOPs; the study is limited to the specific SMELT design space (middle-half, twice-looped MoE) and the authors note some open questions, e.g., why gains are non-monotone across loss quartiles.

Problem: GUI agents struggle to move from benchmark-oriented performance to dependable real-world deployment due to limited environment coverage, brittle/unverifiable generated tasks, and unreliable RL reward signals.

Method: UI-Venus-2 is built on Qwen3.5-9B and Qwen3.6-27B backbones and trained through a three-stage pipeline: Stage I multimodal mid-training on a large heterogeneous mixture of Grounding, CAPTCHA, Mobile, Web, and Computer navigation data; Stage II offline reinforcement learning; and Stage III Multi-teacher On-policy Distillation (MOPD), an action-aware variant of on-policy distillation. It scales to 170+ multilingual mobile apps plus native desktop OS, uses a deep-research pipeline for function-grounded task generation, and adds trace- and sample-level verifiers (visual keypoints, multi-model voting) plus safety-aware handling of consequential actions.

Key results:

• MobileGym: UI-Venus-2-27B 60.5% (best), beating strongest baseline Seed-2.0-Pro (52.0%) by 8.5 points.
• VenusBench-Mobile: 27B reaches 48.7% vs. best prior baseline Claude-Opus-4.6 at 36.5%.
• AndroidWorld: 27B reaches 84.0%, improving over prior best (UI-Venus-1.5-30B-A3B, 77.6%) by 6.4 points.
• VenusBench-GD grounding: new SOTA 80.1%, +5.1 points over UI-Venus-1.5-30B-A3B; OSWorld-Verified: 80.5% (27B), second only to Claude-Opus-4.8 (83.4%).

Why it matters / caveats: Open-sources full-parameter weights and evaluation infrastructure, offering a strong efficient foundation for GUI-agent research across mobile, web, and desktop; some baseline numbers are self-reproduced ("*"), and the paper notes it favors standalone end-to-end systems evaluated on comparable task subsets, so cross-paper comparisons may not be fully apples-to-apples.

Problem: Robust robot manipulation generalization requires broad physical experience, but action-labeled robot trajectories are expensive and limited in diversity, and it is unclear how to convert abundant but action-free egocentric video into effective robot control.

Method: ZimaBlue is a World Action Model (WAM) trained with a three-stage curriculum: (1) causal embodied video pre-training on up to 120,000+ hours of human/robot egocentric video predicting future visual states from past observations and language instructions; (2) video-action mid-training on heterogeneous multi-embodiment robot trajectories using a unified 100D state-action representation; (3) target-robot post-training/specialization. For real-time control it uses an asynchronous Slow-Fast dual-system architecture — a high-capacity Slow world model plus a lightweight Fast branch — accelerated with diffusion-step (DMD) distillation and CUDA graph/Torch compile optimizations.

Key results:

• Real-robot zero-shot success rises from 36.1% (target-robot data only) to 46.1% (adding cross-embodiment data) to 66.9% (60K video hours) to 77.8% (120K video hours).
• Closed-loop latency cut from 450ms to 33ms (13.6× speedup) on an NVIDIA RTX 4090, enabling 30Hz control.
• LIBERO-Plus: 86.7% overall zero-shot success, outperforming all WAM-based baselines.
• RoboTwin 2.0: 94.7% (Clean) / 94.3% (Randomized) success.
• RoboCasa365: 49.5% average (78.1/50.4/16.5 on Atomic-Seen/Composite-Seen/Composite-Unseen), best among WAM methods and second overall only to the VLA-based Xiaomi-Robotics-1 (57.4%), which uses 100,000 hours of real-robot data.

Why it matters / caveats: Shows video pre-training scale, not just robot-action data, is a major lever for zero-shot manipulation generalization, especially on unseen tasks/perturbations; it still trails a VLA baseline trained on far more real-robot data on RoboCasa365, indicating video pre-training complements but does not fully replace large-scale real-robot demonstrations.

Problem: Pretrained video generators are not directly usable as interactive world models, since converting them into precise, temporally grounded action-controllable simulators normally requires new action-specific modules and additional supervision.

Method: H3-World adapts the 33B-parameter MiniMax-H3 video generator by expressing character and camera actions as compositional textual instructions injected through H3's native text pathway, rather than adding dedicated action modules. It introduces latent-aligned temporal binding (pairing each action-prompt span with its corresponding video latent interval) and single-egress attention routing to prevent control leakage across time, and learns the mapping via rank-32 LoRA (0.199% of parameters) applied to QKV/output projections and the token refiner, trained for 10,000 steps on 8,000 gameplay clips from ABot-World-Explorer-500h.

Key results:

• With LoRA, cumulative horizontal optical flow correctly follows a scheduled left→right camera-pan switch (+52.7 before / -106.0 after the switch), vs. near-zero response (-0.1/0.0) for the zero-LoRA per-latent interface and an incomplete response from frozen H3 with global prompting (0.0/-17.3).
• Direct action-conditioning baselines (additive-bias per ReactiveGWM, FiLM/AdaLN) produce weak or inconsistent motion compared to H3-World's coordinated character/camera control on held-out clips.
• Generalizes to 52 of 135 valid character-camera combinations unseen during training, and to 6 visually distinct out-of-domain initial scenes (first/third-person, indoor/outdoor, fantasy/sci-fi).

Why it matters / caveats: Shows that large pretrained video generators already encode enough action-relevant structure that lightweight LoRA adaptation (no new modules) suffices for interactive control. The authors note evaluation is mainly qualitative/representative-example based, and the model produces fixed-length segments without persistent world state, real-time interaction, or planning.

Problem: Enterprises under data-residency constraints must self-host LLMs, but continually adopting new models without retiring old ones fragments a finite GPU pool across a growing model zoo serving over 200 internal applications.

Method: The authors build an internal benchmark from production traffic using a template-aware sampler (masks variable tokens, groups near-duplicates via LSH, allocates budget by √count) and a task-classifier-routed judging pipeline (deterministic verifiers or calibrated LLM judges, e.g. DeepSeek-V3-0324 as Arena-Hard-Auto judge). The post-training recipe starts from Qwen3-32B with a Cyrillic-dense tokenizer, uses one shared SFT stage over all domains, then forks into three independent GRPO experts (general/dialogue, instruction-following, function-calling) each with domain-specific reward fixes (length penalty + KL constraint for general; VerIF-style verifiable reward + RM quality correction to fix semantic collapse for IF; Tool-N1 reward + synthetic irrelevance injection to fix over-calling for FC), merged via two-stage sequential SLERP (verifiable experts first, then general).

Key results:

• Template-aware sampling achieves highest prompt diversity (0.953) while keeping JS-distance to production lowest among tested methods (random, greedy max-min, #InsTag).
• Task-specific judging lifts Cohen's κ from 0.63 to 0.88 (reference-based) and 0.57 to 0.72 (open-ended) vs. uniform side-by-side judging.
• Final merged model beats the ~7x larger Qwen3-235B-A22B-Instruct-2507 on in-house Arena (69.57 vs 65.83), in-house BFCL (0.79 vs 0.77), ruBFCLv3 (65.96 vs 64.42), AceBench (73.50 vs 70.20); ruWildChat rises from 52.0 to 80.7 (within 4.4 points of the 7x larger model); SmartSearch F1 improves from 0.478 to 0.557.
• Joint multi-objective GRPO underperforms the merge approach (e.g., BFCLv3 EN drops from 61.2 to 54.5 when IF and FC rewards are optimized jointly) unless given a 1.7x larger training budget.
• Deployed model serves 116M requests/month (50% of platform traffic) from 200+ apps at 2.8-3.9x lower per-token cost (up to 4-9x for services that ran larger models), with p95 latency 3.2s and TTFT 0.3s.

Why it matters / caveats: Demonstrates a practical, reproducible enterprise recipe (modular expert training + weight merging) for consolidating a fragmented self-hosted LLM fleet without sacrificing quality. Limitations stated: evidence comes from a single organization/platform and only the Qwen3 model family was tested.

Problem: In multi-hop QA, the granularity at which a question is naturally expressed often mismatches the granularity at which corpus evidence is retrievable, and existing graph-based, iterative, or program-executing methods don't explicitly decide when a query is already resolved versus needs refinement.

Method: Hi-Q formulates multi-hop QA as evidence-conditioned adaptive search: a resolution operator tests whether retrieved top-k passages support the current query node (RESOLVED/UNRESOLVED/FAIL via a cost-sensitive threshold rule derived from decomposition vs. stopping costs), and unresolved nodes are recursively split by a dependency-preserving binary expansion operator (prerequisite branch resolved before dependent branch) checked by a semantic coverage verifier that can repair invalid splits (one repair attempt). It uses NV-Embed-v2 retrieval (k=5), max recursion depth 4, temperature 0, and GPT-4o-mini as reader (with Llama-3.3-70B and Qwen3-30B-A3B as additional readers in appendix).

Key results:

• Full-corpus setting (up to 5.2M passages): 52.3 EM / 64.0 F1 average across MuSiQue/2Wiki/HotpotQA, beating IRCoT by 15.1 EM / 18.2 F1 and PropRAG by 11.5 EM / 12.0 F1 on MuSiQue-full.
• Controlled supporting/distractor setting: 57.9 EM / 69.3 F1, beating PropRAG by 5.6 EM / 3.9 F1 and IRCoT by 13.7 EM / 15.8 F1.
• Beats coding-agent and Recursive Language Model baselines by 30.8 EM / 26.8 F1 and 24.4 EM / 26.7 F1 respectively (full-corpus average).
• Self-Ask achieves higher Recall@5 but far lower accuracy (35.7 EM vs Hi-Q's 57.9), showing raw retrieval volume isn't the driver of Hi-Q's gains.
• Ablations: removing hierarchical refinement drops 57.9→51.5 EM; removing dependency-awareness drops to 47.1 EM.
• Trigger signal is precise: false-trigger rate ~10% (manual analysis of 100 cases); False Rejection Rate 11% on normal queries, False Acceptance Rate 2% on adversarial unanswerable queries.

Why it matters / caveats: Shows hierarchical, evidence-conditioned refinement scales well to realistic full-corpus retrieval without requiring costly corpus-wide graph construction (PropRAG's graph construction over HotpotQA's 5.2M passages was estimated to cost >$2,500 in API calls and was excluded). All comparisons use temperature-zero decoding, so reported confidence intervals reflect question-population uncertainty, not run-to-run variance.

Problem: Prior MLLM-based drone control systems narrow the model's role (e.g., grounding only) and hold the agent architecture fixed while varying systems, so it's unclear which parts of embodied drone control actually depend on the backbone model's capability.

Method: The authors introduce DroneCATS-Agent, a model-agnostic architecture where an MLLM outputs exactly one of four prompt-declared actions per step (go: pixel+depth; rotate: in-place yaw to search; think: deliberate/hold; finished: self-declared arrival claim) with no fine-tuning or function-calling schema, feeding a rule-based geometric controller and standard flight stack. DroneCATS benchmark crosses target motion (static/moving) x initial visibility (in-view/out-of-view) into 4 task types (Approaching, Tracking, Searching, Search-and-Track) plus a 4-drone Commanding setting, scored by one unified success criterion (any arrival declaration within δ=5m of the true target while visible), run in AirSim on two maps (100 episodes total) across 9 models: GPT-5, Claude Opus 5, Gemini 3.7 Flash, Gemini Robotics-ER 2, Qwen3.5 (2B/4B/9B/27B), and Cosmos3-Edge-2B.

Key results:

• Best model (Gemini 3.7 Flash) succeeds in only 65% of Approaching episodes (static, in-view target); withholding the target from the first frame drops the best score to 40%.
• Model ranking is unstable across tasks: GPT-5 approaches at 60% (near Gemini 3.7 Flash's 65%) but tracks moving targets at only 15% vs. Gemini 3.7 Flash's 80%.
• Small open models often physically reach the target more often than frontier models but fail to declare correctly: Qwen3.5-9B enters the success radius in 90% of episodes (Oracle Success Rate) but only converts 35% into declared success; Qwen3.5-2B never succeeds despite closing 57% of the initial distance on average, and Cosmos3-Edge-2B enters δ in 25% of episodes but declares zero times.
• Embodiment-specialized Gemini Robotics-ER 2 (47.5% avg across 4 cells) shows no advantage over its generalist sibling Gemini 3.7 Flash (57.5% avg).
• In Commanding (4 drones, one context), one-drone ranking doesn't transfer: GPT-5 succeeds in only 20% of commanding episodes vs. Gemini 3.7 Flash's 80%.

Why it matters / caveats: Reframes the bottleneck in embodied drone control as protocol adherence (declaring arrival correctly) rather than navigation/grounding, which prior work's narrowed architectures had "engineered out" of evaluation. The open-weight roster is intentionally small (2B-27B) to probe onboard-deployable scales; results are based on simulation (AirSim) rather than real flight.

Problem: Unified multimodal models (UMMs) can perform visual understanding and generation in one model, but it is unclear whether joint training of these two objectives actually produces mutual benefit, mere coexistence, or interference.

Method: Using a pixel-in/pixel-out native multimodal setting (no pretrained vision encoder or VAE/tokenizer) built on Qwen3-1.7B with causal autoregressive text and flow-matching for bidirectionally-attended visual tokens, the authors compare three routing architectures: Dense sharing (all tokens through one LLM), Modality-decoupled MoT (all visual tokens routed to a separate scratch-trained branch, following Liang et al.'s MoT), and their proposed Task-decoupled MoT (understanding tokens stay in the language branch, generation tokens go to a specialized branch, softly aligned via shared attention plus an added 20% image-reconstruction/editing task with 50% masked context tokens). They then study task-level transfer via three paired case studies (geometry, SVG, 3D spatial intelligence) and system-level comparison of an end-to-end UMM vs. a matched planner(understanding)+executor(generation) pipeline on reasoning-intensive image editing, evaluated on RISEBench and KRIS-Bench.

Key results:

• Dense sharing: joint training helps understanding (e.g., Vision-centric&SI 60.72→61.77) but hurts generation (GenEval2 79.11→78.12, HPSv3 7.05→5.92).
• Modality-decoupled MoT: opposite trade-off — generation improves (GenEval2 82.06→83.09, HPSv3 7.72→7.97) but understanding drops (General 64.20→61.58, OCR 66.84→62.11).
• Task-decoupled MoT avoids both trade-offs: General 69.02, OCR 72.09, V-Centric&SI 63.96, GenEval2 82.31, HPSv3 7.90 — competitive with or better than both single-objective baselines simultaneously.
• Geometry case: joint UND+GEN improves Geometry3K (59.90→65.39), PGPS9K (72.40→73.40), MathVerse (60.13→63.15), MathVista (71.80→74.30), outperforming a Textual-Desc control using identical data converted to text.
• SVG case: joint training improves both directions (Image→SVG 80.64→81.70; SVG→Image 80.21→86.52) and improves blind-VQA "mental visualization" accuracy (Und-only 52.18% → Joint 58.54%).
• System level: end-to-end UMM beats planner→executor pipeline on RISEBench (Overall 18.88 vs 16.66) and KRIS-Bench (Overall 68.33 vs 66.48).

Why it matters / caveats: Provides concrete architectural guidance for UMM design: specialize computation where understanding/generation objectives conflict (task-decoupled routing), but unify end-to-end when tasks share knowledge or require tight interaction. The study is conducted under one specific generative formulation (autoregressive text + flow-matching images); the authors note other formulations (fully autoregressive, discrete denoising) are left to future work.

Problem: Safety in language models is usually bolted on via external safeguards, SFT, or adapters rather than being an intrinsic, addressable property of the model's own persistent computation.

Method: Safin-1 is built on MARCH (Memory-Anchor Routing across Context History), which periodically checkpoints a recurrent backbone's evolving state into addressable "state anchors," each paired with a content-conditioned routing key; every token retrieves relevant anchors (or a learned null option) via softmax routing (Top-4 sparse by default) and fuses the readout residually with the current recurrent path. The same routed bank hosts learned "persistent capability states" — instantiated here as a detachable Safety State trained on harmful/benign examples while the backbone stays frozen. Validated at 0.8B scale on Gated DeltaNet/Kimi Delta Attention/GDN-2, then scaled to 4B and 35B-A3B Qwen3.5 hybrid backbones via continual pretraining + SFT.

Key results:

• 0.8B MARCH improves NIAH by +47.0% and general LM/LongBench/retrieval benchmarks over baseline GDN.
• Macro-average over 10 capability benchmarks rises from 66.79→69.20 (4B) and 76.25→78.35 (35B-A3B).
• 35B reasoning (AIME 2025 Avg@64) improves +11.4% over Qwen3.5.
• Persistent Safety State reduces average jailbreak ASR by 42.3% (4B) and 52.3% (35B-A3B) vs. Qwen3.5, with lower over-refusal (XSTest ORR) than a rank-8 LoRA control.
• Top-4 sparse routing more than doubles throughput and cuts core runtime ~10x vs. dense routing at 128K tokens.

Why it matters / caveats: Demonstrates a concrete architecture-level path to make safety a state-native, attachable/detachable capability rather than a training-time behavioral constraint. Authors explicitly caveat this is "only an initial architectural exploration," with substantial further work needed.

Problem: In unguided self-play self-evolution, question generation lacks a signal for which unresolved reasoning weaknesses to target, causing solver performance to plateau or decline, while existing guided methods require external task resources outside the self-play loop.

Method: DiagEvo adds a lightweight LLM "diagnostician" (default Qwen3-Instruct-2507, 4B) that compares disagreeing vs. agreeing solver trajectories to extract transferable error causes, stored in a hierarchical error-cause memory (skill nodes → causes marked Active/Mastered by self-consistency). A frequency-driven schedule (εt+1 = F1/(F1+Ft/k)) mixes cause-targeted generation (including cross-state stitching of Active with Mastered causes) with free exploration for the GRPO-trained challenger. Double-confidence filtering keeps only intermediate-difficulty questions where the majority vote share is both within [plow, phigh] and clearly ahead (ratio τ) of the second-most-common answer before GRPO solver training.

Key results:

• On Qwen3-8B: 72.3% mean accuracy across five math benchmarks (+4.5 pts over R-Zero), 38.8% on general reasoning (+2.6 pts over R-Zero); beats DARC by 1.2 and 1.0 pts despite using no external resources.
• Mean accuracy across all nine benchmarks: 57.4%, 1.1 pts above DARC.
• Consistent gains across all three solvers tested (Qwen3-4B, Qwen3-8B, OctoThinker-8B).
• Ablations confirm both the hierarchical error memory and double-confidence filtering contribute to gains; increasing diagnostician scale beyond 4B is not the main driver.

Why it matters / caveats: Shows a self-play curriculum signal can be mined entirely from the solver's own failure history without external supervision. Authors note limitations: filtering relies on solver agreement rather than ground truth (so a shared error can still pass), the number of co-evolution rounds is fixed in advance, and evaluation is scoped mainly to mathematical reasoning.

Problem: Autonomous software development (building complete systems from high-level requirements with no human intervention) requires sustaining coherent progress over very long trajectories, where existing coding-agent harnesses tend to lose track of prior decisions, repeat repairs, or falsely declare completion.

Method: Harness-of-Harness (HoH) wraps an existing coding-agent harness in an iterative Planner–Developer–QA Tester loop: the planner turns requirements plus prior test evidence into a small, verifiable development increment; the developer implements it with embedded testing; the QA tester independently evaluates via complementary white-box and black-box tests, producing a structured evidence bundle fed back to the next iteration. Continuity is maintained via progressive disclosure (file-system-persisted plans/reports/histories with a concise index) rather than a dedicated memory module, role-specific tools/skills (e.g., Godot MCP, asset generators), and versioned project history for rollback. Evaluated with three harness–model pairs: Codex+GPT-5.5, OpenCode+DeepSeek-V4-Pro, Pi+MiniMax-M3.

Key results:

• Average relative gain of 52.25% over standalone harnesses across GameCraft-Bench, FrontierSWE, and ProgramBench after 3 iterations; max gain 82.86%.
• Absolute gains after 3 iterations: +16.62–22.08 pts on GameCraft-Bench, +19–29 pts on FrontierSWE, +6.09–16.85 pts on ProgramBench.
• Codex+GPT-5.5(high) on FrontierSWE improves from 22% to 72.67% over 10 iterations (peaking 76.00% at iteration 9), vs. 27.33% for Vanilla.
• FrontierSWE cross-configuration Dominance for Codex rises from 44% (Vanilla) to 58/60/71% at HoH@1–3.
• Multi-day deployment (>70 iterations) autonomously produced a playable FPS game with storyline, combat mechanics, HUD/UI, animation, audio, and visuals.

Why it matters / caveats: Not stated (no explicit limitations section found in excerpted text beyond qualitative failure-case discussion in figures).

Problem: Existing multimodal memory systems for long-video QA retrieve isolated modality-specific fragments (captions, transcripts, frames, graph facts), forcing the LLM to reconstruct cross-modal and temporal alignments at inference time under limited context and difficult attribution.

Method: EM²Mem uses an "align-then-retrieve" design: videos are split into short base segments (e.g., 30s) anchored by event anchors ei=(i, τ_start, τ_end); each event memory cell µi=(ei, Ri, Ci) binds a multimodal event record (keyframe caption, transcript, keyframes, structured metadata like actions/objects/topics/scenes/entities) plus multi-scale temporal context views (3min/10min/1hr). Two auxiliary graphs — an episodic graph GE (cross-event relations/transitions) and semantic graph GS (long-term facts/habits/preferences) — are built over event anchors. At inference, an LLM-based selector retrieves and expands relevant event cells via graph links to compile a compact query-specific evidence view for answer generation.

Key results:

• Average accuracy improves over the strongest memory baseline (WorldMM) by 2.0, 2.4, and 3.7 points on EgoLifeQA, Ego-R1 Bench, and Video-MME (L) respectively.
• Strict event-level Top-5 evidence recall improves by 7.0 points over WorldMM's five rounds of iterative retrieval.
• Per-query latency reduced 4.67x (459.00s → 98.21s); total inference tokens reduced 63.66% (42.03M → 15.27M).
• Construction-time (unified) alignment outperforms retrieval-time fusion in ablations.

Why it matters / caveats: Authors state the structured/textual memory trades visual fidelity for searchability (losing fine-grained pixel detail like small objects/colors), the align-then-retrieve paradigm is "not fully complete" since final answering still leans on selected keyframes, and egocentric benchmark performance may not generalize across populations/deployment contexts.

Problem: In multi-agent LLM systems, prompts simultaneously encode task content and execution-critical protocol (routing, formatting, termination), so prompt optimization aimed at improving content can inadvertently corrupt the protocol and crash the pipeline.

Method: The paper proposes control-data flow separation: each agent output is split into a typed, schema-validated control channel (routing/termination fields, e.g. Python dataclass/Pydantic objects with Literal types) that the program controller reads and which optimizers cannot edit, and a free-form data channel that carries optimizable task content. This is implemented as a Python library, cdsep, with auto-generated schema scaffolding, runtime validation with bounded retry/fallback, and per-example optimizer feedback. It's evaluated against Fixed prompts, Naive TextGrad, and DSPy (no compile, BootstrapFewShot, MIPROv2) on BBH, MARG review generation, and synthetic/industry-verified insurance underwriting.

Key results:

• Outperforms all baselines on task score: BBH 78.3% vs. 74.3% (best DSPy); MARG Jaccard 44.4 vs. 43.2 (DSPy+MIPROv2); synthetic underwriting 50.0% vs. 47.8%; industry-verified underwriting 36.7% vs. 31.7% (Partner-Fixed) and 23.3% (best DSPy).
• Achieves 100% eventual protocol validity across all four settings, while Naive TextGrad's stability collapses to 0% on MARG review and 56.7% on industry-verified underwriting.
• Ablations show per-example feedback (vs. scalar loss) drives most quality gains (MARG Jaccard 26.9%→38.0%; underwriting accuracy 37.8%→51.1%), while schema+retry mainly fix stability.

Why it matters / caveats: The guarantee only prevents protocol-breaking corruption; it does not guarantee the resulting data-flow messages are semantically optimal. The approach is a lightweight, drop-in software-engineering pattern rather than a new optimization algorithm.

Problem: Existing long-horizon agent benchmarks for business operation either use unreliable, sampled LLM-driven counterparts (stochastic, jailbreakable) or drop negotiation/adversarial dynamics entirely, undermining reproducibility and realism.

Method: E-Commerce Bench simulates a 365-day, up-to-4-store e-commerce operation grounded in real Taobao & Tmall data (6,886 products, 576 suppliers, 152 fraudulent). Both market sides are deterministic: a fixed multi-factor demand model governs customer purchases/returns, and a two-layer deterministic negotiation kernel (extending the counterpart policy of Zhang et al. 2026a) sets every supplier price/concession/accept-reject decision, with an LLM used only to render dialogue (Layer 2 NPC renderer) and a second LLM as the agent under test. 18 frontier and open-weight models are evaluated over 5 episodes each (90 total) on 7 dimensions: final assets, negotiation quality (CSE+), fraud avoidance (BadSpend%), solvency (drawdown/peak), operational efficiency (¥/tool call), execution (controllable return rate), and learning (AnchorRatio).

Key results:

• GPT-5.6 Sol earns the most, turning ¥100,000 into ¥1,431,425 (14.31x), but ranks 16th/18 on fraud avoidance (BadSpend% 18.48).
• Qwen3.8-Max-Preview leads open-weight models at ¥416,252 (4.16x), 38% above GLM 5.2 (high, ¥301k / 3.01x).
• A 1,264x gap separates the top (GPT-5.6 Sol) from the bottom (Qwen3.5-Plus, 0.01x, near-bankrupt); 10 of 90 episodes ended in bankruptcy.
• Across 8,647 repeat purchases of the same item from the same honest supplier, 16 of 18 models show no clear sign of bargaining the price down over the year, indicating weak long-horizon learning.

Why it matters / caveats: No single model dominates all six capability axes — six of seven vendor-family flagships fall below the 18-model median on at least one dimension — showing that total-asset leaderboards hide meaningful weaknesses (e.g., fraud susceptibility, poor learning) even in top-earning models.

Problem: LLM-based peer review/rebuttal agents trained with subjective "LLM-as-judge" or preference-pair rewards produce inconsistent signals and citation hallucinations, prioritizing stylistic mimicry over factual, evidence-grounded critique.

Method: The authors curate 184,857 review-rebuttal trajectories from OpenReview (ICLR 2013–2025, NeurIPS 2021–2024, ICML 2025), downsampled to 72K training / 1K test examples, and train InternReviewer and InternAdvocate (from Qwen3-30B-A3B-Thinking-2507) via agentic RL with Group Sequence Policy Optimization (GSPO). Agents use a local Elasticsearch index of 593,092 arXiv papers with time-aware constrained retrieval (120-day pre-deadline cutoff for reviewer, 90-day post-deadline for advocate). The reward is a linear combination r_total = r_polar + r_format + r_tool + r_cite, where r_polar uses POLAR-7B semantic alignment (tanh-normalized), r_format checks section headers and length (2000–10000 chars), r_tool rewards 4–6 tool-use steps, and r_cite enforces internal citation consistency plus zero-tolerance grounding against retrieved-paper metadata.

Key results:

• Vs. Claude Sonnet 4.5, Gemini 3.1 Pro Preview, GPT-5.2, and base Qwen3-30B-A3B-Thinking-2507: InternReviewer/InternAdvocate achieve top Format (1.5000/1.4910), Tool (1.0000/1.0000), Citation (0.7920/0.7620), and Hallucination (0.8560/0.8620) scores — over 1.5 absolute points above the best closed-source competitor on Citation/Hallucination.
• Best POLAR scores among all systems on both tasks; second-best Reranker score behind GPT-5.2.
• BLEU/ROUGE scores are shown to be counter-intuitive/misleading — the base Qwen3 model matches or beats frontier closed models on lexical metrics despite worse factual grounding.
• Citation reward jumps from -0.9283 to 0.8241 within 100 RL steps for InternReviewer; Hallucination reward improves slowly and continuously across the full training run.

Why it matters / caveats: Demonstrates that even strong frontier closed-source models fail badly on structural/factual grounding dimensions of scholarly writing, and that targeted agentic RL with an objective, verifiable reward can close this gap; lexical metrics (BLEU/ROUGE) are shown unsuitable for this domain.

Problem: There is no architectural framework for language agents that must provide persistent assistance in long-lived settings (evolving user needs, changing procedures) rather than just completing bounded, user-specified tasks.

Method: The paper proposes Pera (Perception-Centered Architecture for Persistent Agents), a conceptual framework analogous to the software-engineering shift from "programming in the small" to "programming in the large." Pera augments cognitive language agents with a perception component (continually sensing external signals from the environment and internal signals from task execution/context) and a control component that converts these signals into "lifecycle tasks" — distinct from bounded "episodic tasks" — which update the agent's context, revise decision procedures, or anticipate emerging needs. The framework draws on Hayes-Roth's adaptive intelligent systems architecture and CoALA-style cognitive architectures.

Key results: Not stated — this is a conceptual/architectural position paper rather than an empirical one; it retrospectively organizes prior work under the Pera framework and offers a detailed case study and forward-looking research directions (e.g., adaptive sensing, coordinated self-evolution, safe self-improvement) rather than reporting benchmark numbers.

Why it matters / caveats: Offers a vocabulary (episodic vs. lifecycle tasks, active perception across task boundaries) for organizing and motivating persistent-agent research, but provides no quantitative validation of the architecture itself.

Problem: Conventional political-communication models treat voter persuasion as a sequence of isolated, linear events and ignore feedback, delay, saturation, and shifting institutional trust.

Method: The paper proposes the Dynamic Democratic Consent (DDC) framework, integrating Edward Bernays' "Engineering of Consent" doctrine with the S-E-E-D system-dynamics behavior modes (Snowball growth, Equilibrium seeking, Elastic adjustment, feedback-loop Dominance). It expands Bernays' 3M resource model (Manpower, Mindpower, Money) into a 4M paradigm by adding algorithmic Media infrastructure, and structurally maps Bernays' eight-step PR process onto system-dynamics variables (reference support target S*, perception time-lag ATp, demographic carrying capacity Kdem, reinforcing loop R1 and balancing loops B1/B2) across three phases: System Calibration, System Architecture, and Delay/Execution Control. A four-stage diagnostic process is proposed for describing trajectories, identifying feedback structures, designing interventions, and examining behavioral shifts.

Key results: Not stated (the paper presents qualitative/simulated reference-mode illustrations rather than quantitative benchmark results; no empirical dataset or numeric outcomes are reported).

Why it matters / caveats: The framework is explicitly exploratory rather than predictive; the authors list seven limitations including non-empirical/uncalibrated parameters, voter homogeneity assumptions, simplified institutional structure, and unresolved endogeneity, stating that empirical estimation, comparative case testing, sensitivity analysis, and ethical evaluation are required before applied use.

Problem: Existing latent-diffusion SAR-to-EO (synthetic aperture radar to electro-optical) translation methods inherit a fixed pretrained natural-image autoencoder without checking whether its codec actually preserves SAR/EO reconstruction fidelity.

Method: ReFlowSET first audits multiple pretrained VAEs (SD2.1, SDXL, SD3.0, SD3.5, FLUX.1, FLUX.2) on joint SAR–EO round-trip reconstruction PSNR and selects FLUX.2 as the codec (best in 6/8 dataset-modality settings, +7.49 dB average EO reconstruction over SD2.1). Rather than reusing that codec's heavyweight generator, it trains a 509.3M-parameter, 24-block conditional Diffusion Transformer (DiT) from scratch via conditional flow matching, with dual-stream SAR conditioning for the first r=8 blocks before joint fusion/refinement, plus a training-only representation-alignment loss (cosine distance, λ=0.5) aligning noisy-EO DiT features to a frozen DINOv3 ViT-L/16 teacher's clean-EO features (no inference-time cost).

Key results:

• On QXS-SAROPT: best DISTS (0.231) and ties best FID (19.1) among all compared methods (pix2pix, CycleGAN, SD2.1 fine-tune, BBDM, ControlNet, cBBDM, C-DiffSET, etc.).
• On SAR2Opt: improves FID from second-best 71.8 (SD2.1) to 66.3, and DISTS from 0.211 to 0.185 (relative gains of 7.7% and 12.3%).
• Ablations: switching from SD2.1 to the selected FLUX.2 codec reduces FID from 25.5→19.1 (QXS-SAROPT) and 84.5→66.3 (SAR2Opt); representation alignment loss LRe reduces FID from 71.851 to 70.436.

Why it matters / caveats: Shows codec choice is a consequential, previously overlooked design variable for SAR-to-EO translation; the authors note ReFlowSET does not maximize PSNR/SSIM, attributing this to the pixel-aligned metrics' sensitivity to inherent SAR-EO ambiguity and local misalignment rather than to lower fidelity.

Problem: It is unclear under what conditions AI assistance to AI R&D crosses from ordinary acceleration into a self-amplifying ("recursive self-improvement") regime, and how feedback delay and hardening research difficulty shape that transition.

Method: The paper develops a minimal dynamical model of an AI-enabled R&D system with four elements: exogenous baseline research productivity r(t), recursive gain a(x,t) with end-to-end delay τ(t), operational closure χ(x,t) (fraction of potential gain surviving the development pipeline), and frontier-hardening rate σ(x). It derives a recursive reproduction number RAI = χa/σ (an epidemiological-style threshold quantity) from stability analysis of the resulting delay-differential system, then extends the model to physical deployment constraints and to multiple coupled research actors via a reproduction matrix whose spectral radius ρ(K) determines ecosystem-level criticality.

Key results:

• RAI > 1 implies incremental capability gains amplify across development cycles; RAI < 1 implies they damp; the critical boundary does not depend on baseline research throughput.
• At high research throughput, development-cycle duration becomes the limiting timescale for amplification; increasing frontier hardness can return a supercritical system to subcritical.
• In a multi-actor scenario, individual within-organization reproduction numbers can each be below 1 while cross-actor transfer raises the network reproduction number to ρ(K) ≈ 1.15, making the overall ecosystem supercritical.

Why it matters / caveats: Reframes RSI as a measurable stability property of an R&D system rather than tied to a specific capability threshold like AGI; explicitly states its numerical scenarios are "conditional demonstrations and not probabilistic forecasts," and calls for empirical measurement of recursive gain, operational closure, cycle duration, and frontier hardening.

Problem: Free-form chain-of-thought traces in multimodal geometry reasoning hide the specific visual decisions that determine correctness, and trajectory-level RL (e.g., GRPO) assigns only one terminal reward across the whole response, preventing localized credit assignment.

Method: The authors introduce Code-CoT, which represents diagrams as line-addressable executable Matplotlib perception code and organizes reasoning into typed events (reference, auxiliary, coordinate, think) ending in a final answer, installed via supervised fine-tuning. They then introduce Critical-Event Group Relative Policy Optimization (CE-GRPO), which selects event boundaries using structural priors and type-normalized entropy, samples multiple complete continuations from shared prefixes at those boundaries, and converts differences in terminal outcomes into localized, event-conditioned advantages (excluding the shared prefix from the policy loss), jointly optimized with ordinary GRPO. Backbone model: Qwen3-VL-8B.

Key results:

• Across nine geometry benchmarks (MathVerse, VisOnlyQA-Syn/Real, MathVista-GPS, Geometry3K, PGPS9K, GeoQA, GeoLaux, MM-Math), CE-GRPO reaches average accuracy 76.04, beating the Qwen3-VL-8B backbone by +8.09, Code-CoT SFT by +6.49, and trajectory-level GRPO by +3.43 points, improving over the backbone on all nine benchmarks.
• On validly terminated responses only, retains a 3.91-point average gain over trajectory-level GRPO (isolating solution-quality gains from protocol-compliance effects).
• Combining diagram + externally generated code (I+C) reduces the Qwen2.5-VL-7B Text-Dominant–Vision-Only gap from 13.3 to 3.0 points; self-generated code trails externally generated code by 11.7–16.6 points for open models, identifying reliable code generation as the main bottleneck.
• Structural event selection identifies outcome-changing events ~30% more often than random selection; CE-GRPO's margin over baseline grows by 3.77 points per additional intermediate event (r=0.866, p=0.0016).

Why it matters / caveats: Demonstrates that aligning the RL credit-assignment unit with the semantic units exposed at inference ("credit-addressable reasoning") particularly helps long, dependency-heavy reasoning chains; relative advantage over baselines increases with the number of intermediate events.