AI papers — 2026-07-21

Problem: Video multimodal LLMs struggle to precisely localize when described events occur, since existing video temporal grounding (VTG) methods rely on holistic single-pass reasoning over uniformly sampled frames and thus fail to pinpoint accurate boundaries in long, information-dense videos.

Method: TimeLens2 reframes VTG as a coarse-to-fine, set-valued interval prediction task built on Qwen3-VL (2B/4B/8B) backbones, trained via a two-stage recipe: (1) long-context supervised fine-tuning on a newly curated 93K-instance dataset (TimeLens2-93K) built through a staged pipeline (hierarchical time-stamped captioning, segment-aware query synthesis, dual-agent independent relocalization by Qwen3-VL-30B-A3B and TimeLens-8B, cross-agent temporal-consensus/semantic verification, and boundary-focused local refinement), combined with instruction/response-format diversity and rollout-guided hard-sample mining; and (2) GRPO reinforcement learning using a reward of tIoU plus a novel temporal Wasserstein reward (exact 1D W1 distance between uniform distributions over merged predicted/target interval supports) minus an invalid-output penalty, which provides dense, matching-free, fragmentation-invariant feedback for disjoint or unequal-cardinality multi-span predictions.

Key results:

• Across seven benchmarks (Charades-STA, ActivityNet-Captions, QVHighlights, VUE-TR, VUE-TR-V2, MomentSeeker, Ego4D-NLQ), TimeLens2-2B/4B/8B reach average mIoU of 44.5, 47.7, and 48.0 respectively, improving over their Qwen3-VL backbones by 14.2, 13.0, and 18.1 mIoU points.
• TimeLens2-4B outperforms Qwen3.5-397B-A17B on every benchmark by 7.5 points mIoU on average, and TimeLens2-8B beats all open-source and most proprietary models, trailing Gemini 2.5 Pro on QVHighlights by only 0.2 mIoU points.
• Progressive label curation (consensus + semantic verification + boundary refinement) improves Qwen3-VL-4B SFT from 42.0 to 45.8 average mIoU using only 93.2K vs. 735.4K raw QA turns (roughly one-eighth the labels).
• The temporal Wasserstein reward improves average mIoU from 47.0 (tIoU-only) to 47.7, reduces constant (uninformative) GRPO reward groups from 13.8% to 3.6%, and rescues 75.8% of all-zero-tIoU reward groups.

Why it matters / caveats: The coarse-to-fine reasoning approach plus set-aware supervision and reward design lets compact 2B–8B models match or beat much larger (up to 397B-parameter) and proprietary systems on precise temporal localization, suggesting data/reward quality can substitute for scale in VTG. Boundary refinement slightly hurts VUE-TR and MomentSeeker scores (-0.5 points) even while improving the overall average.

Problem: Existing role-play and literary simulation systems treat character personas and world states as static or fixed-schema, failing to capture how characters and their fictional worlds should co-evolve together over a long-horizon narrative.

Method: EvolvingWorld couples two open-schema LLM-driven modules — a Character Agent that maintains per-character open-schema profiles plus a "hidden tracker" for accumulating weak/emerging evidence before committing profile updates, and a World Model that tracks an open-schema global state plus location/entity-level physical states. The simulation is decomposed into 7 trainable supervised tasks executed per scene (scene_cast, location_scenario, motivation_update, next_character, interaction_gen, world_update, character_update). Training data was extracted from 57 chronologically-narrated public-domain books using Gemini-2.5-Pro, producing 138,596 supervised samples; open-source backbones (Llama-3.1-8B-Instruct, Qwen2.5-7/14/32B-Instruct, Qwen3-4B-Instruct) were fine-tuned via SFT mixed 1:1 with Tulu3 general instruction data. Evaluation uses a trajectory-level LLM-as-Judge protocol (Claude-4.6-Sonnet) across 10 dimensions and 20 metrics.

Key results:

• Qwen-32B fine-tuned on EvolvingWorld reaches a World Model average of 59.87, surpassing untrained Claude-4.6-Sonnet (57.54) and Gemini-2.5-Flash (59.76).
• Best overall closed-source performance: Claude-4.6-Opus scores 94.97 (Character) and 77.76 (World) average.
• On matched backbones, EvolvingWorld-trained agents substantially outperform role-play-only baselines CoSER and Crab (e.g., Qwen-7B EW Character avg. 45.53 vs. 18.32 for Crab and 18.98 for CoSER).
• Versus BookWorld on shared metrics, EvolvingWorld achieves higher averages across all four tested backbones (e.g., GPT-5.3-Chat Character avg. 85.52 vs. 70.64), and BookWorld's coherence metrics degrade over longer trajectories (dropping from ~40 to ~24 over 20+ scenes) while EvolvingWorld's improve.

Why it matters / caveats: Persistent, coupled character/world state tracking — not just dialogue imitation — enables long-horizon narrative simulation without quality degradation. Caveats: the world is modeled as one objective state shared by all characters (no subjective/differing perceptions), representations are constrained by LLM context length, and the benchmark uses only public-domain Project Gutenberg books.

Problem: Training tool-use web search agents to self-improve is hard because supervised fine-tuning on fixed teacher-distilled trajectories saturates quickly, while sparse-reward reinforcement learning gives weak, trajectory-level guidance for long-horizon multi-step search.

Method: The authors build DeepSearch-World, a deterministic offline Wikipedia environment (~10M-entry corpus, BM25/Pyserini search plus SQLite-indexed visit tool) paired with 420K multi-hop QA tasks from entity-level random walks over the Wikipedia hyperlink graph, plus a 377-instance held-out set (DeepSearch-Val). On top of this they run DeepSearch-Evolve, a self-distillation loop: a scaffold teacher agent (Plan/Act/End phases with working-memory tracking and rule-based reflection on failed tool calls) generates trajectories from the current model (starting from Qwen3.5-9B), filtered via rejection sampling and an LLM quality filter, converted into standard ReAct-style supervision, and used for SFT with importance-sampling-based data mixing across rounds — run for 11 rounds, followed by a GRPO stage on 1,600 real-tool (SerpAPI + Jina) instances.

Key results:

• DeepSearch-World-9B reaches 31.2% on BrowseComp, 61.5% on GAIA, and 93.4% on HotpotQA, competitive with agents distilled from stronger proprietary teachers.
• Against its Qwen3.5-9B-Instruct backbone: +23.8 BrowseComp, +22.9 BrowseComp-ZH, +9.0 HLE, +37.6 GAIA, +29.0 xbench, +48.1 HotpotQA.
• Tool-use behavior improves substantially: average interaction rounds rise from 4.7 to 18.0, and an LLM-judged "Advanced Capability Score" rises from 19% to 70%.
• Ablations: rejection sampling alone raises SearchQA from 46.4% to 54.9%, combined with quality filtering reaches 58.2% (vs. 49.0%/47.8% for OPSD-SDAR/Skill-SD baselines); full pipeline reaches 31.9% on DeepSearch-Val vs. 25.0% for vanilla SFT, dropping to 14.8% without reflection rewriting and state internalization.

Why it matters / caveats: A deterministic, verifiable offline environment can substitute for part of the supervision that otherwise requires stronger proprietary teacher models, enabling scalable self-evolution for long-horizon web agents. The environment is Wikipedia-only and training data is English-only (explaining the smaller BrowseComp-ZH gain).

Problem: Existing context-pruning methods for coding agents reconstruct the pruning signal externally via fixed metrics or a separate scoring model with an explicit goal-hint query, even though the agent's own backbone has already processed the tool output.

Method: SWE-Pruner Pro attaches a lightweight, non-linear pruning head directly to a frozen agent backbone's last-layer hidden states during the existing prefill of each tool response, requiring no extra model call. The head combines a length-aware embedding (keyed to line count, zero-initialized) with a per-token feed-forward classifier (LayerNorm, two Linear-GELU-Dropout blocks, final Linear), with per-line decisions made by majority vote over per-token sigmoid outputs. It is trained on 22,609 samples labeled per-line by Claude Sonnet 4.6, using a per-sample balanced focal loss (γ=2) that averages loss separately over each sample's keep/prune tokens. Backbones tested are Qwen3-Coder-Next (80B MoE, 3B active) and MiMo-V2-Flash (309B MoE, 15B active).

Key results:

• Saves up to 39% of prompt/completion tokens on SWE-QA-Pro and 30% on Oolong, the only method among 7 evaluated that reduces tokens in every setting.
• On SWE-Bench Verified, raises MiMo-V2-Flash's resolve rate by +3.8% (345/500 vs 326/500) with +7.4% input tokens; on Qwen3-Coder-Next it loses only 1.2% resolve rate while cutting input tokens 13.5%.
• Improves MiMo-V2-Flash's Oolong accuracy by +2.2 points (94.6 vs 92.4) while cutting tokens 30.1%.
• Linear probe on frozen hidden states achieves AUC 0.83 and best-F1 0.63 (vs majority-class F1 upper bound 0.46); the balanced focal loss beats plain BCE by +1.13 judge score.

Why it matters / caveats: Coding-agent backbones already encode line-level relevance judgments internally, so pruning can be read out cheaply rather than reconstructed by an external scorer. Evaluation covers only open-weight models exposing hidden states (requiring per-backbone retraining), benchmarks are largely Python-centric, and on Qwen3-Coder-Next API call counts actually increase despite token savings.

Problem: Human-object centric video personalization (HOCVP) struggles to jointly handle inter-subject settings (balancing subject fidelity with accurate human-object interactions, especially for abstract references like logos) and intra-subject settings (leveraging multiple references of the same subject to boost fidelity).

Method: HOMIE builds on a DiT-based text-to-video backbone (Wan2.1-14B / Wan2.2-14B) with a 3D VAE and UmT5 text encoder trained via flow matching, and feeds reference images/prompts through an MLLM (Qwen3-VL-2B-Thinking) whose hidden-state features are aligned to VAE token space via a lightweight MLP and concatenated into the token stream. It introduces Global Multimodal Guidance (GMG), which pools MLLM query/key features to affine-modulate video tokens' query/key projections in self-attention, and Modality-Reference Embedding (MRE), a learnable embedding tagging tokens by modality and reference identity. Training data is curated from OpenS2V-5M, PhantomData, and a self-curated 20K-clip segmentation-mask dataset, trained in a three-stage paradigm consuming ~11K A100 GPU-hours.

Key results:

• HOMIE (Wan2.1-14B) achieves a 21.8% relative OCR accuracy improvement over SkyReels-V3 (0.452 vs. 0.326), with best DINO-I (0.543) and Obj-Sim (0.891).
• On multi-view intra-subject HOCVP, HOMIE (Wan2.2) achieves DINOrec 0.696, best among compared methods (vs. 0.654 for SkyReels-V3).
• Ablations: removing GMG drops Face-Sim from 0.786 to 0.697 and OCR Acc. from 0.452 to 0.430; removing MRE drops OCR Acc. to 0.376.
• User study (40 participants): HOMIE preferred over Kling, Phantom, SkyReels-V3, and UniVideo in 66.1% of subject-consistency judgments and 64.7% of overall-quality judgments.

Why it matters / caveats: Injecting MLLM semantic knowledge via attention-level modulation (rather than replacing/aligning with the text encoder) preserves controllability while enabling reasoning-driven placements and intra-subject fidelity gains. Results rely on a self-curated 200-combination evaluation set and GPT-5.2/DINO-based automatic metrics rather than large public benchmarks.

Problem: Existing video-generation benchmarks judge physical plausibility only from final outputs, so they cannot tell whether a model's apparent physical understanding comes from genuinely applying physical laws or merely producing visually plausible motion.

Method: The authors build Apple-π, anchored on a 400-video dataset called Orchard (243 NVIDIA Isaac Sim simulated cases, 121 self-recorded lab videos, 36 Internet-sourced clips) spanning ten classical-mechanics tasks, using a standardized object vocabulary (sphere/cube/cylinder/cone). Each case is evaluated via a three-stage Perception→Formulation→Deduction protocol (five subtracks), where models are given an infographic-annotated first frame plus a chain-of-frames prompt and must respond with a generated video treated as their visible reasoning trace. Scoring combines MLLM-based subjective rubric scoring (Gemini 3 Flash as judge) with physics-law-grounded objective metrics (segmentation IoU, masked PSNR, spatiotemporal IoU, velocity accuracy).

Key results:

• Across 11 benchmarked models, the best pure video model, Seedance 2.0, scored only 0.473 average (0-1 scale).
• Unified models GPT Image 2 and Nano Banana 2 scored highest overall at 0.704 and 0.699, though even they scored only ~0.40 on Deduction.
• Clear stage-wise "reasoning funnel" (Perception easiest, Formulation harder, Deduction hardest), a pillar-wise gap where multi-law composed tasks score lower than single-law tasks, and a persistent Sim-to-Real gap (e.g., Seedance 2.0: 0.487 simulated vs. 0.459 real).

Why it matters / caveats: Current video models capture useful surface-level physical priors from large-scale training but lack reliable, law-grounded reasoning, especially in composing laws across stages and generalizing to real-world footage — suggesting future world models need explicit understanding modules and reasoning-oriented training.

Problem: Embodied foundation models need visual understanding that is not just accurate but explicitly grounded in physical 3D space and transferable to real-robot action, a requirement general MLLMs and even prior RynnBrain 1.0 did not fully satisfy.

Method: RynnBrain 1.1 is a family of decoder-only vision-language embodied models (2B, 9B, 122B-A10B) built on Qwen3.5 backbones with DeepStack and Interleaved MRoPE, trained via unified autoregressive spatio-temporal pretraining that predicts text alongside discretized spatial tokens (boxes, points, trajectories). New in 1.1: (1) contact-point prediction representing grasps as a center-point-plus-rotation-angle target (replacing 1.0's four-corner grasp rectangle), trained on a 2.6M-sample corpus; and (2) native 3D grounding for the 2B/9B models, predicting 9-DoF 3D bounding boxes. Downstream, RynnBrain-VLA post-trains the base model as a single-stream flow-matching Diffusion Transformer with a unified 81-dimensional cross-embodiment action space, deployed on Unitree G1, Astribot-S1, and Tianji-Wuji robots.

Key results:

• RynnBrain 1.1-122B-A10B outperforms all evaluated models on VSI-Bench (75.0), MMSI (52.0), and RefSpatial-Bench (79.1).
• At 9B scale, MindCube jumps from 56.6 (v1.0) to 86.9 and MMSI from 39.6 to 47.0.
• On SUN RGB-D 3D grounding, the 2B model reaches 34.28 AP@15, beating Seed1.5-VL (33.5) and Gemini 2.0 Pro (32.5).
• Real-robot: RynnBrain-VLA vs. Qwen-Based-VLA raises average process score from 68.33% to 91.28% and success rate from 60.00% to 86.67%; joint multi-task training raises success rate further to 91.67%.

Why it matters / caveats: Embodied-specific pretraining is a prerequisite for reasoning-intensive spatial capability — Qwen3.5 baselines show negative scaling (-39.2%) on such benchmarks from 2B to 122B-A10B, while RynnBrain 1.1 improves (+38.6%). Contact-point prediction is evaluated only qualitatively, and 3D grounding was only added to the 2B/9B models, not the 122B-A10B flagship.

Problem: Existing multimodal memory systems for video streams organize memory around frames/tokens inside a bounded model context rather than around persistent entities, limiting their ability to track recurring people/objects and scale to open-ended (unbounded) streams.

Method: ReflectWorld-MM has three parts: (1) a perception front-end producing entity-resolved observations, where re-identification supplies candidate evidence and a deterministic resolver commits identity decisions; (2) a hierarchical long-term memory grounded in Tulving's memory theory, combining multi-scale episodic memory (entity, trace, schema levels), an evolving entity-centric semantic memory consolidated every N=5 observations via Add/Update/Delete edits with an importance score, and a procedural memory for user rules; (3) a database-backed service (text-embedding-3-small, GPT-5-mini for extraction/consolidation, GPT-5 as query-answering agent).

Key results:

• Achieves best accuracy on all six evaluated benchmarks (VideoMME-Long, LVBench, HippoVlog, EgoLife-QA, M3-bench-robot, M3-bench-web), beating M3-Agent and WorldMM.
• Largest gains on entity-centric M3-bench: +10.4 points over M3-Agent on the web split and +9.1 points on the robot split.
• On EgoLife-QA reaches 46.8 accuracy, exceeding even the original EgoLife paper's 45.5 with manually annotated identity descriptions.
• Only 4.6% video fallback on EgoLife-QA vs. WorldMM's 34.0% fallback on VideoMME-L, while achieving top accuracy.

Why it matters / caveats: Entity-oriented, revisable (non-append-only) memory lets a system track identity and durable knowledge across truly open-ended streams and integrate with arbitrary live sources/assistants. Absolute accuracy on M3-bench remains low for all systems, indicating the benchmark is far from solved.

Problem: Reinforcement learning of LLMs over heterogeneous task mixtures suffers because GRPO's per-group reward normalization induces an entropy-dependent structural bias, making advantage signals across prompt groups statistically non-comparable and causing imbalanced optimization across tasks.

Method: The authors propose Group Entropy-Controlled Policy Optimization (GEPO), computing a per-prompt "group entropy" estimator from the existing K sampled responses (no extra sampling), then applying entropy-conditioned asymmetric advantage shaping: it attenuates positive advantages by α_low in low-entropy groups and attenuates negative advantages by α_high in high-entropy groups (α_high < α_low). Entropy thresholds are set adaptively from the batch's group-entropy mean/std, smoothed via an EMA. Evaluated on Intern-S1-mini and Qwen3.5-9B, trained on an in-house multimodal dataset (math, physics, science, code, instruction following), against GRPO, AEPO, Clip-Cov, and KL-Cov baselines.

Key results:

• On Intern-S1-mini, GEPO achieves the best average score across 13 benchmarks (54.2) vs. GRPO (51.5), winning 7 of 13 individual benchmarks (e.g., AIME25 82.0 vs. GRPO's 75.7).
• On Qwen3.5-9B, GEPO reaches the highest average (71.2) vs. GRPO (70.7).
• GRPO exhibits a catastrophic training collapse on Qwen3.5-9B around step 170 (AIME25 accuracy dropping from ~85% to ~40%), while GEPO shows smooth, monotonic validation curves.
• Ablations: removing high-entropy control drops the average to 51.3, removing low-entropy control to 52.6, removing asymmetric shaping to 53.0 (vs. full GEPO's 54.2).

Why it matters / caveats: GEPO offers a zero-extra-sampling-cost, model-agnostic mechanism for balancing exploration-exploitation across heterogeneous RL task mixtures, improving both accuracy and training stability. Results are from a single in-house dataset and only two base models.

Problem: Multilingual ASR quality is highly uneven across languages, with underrepresented Central Asian languages (Kazakh, Kyrgyz, Uzbek) suffering severe data scarcity and poor recognition performance compared to high-resource languages.

Method: The authors extend GigaAM, a Conformer-based self-supervised encoder, to a 600M-parameter, 24-layer multilingual Conformer with rotary position embeddings, pre-trained via HuBERT-style masked unit prediction (K-means, K=1000) on 2 million hours of audio spanning 70+ languages. To counteract head-language dominance, they build a weighted language co-occurrence graph, apply community-detection clustering into five language clusters, and use cluster-level sampling weights during pre-training. For fine-tuning, they train a CTC-based ASR model with a shared character vocabulary across English, Russian, Kazakh, Kyrgyz, and Uzbek using domain-aware sampling.

Key results:

• On internal in-the-wild test sets, GigaAM Multilingual achieves WER of 15.8% (Kazakh), 9.8% (Kyrgyz), and 12.7% (Uzbek), versus Whisper large v3 at 65.2%, 102.2%, and 120.6%.
• Pre-training cluster reweighting improved Kyrgyz WER from 9.4% to 8.5% and Uzbek from 10.5% to 9.7%.
• Under matched CTC fine-tuning, even a smaller 240M encoder beat Whisper Large v3 CTC and Omnilingual SSL 1B CTC on average WER (12.2% vs. 14.1% vs. 16.6%).
• On tail languages not targeted by fine-tuning data (Bashkir, Georgian), GigaAM achieved 3.6%/3.8% WER vs. Whisper's 11.1%/13.4%.

Why it matters / caveats: Explicit cluster-level data balancing during pre-training and domain-aware sampling during fine-tuning matter more than raw model scale for low-resource language adaptation. Comparisons are based on the authors' own controlled fine-tuning pipeline rather than baselines' original released checkpoints.

Problem: Audio-conditioned LLMs struggle to produce reliable, verifiable timestamped answers when reasoning over long recordings (up to two hours), since time is not natively represented in audio token streams.

Method: GigaChat Audio attaches an encoder→subsampler→projector audio front-end (HuBERT-style pretraining on 2M hours of unlabeled multilingual audio) to a 10B-A1.8B MoE text checkpoint (GigaChat3-10B-A1.8B, 256k context), then fine-tunes on audio SFT data interleaving continuous audio tokens with periodic "inter-timing" markers (plain-text hh:mm:ss or dedicated timing tokens) every 60 seconds. Temporal supervision is generated via a cascaded synthetic pipeline: YODAS2 English audio transcribed/word-aligned with WhisperX, filtered down to 14k hours, sliced into ~10-minute fragments, and used by GPT-OSS-120B to generate QA/description/summarization pairs with a separate verifier checking consistency.

Key results:

• On 20–40 min temporal grounding, reaches 53.8 mIoU (60s anchors) vs. Qwen3-Omni-30B-A3B at 3.6 mIoU and Gemini 3 Flash at 56.1 mIoU; 7s anchors improve this to 65.2 mIoU.
• Removing periodic inter-timing anchors collapses long-form grounding from 53.8 to 14.2 mIoU.
• Anchor frequency trades accuracy for compute: mIoU/MAE go from 63.0/1.5s (7s anchors) to 31.0/8.5s (240s anchors).
• On AMI Corpus meeting understanding, achieves 3.50s MAE vs. 290.5s for Qwen3-Omni-30B and 1.00s for Gemini 3 Flash.

Why it matters / caveats: Periodic temporal anchors and duration-mixture training are necessary for generalizing temporal grounding across short and long audio. Results rely on LLM-generated synthetic supervision and LLM-as-a-judge evaluation rather than fully human-annotated ground truth.

Problem: Video editing models currently need labour-intensive, poorly scalable data pipelines (mask annotation, error-introducing I2V/ControlNet synthesis, VLM-based filtering), severely limiting task diversity compared to image editing.

Method: FlowMimic post-trains the pretrained Wan2.1-T2V-1.3B DiT model, adding a reference-inject self-attention mask, separate 3D RoPE with frame correspondence, and element identity encoding. Its core mechanism is a "pixel-pair temporal warped flow field": a shared, randomized, time-varying deformation grid (pan, zoom, rotation, stretch, elastic warps) is applied identically to the source and target images of any existing image-editing pair, generating temporally-coherent online video-editing pairs without curated video data. Training also uses a first-frame "modality mimic generation loss" (KL divergence between T2I and T2V-first-frame latents) and a "modality mimic editing loss" (mutual imitation between V2V and I2I).

Key results:

• No quantitative benchmark metrics (CLIP score, FID, PSNR, user-study win rates) are reported in the extracted text; evaluation on FiVE-Bench and UNIC-Bench is presented only qualitatively via figure comparisons.
• Training used a low learning rate (5×10⁻⁶ to 1×10⁻⁵ over 5000 steps) on 32 mid-range 80GB GPUs.
• Qualitatively shown to produce more camera-authentic T2I outputs, more accurate referring-expression localization on UNIC-Bench, and emergent SAM3-like referring-expression segmentation/tracking despite never being trained on curated video editing data.

Why it matters / caveats: Video editing capability can potentially be learned purely from image-editing pairs plus a cheap synthetic warping procedure, sidestepping expensive video data curation. However, the paper's own reported evidence is entirely qualitative rather than quantitative, and it uses a relatively small 1.3B model.

Problem: Training API-calling LLM agents requires massive multi-step trajectory data, but existing synthetic data generation methods depend on fully implemented executable environments and populated backend databases, creating a major scalability bottleneck.

Method: The paper proposes ESAT (Environment-free Synthetic Agentic Trajectory generation), a three-stage pipeline needing only API specifications as input: (1) task synthesis via LLM generation across a bucketized difficulty/action-type/task-focus/app-count grid with inverse-frequency sampling; (2) trajectory synthesis, where a teacher agent (GLM-5.1-FP8) issues API calls while an LLM-based simulator (also GLM-5.1-FP8) generates schema-validated, state-consistent synthetic API responses; and (3) trajectory filtering via an LLM judge (Gemini-3.1-Pro) discarding low-quality trajectories via majority-vote correctness. Resulting trajectories are used for standard SFT on Qwen3 (1.7B-14B) and Qwen3.5 (2B-27B) models.

Key results:

• On AppWorld, fine-tuning on synthetic ESAT-S52 data (from apps unrelated to AppWorld) yields gains of 8.4–47.0% TGC over zero-shot base models; on OfficeBench, gains of 5.2–60.5% pass rate.
• Combined ESAT data outperforms AWT (634 trajectories from the real AppWorld environment) by 0.7–15.3 points across eight models.
• The trajectory-filtering LLM judge achieves 95.2% precision against ground-truth AppWorld verifiers and 95% agreement with human annotators.
• The API simulator produces valid responses for 93.7% of ~27K simulated calls, but failure rate rises from 2-3% for short responses to 23% for responses over 2K tokens.

Why it matters / caveats: Effective agent-training supervision can be generated purely from API specifications without building or executing real environments, making data generation scalable to new API ecosystems. The simulator's failure rate rises notably for long outputs, attributed to limitations of the specific simulator LLM used.

Problem: Existing egocentric video resources for embodied learning rarely combine low-cost continuous capture, manipulation-level structured annotations, and reusable tools for downstream robot learning in one open pipeline.

Method: Open-AoE is data infrastructure built on the AoE consumer-smartphone capture framework, using a four-stage pipeline (edge-side gating, offline quality check/scene labeling via Qwen3.7-Plus, reconstruction/annotation, three-gate quality inspection) to convert raw smartphone video into structured samples. Camera trajectories are recovered with DROID-W (SLAM), hands reconstructed as 3D MANO meshes via a two-hand detector plus HaWoR with global bundle adjustment, and videos segmented into atomic actions with human-in-the-loop correction. A downstream toolchain converts aligned signals into multiple action representations feeding VLA policies, World Action Models, and World Models.

Key results:

• The first release contains ~2,000 hours of egocentric manipulation video from 500+ contributors using 400+ smartphone models, covering 400+ scenes and 8,000+ tasks.
• On a 100-hour audit: 32,407 distinct natural-language action descriptions, 175 action verbs, 8,030 object strings, 99.99% temporal annotation coverage.
• CLIP-embedding diversity analysis shows Open-AoE has the highest mean on all six metrics vs. OpenEgo, EgoDex, EgoXtreme, e.g., Effective Rank 97.43 vs. 76.9/85.2/65.3.
• Training-window yield reaches ~1,760 candidate history-future windows per hour (97.8% of theoretical ceiling), the only dataset with all five supervision modalities near-universally available.

Why it matters / caveats: Open-AoE is positioned as reusable, community-extensible open infrastructure spanning capture-to-training, potentially lowering barriers for VLA/world-model research. The authors note the benefit of camera-domain diversity for downstream transfer remains an untested training hypothesis requiring controlled ablations.

Problem: Existing 3D generative approaches either force watertight topology (implicit fields) or rely on resolution-dependent binary occupancy maps (geometry images), causing staircase artifacts and non-differentiable mesh reconstruction unsuitable for thin-shell, non-manifold geometries like garments.

Method: DiffGI replaces binary occupancy maps with a continuous 2D Truncated Signed Distance Function (TSDF) computed on a 1024×1024 UV grid, and introduces Differentiable Marching Squares (DMS), a vectorized PyTorch algorithm using analytical linear interpolation to extract mesh vertices while preserving gradients through ambiguous saddle-point cases. A DiffGI-VAE, initialized from Stable Diffusion 1.5 weights, compresses geometry images into a 32×32×4 latent using position/TSDF losses plus a geometry-aware normal rendering loss. A DiT-based or UNet-Tiny latent diffusion model with flow-matching then performs label-, image-, or occupancy-conditioned 3D generation.

Key results:

• On GarmageSet VAE reconstruction, DiffGI achieves CD 0.46×10⁻³ and NC 0.96, versus GarmageNet's 1.31–2.19×10⁻³ CD.
• Image-to-3D on GarmageSet: DiffGI achieves CD 1.35×10⁻² with only ~23K vertices, versus TRELLIS (109K vertices, CD 3.44×10⁻²).
• Label-conditioned generation on ABO: mean P-FID improves from 31.14 (Omages) to 19.91 (DiffGI).
• Inference: image-conditioned generation runs in ~1.2s on an RTX 4070 (3.22GB VRAM) vs. TRELLIS requiring 16.28GB VRAM.

Why it matters / caveats: The continuous TSDF plus differentiable mesh extraction enables real-time, edge-deployable 3D generation of thin-shell/non-manifold surfaces with UV-friendly output. Limitations include rounding artifacts on sharp mechanical edges, no RGB/PBR texture generation, and visible seams at UV chart boundaries.

Problem: Reinforcement learning post-training on non-verifiable, open-ended tasks compresses an LLM-as-a-Judge's rich rubric-based evaluation into a single scalar reward, discarding textual feedback and making distinct high-quality responses indistinguishable to the optimizer.

Method: The paper proposes Experiential Learning (EL), repurposing the feedback model as an "LLM-as-a-Coach" that analyzes an on-policy response against prompt-specific rubrics and distills the assessment into transferable "experiential knowledge." This knowledge is prepended as context to a teacher model, and the policy is trained via on-policy context distillation minimizing token-level reverse KL divergence between the policy and the context-conditioned teacher. Training data is 7,500 WildChat-IF prompts with GPT-4o-generated rubrics; baseline is Rubric-as-Reward via GRPO, evaluated using Qwen3-8B and OLMo-3-7B-Instruct as policies.

Key results:

• On held-out WildChat test, EL improves scores over RL for all four policy+feedback combinations (e.g., OLMo-3-7B+OLMo-3-7B: RL 76.0 to EL 77.1).
• On unseen benchmarks, larger gains: AlpacaEval v2.0 win rate for OLMo-3-7B rises from RL 45.9% to EL 50.8%.
• The theoretical feedback bandwidth of experiential knowledge (~17,600 bits/sample) is estimated at over 5,000x that of a discrete 1-10 scalar reward (3.3 bits).
• EL shows smaller gains than RL on the training set but larger gains on out-of-distribution benchmarks, indicating RL overfits/reward-hacks the training distribution while EL transfers better.

Why it matters / caveats: Higher-bandwidth, distributional textual feedback can provide denser supervision and mitigate reward hacking compared to scalar-reward RL. The bandwidth calculation is a theoretical upper bound, not a measure of usable supervision, and EL doesn't address the underlying calibration/accuracy of the feedback model itself.

Problem: Synthesizing hand-object interaction (HOI) videos that stay visually and geometrically consistent across multiple synchronized viewpoints from a single reference image is challenging because existing video generation methods lack explicit 3D geometry/motion modeling.

Method: HarmoHOI builds on pretrained WAN 2.1-1.3B-T2V and adds a Mixture of Multi-view Diffusion Transformer (M2DiT) that jointly denoises RGB video latents and "motion pseudo-video" latents (3D point tracks, depth-colormapped) using intra-view, inter-view, and cross-branch attention, plus a learnable scale token; and Global Motion Aligning Diffusion (GloMAD), built on Point Transformer V3, that refines coarse point tracks into globally aligned metric-scale 3D trajectories. Training uses a three-stage progressive curriculum: warm-up on single-view HOI videos (HOIGen1M), synchronized multi-view videos rendered in Unreal Engine 5 (SynCamVideo), then fine-tuning on the lab-captured multi-view TACO dataset.

Key results:

• HarmoHOI achieves Matching Pixels 535.8 and CLIP-Views 83.18, both best among compared methods (vs. SynCamMaster 410.4/78.97).
• 3D motion accuracy: Chamfer distance 0.0097 and multi-view RPE/PI of 34.7/34.24, versus the next-best baseline (Depth Anything 3) at Chamfer 0.0241 and RPE/PI of 61.9/11.05.
• Ablations show removing simultaneous multi-view generation drops Matching Pixels to 335.1; removing GloMAD yields 503.7/47.6 vs. full model's 535.8/34.6.

Why it matters / caveats: Modeling 3D point tracks and RGB video jointly within one diffusion pipeline improves both visual and geometric multi-view consistency for HOI synthesis, useful for animation and embodied AI. Even the TACO training dataset offers only 12 viewpoints, a key data scarcity limitation.

Problem: Existing LLM post-training methods — RL with coarse-grained outcome rewards, and on-policy distillation (OPD) with unconditional KL-based imitation — each have limited ability to efficiently transfer new knowledge from a teacher to a student, especially across different model families.

Method: The paper proposes Distilled RL, folding teacher supervision directly into the RL (GRPO-style) objective. It computes a reverse (teacher-to-student) importance ratio on student-sampled tokens, clips it for stability, applies "negative sample reset" so teacher reweighting only affects positive-advantage trajectories, and applies sequence-level geometric normalization to prevent systematic over/under-suppression. Trained on DAPO-17K using EasyR1/VeRL, with Qwen3-8B-GRPO as teacher and DeepSeek-R1-Distill-Qwen-1.5B, Qwen3-1.7B, and Qwen3-4B as students, against GRPO, OPD, and OPD+RL baselines.

Key results:

• On DSQW-1.5B (cross-family), Distilled RL raises average Pass@1 across 10 math benchmarks from 31.70 (base) to 40.00, beating OPD by +4.73 and RL by +3.14 points.
• On Qwen3-4B (within-family), Distilled RL improves average Pass@1 from 46.33 to 58.96, outperforming OPD (+2.99) and RL (+1.56).
• Ablations show removing "negative sample reset" causes the largest drops (-8.81 points average Pass@1 on Qwen3-4B).
• Also achieves the strongest Pass@16 scores on most competition benchmarks (e.g., AIME24 Pass@16 of 80.00 for Qwen3-4B, tying RL).

Why it matters / caveats: Selectively injecting teacher preference only into successful trajectories and normalizing its scale avoids OPD's premature convergence and RL's coarse credit assignment, yielding especially strong gains in cross-family distillation. The teacher is only queried to score student-generated trajectories, so training cannot verify the teacher can actually solve a given problem.

Problem: The paper seeks a rigorous continuous mathematical vocabulary — grounded in differential geometry, measure theory, and stochastic calculus — that explains why specific Transformer components (RMSNorm, RoPE, Attention, FFN, SGD, weight decay) produce the stability limits, context bounds, and optimization dynamics observed empirically.

Method: The authors model the token sequence as a discrete 1-manifold with a canonical measure lattice and define a "semantic fiber bundle," translating each Transformer component into a geometric operator: RMSNorm as a diffeomorphic radial embedding, RoPE as the gauge connection of a principal torus bundle, Softmax Attention as an entropic-optimal-transport Schrödinger bridge, and SGD with weight decay as an Itô diffusion (Ornstein–Uhlenbeck process). They validate derived predictions via a six-part falsification campaign on five frozen pretrained architectures (Qwen3-0.6B/4B, LLaMA-3.1-8B, Gemma-3-1B, GPT-2, Mistral-7B-v0.3), plus controlled training experiments with both AdamW and plain SGD.

Key results:

• The RMSNorm ε-scaling law matches the predicted exponent exactly: α = −0.5000 with R² = 1.000000 across 4,500 independent measurements, holding over 13 orders of magnitude in ε.
• The Lie–Trotter "torsion interferometer" test shows systematic positive alignment far above baseline: median cosine similarity of 0.792 (Qwen3-0.6B) vs. isotropic random baseline of 0.031.
• A parameter-efficiency hypothesis (low-rank skew-symmetric FFN correction) motivated by the framework was tested and refuted: it changed loss by only ~0.003 nats (null within noise).
• A companion experiment found pre-softmax attention logits significantly correlated with exact pairwise mutual information in a random tensor network (Spearman ρ = 0.51, p = 5×10⁻²³).

Why it matters / caveats: The authors explicitly frame this as an "isomorphic descriptive lens" rather than a claim that Transformers are literal physical systems ("the geometry is the map, not the territory"). Not every geometrically motivated hypothesis holds up — the FFN low-rank efficiency idea was tested and refuted by the authors' own follow-up experiments.

Problem: The paper investigates whether general-purpose large language models can generate 3D ligand structures that satisfy multiple heterogeneous spatial constraints (protein pockets, anchor fragments, pharmacophore points, mandatory interactions) as well as specialized pocket-conditioned diffusion models.

Method: The authors introduce 3D-Fit, a benchmarking framework on CrossDocked2020 (948 test complexes) and PLINDER (505 test complexes), defining compact textual representations for protein pockets, mandatory interactions (via ProLIF), anchor fragments (BRICS-fragmented), and pharmacophore points (via Pmapper). Generated molecules are output in a "Simplified SDF" format (found to outperform SMILES+XYZ for general-purpose LLMs). Eleven proprietary and open-weight LLMs (GPT 5.5/5.4, Claude Opus 4.6/4.7/4.8, Gemini 3.1 Pro, Grok 4.1/4.3, Qwen-3.5, GLM-5) are compared against specialist diffusion baselines (PocketXMol, DiffSBDD, PMDM, MolSnapper, and others) via UniDock docking score, PoseBusters validity, and condition-satisfaction success rate.

Key results:

• Raw (pre-optimization) LLM UniDock scores were consistently above the −6 kcal/mol threshold indicating severe steric clashes; after local pose optimization, most LLM scores improved to ~−6.0 to −7.0, remaining weaker than top diffusion models like MolSnapper (−6.8/−8.6).
• Some LLMs achieved high condition-satisfaction success after optimization (e.g., Opus 4.7 anchor SR up to ~95%), but performed much worse on mandatory interaction points (often single digits to ~30%).
• Adding more spatial conditions improved pocket-related metrics for most LLMs, suggesting extra seed-ligand information helps placement accuracy.

Why it matters / caveats: LLMs show emerging, promising capability to follow complex 3D structural instructions and combine constraint types that no existing diffusion model natively supports simultaneously, but still lag specialized diffusion models in physical plausibility. The authors note limitations including lack of statistical significance testing and use of only one pharmacophore/interaction point per example.

Problem: Self-hosted AI agents must continuously read and rewrite their own persistent state (memory, instruction, configuration files) to function, creating "self-state attacks" where a compromised agent corrupts its own state through file operations indistinguishable from its own legitimate updates, and it is unclear how far OS-level defenses can go in closing this gap.

Method: The authors formalize self-state attacks and define a four-axis attack space (Target, Mechanism, Granularity, Temporal pattern), yielding a 23-cell canonical attack matrix expanded into 43 concrete attack-script instantiations. They build "openclaw-core," an OpenClaw harness driven by Gemini 3 Flash, collect live inotify VFS traces across four workload profiles (30 tasks each), and evaluate three defense dimensions: prevention (six cumulative Unix-permission levels L0–L5), detection (Wazuh-style file-integrity monitoring, workload-blind statistical anomaly, and their proposed workload-conditioned statistical anomaly using per-(file,op_type) z-scores), and recovery (periodic backups).

Key results:

• Under their workload-conditioned detector at threshold τ=2.0, the 23-cell matrix splits into 11 Visible, 8 Conditioned, and 4 Indistinguishable cells (all four concentrated on Memory-layer writes).
• The Wazuh-style FIM baseline achieves TPR=1.00 on every attack but at the cost of a per-event false-positive rate of 0.76–0.99 on legitimate writes.
• The recommended prevention level L3 (locking Instruction+Configuration) yields protection 0.61 and functionality 0.70 but collapses one workload profile's functionality to 11%.
• Recovery achieves 100% remediation across attack classes given detection fires, but a backup-destruction "sentinel" attack achieves 0% remediation by construction.

Why it matters / caveats: A layered OS defense stack closes most self-state attack cells but leaves a small, structurally unclosable residual surface — small/minimal-delta writes to Memory files falling within an agent's own legitimate write distribution — implying semantic-layer defenses remain necessary. Results are based on a single self-hosted agent architecture and four author-defined workload profiles.

Problem: Existing GPU rental or batch-workload cloud services allocate exclusive, fixed card-hour resources per tenant for Vision-Language-Action (VLA) model post-training, forcing users to manage complex heterogeneous infrastructure and wasting accelerator capacity on small, bursty, or idle-heavy workloads.

Method: JoyNexus is a Tinker-style, service-oriented multi-tenant framework decoupling a Training Model Service, an Inference Model Service, and an Environment Service, each keeping a resident shared vision-language backbone while mounting tenant-specific action modules and policy versions in isolated slots. A Master Service compiles high-level SFT/RL/evaluation API calls into tenant-scoped workloads dispatched via separate scheduling paths, and a group-batching mechanism canonicalizes heterogeneous VLA data schemas so a single shared-backbone forward pass can serve multiple tenants before splitting outputs. Tested on StarVLA (QwenGR00T, QwenOFT, Qwen3-VL-4B backbone) and OpenPI (π0.5).

Key results:

• In a 4-tenant, 8-GPU mixed RL/SFT workload simulation, JoyNexus reduced aggregate GPU time by 28.3%, a 1.39x GPU-time efficiency improvement over isolated single-tenant execution.
• Cross-tenant scheduling nearly doubled Training Model Service utilization (20.8% → 41.3%).
• Group batching yielded speedups of 2.01x-2.21x for StarVLA-GR00T, 1.71x for OpenPI, and 3.24x for StarVLA-OFT, with benefits increasing as tenant count rises.
• Per-tenant training-loss trajectories under group execution closely tracked serial-execution baselines, confirming grouping preserves separate tenant optimization behavior.

Why it matters / caveats: A service-oriented, shared-backbone architecture can substantially raise GPU utilization and cut aggregate compute cost for multi-tenant VLA post-training without disrupting per-tenant training correctness. The group-batching speedups measure only the shared-forward stage, not end-to-end training throughput, and the system lacks dynamic runtime resource reallocation.

Problem: Existing LLM benchmarks either test knowledge of already-known outcomes or coarse single-answer forecasts, leaving open whether language models and deep-research agents can gather current evidence and produce consistent, fine-grained football predictions before a match's outcome is known.

Method: WorldCupArena is a dynamic benchmark applied to all 104 matches of the 2026 FIFA World Cup, comparing 13 systems under two evidence settings: a common prepared evidence package (no tool use) versus self-directed web search/deep research. Each system submits, 24 hours before kickoff, a structured JSON prediction covering five layers (result/score, players/lineups, events/timeline, tactics/statistics, competition outcome), scored via Brier score, exact-score accuracy, a partial-credit "Scoreline" score, and per-layer metrics combined via availability-aware weighted aggregation.

Key results:

• Claude Opus 4.7 (Thinking + Search) had the best result accuracy (70.7%) and exact-score accuracy (17.2%), only 1.0 point above the human-fan baseline (69.7%).
• Exact-score accuracy across all systems ranged only 10.3-17.2%, and adding search hurt overall Composite score for all three tested model pairs.
• Four systems predicted champion Spain, but only the two Claude configurations also predicted the exact Spain-Argentina final.
• All 13 systems failed together on the same upsets: e.g., Spain-Cape Verde ended 0-0 despite Spain being an 89.1% market favorite and all 13 systems predicting 3-0 or 4-0.

Why it matters / caveats: Result accuracy alone masks meaningful differences in prediction quality, web search does not reliably improve match forecasting for these commercial systems, and models share correlated blind spots on the same upsets. These are descriptive comparisons of complete commercial products, not controlled architecture comparisons, and coverage varies across systems (58-104 matches).

Problem: Off-policy post-training methods like DPO are simple but generalize poorly to out-of-distribution inputs when trying to reduce hallucinations and improve factual faithfulness in text generation.

Method: The paper proposes Token-Level Off-Policy Labeling (TOPL), reframing post-training as a token-level binary correctness classification task rather than next-token prediction: a lightweight reward head is attached to an intermediate layer and trained with binary cross-entropy to predict whether each token is "factual" or "corrupted," using LoRA (rank r=4, α=8) in a selected range of transformer layers. Training data comes from the FAVA dataset (token-level perturbations simulating hallucinations); after training, the reward head is discarded and the LoRA update merged into the base model for standard generation. Evaluated on Qwen3-8B, Llama-3.1-8B, and Gemma-3-4B against SFT, DPO, SOPL, Token-Level DPO, TLDR, and Token-Level Unlikelihood Training.

Key results:

• On out-of-distribution evaluation across 11 AggreFact datasets, TOPL achieves the best average factuality scores: 0.8726 (Qwen3-8B), 0.8619 (Llama-3.1-8B), 0.8398 (Gemma-3-4B), versus SFT baselines of 0.8511, 0.8555, 0.8221.
• On machine translation (Qwen3-8B, XCOMET), TOPL scores 0.9300 OOD, the best among compared methods.
• TOPL's LoRA-A subspace shows much stronger separability between factual/non-factual tokens (AUROC 0.7541) than SOPL (0.6064) or DPO (0.6026).

Why it matters / caveats: TOPL offers a simple, off-policy alternative to RL-style fine-tuning with better OOD generalization on faithfulness tasks and an interpretable mechanism. Random-label ablations show Qwen3-8B is unusually robust to spurious supervision unlike Llama/Gemma, and the LoRA-A/OOD-performance relationship is correlational, not established as causal.

Problem: Deploying diverse real-time multimodal application pipelines (e.g., voice agents, video world models) efficiently across multiple GPUs currently requires hand-crafted, application-specific systems engineering because existing serving frameworks only support limited, fixed deployment policies.

Method: FlashRT is an agent harness that guides Claude Code (running Claude Opus 4.8, effort=max) to transform a developer-written sequential single-GPU reference implementation into an optimized multi-GPU deployment. It uses a "chain-of-program" paradigm where the agent lowers the reference into a hierarchical DAG intermediate representation with node-level state annotations and edge-level blocking/streaming annotations, validates it with a sequential interpreter, and then enters a self-driven, measurement-gated validation loop that hypothesizes, implements, verifies correctness, and benchmarks candidate deployments.

Key results:

• On the face-to-face conversational agent (LiveAvatar) case study, FlashRT reduces latency from a 107.92s sequential baseline to 1.57s (3 GPUs, ~70x reduction) while reaching 173.67 FPS at 8 GPUs.
• For Qwen3-Omni text-to-audio, FlashRT achieves 0.323s latency vs. 0.433s for hand-engineered vLLM-Omni (25% reduction) on NVIDIA B200; 65% latency reduction on AMD MI355X.
• Across applications, delivers up to ~70x latency reduction and 2.8x throughput improvement on NVIDIA B200, and 3.6x throughput improvement on AMD MI355X.

Why it matters / caveats: A generic coding agent with a structured IR-and-validation workflow, requiring no human intervention beyond a reference implementation, can autonomously discover deployment strategies that match or beat expert-engineered systems, generalizing across NVIDIA and AMD hardware. The authors note FlashRT has not been integrated with LLM kernel-optimization agents and only one fixed agent configuration was tested.

Problem: Existing video diffusion models generate realistic single-shot clips but lack explicit shot-planning mechanisms needed for coherent multi-shot cinematic video generation with narrative structure.

Method: ShotPlan is built on Wan2.1-T2V-14B and introduces learnable planning tokens (a shared hard-cut token replicated per user-specified cut, or start/end tokens for gradual transitions) concatenated with patchified video tokens and jointly processed through the DiT's attention/FFN blocks, then discarded before unpatchification. To align these tokens with frame-level timestamps despite VAE temporal compression, the paper introduces Fractional Temporal Rotary Position Embedding (FRoPE), assigning planning tokens continuous fractional latent coordinates. Training uses a curated dataset built from VideoEvent (TransNet V2 shot-boundary detection, Gemini 2.5 filtering), fine-tuning for 3,500 steps on 8 H100 GPUs.

Key results:

• On a 100-prompt multi-shot benchmark, ShotPlan achieves the lowest Transition Deviation of 0.64 frames, versus 1.12 for MultiShotMaster and 2.71 for HoloCine.
• Improves inter-shot character consistency to 0.46 and scene consistency to 0.37, versus 0.39/0.32 for the strongest baseline (HoloCine).
• Ablations show removing Fractional RoPE raises transition deviation from 0.64 to 2.13; a noised-token training variant needs ~11,000 steps to match the clean-token design's performance at 3,500 steps.

Why it matters / caveats: ShotPlan offers a lightweight, architecture-preserving way to add precise, frame-level shot-transition control to a pretrained video diffusion backbone, and the same mechanism generalizes to continuous camera-motion control. Evaluation relies partly on Gemini-2.5-judged narrative coherence and a user study for camera motion.

Problem: In multi-teacher on-policy distillation for agentic tool-use models, vanilla generalized knowledge distillation (GKD) improves tool-call recall but also drifts the model's should-call/should-respond decision boundary, causing over-calling of tools, and this drift is invisible to aggregate loss statistics.

Method: The authors train a Qwen3.5-9B student via GKD-style on-policy distillation against two frozen teachers (a tool-call teacher and a response teacher) using token-level Jensen-Shannon divergence losses on student rollouts, over data derived from APIGen-MT. They diagnose boundary drift via token-level JSD concentration metrics, then compare four calibration strategies: Hard Clip, Global Reweight, Soft Clamp (their proposed method, using a detached batch-adaptive threshold to cap extreme token divergences while rescaling rather than zeroing gradients), and an inference-time entry bias on the <tool_call> logit.

Key results:

• Vanilla GKD raises over-calling to 14.2±2.1% (from 4.9% for Base SFT) while call recall reaches 91.5±1.7%; Soft Clamp reduces over-calling to 9.0±0.2% with call recall at 86.5±1.4%.
• In vanilla GKD, the top 1% of supervised tokens account for 41.2±0.4% of total JSD; Soft Clamp reduces this to 15.4±3.2%.
• In BFCL multi-turn loop diagnostic, vanilla GKD makes 1.515 tool calls/turn with 15.1% Loop@3, versus Soft Clamp's 1.289 calls/turn and 10.3% Loop@3.

Why it matters / caveats: The paper reframes multi-teacher on-policy distillation auditing around behavior-boundary calibration rather than aggregate loss monitoring. No method dominates on every metric, the mechanism evidence is diagnostic rather than a full causal proof, and results come from one primary model family/teacher pair.

Problem: Autonomous driving policies need robust performance on rare long-tail events, but such events are scarce in curated multi-view datasets and mostly exist as heterogeneous, unposed monocular (e.g., dashcam) videos that cannot be directly used as multi-view training data.

Method: OpenLongTail is a generative data engine that recovers a metric-scale ego-trajectory from monocular video using MapAnything (stabilized with a Kalman filter plus RTS smoother), then synthesizes the five missing non-front views using a LoRA-finetuned Wan2.1-VACE-1.3B DiT diffusion backbone conditioned on Plücker-ray camera geometry, a temporal depth-warping module (using frozen DepthCrafter depth), and a directed cross-view memory bank. Training uses a flow-matching objective on ~200K clips from ~50K scenes (NVIDIA PhysicalAI-AV, PandaSet, nuScenes) using 32 H200 GPUs for ~96 hours.

Key results:

• Closed-loop AlpaSim evaluation (53 long-tail events): fine-tuning with synthesized data raises average AlpaSim Score from 0.534 (baseline) to 0.748, close to ground-truth multi-view fine-tuning (0.764).
• Unseen-scene view synthesis outperforms baselines (TrajectoryCrafter, Gen3C, ReCamMaster, Vista4D) on PSNR/LPIPS across all 5 target views, achieving GeoKPMµ of 82.41 vs. next-best 18.86.
• Pose recovery matches MapAnything's best metric-scale ATE (2.212) while cutting jerk from 4737.4 to 283.9.

Why it matters / caveats: Ubiquitous monocular dashcam footage can be converted into a scalable, near-ground-truth-quality training source for long-tail driving robustness. The authors note gains are "alignment-conditioned" (out-of-distribution source data helps only when aligned with the target long-tail slice).

Problem: Standard diffusion model samplers treat each denoiser prediction as an exact, noise-free estimate for numerical integration, ignoring the model's inherent statistical estimation uncertainty and temporal redundancy across the reverse trajectory, causing prediction drift especially in few-step (low-NFE) inference.

Method: DiFA is a training-free, plug-and-play inference-time wrapper compatible with existing solvers (DDIM, DPM-Solver++, UniPC, Heun, Euler). Motivated by Best Linear Unbiased Estimation/Kalman filtering, it builds a forward-aligned causal temporal consensus by fusing a sliding window of historical clean predictions weighted by structural consistency and logSNR-based noise-level compatibility, computes the deviation between the current prediction and this consensus, applies noise-level-conditioned deviation guidance, and adds the scaled guidance back before passing to the unchanged downstream solver — adding only O(Kd) overhead with no extra network evaluations.

Key results:

• On CIFAR-10 with DPM-Solver++, DiFA reduces FID at NFE 8 from 8.40 to 4.15 (over 50% relative improvement); at NFE 20 from 2.33 to 1.96.
• On class-conditional ImageNet-64 with Heun, DiFA reduces FID from 230.05 to 110.20 at NFE 5.
• Also improves FD-DINOv2 perceptual scores across NFE settings on ImageNet-64, and FID/IS on ImageNet 256x256 with SiT-XL/2.

Why it matters / caveats: DiFA offers a training-free way to substantially close the gap between fast, few-step sampling and high-fidelity generation, competitive with costly distillation methods without retraining. The method relies on fixed hyperparameters and an idealized static-anchor assumption; extensions to text-to-image, video, and inverse problems are left to future work.

Problem: Existing benchmarks reduce OCT (optical coherence tomography) image understanding to coarse disease classification or isolated VQA, failing to evaluate whether multimodal LLMs can perform the full clinical workflow from visual perception through medical cognition to clinical reasoning.

Method: The authors build OCT-Bench, 10,076 expert-verified multiple-choice questions constructed from 4,137 OCT images drawn from seven public datasets (OCT5k, OIMHS, OCT-C8, AMD-SD, OCTDL, MMC-AMD, GOALS), organized into a hierarchical taxonomy of three dimensions (Perception, Cognition, Reasoning), nine capability groups, and 20 fine-grained tasks. Construction follows a five-stage pipeline including GPT-4o-based question generation conditioned on medical knowledge from authoritative sources, and two-stage quality control (automatic GPT-4o checks plus manual expert review). 20 MLLMs are evaluated zero-shot.

Key results:

• Best overall accuracy is 62.0% (GPT-5.4-mini), followed by Gemini-2.5-flash (60.5%); no model exceeds 62% overall.
• Accuracy drops sharply across dimensions: best Perception 75.8% vs. best Cognition 64.2% vs. best Reasoning only 42.9%, a 32.9-point decline from perception to reasoning.
• Near-saturation on easy tasks (modality perception ~97.9% avg) but severe fragility elsewhere: morphological description averages only 23.4%.
• Disease-category diagnosis is highly imbalanced: models do relatively well on MH (58.4%) but poorly on Glaucoma (7.8%) and RAO (5.5%).

Why it matters / caveats: Current MLLMs, including medical-specialized ones, remain far from reliable clinical OCT interpretation, with the bottleneck concentrated in higher-level reasoning and fine-grained visual/anatomical discrimination rather than basic image recognition. OCT-Bench is offered as a diagnostic tool to localize these capability gaps.