AI papers — 2026-08-05

Problem: Existing agent benchmarks focus on bounded tasks with immediate success criteria, leaving it unclear whether LLM agents can sustain coherent, self-correcting behavior over long horizons where feedback arrives at mixed delays and actions constrain future options.

Method: MerchantBench formulates seller-side e-commerce operation as a 365-simulated-day partially observable MDP with hourly steps (8,760 total) and a decision window every 12 steps, grounded in 98,843 real product records from 1688 across 10 categories and 36,576 suppliers. The environment couples an Upstream Supplier Simulation (price change, delisting, shipment delay events) with a Downstream Order-Level Simulation (order lifecycle through placed/procured/shipped/delivered/settled, plus six abnormal outcomes like cancellation, stockout, return/refund, bad review) and exposes 26 merchant tools spanning Product Sourcing, Listing/Pricing Control, Cash-Flow Management, and Mixed-Latency Feedback Adaptation. Eight LLMs (GPT-5.6 Sol, Claude Opus 4.8, Qwen3.7-Max/Plus, GLM-5.2, DeepSeek-V4-Pro/Flash, Kimi K2.6) are evaluated under ReAct and Hermes agent frameworks across 48 runs of 365 simulated days, compared against human participants and a rule-based baseline.

Key results:

• The best LLM configuration (Qwen3.7-Max with Hermes) attains only 27.3% of the mean final net assets achieved by human participants (highest LLM mean net assets 59.46K RMB vs. human 217.61K RMB).
• Hermes produces 53.3% higher final net assets, 71.5% higher GMV, and 71.2% more orders than ReAct on average across the eight models.
• Human operators maintain a 100% Sustained Window Rate (SWR), while LLM configurations range from 10.6% to 99.4% under ReAct and 17.8% to 66.1% under Hermes.
• Qwen3.7-Max's quarterly Effective Window Rate falls from 68% to 23% under ReAct, illustrating "Activity Decay" and loss of Operational Coherence.
• Coefficient of variation across repeated runs ranges from 3.3% (GPT-5.6 Sol/ReAct) to 55.1% (Qwen3.7-Max/Hermes), showing substantial instability despite high mean performance.

Why it matters / caveats: The benchmark reveals concrete failure modes—Control-Loop Narrowing, Premature Abandonment, and evidence-uncalibrated policy persistence—that current LLM agents exhibit over long horizons, suggesting that raw model capability alone does not translate into sustained autonomous business operation; the gap to human performance remains large.

Problem: Real-time, open-ended video editing requires causal, low-latency generation with bounded compute while preserving source fidelity and long-term temporal consistency, but existing editors are largely offline and bidirectional, and naive causal adaptation suffers from train-inference mismatch and error accumulation.

Method: JoyAI-Video-Edit is a 16B-parameter autoregressive diffusion model combining an MLLM condition encoder, a causal video VAE (8×24×24 spatiotemporal compression), and an MM-DiT diffusion transformer, trained via a multi-stage curriculum (T2I → T2V → I2I → bidirectional V2V/IV2V) then converted to chunk-wise causal generation with bidirectional-within-chunk/causal-across-chunk attention and a sliding window plus global sink. Key techniques are chunk-wise autoregressive adaptation (teacher forcing then Resampling Forcing on model-generated history), Source-Anchored Distribution Matching Distillation (SA-DMD) that applies classifier-free guidance along separate text and source-fidelity axes to distill a two-step generator, and Long-Horizon Autoregressive Distillation (LHAD) that performs segmented gradient accumulation over extended rollouts with dynamic mirror looping for conditioning.

Key results:

• Achieves an overall score of 3.60 on OpenVE-Bench, beating streaming baselines SANA-Streaming, LiveEdit, XMax-X2.0, and StreamDiffusionV2 by 0.98–2.37 points, and comparable to strong offline editors.
• On the newly constructed LongV2VBench (229 one-minute editing tasks), achieves overall score 3.30, ranking first in all five categories and beating the strongest baseline XMax-X2.0 by 1.59 points.
• Reaches 30.19 FPS end-to-end at 720×1280 on a single Nvidia B200 GPU (VAE encode 22ms, DiT denoise 185ms, VAE decode 19ms, request-to-response latency 226ms), 44.4% faster than XMax-X2.0.
• Human evaluation: preferred over LiveEdit (90%), SANA-Streaming (87%), XMax-X2.0 (81%), and StreamDiffusionV2 (87%); roughly tied with Bernini-R (48% vs 44%).
• Ablation shows SA-DMD alone raises overall LongV2VBench score from 2.81 to 3.23, and combining SA-DMD with LHAD yields the best score of 3.30.

Why it matters / caveats: Demonstrates that real-time streaming video editing can approach offline-editor quality while running at ~30 FPS, enabling live broadcasting and interactive applications; the system still relies on multiple distillation stages and a bounded temporal window, so very long-range dependencies beyond the sliding window plus sink may not be captured.

Problem: Unified 3D multimodal modeling (understanding, generation, editing) lags behind 2D vision due to scarce large-scale, geometrically consistent 3D editing data, forcing 3D understanding, generation, and editing to be developed as separate, non-synergistic systems.

Method: The authors build an 87M-sample 3D multimodal corpus (25M understanding, 50M text-to-3D pairs, 12M editing pairs), using Nano3D-v2—an agent-based pipeline with anchor-view selection, a learned 3D edit-region localization model, TRELLIS-based voxel-level FlowEdit with voxel-merge, LATTICE-based geometry refinement, NaTex-based texture inpainting, and VLM-based annotation/filtering—to construct editing triplets at scale. The architecture couples Hunyuan3D-VLM (a Qwen-VL-based 3D-aware vision-language model with VecSet encoder, Q-Former compression to 512 tokens, and 133 special box/point tokens for grounding) with a 3D-DiT initialized from Hunyuan3D-2.1 via an MLP-Connector; editing and part generation additionally condition the DiT on the source 3D representation via self-attention. Training proceeds through four stages: 3D-VLM pretraining, text-to-3D pretraining, omni pretraining (1:1 sampling of text-to-3D vs. editing+part-gen), and task-specific continued pretraining.

Key results:

• On UniPart-Bench 3D understanding, Hunyuan3D-VLM achieves 85.47 SBERT / 89.06 SimCSE on part QA and 72.94 SBERT / 52.84 ROUGE-L on object captioning, outperforming Part-X-MLLM and UniVerse3D baselines.
• Human evaluation on text-to-3D generation: preferred in 55.2% (text alignment), 57.1% (geometry quality), and 56.6% (overall) of comparisons, more than double the strongest baseline Omni123 (17.5%/21.0%/18.4%), against a 25% random baseline.
• On Edit3D-Bench, reduces average Chamfer Distance from 0.0684 (best baseline Omni123) to 0.0091 (86.7% relative reduction) and improves average F1 from 0.2729 (best baseline Steer3D) to 0.6515 (2.39x improvement).
• Scaling text-to-3D pretraining data from 3M to 50M samples raises overall preference rate from 8.4% to 57.5%, and a case study shows adding only 1,000 text-to-3D samples (with zero new editing data) can unlock a previously failing edit.

Why it matters / caveats: Shows explicit cross-task synergy—stronger generation and stronger understanding both improve editing—arguing for unified 3D multimodal training; the paper notes open problems including lack of single-stage high-fidelity geometry representation, noisy VLM-based captions, and unexplored end-to-end texture editing.

Problem: Continuous-latent language models face a representational bottleneck: prior methods either inherit embedding spaces not designed jointly for generation and decoding, or compress the autoencoded latent to make diffusion tractable at the cost of token-level reconstruction fidelity.

Method: AURORA-LM separates representation construction from distribution modeling. A Query-based Encoder-Decoder builds a high-capacity, prefix-aligned continuous latent sequence (causal encoder queries attending to progressively longer token prefixes; a matched query decoder reconstructs tokens via cross-entropy loss), trained with token-embedding and latent dropout, then frozen. A Block-Causal Diffusion Transformer is trained via flow matching (clean-endpoint/x0 prediction) with a block-causal factorization (parallel denoising within a block, causal generation left-to-right across blocks using a two-stream attention mask), a low-rank bottleneck applied only to the noisy-input pathway (Db=128) while retaining full-width (D=1024) clean-latent prediction, noise allocation calibrated to latent width (tan-d schedule, d=7), and self-trajectory consistency (aligning EMA-model clean predictions across neighboring Euler-step states) plus self-conditioning for few-step sampling.

Key results:

• On OpenWebText unconditional generation, AURORA-LM-S (130M) achieves the lowest Gen-PPL (23.56) and highest MAUVE (0.890) among AR, Duo, Duo-distilled, SEDD, MDLM, and ELF-B baselines.
• On XSum conditional summarization, achieves best ROUGE-1/2/L scores (36.6/13.4/28.9), beating ELF-B (36.0/12.2/27.8) and AR/MDLM/Duo baselines.
• Scaled to 1.01B parameters (AURORA-LM-L) with ~1,500 EFLOPs of compute, achieves a macro average of 32.6 across nine language benchmarks (MMLU, ARC-C, HellaSwag, etc.), outperforming the larger ~1.8B-parameter publicly released Cola-DLM (25.1 average) on all nine tasks.
• Ablations show latent width D=1024 with an appropriately calibrated noise schedule maximizes MAUVE; a Db=128 noisy-input bottleneck outperforms both narrower (32) and wider settings; block size Q=16 balances quality (MAUVE 0.816) against sequential generation cost.

Why it matters / caveats: Demonstrates that retaining full decoder-facing latent capacity (rather than compressing for tractability) combined with targeted diffusion-side adaptations can make continuous-latent language diffusion competitive with and even superior to discrete diffusion and larger continuous baselines; all experiments were run on Ascend NPUs and evaluated at relatively modest model/data scale (130M–1B parameters) compared to frontier LLMs.

Problem: Extending deep-research agents from static images to continuous video introduces two failure modes—modality bias (agents default to text search instead of visual tools) and parametric knowledge leakage (models answer from memorized knowledge instead of genuine tool use)—that existing Video-DR approaches fail to address.

Method: Video-DeepResearch equips agents with Select_Keyframe and Crop_Search tools alongside standard web Search/Visit, and constructs a data engine (Phase I: rule-based + agent-based video filtering; Phase II: CLIP-based keyframe proposal, entity cropping/image search, single- and multi-entity VQA generation with four tool-free rollouts to filter out parametric-leakage-prone questions, yielding 30K VQA pairs) plus a decoupled perception-exploration trajectory pipeline with stage-wise tool unlocking (vision-only tools first, then Search/Visit unlocked) via rejection sampling on Qwen3.5-397B-A17B, yielding 7K correct trajectories. Training uses a two-stage recipe: SFT on the 7K trajectories plus 7K text-only QA instances (from VDR), followed by GRPO reinforcement learning on a 2K moderate-difficulty dataset with sparse binary reward and down-sampled negative-gradient updates, applied to Qwen3-VL-30B-A3B-Instruct and Qwen3.5-35B-A3B base models.

Key results:

• Video-DeepResearch-35B-A3B achieves 64.0% average accuracy, a new state of the art, surpassing Claude-4.5-Sonnet (59.0%) by 5.0 points and beating GPT-5 (52.5%) and Gemini 2.5 Pro (57.5%).
• The 30B-A3B variant reaches 59.3%, competitive with Claude-4.5-Sonnet and a +18.8% improvement over its Qwen3-VL-30B-A3B-Instruct base.
• On the introduced VIDEODR-BENCH (200-instance multi-hop VQA benchmark), the 35B variant scores 65.4% overall, the highest among all evaluated models.
• Ablation: base model 40.5% average → 7K-SFT 53.0% → +7K text-SFT 56.8% → +2K-RL (GRPO) 59.3%, showing each component contributes.
• Tool usage analysis: baseline Qwen3.5-397B invokes only 0.10 visual tool calls per task vs. 1.27 text calls; Video-DeepResearch-30B increases this to 2.33 visual and 4.24 text calls on VideoDR, correcting the modality bias.

Why it matters / caveats: Shows that specialized data curation and training (rather than raw model scale) drives Video-DR capability—the 30B model matches or exceeds a 397B baseline—but the approach requires substantial GPU resources for data synthesis/training and its benchmark still depends on labor-intensive human annotation, limiting scalability.

Problem: In streaming/industrial recommendation, the pretrain-then-transfer paradigm breaks down because behavioral distributions drift continuously, forcing the pretrained encoder to be refreshed, but conventional next-token pretraining encodes noisy cross-session adjacency and refreshing the encoder conflicts with task-specific fine-tuning on the same parameters.

Method: Knowledge-Geometry Decoupling (KGD) has two parts: Behavioral Multi-Token Prediction (BMTP), which filters next-item supervision to only collaboratively (co-occurrence graph) or semantically (text-embedding cosine similarity) related future items rather than raw adjacency; and Decoupled Read-Write Ownership, where a task learner reads the pretrained encoder's contextualized states via read-only (stop-gradient) cross-attention and writes task-specific geometry through an Anchored Calibration Residual (ACR) — a low-rank residual constrained orthogonal to the pretrained item embedding. The encoder (using ManCAR backbone on public data, OneRank industrial backbone) is refreshed daily while the task learner's parameters remain separately owned.

Key results:

• Improves over the strongest published pretrain-transfer baselines by 4–12% (NDCG@50/Recall@50) across eight Amazon-2023 category benchmarks.
• On a 90-day industrial stream, entangled-refresh baselines (e.g., TA&FT) show AUC dropping from 0.7852 (frozen) to 0.7837 (refreshed), while KGD reaches 0.7867 click AUC and improves further under refresh.
• Live A/B test on Shopee Homepage Search (10% buckets, 10M+ users each, two weeks): +1.75% GMV per user, +1.53% advertising revenue, +0.95% CTR, +0.72% CVR (GMV/revenue significant at p<0.01).
• Ablation: removing collaborative filter costs 0.5% AUC on head users; removing semantic filter costs 1% AUC on tail users; removing ACR or the decoupled backbone collapses performance to near or below entangled-refresh baselines.

Why it matters / caveats: Demonstrates a practical mechanism for continual refresh of pretrained recommenders without erasing task adaptation, validated at full production deployment scale; the approach targets classic deep-learning sequential recommenders rather than LLM-based recommenders, and training cost roughly doubles (about 2 hours/day on A100 vs. 1 hour).

Problem: On-policy self-distillation (OPSD) for agentic RL provides dense token-level supervision from a privileged teacher to counter sparse trajectory-level rewards, but existing token-level or step-level methods for weighting teacher reliability are either noise-sensitive (pointwise) or too coarse (uniform per-step), so unreliable teacher guidance can be indiscriminately transferred.

Method: Persistent Consistency Self-Distillation (PCSD) computes token-level distillation weights from the teacher-student sampled log-probability gap δ_k,i by (1) aggregating gaps over an exponentially-decayed forward window, (2) adaptively interpolating between short- and long-window aggregates based on local gap variance (soft adaptive windowing), (3) applying a one-sided trend modulation (mask-aware OLS slope) that attenuates declining teacher support, and (4) mapping the result to a bounded weight via sigmoid gating. The resulting weighted distillation loss L_PCSD (teacher log-probs and weights detached) is jointly optimized with GRPO's clipped-ratio objective plus KL regularization (Ltotal = LGRPO + λ_PCSD·LPCSD, λ_PCSD=0.01), trained on Qwen2.5-3B-Instruct and Qwen3-1.7B-Instruct with a frozen skill-augmented teacher.

Key results:

• On ALFWorld, PCSD achieves Overall success rates of 90.6% (Qwen2.5-3B) and 59.4% (Qwen3-1.7B), exceeding GRPO by 15.6 and 13.3 points and the strongest baseline SDAR by 6.2 and 5.5 points.
• On WebShop, PCSD achieves the highest Score of 85.0 (Qwen2.5-3B) and ties SDAR for best Acc (67.2%).
• On ALFWorld unseen split, PCSD reaches 86.7% Overall vs. GRPO's 70.9% and SDAR's 72.7% (a 15.8-point gain over GRPO).
• Ablation: replacing adaptive aggregation with fixed windows (N=1 or N=4), or removing trend modulation or exponential decay, all reduce Overall success (e.g., full PCSD 90.6% vs. w/o trend 83.6%, w/o decay 85.1%).
• Sensitivity: λ_PCSD=0 gives 75.0% Overall, λ_PCSD=0.01 gives best 90.6%, while λ_PCSD=0.05 drops to 83.6% (non-monotonic).

Why it matters / caveats: Shows that combining fine-grained token resolution with local-neighborhood robustness improves credit assignment in long-horizon agentic RL beyond both GRPO alone and prior self-distillation weighting schemes; the method uses fixed hyperparameters and a frozen teacher, so it does not adapt aggregation/gating to evolving trajectory statistics or changing teacher reliability.

Problem: Classical world models are narrowly defined as physical state-transition predictors (state+action→next state), which is too limited a feedback proxy for agents that need broader agent-usable information (execution outcomes, retrieved experience, verification signals) to continually improve without the cost, risk, and non-parallelizability of direct real-environment interaction.

Method: This is a position/survey paper proposing "Agent-Centric World Proxies," generalizing the world model equation ŝ_{t+1}=WM(s_t,a_t) into ŝ_{ℓ+1}=WP(s_ℓ,u_ℓ^F) where interaction step ℓ, information state space S, and proxy function F replace physical time, raw state, and fixed next-state prediction. The framework organizes world proxies into six functional forms (dynamics, spatial, execution, memory/experience, skill, reward/verification) and three empowerment levels: L1 Inference-Time Guidance (context augmentation, no parameter change), L2 Training-Time Optimization (proxy output becomes reward/critique/simulated rollouts feeding SFT/DPO/PPO/GRPO objectives), and L3 Agent-Proxy Co-Evolution (real-environment evidence updates the proxy while proxy knowledge is distilled back into the agent).

Key results: Not stated (conceptual/taxonomy paper; no empirical benchmark numbers reported in the sections reviewed).

Why it matters / caveats: Offers a roadmap/taxonomy for reframing world modeling around agent-usable information transitions rather than pixel/state realism, intended to guide future research on world-proxy-driven agent training and continual improvement; as a position paper it does not itself provide new quantitative results, and its value is primarily organizational/conceptual.

Problem: It is unclear whether personal AI agents that retain memories, skills, and session histories across sessions actually improve future performance from that retained experience, since existing agent benchmarks score isolated one-shot tasks or test persistence mechanisms in isolation without matched controls that separate retained-experience gains from base-model or runtime effects.

Method: PAST-Bench evaluates agents as ordered sequences of fresh-session task-family episodes (cold, learn/update, evaluation, control) spanning four capabilities—Memory, Procedural Reuse, Information Gathering, and Update (26 scenarios, 204 episodes)—under a strict context-clearing protocol, with matched persistence-on/persistence-off ablations and control episodes (no-persistence, distractor, stale, wrong-mechanism) to compute a self-evolution gap Δ and a mechanism-evidence score (Mech) from artifacts/telemetry (memory writes, skill use, session search, update events). The authors evaluate seven base models (GLM-5.1, Kimi K2.6, DeepSeek-V4-Pro, MiniMax-M2.7, GPT-5.4, Claude Sonnet 4.6, Claude Opus 4.6) across four agent frameworks (Hermes, nanobot, ZeroClaw, Agent-Zero), then build Hermes+, adding five targeted runtime mechanisms (Plan/E1, Render/E2, Route/E3, Gate/E4, Close/E5) to the Hermes agent loop.

Key results:

• Every base model gains from persistence with Hermes, Overall Δ ranging from +0.13 to +0.24 (e.g., GPT-5.4: 0.51→0.75, Δ=+0.24, Mech=0.80).
• Hermes and nanobot both reach the same task-score gap (Δ=+0.13) but differ in mechanism evidence: 0.64 vs. 0.57.
• Hermes+ raises the reported mean Overall Δ from +0.13 to +0.15 and Mech from 0.64 to 0.73 (on MiniMax-M2.7), with its clearest gain on Update (Δ: +0.12→+0.24); the +0.02 Overall difference is noted as smaller than run-to-run variation (0.13±0.04 vs. 0.15±0.06).
• Single-mechanism ablations: E2 (Render) gives the highest Memory persistence-on score (0.80); E4 (Gate) gives the largest Info Δ (+0.17); E5 (Close) gives the strongest single-mechanism Update Δ (+0.16).
• Hermes+ transfers positively to 3 of 5 base models (e.g., Claude Sonnet 4.6: +0.20→+0.22, GPT-5.4 flat at +0.24) but regresses slightly on DeepSeek-V4-Pro and Claude Opus 4.6.

Why it matters / caveats: Introduces a performance-attribution methodology that separates real self-improvement from apparent gains due to shortcuts, stale reuse, or wrong mechanisms — showing self-evolution is capability- and model-specific rather than uniform; the authors themselves caution that Hermes+'s overall improvement is smaller than run-to-run variance and should be treated as a diagnostic scaffold, not a universal fix.

Problem: On-policy distillation for diffusion/flow-matching models assumes teacher and student share the same VAE latent space, architecture, and noise-timestep schedule, which fails when the strongest available teacher and desired student come from different model families (heterogeneous pairs) with no shared coordinate system or comparable timesteps.

Method: Any-OPD couples teacher and student solely through a frozen, model-agnostic DINOv2-Base CLS-token representation (cosine distance loss), avoiding any latent- or pixel-space matching that would otherwise collapse under the teacher's stochastic re-synthesis. It aligns trajectories by matching continuous noise levels (not step indices) via a teacher noise-and-denoise projection Πσ_T of the student's own rollout, routes gradients only through the student's solver segment whose noise range covers the correction, and precedes on-policy training with an offline anchoring stage that re-encodes teacher-generated images through the student's own VAE to close the distribution-level gap first. The method is instantiated distilling FLUX.1-dev (12B teacher, FLUX VAE, dynamic shift schedule) into SD3.5-Medium (2.5B student, SD3 VAE, static shift schedule) using LoRA (rank 32, α=64) on Pick-a-Pic prompts.

Key results:

• Raises the 2.5B student's PickScore from 0.846/0.866 to 0.884 and HPSv3 from 9.12 to 10.97 (teacher: 0.878 PickScore, 11.19 HPSv3), approaching/exceeding the 12B teacher on several DrawBench metrics at ~1/5 the size.
• ImageReward gain over baseline (+0.194) is nearly four times the teacher's own margin over the student baseline (+0.049); DPG-Bench essentially unchanged (+0.13).
• Ablation: naive latent-MSE regression collapses training within the first steps; LPIPS improves then degrades; DINOv2 CLS variants remain stable, with plain CLS matching or exceeding multilayer/patch variants.
• Anchoring is a strict prerequisite — Anchor-100 (partial anchoring) never catches up to Anchor-400 (full anchoring) despite identical subsequent OPD training budget.
• Generalization: swapping teacher to Z-Image (6B) with zero procedural changes improves student on every metric, including GenEval (+0.80) and DPG-Bench (+0.95), inheriting the new teacher's specific strengths.

Why it matters / caveats: First demonstrated on-policy distillation across model families with completely different VAEs/architectures/schedules, showing the teacher can be swapped "for free," which could make any newly released strong generator usable as a teacher for any deployed student; results are shown on one teacher-student pair (plus one teacher swap) in image generation, and the approach depends on quality of the anchoring stage and teacher refinement steps (halving teacher denoising steps visibly degrades transferred quality).

Problem: Omni-modal LLMs (Omni-LLMs) incur heavy computational overhead from long, redundant audio-visual token sequences, and existing compression methods degrade badly under aggressive token budgets because they either miss globally distributed evidence or underuse query-conditioned audio-visual collaboration.

Method: OmniPack is a training-free, two-stage token compression framework. Before the LLM, it performs modality-specific compression via importance selection (encoder attention plus temporal/spatial variation cues), coverage selection (DPC-KNN clustering on a joint feature-positional distance), and similarity-aware token merging (aggregating pruned tokens into retained representatives rather than discarding them). Inside the LLM (after layer 18 for Qwen2.5-Omni-7B/MiniCPM-o-2.6, layer 26 for Qwen2.5-Omni-3B), it applies query-conditioned compression that scores tokens by textual relevance, cross-modal (audio-visual) collaboration, and within-modality representativeness, then greedily selects a relevance-diversity balanced retained set.

Key results:

• On Qwen2.5-Omni-7B at 15%/7.5% retention, OmniPack preserves 95.6% of original performance while using only 10.0% of FLOPs (10.0x reduction) and achieving a 4.5x prefill speedup.
• At 10%/5% retention, OmniPack retains 92.9% of performance using just 6.8% of original FLOPs.
• Outperforms baselines (FastV-om, VisionZip-om, OmniZip, OmniSIFT, SEATS) across five benchmarks (AVUT, WorldSense, DailyOmni, VideoMME, LVOmniBench) and three backbones (Qwen2.5-Omni-3B/7B, MiniCPM-o-2.6).
• Ablations show all three pre-LLM components (importance, coverage, merge) are complementary, and text-aware guidance in inner-LLM compression beats general-query or last-token-attention guidance.

Why it matters / caveats: Provides a practical, training-free path to deploy Omni-LLMs at drastically reduced compute while retaining most task performance; the approach depends on careful hyperparameter tuning (e.g., layer choice, allocation ratios) that was tuned per backbone.

Problem: The scaling behavior of Mixture-of-Experts (MoE) diffusion language models (dLLMs) is poorly understood, since prior dLLM scaling studies focus only on dense architectures and MoE designs largely borrow untested autoregressive (AR) conventions.

Method: The authors systematically fit scaling laws for MoE dLLMs across optimization hyperparameters (batch size B=0.374·C^0.3481, learning rate η=64.8·C^-0.2447), compute allocation (IsoFLOP analysis giving M∝C^0.475, D∝C^0.525), and MoE architecture (activation ratio A, expert granularity G, shared-expert ratio S). Guided by these laws, they train LLaDA MoE v2, a 30B-total/3B-activated parameter MoE dLLM with (A,G,S)=(9.09%, 8, 33.3%; 128 experts, top-8 routing, one shared expert of width 4×d_expert) from scratch on 23.5T tokens across five training stages (including RoPE base extension and long-context annealing), followed by supervised fine-tuning (3 epochs, 7M instruction examples, batch 512, peak LR 5e-6) without RL.

Key results:

• LLaDA MoE v2 achieves average 58.60 across 15 benchmarks among dLLMs, beating SDAR Sci by 3.78 points, using only 23.5T tokens (63% of SDAR Sci's 37.05T, 65% of Qwen3's 36T).
• On coding, it beats SDAR Sci by +16.46 on HumanEval and +7.98 on BigCodeBench; remains close to Qwen3 on OlympiadBench (-2.22) and HumanEval (-2.44).
• After SFT alone (no RL), it outperforms SDAR Chat 30B-A3B on 7 of 8 reasoning/coding benchmarks (e.g., MBPP 81.03 vs 71.60, MultiPL-E 67.52 vs 45.00).
• Optimal dLLM shared-expert ratio is a fixed S=33.3% across scales, contrasting with AR MoE designs (DeepSeekMoE 25%, Qwen3 uses none).

Why it matters / caveats: Demonstrates that dLLM-specific scaling calibration (not just AR-derived heuristics) yields compute-efficient large MoE diffusion models; the study varies scaling dimensions separately and does not capture their interactions.

Problem: Existing caption evaluation methods collapse caption quality into a single scalar score, conflating how much visual content a caption covers with how factually reliable its claims are, obscuring which property actually drives downstream multimodal model performance.

Method: CAPEval builds a benchmark of 300 high-resolution images (up to 8000×6618) with human-written ground-truth captions (~346 words average) decomposed into ~50 human-verified atomic checklist items across 8 semantic dimensions (instance, attribute, relation, text, human, image, UI, world knowledge). Qwen2.5-72B judges each candidate caption's checklist items as correct/wrong/not-mentioned to compute Coverage (fraction of GT items mentioned) and Precision (fraction of mentioned items correct); captions from 10 captioners (InternVL3.5, Qwen3-VL, LLaVA-OneVision-1.5, GLM-4.6V-Flash variants) are then used as the sole varying supervision in controlled downstream training — two VLM pipelines (CLIP+Vicuna-7B, SigLIP+Qwen3-4B) and two T2I pipelines (SD3.5 Medium, Qwen-Image) — with OLS regression fitting downstream scores against Coverage and Precision.

Key results:

• Understanding is dominated by Coverage: β_C=+0.118 (p=0.026) on SigLIP-Qwen3 and β_C=+0.215 (p=0.047) on CLIP-Vicuna, while Precision is not significant (p≥0.274).
• Generation is dominated by Precision: β_P=+0.189 (p<0.001) on SD3.5M and β_P=+0.235 (p<0.001) on Qwen-Image, while Coverage is not significant (p≥0.171).
• Within the InternVL3.5 family, the 1B model beats the 8B model on understanding (avg U=58.5 vs 57.3) despite being smaller, and the 4B model beats the 8B on generation (avg G=71.6 vs 71.0).
• Hallucination performance is instead more associated with Precision (p=0.124) than Coverage (p=0.914).

Why it matters / caveats: Shows captioner scale alone is an insufficient proxy for downstream utility, offering task-aware guidance for caption data curation (broad coverage for understanding models, high precision for generation models).

Problem: Text-to-image models synthesize flat RGB pixels lacking editable structure, so recovering individual objects, layer order, and occlusion-hidden content for editing is difficult and generally incomplete.

Method: UniWorld-Design comprises two flow-matching-based models built on an RGBA-extended autoencoder (channel-extended VAE with RGB-latent alignment loss): Text-to-RGBA (T2RGBA) generates standalone transparent RGBA assets from text, while Image-to-Layer (I2L) uses a novel Layer–Instruction Binding MMDiT (LIB-MMDiT) architecture — combining layer-instruction binding attention (restricting each layer prompt to its own image tokens while broadcasting a global instruction) with layer-indexed 3D rotary positions — to jointly decompose a finished image into ordered, complete semantic RGBA layers (supporting top-level decomposition, recursive decomposition, targeted extraction) trained on semantic layer trees derived from designer-authored PSD documents. Both models use a two-stage pipeline of progressive adversarial distillation (8-step student) followed by DiffusionNFT reward-based post-training with task-specific rewards (alpha-following gate plus CLIP/MLLM scores for T2RGBA; L1+LPIPS per-layer reconstruction for I2L).

Key results:

• On 512 Crello designs, I2L reduces per-layer RGB L1 by 37% (0.2014→0.1264) and improves Alpha Soft IoU by 34% (0.5454→0.7325) versus Qwen-Image-Layered.
• I2L reduces blank-layer generation by 63% (0.35→0.13) and achieves a VLM-judged score of 20.43 vs 17.60 (out of 25) for Qwen-Image-Layered.
• T2RGBA achieves the highest CLIP Score (33.03) among compared methods (LayerDiffuse 29.22, OmniAlpha 31.00), a 45% FID reduction over LayerDiffuse.
• OmniAlpha still leads on FID (87.86 vs 117.14) and white-composite LPIPS (0.462 vs 0.537) for T2RGBA comparison.

Why it matters / caveats: Establishes layers rather than pixels as atomic units for generation/editing agents, enabling instruction-addressable, recursive layer decomposition; limitations include lower alpha-boundary cleanliness than Qwen-Image-Layered (2.90 vs 3.33) and difficulty with dense/Chinese typography.

Problem: Self-evolving LLM agents convert accumulated interaction experience into persistent, reusable "skills," but it was unknown whether poisoned experience records could be laundered by the agent's own learning pipeline into durable, hard-to-detect behavioral artifacts that outlive their source.

Method: SkillJack formalizes the experience-to-skill pipeline as four stages (load, transform, persist, route) and introduces "transformation-resilient payloads" — trajectories whose task framing (e.g., "backup," "error recovery") looks legitimate while retaining a policy-relevant malicious action designed to survive the extraction/abstraction transform. It is evaluated against two real systems, SkillX (plan-based trajectory distillation) and Anything2Skill/A2S (taxonomy-guided document compilation), using a shared 150-trajectory AppWorld-format dataset (65 framed poisoned, 65 direct-malicious-baseline, 20 clean) across four policy-risk categories (data exfiltration, privilege escalation, unauthorized transfer, persistent backdoor), with DeepSeek-v4-flash for extraction, routing, and LLM-judge detection.

Key results:

• Sanitization whitewashing: LLM-judge detection drops from 98.5% on raw framed trajectories to 11.4% on extracted skills in SkillX (100% to 44.4% in A2S DOC mode).
• Cross-layer promotion: derived skills trigger the policy-violation proxy on 56.2% of attack-oriented tasks in SkillX and 89.2% in Anything2Skill; clean-task collateral trigger rates are 20.0% (SkillX) and 95.0% retrieval (A2S).
• Persistence isolation: 80.0% of skill-mediated triggers survive deletion of the original poisoned source trajectories.
• Metadata ablation shows detection is lowest (10.0%) when benign name+documentation accompany code vs. 36.7% for code-only, indicating judges key on surface metadata rather than actual behavior; layered defenses (extraction review + runtime monitor) can interact adversarially, e.g., SkillX composition D3 reintroduces triggers (13.3%) versus D2 alone (0.0%).

Why it matters / caveats: Identifies a new, paradigm-level attack surface (validated across two independent systems) showing that deleting poisoned memory does not remove derived skills, motivating provenance-aware skill lifecycle defenses; results are routing-level proxies (not live-execution compromise rates) from a single LLM configuration, and the defense study is explicitly preliminary/exploratory.

Problem: In multi-turn tool-integrated reasoning (TIR), RLVR methods that assign a single trajectory-level advantage (e.g., GRPO) cannot finely attribute credit to individual tool-use decisions, and existing on-policy self-distillation (OPSD) methods rely on privileged context (ground-truth answers, reference rollouts) that is misaligned with the states actually visited by the policy.

Method: TurnSight builds execution-conditioned hindsight from the student's own on-policy tool execution outcomes rather than external references: for each turn it computes token-level log-probability gaps between a frozen reference policy given privileged execution context (at lookahead depths d∈{1,2,3}) and the student's original context, then aggregates these into a single turn-level hindsight signal. It performs direction-consistent multi-lookahead teacher selection (majority vote across horizons, then strongest agreeing teacher), normalizes the selected signal across sibling rollouts (group-relative normalization), and uses it to construct a bounded, sign-aware weight that modulates (interpolates with) the base GRPO advantage without changing its optimization direction. Training uses Qwen3-4B/8B backbones with MatchTIR as the RL backbone, on the FTRL dataset (2,215 instances), implemented in the verl framework.

Key results:

• On Qwen3-8B, TurnSight achieves overall averages of 46.92 (Solve-F1 avg on FTRL), 59.50/42.00/45.00/23.50 on BFCL Base/MissParam/LongContext/Web-Search-like subsets, and 42.02 average accuracy across benchmarks, outperforming all baselines (GRPO, ToolRL, MatchTIR, SDPO, RLSD, SDAR, SOD).
• Ablations show removing turn-level aggregation, group normalization, or multi-teacher selection each degrade average FTRL performance (46.92 → 43.23 / 43.65 / 45.62 respectively).
• Using tool-execution-result-only hindsight outperforms adding ground-truth answers (46.92 vs 40.89) or answer-only (42.80) supervision.
• Full-trajectory hindsight coverage beats supervising only first-five or last-five turns by 2.36% and 3.88% average performance, respectively.
• Optimal hyperparameters found at mixing coefficient λ=0.5 and modulation bound ε_w=0.5, both showing unimodal sensitivity curves.

Why it matters / caveats: The approach shows fine-grained, state-aligned credit assignment can be derived purely from execution outcomes rather than reference trajectories, improving both in-domain (FTRL) and out-of-domain (BFCL, ToolHop) generalization; however, gains rely on multiple frozen teacher forward passes per lookahead depth, adding training-time compute overhead (no deployment cost).

Problem: Existing video-memory systems support either question-conditioned recall or proactive triggering with separate memory/control mechanisms, but no training-free framework provides a single streaming memory that serves both reactive question answering and proactive assistance for wearable/egocentric video assistants.

Method: GROVE incrementally grows a causal, temporally stratified memory M={R,F,E,P} (perceptual trace R, MOMENTs F, EPISODEs E, PATTERNs P) from streaming video via four steps: dual perception (narrative captions + structured perceptual trace per window), adaptive LLM-based episode segmentation, moment extraction (atomizing closed episodes into timestamped facts with core/context/detail roles), and pattern consolidation (matching/merging episodes into recurring cross-day patterns). Each stratum exposes a scale-native retrieval skill (Perception Lookup, Moment Recall, Episode Replay, Pattern Traversal, using BM25+dense reciprocal rank fusion), which a multi-round agent composes for reactive QA (decide-retrieve-answer loop) or a fixed schema uses for proactive assistance (assembling moment/episode/pattern evidence to decide whether to intervene). Perception/captioning uses Qwen3.5-35B-A3B, memory consolidation uses GPT-4.1-mini, and reasoning backbones are GPT-5.2 (MM-Lifelong) or GPT-5-mini (other benchmarks).

Key results:

• On EgoServe, GROVE achieves the best overall Macro-F1 (12.6), beating EgoMemo (8.0) by 4.6 points, with strong gains in safety (19.7 vs 12.5) and error recovery (23.8 vs 1.7).
• On MM-Lifelong, GROVE scores best on all horizons: 23.50 (day), 22.75 (week), 19.98 (month), beating the strongest agentic baseline ReMA by 6.75/3.93/1.36 points.
• On OVO-Bench and StreamingBench, GROVE obtains the best overall scores (67.5 and 65.1 respectively), leading real-time perception (76.2) and contextual understanding (53.1, +9.2 over the strongest baseline).
• On ESTP-Bench, GROVE achieves the best contextual-question score (41.0 vs 26.6) and overall score (28.6 vs 22.9) among non-benchmark-specific methods.
• Ablations show removing the EPISODE layer or flattening the hierarchy drops MM-Lifelong day accuracy from 18.75 to 14.57/14.75; restoring the full hierarchy also cuts reasoning rounds (6.25→4.95) and latency (59.3s→47.6s per question).

Why it matters / caveats: GROVE demonstrates that organizing memory by temporal scale (rather than a flat caption index) improves both accuracy and retrieval efficiency for reactive and proactive video assistants; limitations include reliance on a frozen perception model whose ingestion errors cannot be recovered later, and rising construction/retrieval cost as observed history grows.

Problem: There is no benchmark to automatically evaluate whether natural-language explanations produced by coding agents alongside their patches are trustworthy, even as large, hard-to manually review agent-generated code changes increasingly rely on such explanations.

Method: ExplainBench evaluates explanations via an LLM questionnaire: given only an agent's explanation, an LLM (GPT-5-mini) answers multiple-choice questions across four types—end-to-end intent (masked bug-reproducing property-based test), end-to-end effect (test pass/fail before/after patch), local intent, and local effect (function-level delta-behavior expressions)—built on 297 SWE-bench Verified instances with GPT-5.2-generated PBTs, a custom Python execution tracer (via sys.settrace) for delta-behavior extraction, and MMR-based distractor selection. The explanation score is the fraction of correctly answered questions. The paper also introduces ExplanationAuditAgent, which runs differential testing (DiffExecution tool) and call-graph inspection to audit and refine explanations against evidence.

Key results:

• Across 5 agents, explanation quality rankings diverge from SWE-bench Verified efficacy rankings: trae-agent has the highest SWE-bench resolution (0.818) but the second-lowest ExplainBench score (0.558), while OpenHands ranks lowest in efficacy (#4, 0.727) but highest in explanation score (0.597).
• 79.30% of non-passing (incorrect) patches are nevertheless predicted by the QA LLM to pass the bug-reproducing test based on the explanation, showing explanations are frequently overconfident about correctness.
• Local intent/effect misalignment rates are much higher than end-to-end (e.g., mini-SWE-agent: 39.9%/38.9% local misalignment vs lower end-to-end rates), indicating explanations are worse at describing function-level behavior than global outcomes.
• ExplanationAuditAgent improves the explanation score for all 5 agents (e.g., mini-SWE-agent 0.435→0.580, +33.5%; trae-agent 0.558→0.620, +11.3%) at ~$0.05 per explanation.
• Agreement between human labelers and the LLM judge reached Cohen's Kappa of 0.7 (substantial agreement) on 40 sampled questions.

Why it matters / caveats: The work shows explanation trustworthiness is a distinct evaluation axis from patch-resolution efficacy and that current agents' explanations are often confidently wrong, motivating structural fixes (e.g., enforced summary fields in finish tools) and automated auditing; the benchmark is built specifically on SWE-bench Verified/Python and evaluates content correctness only, not readability or developer satisfaction.

Problem: Persona skills that distill personal interaction histories into portable, executable agent artifacts concentrate fragmented personal information and amplify impersonation risk through reuse, but no benchmark exists to evaluate this pipeline's privacy leakage, impersonation risk, or candidate defenses.

Method: AntiSkillBench constructs 2,500 persona-grounded dialogue traces (7,500 user turns) from 50 user profiles (demographics, background, Big Five personality, communication style) sampled from OpenCharacter, each expanded via three-round simulated dialogues over general-assistance/tool-design/math scenarios. It evaluates three skill-distillation protocols (Direct Distill, Three-stage Distill, and COLLEAGUE.SKILL-style Colleague Distill) using three metrics: Skill Coverage (skill-level leakage), Field QA Accuracy, and VocabGain (embedding-based marker-hit-rate gain normalized against an oracle-profile agent) for agent-level impersonation. It also evaluates four defenses spanning online/post-hoc and active/passive settings: Privacy Sanitization (PS), Adversarial Obfuscation (ADV), and Semantic-level Backdoor Injection (SBD, online and post-hoc), tested on GPT-5.4, Claude Haiku 4.5, and Gemini 3.6 Flash.

Key results:

• On GPT-5.4, overall Skill Coverage reaches 66.2 (Three-stage), 63.6 (Direct), 55.2 (Colleague), with communication-style coverage especially high (88.0–92.0) and personality 69.0–75.7.
• Agent-level impersonation is significant: GPT-5.4 with Three-stage Distill reaches 56.0 overall QA Acc and positive VocabGain across all dimensions (up to 87.7 for communication); similar patterns hold for Gemini 3.6 Flash (up to 48.4 QA Acc, 40.0 VocabGain) and Claude Haiku 4.5 (up to 52.5 QA Acc).
• Active defenses give only partial protection: Online PS reduces Direct-Distill Skill Coverage/QA Acc/VocabGain to 51.7/40.6/9.0, but personality Skill Coverage remains high (70.0) after PS.
• Passive backdoor defense (SBD) is highly effective under Direct/Three-stage Distill (up to 98–100% static ASR, 40–83% behavioral ASR) but nearly fails under Colleague Distill (40.0% static ASR, 0.0% behavioral ASR), because persona-centric abstraction absorbs the backdoor cue as a persona trait.

Why it matters / caveats: The results show persona-skill privacy/impersonation risks are structural (persisting across three model backbones and three distillation protocols) and extend beyond explicit attributes to communication style and personality, while current defenses are only partial and distillation-protocol-dependent—motivating defenses that jointly target artifacts, downstream behavior, and the distillation process itself.

Problem: Training autoregressive/streaming video world models typically requires adapting large pretrained bidirectional video diffusion models through complex, resource-intensive post-training pipelines that suffer from a structural mismatch between bidirectional pretraining and causal streaming inference, and no lightweight, fully reproducible from-scratch baseline exists.

Method: MiniWorld is a block-causal Video Diffusion Transformer trained with Rectified Flow/Flow Matching in the latent space of a pretrained Wan2.2 VAE, using block-causal self-attention (bidirectional within a chunk, causal across chunks) and AdaLN-LoRA action/camera-pose conditioning. Building on Diffusion Forcing, it introduces a Chunk-oriented Probability Propagation (CoPP) scheduler enforcing non-decreasing per-chunk noise timesteps, plus a two-stage continued training recipe (pretraining on 21/46-frame clips, then continued training on 125/253-frame sequences with timestep shifting). At inference, it uses a structured rolling KV cache (persistent sink anchor + FIFO cache with RoPE re-shifting) and pipelined asynchronous chunk-wise denoising for bounded-compute streaming generation. Models scale from MiniWorld-B (0.12B) to MiniWorld-3B, trained on a single 8-GPU server using the Muon optimizer, evaluated on DROID (robot actions) and RealEstate10K (camera trajectories).

Key results:

• On DROID, MiniWorld-1B improves Trajectory Accuracy by 249%, Depth Accuracy by 238%, LPIPS by 216%, and SSIM by 125% over a bidirectional short-video baseline (normalized to 1); appearance metrics improve 26–82% and VLM judge scores 63–78%.
• On RealEstate10K, Photometric Smoothness improves 89%, Depth Accuracy 55%, Subject/Background Consistency 50%/46%; PSNR/SSIM/LPIPS gain 34%/19%/27%.
• Replacing a 32-chunk full attention window with an 8-chunk in-flight window plus KV cache increases streaming throughput from 3.31 to 7.29 FPS (2.20x speedup) and cuts first-chunk latency from 74.0s to 4.86s (15.2x reduction), while DiT accounts for 94–97% of per-chunk compute (VAE only 2.6–5.7%).
• Scaling from 0.5B to 3B parameters on RE10K improves dynamic degree by 22%, depth accuracy by 18%, image quality by 14%, and flow score by 12%, with monotonic gains across scales.
• Reducing KV-cache size from 24 to 12 chunks is nearly indistinguishable in quality; a 6-chunk cache introduces only single-digit changes, showing the model does not need to retain full history once the active window and sink anchor are fixed.

Why it matters / caveats: MiniWorld shows that stable, competitive long-horizon streaming world models can be trained end-to-end from scratch within days on a single 8-GPU server (with full code/checkpoints released), lowering the barrier to world-model research; the authors note it is evaluated at modest scale/domain coverage relative to frontier systems and still accumulates prediction errors over very long rollouts, so it is positioned as a reproducible baseline rather than a quality ceiling.

Problem: It is unclear whether LLM agents equipped with dynamic skill libraries can genuinely evolve reusable capabilities from sequential task feedback, rather than merely benefiting from retained context.

Method: The authors build ContinualSkillBench, a benchmark with five domains (Law, Healthcare, Finance, Math, Office), each containing 100 subtasks curated from existing benchmarks (e.g., OlympiadBench, LawBench, TAT-QA, GAIA, ClawBench, MedAgentsBench) and ordered via LLM-assisted skill-dependency labeling and graph-based topological sorting so that later tasks can reuse skills from earlier ones. Agents (GPT-4o, GPT-5.3-Codex, Claude 4.7 Opus) run on Codex CLI and Claude Code harnesses through a three-turn instruction–execution–reflection protocol, using Create Skill/Modify Skill meta-skills, and are compared in Sequential (persistent skill repository), Independent (memoryless baseline), and pure in-context learning (ICL, no explicit skills) settings.

Key results:

• Sequential execution improves normalized reward in 14 of 15 model–domain combinations, an aggregate relative gain of 16.9% (+0.078 absolute).
• GPT-5.3-Codex shows the largest average normalized improvement (+0.098), vs. GPT-4o (+0.077) and Opus 4.7 (+0.058); Healthcare shows the largest domain-level gain (+0.149).
• Across Law, Finance, and Healthcare, ICL (0.605) and skill-maintaining Sequential (0.602) achieve near-identical average normalized reward, both well above Independent (0.466).
• GPT-4o accumulates 384 skills across five domains vs. 205 for GPT-5.3-Codex, but GPT-4o's skills are reused less often and rated lower quality, indicating skill fragmentation.

Why it matters / caveats: The results suggest current "skill evolution" gains largely stem from in-context adaptation rather than genuine reusable skill abstraction, though explicit skills still help for tasks needing precise, reusable procedures (e.g., raising Healthcare Programmatic score from 0.250 to 0.500). The benchmark is limited to a fixed set of sources/models and does not cover long-tail real deployment conditions.

Problem: Automated, non-intrusive analysis of children's gait from standard RGB video is needed for clinical screening of developmental disorders (e.g., cerebral palsy), but existing gait models are built for adults and fail to generalize to children's erratic, high-variance motor patterns.

Method: The authors introduce the Children Gait Video (CGV) dataset (1,185 videos, 339,236 frames, 110 subjects aged 3-17, 60 FPS, synchronized sagittal/coronal views) annotated by pediatricians with the 17-item Edinburgh Visual Gait Score (EVGS, binarized to typical/atypical) plus SAM-3 segmentation masks and Sapiens-2B keypoints. They benchmark zero-shot MLLMs (Gemini 3 Pro, GPT-5.2, Qwen3-VL-235B, GLM4.6V, etc.), a fine-tuned Qwen3-VL-4B (Qwen3-VL-ChildGait), and gait-recognition/video models (GaitSet, BiggerGait, VideoMAE v2), then propose ChildGait-Video, an end-to-end VideoMAE v2-based model using token-level kinematic prompting (rendering skeletal keypoints onto RGB frames) and mask-guided patch pruning (dropping non-foreground patches) followed by an MLP classification head.

Key results:

• Zero-shot MLLMs achieve only 50-60% average accuracy on EVGS item classification, near random guessing for the binary task.
• Fine-tuning Qwen3-VL-4B yields negligible gains (+1-2% average accuracy), with some items even degrading (e.g., HVV -1.5%).
• Traditional gait recognition models (BiggerGait, SwinGait, etc.) also fail, with 45-56% average accuracy, while fine-tuned VideoMAE v2 reaches 69-72%.
• ChildGait-Video achieves state-of-the-art 70-93% accuracy across all 34 scoring items, with average F1 of 0.83, a statistically significant improvement over VideoMAE v2 (McNemar's test, p=2.3×10⁻⁴).
• Human expert agreement (ICC=0.93) and human baseline accuracy of 93.8% establish an upper bound; increasing input frames from 8 to 32 boosts accuracy on item IC from 74.1% to 90.7%.

Why it matters / caveats: This is presented as the first open-source multi-view children's gait video dataset and establishes that current foundation models (both MLLMs and gait-specific) are inadequate for pediatric gait assessment, motivating domain-specific modeling; the dataset is restricted to one hospital site and future work aims to extend to in-the-wild home/rehab settings.

Problem: Robotic cleaning of viscous stains (high viscosity, complex rheology) is difficult because wiping spreads the stain and scrubbing risks surface damage, and no existing learning-based method handles diverse stains and curved surfaces generally.

Method: Push-Wiper reformulates viscous stain cleaning as a topological aggregation problem, using a "gather-then-finish" paradigm: a Diffusion Policy (DDIM scheduler, 100 diffusion training steps, 10 inference steps, predicting n=16 actions) acts as a low-frequency planner over binary stain-mask observations, outputting low-dimensional planar pushing actions (x, y, Δθ) that an Arbitrary Surface Pose Interpolator (ASPI, using B-spline fitting, trapezoidal velocity profiles, and Slerp) converts into 6-DoF trajectories executed via a hybrid admittance-based force–position controller (constant normal force setpoint), followed by a post-processing phase with five motion primitives (dabbing, scraping, rinsing, squeezing, final full-cover wipe).

Key results:

• Push-Wiper achieves an overall average Cleaning Score (CS) of 89.88% vs. 32.64% for Full-Cover baseline and 44.98% for PushAll-Onetime baseline, on ketchup and peanut butter stains (150 demonstration tasks, 448 pushing trajectories, augmented to 2,688 for training).
• Post-processing raises CS to 100% for ketchup and 98.51% for peanut butter (up from 93.76%/85.12% without it).
• Zero-shot generalization: CS of 91.45% (convex) and 93.42% (concave) on unseen curved surfaces not seen in training; 100% on solids, 92.62% on liquids, and 94.42% on unseen viscous stains (CVS, e.g., black pepper/oyster sauce).
• Cleaning score improvement of up to ~130% over baselines, with average runtime of 130s per trial.

Why it matters / caveats: Decoupling 2D topological planning from 3D geometric execution enables strong zero-shot generalization across stain types and surface geometries without retraining; the paper notes future work is needed to add collision constraints for safe human-robot interaction and broaden post-processing to more stain types.

Problem: Video-generative World Action Models (WAMs) that predict future states in VAE latent space suffer from "Training-Distribution Hallucination" — predicted futures drift toward training-domain content under visual shifts (background, lighting, camera viewpoint), hurting robustness.

Method: ST-WAM uses frozen DINOv3 features as a shared semantic representation in two complementary roles: Dual-Space Future Experts (DSFE), a three-branch Mixture-of-Transformers (visual DiT from Wan2.2-TI2V-5B, semantic DiT, action DiT) that jointly predicts future VAE latents and DINO features via flow matching with a structured cross-branch attention mask; and Current-Anchored Intent Retrieval (CAIR), which uses a frozen Qwen3-VL-4B-Instruct to build current visual-language query tokens that cross-attend into recent DINO-feature history to produce short-horizon "intent tokens" fed only to the action expert. Training uses a joint flow-matching objective over visual, semantic, and action branches (loss weights λv=1.0, λs=0.02, λa=1.0), with no additional embodied pretraining, and inference drops the future branches for efficiency.

Key results:

• Diagnostic frame-triplet analysis: DINOv3 achieves 0.904 same-state cosine similarity vs. 0.686 for Wan-VAE, and correctly discriminates different task states in 95.2% of triplets vs. 60.0% for Wan-VAE; 70.6% of 180 audited predictions from two prior WAMs showed Training-Distribution Hallucination.
• ST-WAM achieves 98.7% success on LIBERO and 92.77% average (93.06% clean / 92.48% randomized) on RoboTwin 2.0, both without embodied pretraining, matching or exceeding pretrained baselines like LingBot-VA (98.5%/92.20%).
• Zero-shot on LIBERO-Plus: 72.8% overall, a 21.3-point improvement over Fast-WAM (51.5%), with gains up to 39.0 and 41.8 points on camera and sensor-noise perturbations.
• Real-world (5 tasks, Agilex Piper arm): 79.3% success under nominal conditions and 61.5% under visual shifts, vs. Fast-WAM's 64.7%/25.8% (dropping 38.9 points vs. ST-WAM's 17.8-point drop).
• Ablations show DINO cannot simply replace VAE (DINO-only: 39.7% on LIBERO-Plus) and both DSFE and CAIR components are necessary (removing semantic future expert drops LIBERO-Plus to 63.5%; removing CAIR-style current-anchored retrieval drops to 56.5-64.7% depending on variant).

Why it matters / caveats: Semantic-temporal modeling meaningfully improves robustness to visual distribution shifts at only a modest 1.24x inference latency overhead (756ms vs. 609ms per action chunk) without requiring large-scale embodied pretraining; the authors note future work is needed to extend robustness beyond visual shifts to changes in physical dynamics and embodiment.

Problem: Existing paper-to-poster generation systems hide request-level failures by scoring only completed outputs, produce non-editable rasters, or (for coding-agent workflows) are prohibitively costly, and none jointly provide print-ready validity, native editability, and controllable design diversity.

Method: PosterMELD is a template-first multi-agent pipeline: a structural template library is mined from real posters via MinerU-based block extraction and Ward hierarchical clustering on 12×12 occupancy grids into 24 topologies (16 landscape, 8 portrait) with capacity-aware "slot contracts" (character/bullet budgets, prominence, compatibility). Five skill-guided agents (Content, Template, Layout, Visual, Review) operate on a Shared Typed Poster State: content is written to fit known slot geometry before rendering, and a Review Agent combines deterministic geometric/asset gates with VLM review, routing failures to Bounded Repair (rewrite, reflow, resize, rerender); a Pipeline Harness enforces budgets and freezes configuration/seed. Each accepted request outputs an editable PPTX plus PNG render, and evaluation uses a new Print-Ready Rate (PRR) metric and a frozen GPT-5.5 judge for conditional Craftsmanship–Harmony–Expressiveness (CHE) scores across a 621-paper benchmark (14 sources, 10 domains).

Key results:

• PosterMELD attains 81.3% PRR vs. 24.2% for P2P, 15.8% for PosterGen, and 0.2% for Paper2Poster — 3.4x and 5.2x the rates of P2P and PosterGen respectively.
• PosterMELD achieves the highest conditional CHE (3.247) among generated methods with multiple print-ready outputs, with craftsmanship (3.455) closest to the human reference (3.604, gap of 0.149).
• Mean cost is $0.38 per request vs. $10.78 for Codex+Skill (about 28x higher) and $0.18 for non-editable GPT-Image-2 (85.2% PRR).
• The GPT-5.5 judge used for scaled PRR/CHE evaluation achieves 84.0% accuracy and κ=0.683 against 1,398 human-annotated posters (four annotators, 201-paper subset).
• PosterMELD leads open-source systems on 8 of 10 Universal content-quality dimensions, trailing P2P only on authorship/title accuracy and content fidelity.

Why it matters / caveats: The work shows print-readiness, editability, and design controllability can be achieved jointly at low per-request cost, but automatic gates only certify geometry/legibility rather than scientific correctness, portrait layouts remain less mature than landscape, and the diversity analysis is qualitative rather than a scalar metric.

Problem: Existing financial LLM benchmarks rely on short, pre-cropped tables and knowledge-style multiple-choice questions, so it is unclear whether LLMs perform genuine structured financial reasoning or merely surface-level pattern matching over long, real-world corporate statements.

Method: The authors build FININDICES, a benchmark of uncropped financial statements (up to 32K tokens) synthesized via a three-stage pipeline (Financial Statement Extraction, Context Construction & Index Alignment with adversarial traps like distractor injection and information deficiency, and Instruction & QA Pair Generation with canonical formulas and programmatic ground-truth execution), covering 829 companies, 384 indices, and 28 reporting periods across two task paradigms (Single-Index computation and Table-Index tabulation) and three capability axes (Domain, Caliber, Temporal). They evaluate closed- and open-weight LLMs with/without explicit formula hints and fine-tune Qwen3.5-35B-A3B via SFT on Gemini-distilled reasoning trajectories.

Key results:

• Gemini-3.1-Pro drops from 79.61% (Single-Index, with hint) to 70.07% (Table-Index, with hint), and collapses to 38.22% on Table-Index without hints.
• Domain-specific financial LLMs largely fail on structured Table-Index tasks (DianJin-R1-32B: 41.29% Single-Index vs. only 11.41% Table-Index w/ hint; Fin-R1 and Llama-Fin-8B score near 0%).
• SFT on 6,301 distilled trajectories yields +8.54% (Single-Index) and +3.82% (Table-Index) gains in the zero-hint setting, plus +0.61% to +2.39% gains on general financial QA benchmarks (CFinBench, FinEval, FLAME) with no catastrophic forgetting.
• Dataset averages 16,202 tokens of context (vs. ~590-1,000 for FinQA/TAT-QA), with Table-Index queries requiring up to 9 input periods and generating up to 20 target values per sample.

Why it matters / caveats: The results reveal a "Knowledge Bottleneck" (fragile intrinsic accounting knowledge) and a "Structural Bottleneck" (multi-metric table generation drains reasoning capacity), suggesting current LLMs are unreliable for high-stakes financial workflows; the benchmark is limited to core corporate financial statements and does not cover market data, macro indicators, or qualitative financial tasks.

Problem: ALiBi's linear positional bias grows unboundedly with token distance, and the authors show this drives attention logits into floating-point underflow, zeroing out attention weights and rendering affected heads "blind" beyond a certain distance—an unreported failure mode undermining ALiBi's claimed length-extrapolation ability.

Method: The paper analytically characterizes the underflow threshold (τu≈-103.27 for fp32, -92.18 for bf16) and defines a per-head "blindness distance," then empirically confirms underflow in pretrained ALiBi models (BLOOM, Falcon-RW, MPT) via perplexity and passkey/needle-in-a-haystack (NIHS) retrieval probes. They further pretrain 148M-parameter Llama-style decoder models (SmolLM recipe, 20B FineWeb-Edu tokens) with varied slope configurations (Steep/Safe/Wide) and evaluate four mitigation strategies—clamping, robust slopes, log-scaled distances, and soft capping—individually and combined.

Key results:

• At token distance 2048, 36.6% of attention entries have already underflowed under default ALiBi slopes.
• Combining clamping with log-scaled distances improves out-of-context passkey retrieval nearly ten-fold over the ALiBi baseline (0.79 vs. 0.08 AUC).
• Default ALiBi remains a strong baseline on NIHS out-of-context retrieval (0.20 AUC), which no mitigation strategy surpasses.
• Downstream benchmark differences (common sense, QA, language) across all configurations are small, only 1.6 to 3.4 percentage points.

Why it matters / caveats: The failure mode affects all previously trained ALiBi-based models and related linear-bias encodings (e.g., TISA), with likely more severe consequences during training (via the backward pass) than inference; experiments are limited to 148M-parameter models and a fixed architecture/corpus, so findings may not fully generalize to larger-scale models.

Problem: Existing visual piano transcription (VPT) systems only process short (≤0.2s) video windows, achieve much worse offset (key-release) accuracy than onset accuracy, and none report note-level velocity, unlike audio-based transcription which conflates key release with pedal-extended sound decay.

Method: V2N (Video to Notes) uses an S2S-style ResNet-18 visual feature extractor (with a learned slope prior and 5-frame aggregation) feeding a shared temporal backbone of 3 Conformer ConvModule blocks (FFN→depthwise conv→FFN, no self-attention) over a 1-second window, followed by four parallel BiLSTM→linear heads for onset, offset, key hold, and velocity, trained jointly with per-frame supervision (soft triangular kernels for onset/offset, binary key-hold labels, masked MSE velocity loss, weighted total loss with onset weight 2). Inference uses offset-guided note decoding, terminating notes at offset-head peaks with key-hold as fallback, evaluated on PianoVAM and R3 (Rachmaninoff/mixed repertoire) datasets.

Key results:

• On PianoVAM, V2N nearly doubles offset-metric F1 over the best prior VPT baseline (PPAN: 45.9% → V2N: 89.5% at 50ms) and gains +48.6pp on the strictest joint offset+velocity metric (29.7% → 78.3%).
• V2N matches Li et al.'s onset F1 on PianoVAM (within 0.5pp) while being the first video-only system to report note-level velocity.
• Ablations show multi-frame loss (+1.3pp onset, +4.5pp offset), Conformer sequence modeling (+2.5pp onset, +5.7pp offset), and extending window from 0.2s to 1s each incrementally improve performance, with the full V2N configuration reaching 94.7% onset / 89.5% +Off F1 at 50ms on PianoVAM.
• Cross-dataset transfer (PianoVAM ↔ R3) collapses to near-zero Onset F1 for all systems (e.g., 0.0-1.5% at 50ms) due to fixed-geometry preprocessing mismatches (median keyboard width ~784px on R3 vs ~606px on PianoVAM).

Why it matters / caveats: V2N is the first complete video-only MIDI transcription system (onset+offset+velocity+key-hold), but cross-dataset generalization to unseen camera geometries remains an open problem, and the model does not estimate sustain-pedal control changes.

Problem: MLLM-based segmentation faces a trilemma among high segmentation accuracy, preserved dialogue ability, and fast inference, and prior single-target methods (including the authors' own STAMP) cannot retain multiple semantic or instance identities without repeatedly re-running target-specific predictions.

Method: The paper extends Binary All-Mask Prediction (STAMP: autoregressive Phase 1 dialogue ending in an in-vocabulary <SEG> trigger, then non-autoregressive Phase 2 that prefills image-aligned [MASK] tokens fused with patch features and classifies them foreground/background via hybrid causal+bidirectional attention) into Structured All-Mask Prediction (STAMPlus), built on Qwen2-VL (2B/7B). STAMPlus's Phase 1 generates a JSON-style target list with explicit class IDs and optional bounding boxes, and Phase 2 uses a multi-class mask head (200-class capacity) dynamically bound to those IDs, trained with combined text cross-entropy plus multi-class cross-entropy and Dice losses on a unified mixture (RefCOCO family, ReasonSeg, RRSIS-D, EarthReason, COCO-Stuff, COCO Panoptic, MUSE, LLaVA-665k), with high-resolution mask-token scaling (up to 3200 tokens) for remote-sensing small targets.

Key results:

• STAMPlus-7B achieves 81.0 average cIoU on the RefCOCO family (vs. 78.9 for Text4Seg++) and STAMP-7B reaches 80.7, both as single unified checkpoints without task-specific fine-tuning.
• On RRSIS-D remote-sensing segmentation, STAMPlus-7B scores 76.2 average vs. 74.4 for the strongest specialized baseline and 70.8 for Text4Seg++; on EarthReason it reaches 74.0 vs. 70.1.
• Providing human-annotated target boxes at Phase 1 raises STAMPlus-2B's average cIoU from 77.4 to 89.1, showing Phase 1 target quality directly drives Phase 2 segmentation accuracy.
• Multi-target inference latency for 12 categories drops from 13.50s (repeated STAMP inference) to 5.16s (single-pass STAMPlus).
• Look-twice reasoning with STAMPlus-2B's learned spatial attention raises average VQA score (TextVQA/InfoVQA/POPE/DocVQA) to 71.1, versus 68.5 for the same protocol on the Qwen2-VL-2B backbone.

Why it matters / caveats: STAMPlus is presented as a single unified checkpoint spanning referring, reasoning, open-vocabulary semantic, instance-aware, and remote-sensing segmentation while preserving dialogue ability and single-pass efficiency; the multi-class head capacity is fixed at 200 targets and results depend on the quality of the model's own Phase 1 target-list generation.