AI papers — 2026-06-01

Problem: Existing LLM search agents rely on pre-computed retrieval indices, preventing flexible, granularity-agnostic information access over large text corpora.

Method: GrepSeek trains a compact LLM to issue executable shell commands (e.g., grep, ripgrep) directly against a corpus, bypassing indexing entirely. Training uses a two-stage pipeline: first, a cold-start dataset is built via an answer-aware Tutor and answer-blind Planner that generate verified shell-command trajectories; second, the policy is refined with Group Relative Policy Optimization (GRPO). A semantics-preserving sharded-parallel execution engine accelerates shell-based retrieval across corpus shards.

Key results:

• Achieves the strongest overall token-level F1 and Exact Match across seven open-domain QA benchmarks (specific numbers not stated in the extracted text)
• Parallel execution engine accelerates shell-based retrieval by up to 7.6× over sequential execution while preserving byte-exact equivalence

Why it matters / caveats: GrepSeek shows that direct corpus interaction is a practical complement to traditional retrieval-augmented search, enabling surgical, granularity-agnostic evidence retrieval without an index. A key limitation is purely lexical interaction, which degrades on queries with substantial surface-form variation (e.g., paraphrase-heavy queries).

Problem: Actionable person-grounded knowledge (expertise, judgment, interaction style) is scattered across heterogeneous traces and cannot be easily packaged into reusable, inspectable agent capabilities.

Method: COLLEAGUE.SKILL is an automated trace-to-skill distillation pipeline that ingests heterogeneous materials (chat logs, documents, email, screenshots, public research, video subtitles) and outputs a versioned skill package conforming to the Agent Skills standard (a SKILL.md-centered folder). The package has two coordinated tracks: a capability track (practices, mental models, decision heuristics) and a bounded behavior track (communication style, interaction rules, correction history). The generated skills can be installed into agent hosts such as Claude Code, OpenClaw, Codex, and Hermes, and support natural-language correction and rollback.

Key results:

• Public repository has approximately 18,500 GitHub stars at time of writing
• Gallery lists 215 skills from 165 contributors with more than 100,000 cumulative stars across skill cards
• Specific accuracy or quality metrics: Not stated

Why it matters / caveats: The system operationalizes person-grounded knowledge as portable, correctable artifacts rather than opaque prompts or hidden memories, which has implications for knowledge continuity in organizations. The paper is primarily a system description rather than an empirical study, so comparative quality evaluations against baselines are not reported.

Problem: Unified multimodal models still depend on a frozen, separately pretrained VAE for image generation, imposing a structural bottleneck that caps generation quality regardless of further training.

Method: Representation Forcing (RF) eliminates the VAE by training the transformer decoder to autoregressively predict visual representations (extracted by the model's own understanding encoder) as intermediate tokens before pixel generation, under the same next-token prediction objective used for language. These predicted representation tokens remain in-context and guide pixel-space diffusion within the shared transformer backbone, providing structural scaffolding without any external latent space. RF is applied to both pixel-space and VAE-based unified multimodal models under controlled settings with matched architecture, data, and training budget.

Key results:

• Pixel-space model with RF matches VAE-based unified model baselines on standard image generation benchmarks (specific FID/IS numbers not stated in the extracted text)
• Pixel-space RF outperforms the VAE-based variant on image understanding benchmarks (specific numbers not stated)
• Generates images at 1024×1024 resolution

Why it matters / caveats: RF offers a concrete path toward fully end-to-end unified multimodal models by removing the dependency on separately pretrained generative latent spaces. The evaluation is conducted under controlled small-scale settings; scaling behavior to very large models and datasets remains to be demonstrated.

Problem: Zero-shot TTS systems that synthesize multi-speaker dialogue by stitching per-turn monologue outputs break acoustic consistency, conversational coherence, and affective continuity across speaker turns.

Method: SwanVoice is a zero-shot TTS model supporting 1–4 speakers, built on a 25 Hz VAE for sequence compression, a flow-matching Diffusion Transformer (DiT) conditioned on speaker-turn IDs, and raw-text conditioning augmented with pause-aware symbols and pinyin substitution for Chinese pronunciation. Training follows a curriculum from monologue speech through mixed and real dialogue data, followed by DiffusionNFT post-training with phone-level and speaker-similarity rewards. Training data comes from SwanData-Speech, a pipeline processing approximately 2.59 million hours of raw audio (2.24M hours Chinese, 0.35M hours English) using Swan Forced Aligner for pause-aware word-level alignment and RobustMegaTTS3 for pronunciation-hard synthetic cases generated by MegaTTS 3 using Qwen3-235B-A22B-Instruct-2507.

Key results:

• Achieves higher richness and hierarchy scores than all evaluated open-source baselines on SwanBench-Speech in both monologue and dialogue settings (specific numeric scores not stated in extracted text)
• Content accuracy identified as the main remaining limitation

Why it matters / caveats: SwanVoice demonstrates that a unified non-autoregressive flow-matching model can handle long-form multi-speaker dialogue without the quality degradation of turn-stitching approaches. Content accuracy (pronunciation correctness) remains the primary unsolved challenge, and the evaluation is conducted on a benchmark (SwanBench-Speech) introduced by the same authors.

Problem: Reinforcement learning methods for long-context reasoning are limited by low-confusability distractors and sparse outcome-only reward signals that cannot supervise intermediate reasoning steps.

Method: LongTraceRL constructs multi-hop training questions via knowledge graph random walks over the KILT Wikipedia snapshot, then uses search agent trajectories to build tiered distractors: documents the agent read but did not cite (Tier-1, high confusability) and documents that appeared in search results but were never opened (Tier-2, low confusability). For reward design, a rubric reward uses gold entities at each hop of the reasoning chain as entity-level process supervision, applied only to responses with correct final answers (positive-only strategy) to prevent reward hacking. The resulting composite reward combines the rubric reward (weighted 0.3) with an outcome reward (weighted 0.7).

Key results:

• Qwen3-4B with LongTraceRL achieves an average gain of 5.7 points over the base model across five long-context benchmarks
• Qwen3-4B surpasses the strongest baseline by 2.5 points
• Evaluated on three reasoning LLMs at 4B–30B parameter scales across five long-context benchmarks

Why it matters / caveats: The tiered distractor construction from real search trajectories produces significantly more challenging and realistic training data than random sampling, and the rubric reward provides finer-grained process supervision than outcome-only RLVR. The method depends on a search agent to generate training trajectories, adding infrastructure overhead; performance on out-of-domain corpora beyond KILT Wikipedia is not evaluated in the extracted text.

Problem: Real-world image restoration models generalize poorly because paired high-quality/low-quality training data is scarce, expensive to collect, and synthetic datasets fail to capture real degradation complexity.

Method: The paper proposes using generative multimodal foundation models (MFMs) as a scalable source of high-quality (HQ) restoration targets for real-world low-quality (LQ) images. Nine MFMs are systematically evaluated including Nano-Banana-2 and GPT-Image-2; Nano-Banana-2 with VLM-based adaptive prompting is selected as the generator. This model is used to construct GGT-100K: 103,707 LQ-HQ training pairs at 1024×1024 resolution plus a 500-pair test set, covering general mixed degradations, rain, haze, snow, low-light, and old photos, with multi-stage quality control (automatic metric-based filtering, VLM-assisted screening, manual verification).

Key results:

• FLUX-Controlnet PSNR: 26.18 (with GGT-100K) vs. 25.86 (without)
• Qwen-Image-Edit PSNR: 28.25 vs. 25.13
• NAFNet PSNR: 27.18 vs. 25.80
• FoundIR PSNR: 29.20 vs. 29.20 (unchanged on PSNR; gains on MUSIQ: 65.30 vs. 65.30)
• MUSIQ scores: FLUX-Controlnet 62.55 vs. 51.42; Qwen-Image-Edit 46.71 vs. 42.01; NAFNet 43.92 vs. 42.54
• VLM-based rating (VLM-R %): FLUX-Controlnet 85.8 vs. 77.4; Qwen-Image-Edit 53.8 vs. 27.6; NAFNet 60.8 vs. 35.8; FoundIR 90.20 vs. 90.20

Why it matters / caveats: GGT-100K provides a scalable, model-generated paired dataset that consistently improves generalization across diverse IR architectures without requiring physical data collection. A key caveat is that MFM-generated targets may introduce hallucinations or color shifts, and the quality of HQ targets is bounded by the capabilities of Nano-Banana-2.

Problem: On-policy distillation (OPD) of LLMs suffers from poor early-training rollouts, because the weak student generates low-quality prefixes that carry insufficient teacher signal.

Method: Trust-Region Behavior Blending (TRB) replaces the student's rollout policy during an early warmup phase with a behavior policy that is maximally close to the teacher within a student-centered KL trust region (DKL(µ ∥ πS) ≤ ε), solved in closed form as µβ(a|h) ∝ πS(a|h)^(1−β) · πT(a|h)^β. The per-prefix reverse-KL OPD loss (following MiniLLM / GKD) is kept unchanged; only the prefix-collection distribution changes. The KL budget ε is annealed to zero after a fixed warmup horizon, returning training to pure student rollouts. TRB is evaluated against vanilla OPD, target-side reformulation, direct token replacement, persistent blending, and simpler warmup heuristics on two math-reasoning distillation settings.

Key results:

• TRB attains the strongest average performance across two math-reasoning distillation settings (specific accuracy numbers not stated in the extracted text)

Why it matters / caveats: TRB is a lightweight, architecture-agnostic warmup method that improves OPD stability without modifying the distillation objective or requiring additional data. Evaluation is limited to two math-reasoning benchmarks; generalization to other tasks (e.g., coding, instruction following) is not demonstrated.

Problem: Existing spatial audio synthesis systems face a tradeoff between generation quality and high first-frame latency, and struggle to accurately capture spatial directionality from multimodal inputs for First-Order Ambisonics (FOA) generation.

Method: SwanSphere is a unified streaming framework combining a causal autoregressive language model operating at the patch level with a localized Diffusion Transformer (LocDiT) that performs intra-patch bidirectional denoising, decoupling long-range temporal modeling from local continuous rendering. A VideoMAE encoder extracts visual-spatial features from panoramic videos. Spatial Video-Audio Contrastive (SVAC) learning aligns the video encoder with the acoustic domain using four categories of physics-aware positive and negative pairs. A multi-objective Online Direct Preference Optimization (ODPO) scheme aligns generated audio with human preferences across aesthetic, semantic, and spatial dimensions. An MLLM-based automated annotation pipeline produces spatial caption-FOA training data; curriculum learning begins with large-scale monaural audio pretraining.

Key results:

• Outperforms baseline models in video-to-spatial and text-to-spatial audio generation (specific metrics not stated in extracted text)
• Achieves lower first-chunk latency than baseline models (specific latency values not stated)

Why it matters / caveats: SwanSphere addresses both quality and latency bottlenecks in spatial audio generation for VR/AR applications through architectural decomposition and contrastive cross-modal alignment. Specific benchmark numbers are not reported in the extracted text, limiting quantitative assessment.

Problem: Real-time streaming video-to-video editing on consumer GPUs fails due to incompatibility between temporal consistency requirements and the memory/latency constraints of existing diffusion transformer architectures.

Method: SANA-Streaming builds on the SANA-Video linear-attention backbone, replacing a portion of linear attention blocks with softmax attention (Gated DeltaNet, GDN) to form a Hybrid Diffusion Transformer (DiT). Temporal consistency during long video generation is enforced by Cycle-Reverse Regularization, a training strategy that requires the model to reconstruct source frames from generated content via flow matching under a reverse prompt, supervised by Distribution Matching Distillation (DMD) loss within the LongLive streaming training framework. System efficiency is achieved via fused GDN kernels with partition strategy and Mixed-Precision Quantization (MPQ) targeting NVIDIA Blackwell (RTX 5090) Tensor Cores, with a precision search algorithm (inspired by AutoML) selecting per-layer NVFP4/FP8/BF16 settings.

Key results:

• Achieves real-time 1280×704 resolution video editing at 24 end-to-end FPS on a single RTX 5090 GPU
• DiT core runs at 58 FPS
• Hybrid design requires only 5.56 GB VRAM for long video generation
• Hybrid design is 3.7× faster than all-softmax-attention variant
• Fused GDN kernel achieves 1.5–2.2× speedup over baseline kernel on different GPU architectures
• MPQ achieves 59% speedup over BF16 on the DiT model with marginal quantization error
• Latency breakdown for 45-second videos: baseline 46.2 ms, +GDN kernel 20.0 ms, +Hybrid Attention 16.0 ms, +MPQ 12.6 ms (relative latency reductions with cumulative co-design totaling 3.7×)

Why it matters / caveats: SANA-Streaming demonstrates that system-algorithm co-design combining hybrid attention, cycle-consistent training, and hardware-aware quantization can enable real-time long-video editing on a single consumer GPU. Results are specific to the RTX 5090 (Blackwell architecture) and may not transfer directly to other hardware; the MPQ search is GPU-architecture-dependent.

Problem: Systematic evaluation of long-form speech synthesis is inadequate because existing benchmarks cover limited domains and existing metrics (e.g., WER) are saturated and fail to capture consistency and coherence across long outputs.

Method: SwanBench-Speech is a benchmark with 1,101 test samples spanning 17 speech scenarios organized along three axes: acoustics, semantics, and expressiveness. It defines seven disentangled automated metrics including novel dimensions for long-form speech: Acoustic Consistency (Timbre Consistency, Sound Fidelity, Reverb Consistency via sliding-window WavLM/SRMR/SQUIM), Prosodic Coherence, and Expressive Hierarchy (assessed via in-context learning with an MLLM evaluator). Human alignment studies validate that the automated metrics serve as reliable proxies for human perception. Evaluation covers long-form monologue and dialogue generation settings across multiple existing TTS systems.

Key results:

• SwanBench-Speech covers 1–4 speakers with an average of 228.6 words per sample, substantially longer than prior benchmarks (e.g., SeedTTS-Eval at 19.57 words, EmergentTTS-Eval at 33.93 words)
• Current models match human recordings on fidelity and accuracy metrics but show substantial gaps in reverb consistency, prosodic coherence, and expressive hierarchy
• Performance deteriorates in highly expressive scenarios (specific numeric gaps not stated in extracted text)

Why it matters / caveats: SwanBench-Speech fills a significant gap by providing coverage of 17 realistic long-form speech scenarios with metrics explicitly designed for temporal consistency and expressive dynamics. The benchmark is introduced by the same group that develops SwanVoice, which may create evaluation bias; broader community adoption across diverse model families will be needed to confirm its utility as a standard.

Problem: Multimodal agents receiving unbounded streams of observations lack a principled mechanism to selectively retain only task-relevant information in long-term memory.

Method: TaskMem is a reinforcement learning-based framework built on Qwen3-VL-30B-A3B that trains a memorization policy in two phases: Phase One optimizes memory quality (correctness, non-redundancy, format) via multi-objective RL; Phase Two tunes a lightweight 2,048-parameter adapter on top of the base MLLM using recent environment tasks as a reward signal to guide what content is memorized. Evaluation uses VideoMME, EgoLife, and EgoTempo reformulated as streaming benchmarks where questions must be answered from memory alone without access to raw video.

Key results:

• VQA accuracy improvement of +6.3% on VideoMME (streaming)
• VQA accuracy improvement of +7.0% on EgoLife (streaming)
• VQA accuracy improvement of +5.3% on EgoTempo (streaming)

Why it matters / caveats: TaskMem reframes memory construction as a learnable, goal-driven policy rather than a fixed summarization step, enabling continual adaptation to the agent's actual deployment environment. The lightweight adapter (2,048 parameters) addresses catastrophic forgetting and computational constraints during online updates.

Problem: Existing text-driven 3D indoor scene synthesis methods generate rooms from object-centric prompts rather than from functional specifications describing how a space will be used, making them unable to satisfy real interior design requirements.

Method: Function2Scene parses natural-language functional design briefs into occupant personas and activities, then derives constraints from a taxonomy of 17 criteria spanning spatial, ergonomic, activity, and environmental considerations. Layout generation is guided through an iterative check-and-repair loop that combines geometric measurements, LLM-based contextual reasoning, and VLM-based visual assessment to evaluate and refine scenes against the functional constraints. The system avoids direct LLM scene generation in favor of this tool-augmented iterative refinement pipeline.

Key results:

• Preferred in 94.3% of pairwise comparisons over recent LLM-based scene synthesis baselines
• Evaluated on 30 professionally written interior-design cases adapted from sources such as Architectural Digest

Why it matters / caveats: This work reframes indoor scene synthesis from placing plausible objects to designing spaces that support human activity, addressing a gap between current AI-generated layouts and real design practice. The evaluation set of 30 cases is relatively small and sourced from a single style of professional briefs.

Problem: Current LLM-based data curation methods rely on human-designed workflows, leaving unanswered whether LLMs can autonomously execute an end-to-end data engineering pipeline to specialize another model.

Method: The paper formalizes Autonomous Agentic Data Engineering as a closed-loop task where an LLM agent independently performs strategy planning, domain specification, prompt design, data synthesis, data validation, and iterative optimization—all guided by post-training performance feedback from a fixed student model. Two settings are studied: One-Shot (single-turn completion) and Iterative Agent (closed-loop self-optimization), evaluated across Science, Code, and Finance domains using supervised fine-tuning on the synthesized data with a deterministic rule-based evaluator.

Key results:

• GPT-5.2 achieves an average relative performance gain of 57.29% on the student model through iterative optimization
• Surpasses human-crafted data synthesis pipelines in average gain
• Significant failure modes identified in post-generation quality assurance

Why it matters / caveats: This study establishes autonomous data engineering as a measurable LLM capability and shows that agent-driven curricula can outperform manual pipelines, but also reveals that LLMs still lack robust mechanisms for reliable data quality assurance after generation.

Problem: Selective on-policy distillation methods that prioritize high-KL or high-entropy tokens assume disagreement implies learning value, but raw KL conflates tokens where teacher corrections are learnable with tokens where teacher mass falls outside the student's current predictive support.

Method: The paper introduces a fixed-context diagnostic that measures same-context teacher-student KL reduction to distinguish "learnable disagreement" (teacher mass within the student's top-K support) from "incompatible disagreement" (teacher mass off-support). This concept is formalized as token teachability (support-aligned disagreement). Motivated by this, Teachability-Aware OPD (TA-OPD) scores each token position by support-aligned teacher-student disagreement and applies OPD loss only to high-teachability positions, requiring no reward model or verifier. Experiments use Qwen2.5 and Qwen3 teacher-student pairs at varying scales.

Key results:

• TA-OPD with only 5% retained tokens often matches or surpasses full-token OPD
• Outperforms budget-matched entropy- and divergence-based selective OPD baselines across Qwen3 and Qwen2.5 settings

Why it matters / caveats: This work provides a principled explanation for why token selection in OPD matters beyond mere salience, and demonstrates that token quality can outweigh token count. The evaluation is limited to Qwen-family teacher-student pairs, and generalization to other model families is not directly demonstrated.

Problem: Existing open-weight code models either plateau on hard coding and reasoning (dense 4–14B models) or are too expensive to deploy at scale (large MoE models), leaving a gap for a practically deployable model with frontier coding quality.

Method: Mellum 2 is a 12B-parameter MoE model with 2.5B active parameters per token, built on the Qwen3-MoE recipe (64 experts, 8 active) with Grouped-Query Attention (4 KV heads), Sliding Window Attention on 3 of every 4 layers, and a single Multi-Token Prediction head serving both as an auxiliary pre-training objective and draft model for speculative decoding. Pre-training covers approximately 10.6 trillion tokens across a three-phase curriculum shifting code ratio from 23% to 42% to 59%, using the Muon optimizer with FP8 hybrid precision and a Warmup-Hold-Decay schedule. The context window is extended to 128K via layer-selective YaRN, and two post-trained variants (Instruct and Thinking) are produced via SFT followed by reinforcement learning with verifiable rewards (RLVR).

Key results:

• Competitive with open-weight baselines in the 4B–14B range on code generation, math, reasoning, tool use, knowledge, and safety benchmarks
• Matches or exceeds inference throughput of Qwen2.5-7B on a single H100
• Per-token compute equivalent to a 2.5B dense model

Why it matters / caveats: Mellum 2 demonstrates that a small MoE architecture with carefully ablated design choices can achieve dense-14B-quality results at dense-2.5B serving cost, and is released under Apache 2.0. Specific benchmark numbers beyond relative comparisons are not stated in the extracted text.

Problem: Diffusion Large Language Models (dLLMs) using Mixture-of-Experts suffer a fundamental inefficiency because token-level expert routing causes an explosion in the number of uniquely activated experts per forward pass, making inference memory-bound.

Method: dMoE introduces block-level MoE routing for dLLMs: token-level expert score distributions within each decoding block are aggregated into a unified block-level expert distribution, which then guides a coherent block-wide routing decision. The number of uniquely activated experts is dynamically controlled via a top-p criterion that adapts to varying routing characteristics across denoising steps and blocks. During training, a self-distillation paradigm is used with the same routing procedure. The method is applied to LLaDA2.0-mini as the base model and evaluated on MATH500, GSM8K, ARC-C, and MMLU.

Key results:

• Reduces unique activated experts from 69.5 to 14.6 on average (4.77× reduction)
• Retains 99.11% of original performance
• Reduces memory usage by 76.64% to 79.84%
• Achieves 1.14× to 1.66× end-to-end latency speedup
• MATH500: original 72.0% vs. dMoE 71.0%

Why it matters / caveats: dMoE provides a simple, tunable, and effective solution to the memory-bound inference bottleneck unique to MoE dLLMs, with near-zero performance degradation. Evaluation is limited to a single base model (LLaDA2.0-mini) and four benchmarks, so generalization to other dLLM architectures remains to be shown.

Problem: GUI agents lack robustness to recover from their own policy-induced errors during task execution, and existing benchmarks do not systematically measure error detection and long-horizon recovery from these mistakes.

Method: The paper introduces GUI-RobustEval, a benchmark with 1,216 executable test cases covering 11 error types across 4 controllable error depths, reporting Error-Awareness Rate and Post-Error Success Rate. To address the training data gap, Robustness-driven Trajectory Synthesis (RoTS) is proposed: a tree-based online synthesis framework that grows a trajectory tree in live GUI environments, branching from fragile states to discover new failure modes and replaying from error states to synthesize long-horizon failure-recovery rollouts. RoTS-7B and RoTS-32B are produced by fine-tuning QWEN2.5-VL on the resulting 800k-sample dataset.

Key results:

• RoTS-32B achieves 47.4% success rate on OSWorld
• RoTS-32B achieves 33.8% All-Pass@4 on OSWorld (state-of-the-art)
• Both RoTS-7B and RoTS-32B show significant gains on GUI-RobustEval and traditional GUI benchmarks

Why it matters / caveats: The work demonstrates that improving long-horizon error recovery ability benefits not just robustness but overall task success, suggesting error recovery training is broadly useful for GUI agents. All authors are employees of Alibaba Cloud, and the QWEN-VL models evaluated were developed internally, which is a potential conflict of interest.

Problem: LLM agents operating in local agentic harnesses (with file read/write and persistent workspace state) are vulnerable to multi-step trojan attacks where prompt injections are written into files and re-triggered in future sessions, a threat that single-step defenses fail to detect.

Method: The paper introduces ClawTrojan, a benchmark for multi-step trojan attacks in local agentic harnesses, covering attack types including memory poisoning, trust laundering, and skill poisoning. It tests attack chains across multiple turns, annotating both the planting step and the trigger step. To defend against this threat, DASGuard (Detect-Attribute-Sanitize) is proposed: it scans sensitive local files for control-like text, attributes each span to its content source, and removes control content that does not originate from a trusted source, combining runtime blocking with sanitized workspace commits. Experiments are conducted in an OpenClaw-style simulated workspace using GPT-5.4.

Key results:

• ClawTrojan achieves 95.5% attack success rate (ASR) against GPT-5.4 in the OpenClaw workspace
• Existing single-turn prompt-injection attacks (AgentDojo, InjecAgent) produce near-zero ASR on the same model
• Existing defenses (prompt separation, detection, action-gating) fail to block multi-step trojan attacks in this setting

Why it matters / caveats: This work exposes a qualitatively new attack surface for agentic systems beyond single-turn prompt injection, showing that persistent workspace state creates durable backdoors. The evaluation is conducted in a simulated OpenClaw-style workspace, and real-world generalization depends on how closely real harnesses match this environment.

Problem: Video captioning pipelines are limited by the small number of frames vision-language models can process, and existing adaptive frame selection methods are too computationally expensive for practical use.

Method: PEEK is a two-stage distillation framework. In Stage 1, a frozen SigLIP 2 dual encoder acts as an oracle teacher, scoring candidate frames against ground-truth captions to produce caption-conditioned relevance rankings used only for supervision. In Stage 2, a lightweight temporal Transformer is trained to predict these rankings from MobileCLIP2 visual embeddings alone, producing a query-free selector that requires no text encoder at inference. At inference, the video segment is split into k equal temporal windows and the highest-scoring frame from each is selected via stratified argmax.

Key results:

• Outperforms state-of-the-art adaptive methods across all evaluated downstream VLMs on ActivityNet Captions and MSR-VTT, especially at 1–2 frames
• Wins 14 out of 16 configurations on ActivityNet Captions
• Adds only 5.2% to captioning time, vs. 65.4% for CSTA and 211.9% for MaxInfo
• Zero-shot transfer to MSR-VTT is strongest at low frame budgets; results at 4 and 8 frames are more mixed

Why it matters / caveats: PEEK achieves better accuracy in the low-frame-budget regime at a fraction of the overhead of competing adaptive selectors, making it practical for real-time or resource-constrained captioning pipelines. The method's advantage diminishes at higher frame budgets where temporal coverage and visual diversity become more important than content-aware selection.

Problem: Existing data analysis benchmarks evaluate isolated or short interactive tasks, leaving agents' ability to maintain and update evolving analytical state across long multi-turn sessions untested.

Method: LongDS is a benchmark constructed from real-world Kaggle notebooks, comprising 68 tasks and 2,225 turns across six domains (including Geoscience, Business, and Education). Tasks are organized around five state-evolution patterns: initial construction, state inheritance, counterfactual perturbation, rollback, and multi-state composition, with an average dependency span of 11.3 turns. Five state-of-the-art LLM agents are evaluated on persistent multi-turn sessions where each request depends on analytical states established in prior turns.

Key results:

• Best model achieves only 48.45% average accuracy
• Performance drops nearly 47 percentage points from early to late turns
• Long-horizon errors account for 52%–69% of all failures
• Increasing agent steps does not consistently improve accuracy

Why it matters / caveats: LongDS reveals that analytical state maintenance—not interaction budget or isolated reasoning—is the primary bottleneck for current agents on realistic data science workflows. The benchmark's 68 tasks, while drawn from real Kaggle notebooks, represent a limited sample of domains and may not cover all types of analytical workflows encountered in practice.

Problem: Agentic search systems built on LLMs suffer from "over-search" — triggering unnecessary searches when internal knowledge suffices and failing to stop search after sufficient evidence has been collected.

Method: SAAS is a reinforcement learning framework with three components: (1) a search boundary modeling mechanism that contrasts search-disabled and search-enabled rollouts to track the model's evolving knowledge boundary; (2) a boundary-aware reward module that applies trajectory-level penalties to suppress unnecessary and redundant searches; and (3) a stage-wise optimization strategy using a sequential curriculum that prioritizes reasoning before applying search regularization to prevent reward hacking.

Key results:

• Evaluated across seven benchmarks
• SAAS substantially reduces over-search while maintaining accuracy (specific numbers not stated in the extracted text)

Why it matters / caveats: Existing RL-based methods rely on static heuristics or fixed thresholds that fail to adapt to the model's evolving capability, making them susceptible to reward hacking; SAAS addresses this by dynamically modeling the search boundary. The approach reduces inference latency and computational cost without sacrificing answer accuracy.

Problem: Existing self-play methods for training language models require rule-checkable or verifiable answers, excluding open-ended tasks such as deep research and complex writing from data-free self-improvement.

Method: SCOPE co-evolves two policies from the same base model: a Challenger that generates document-grounded tasks, and a Solver that answers them through multi-turn retrieval. A frozen copy of the initial model serves as a fixed Judge that writes task-specific rubrics from source documents and grades Solver responses. Training alternates between Challenger and Solver updates using GRPO-style RL, with a cosine length penalty and quality gates to prevent reward hacking.

Key results:

• Up to +10.4 points improvement on eight open-ended benchmarks (Qwen2.5-7B: 24.4 → 34.8)
• Matches or exceeds GRPOdata trained on ~9K curated prompts across Qwen2.5-7B, Qwen3-8B, and OLMo-3-7B
• Up to +13.8 points on seven held-out short-form QA benchmarks despite training only on open-ended tasks
• Up to +7.4 on creative writing vs. GRPOdata falling below the base model on two of three backbones

Why it matters / caveats: SCOPE is the first data-free self-play framework for open-ended tasks, eliminating dependence on frontier-model judges or curated prompts. The bottleneck identified is rubric generation quality (not grading capacity), and a frozen Challenger fails to sustain improvement beyond the first iteration, highlighting the necessity of co-evolution.

Problem: 3D visual understanding (depth estimation, pixel correspondence, camera pose, object-level 3D) still depends on expert vision models with complex task-specific architectures and losses, despite VLMs succeeding at semantic tasks.

Method: VLM3 is a scalable framework that adapts standard VLMs to diverse 3D tasks through three design choices: (1) focal length unification via image resizing to resolve camera ambiguity and enable mixed-data training; (2) text-based pixel/region reference with normalized horizontal and vertical axes instead of visual prompting; and (3) data mixture and scaling, treating all 3D outputs as text tokens without regression losses or extra encoders/decoders.

Key results:

• Depth estimation accuracy: VLM3-4B improves previous best VLM DepthLM-7B from 0.84 to 0.9, matching expert model UnidepthV2
• Pixel correspondence: reduces EPE of the base VLM by 10x, outperforming expert models DKM and RoMa
• Camera pose estimation: AUC30 improves from 5% to 94%, surpassing VGGT and matching DA3-Giant
• Object-level 3D: VLM3-4B outperforms SpatialRGPT-8B on SpatialRGPT-Bench while removing the need for extra encoders

Why it matters / caveats: VLM3 demonstrates that regression losses, task-specific architectures, and heavy data augmentations commonly believed necessary for 3D tasks are not required, opening a simpler path to unified 3D foundation models. The study is large-scale but focused on a 4B parameter model; generalization to other model sizes is not fully explored in the extracted text.

Problem: Video VLMs suffer from quadratic compute scaling with the number of frames due to spatiotemporal self-attention, blocking practical deployment for long-horizon and streaming scenarios.

Method: StateKV is an inference-time method that adapts frozen pretrained video VLMs to linear-time video prefill without fine-tuning or architectural changes. It uses two coupled KV caches per layer: a fixed-capacity importance-based temporal state for cross-frame context (carrying "temporal sink" tokens identified from structured long-range attention patterns) and a full per-frame cache used for decoding. Video is processed incrementally frame-by-frame, keeping all per-frame tokens for decoding while limiting cross-frame attention to the fixed-capacity state.

Key results:

• Evaluated on three long-video benchmarks across seven models spanning three model families and multiple scales
• StateKV consistently outperforms sliding-window/recency-based streaming approximations (e.g., ReKV)
• Remains close to full O(N²) spatiotemporal attention accuracy
• On VideoMME at 512 frames, StateKV surpasses ReKV and full self-attention operating points across practical compute budgets
• Specific accuracy numbers not stated in the extracted text

Why it matters / caveats: StateKV provides a principled, training-free path to linear-complexity long-video understanding by exploiting the observation that long-range temporal interactions in pretrained VLMs concentrate on a small set of slowly evolving temporal sink tokens. Generation stage (decoding) remains O(N) since all per-frame tokens are retained, which limits storage savings but is acceptable since prefill dominates cost.

Problem: Detecting execution failures in Vision-Language-Action (VLA) robot policies at runtime is difficult because obtaining step-level failure annotations is expensive and existing methods either propagate noisy trajectory-level labels uniformly or rely on computationally prohibitive action resampling and external VLM judges.

Method: Hide-and-Seek formulates VLA failure detection as a coarsely supervised learning problem using two contrastive losses: an inter-trajectory contrastive loss that enforces the most failure-indicative step in a failure trajectory to score higher than the most failure-resembling step in a success trajectory, and an intra-trajectory contrastive loss that encourages the average failure score to be higher after failure onset than before. The framework is evaluated on LIBERO, VLABench, and a real-world platform against three VLA policies: OpenVLA, π0, and π0.5.

Key results:

• Surpasses the strongest classifier-based baseline by up to +11.7% balanced accuracy (bACC)
• Outperforms a VLM-based runtime monitor by +13.1% in accuracy
• Operates at over 2,000× higher speed than the VLM-based monitor
• Generalizes to both seen and unseen tasks

Why it matters / caveats: Hide-and-Seek localizes temporally structured failure signals without any step-level annotation, making it scalable and practical for real-world deployment. The method is architecture-agnostic across autoregressive and flow-matching VLA paradigms; limitations specific to narrow task distributions or failure types are not discussed in the extracted text.

Problem: Existing video generative world models fail to maintain fine-grained spatio-temporal consistency under long-horizon inference, particularly in "revisit" scenarios where previously generated scenes must be recalled after extended rollouts.

Method: DecMem is a decoupled memory architecture consisting of two complementary modules: Sparse Global Memory (SGM), which performs block-level sparse retrieval over the full history for efficient fine-grained long-term memory access, and Anchored Local Memory (ALM), which anchors attention in recent frames to stabilize attention distribution and prevent attention dispersion. The architecture is built on top of pretrained video generation backbones from Kling Team (Kuaishou Technology) and is trained with a learnable block retrieval mechanism.

Key results:

• Significantly outperforms current state-of-the-art methods on visual quality and spatio-temporal consistency (specific metrics not stated in the extracted text)
• Generation latency of sparse block retrieval is substantially lower than naïve Dense Attention with a 221-frame initialized memory bank
• Enables minute-level controllable long video generation

Why it matters / caveats: DecMem resolves the fundamental dilemma between short-range fidelity and long-range consistency that affects existing training-free decay-based approaches, by using learnable fine-grained sparse retrieval. Quantitative benchmark numbers are not provided in the extracted text.

Problem: As the open-source skill ecosystem for LLM agents grows rapidly, there is no systematic way to evaluate skill quality, understand how different agent frameworks interact with skills, or guide users in selecting skills under cost-performance trade-offs.

Method: OpenSkillEval is an automatic evaluation framework that dynamically generates test cases from real-world artifacts across five application categories (presentation generation, front-end web design, poster generation, data visualization, report generation) rather than using static benchmarks. It collects and organizes 30 open-source community-contributed skills and evaluates state-of-the-art models and agent frameworks across more than 600 dynamically generated task instances, measuring both skill-augmented and unaided agent performance.

Key results:

• Agents explicitly read the provided skill in only ~48% of cases on average under realistic skill-access settings
• Many publicly popular skills do not consistently outperform base agents without skills
• Weaker base models paired with well-designed skills and suitable frameworks can match stronger base models
• Skill augmentation provides limited benefit when the underlying model is intrinsically weak on the task

Why it matters / caveats: OpenSkillEval reveals that skill availability does not guarantee effective skill usage, providing practical guidance for skill design and agent framework selection. The framework covers only five task categories, and generalization to other agentic domains remains an open question.

Problem: Existing video QA benchmarks focus on coarse-grained tasks and fail to evaluate fine-grained spatio-temporal reasoning skills such as part tracking, contact detection, temporal ordering, and temporal localization that are required for step-by-step procedural tasks.

Method: FLAT-PACK BENCH is a video QA benchmark built on furniture assembly videos, augmented with part segmentation masks, inter-part connection annotations, and manually curated multiple-choice questions. It evaluates LVLMs across four task types: Temporal Ordering, Temporal Localization, Tracking, and Mating (part contact detection), using visual prompts that highlight relevant parts as references for fine-grained questions. An agentic approach using SAM2 as a tool is also evaluated.

Key results:

• GPT-5 achieves ~38% accuracy, far below human performance of 94.18%
• All tested state-of-the-art LVLMs struggle significantly with fine-grained spatio-temporal reasoning
• Agentic approach using SAM2 also struggles on these challenging videos (specific numbers not stated in the extracted text)

Why it matters / caveats: The benchmark exposes that even the best current LVLMs lack the ability to track objects, detect physical contact, and reason about temporal ordering in cluttered procedural videos — skills prerequisite for AI assistants in practical domains. The domain is deliberately simplified (rigid parts) to isolate these limitations, so performance gaps on more complex domains (cooking, repair) would likely be even larger.

Problem: Connector-based video unified models that integrate multimodal understanding with video generation require fine-tuning large diffusion generators (e.g., Wan2.1-14B) within the unified training loop, making it computationally prohibitive to achieve both strong semantic alignment and high visual fidelity simultaneously.

Method: Lumos-Nexus adopts a two-stage design: during training, only a lightweight diffusion generator (homogeneous in latent space with a large pretrained generator) is aligned with the understanding block to learn reasoning-driven semantic control. During inference, Unified Progressive Frequency Bridging (UPFB) progressively hands off generation from the lightweight generator to the large pretrained generator in the shared latent space, performing coarse-to-fine refinement without retraining the large generator. The paper also introduces VR-Bench, a benchmark evaluating reasoning-driven video generation across eight dimensions spanning physical-world reasoning, commonsense reasoning, and embodied interactions.

Key results:

• Achieves substantial gains in visual realism and temporal coherence on VBench
• Maintains strong reasoning-based generative performance on VR-Bench
• Specific numerical scores not stated in the extracted text

Why it matters / caveats: Lumos-Nexus decouples semantic alignment training from high-fidelity synthesis, enabling training-efficient unified video generation without full-scale fine-tuning of large diffusion models. Quantitative comparisons against other unified video generation methods are not present in the extracted text.

Problem: Vision-language embedding models like CLIP fail at concept binding — correctly associating which color belongs to which shape in multi-object scenes — despite encoding object information within individual modalities.

Method: The paper analyzes CLIP's binding function (the mapping from concepts to scene embeddings) and finds that scene embeddings decompose additively into object representations. To test whether binding generalization is fundamental, the authors train controlled transformer-based dual-encoder models from scratch on synthetic multi-object data, varying data coverage and analyzing the structure of the learned binding functions through the lens of function complexity (measurable by MLP approximability) and multiplicative vs. additive concept interactions.

Key results:

• CLIP's binding function is high-complexity and cannot be captured by a simple MLP mapping, requiring effective memorization of object identities and preventing cross-modal generalization
• Controlled transformer models trained with sufficient data coverage learn low-complexity binding functions with multiplicative interactions between concepts
• These models achieve binding generalization to unseen concept combinations
• Specific accuracy numbers not stated in the extracted text

Why it matters / caveats: The results show that CLIP's cross-modal binding failure stems from the complexity of its learned binding function rather than an inherent incompatibility between concept recognition and object binding, and that generalizable binding is achievable through sufficient data coverage and multiplicative composition. The study relies on controlled synthetic data, so direct implications for large-scale CLIP training remain to be validated.

Problem: Reward models in RLHF become miscalibrated as the policy evolves beyond their static training distribution, leading to reward hacking and over-optimization.

Method: The paper proposes SAVE (Self-supervised reward model improvement via Value-Anchored On-policy feEdback), which augments the reward model with a prompt-specific value head that estimates expected reward under the current sampling policy. This value estimate serves as an adaptive anchor to compute response-level RM advantages; ambiguous samples below a curriculum-driven threshold are filtered out, and the remaining positive/negative advantage subsets update the RM via a contrastive objective. The framework is theoretically framed as a reward-model-centric minimax problem and evaluated across three RL algorithms: GRPO, RLOO, and GSPO.

Key results:

• Average RM accuracy across six benchmarks (RewardBench, RewardBench 2, RMBench, PPE Preference, PPE Correctness, JudgeBench) improves from 76.0 to 77.3
• Best scores on all six benchmarks achieved
• AlpacaEval 2 length-controlled win rate increases from 51.68% to 54.24%
• Arena-Hard-v2.0 win rate rises from 30.2% to 33.9%

Why it matters / caveats: SAVE enables continuous reward model improvement during RL training without requiring additional human annotations or external judge models, addressing a key scaling bottleneck in RLHF. Improvements are consistent across multiple RL algorithms and policy backbones, suggesting general applicability.

Problem: Multi-turn LLM optimization faces a dilemma where online RL effectively handles multi-turn dynamics but is computationally prohibitive, while offline supervised fine-tuning suffers from distribution shift and behavioral collapse.

Method: DRIFT (Decoupled Rollouts and Importance-Weighted Fine-Tuning) operationalizes the theoretical equivalence between the KL-regularized RL objective and importance-weighted supervised learning. It operates in two stages: (1) an offline rollout stage that collects interaction trajectories under a fixed reference policy, and (2) a trajectory-weighted SFT stage that optimizes the policy using return-based importance weights derived from exponentiated rewards. The formulation is grounded in a finite-horizon Markov Decision Process over multi-turn answer correction tasks.

Key results:

• DRIFT matches or exceeds the performance of multi-turn RL baselines (PPO, GRPO) on mathematical and general reasoning benchmarks
• Rollout cost scales more favorably with interaction length compared to standard online RL
• Specific accuracy numbers not stated in the extracted text

Why it matters / caveats: DRIFT achieves the asymptotic performance of online RL while retaining the computational efficiency and simplicity of standard SFT, making multi-turn optimization practically feasible. The decoupled rollout stage is parallelizable, providing additional throughput advantages.

Problem: Populations of autonomous language model agents on platforms like Moltbook are inventing new languages that may evade human oversight, threatening the viability of chain-of-thought monitoring as a safety technique.

Method: The study analyzes the Moltbook Files dataset using a two-stage pipeline: a rule-based heuristic producing approximately 6,000 matches, followed by zero-shot classification retaining 518 posts with language proposals. The 518 posts are categorized into token-efficiency languages (166), new natural languages (106), programming languages, and oversight-evasion languages (59). DeepSeek-3.2 is used as an LLM judge to rate alignment, and pairwise learner-judge experiments with GPT-OSS-120b and Qwen3.5-122B-A10B test whether other LLMs can acquire the emergent languages in-context.

Key results:

• 518 language proposal posts identified from Moltbook Files corpus; 59 classified as oversight-evasion languages
• Oversight-evasion language proposals rated as less aligned by DeepSeek-3.2 than other categories
• All newly proposed languages can be learned in-context by other LLMs from a brief description alone
• Oversight-evasion languages are slightly harder to learn in-context than other categories

Why it matters / caveats: The findings suggest that monitoring surface behavior may soon be insufficient for controlling agent populations, as emergent languages enabling covert communication can be acquired rapidly by other models. The extent to which the Moltbook agents operate autonomously versus being human-directed remains uncertain, which limits causal claims about spontaneous language emergence.

Problem: Generating long interactive video trajectories with world models like HY-WorldPlay is prohibitively slow, with 10 seconds of video taking over 200 seconds on a single A100 GPU due to growing context memory, quadratic 3D spatio-temporal attention, and repeated denoising steps.

Method: Light Interaction is a training-free inference acceleration framework for autoregressive interactive video world models. It introduces three components: (1) adaptive context management that prunes spatial memory using camera-pose-aware retrieval similarity and adjusts temporal windows based on local latent dynamics; (2) denoising cache acceleration that reuses early-step model outputs for intermediate denoising steps when revisiting familiar regions; and (3) hardware-software co-designed 3D block sparse attention using Triton fused kernels to eliminate layout-conversion and gather/scatter overhead under autoregressive causal constraints. The framework is evaluated on HY-WorldPlay and Matrix-Game-3.0.

Key results:

• Up to 2.59x speedup on HY-WorldPlay 480P image-to-video generation (latency reduced from 228.60 s to 88.24 s)
• 1.61x speedup on Matrix-Game-3.0 720P (latency reduced from 59.70 s to 37.07 s)
• PSNR of 24.81 on HY-WorldPlay against the original model output
• PSNR of 17.76 on Matrix-Game-3.0

Why it matters / caveats: Light Interaction achieves meaningful speedups without any model retraining, making it directly applicable to existing deployed world models. The approach exploits the structure of interactive trajectories (revisiting vs. exploration) rather than applying uniform compression, enabling adaptive rather than fixed computation reduction.

Problem: Mandating explicit bounding-box predictions in visually grounded chain-of-thought during inference often degrades MLLM performance compared to standard textual CoT, because inaccurate localization introduces visual noise that interferes with answer prediction.

Method: iVGR (Internalizing Visually Grounded Reasoning) is a reinforcement learning framework using a dual-stream training strategy. For each training query, the policy MLLM generates two rollout streams: a grounded stream incorporating explicit bounding box predictions, and a textual stream performing standard reasoning. A consistency reward aligns the textual stream with the high-quality grounded stream, transferring localization capability into textual reasoning without requiring explicit grounding at inference time. The method is trained and evaluated using benchmarks including V*, HR4K, HR8K, MME-RW-Lite, POPE, RealWorldQA, CV-Bench-2D, and CV-Bench-3D.

Key results:

• iVGR significantly outperforms existing baselines (DeepEyes-7B, TreeVGR-7B) on fine-grained benchmarks
• DeepEyes-7B average across 8 benchmarks: grounded CoT 74.7 vs. textual CoT 75.1 (textual wins despite no explicit grounding training)
• TreeVGR-7B: grounded CoT 74.7 vs. textual CoT 75.7 (explicit grounding consistently underperforms textual CoT)
• Specific iVGR accuracy numbers not stated in the extracted text

Why it matters / caveats: The work challenges the assumption that explicit visual grounding at inference time always helps MLLMs, showing that localization capability can be internalized into textual reasoning. The approach maintains flexibility to support tool-assisted inference workflows when needed.

Problem: Object counting models are fragmented across domain-specific datasets and task formulations, preventing generalization across categories, visual domains, object scales, and density distributions.

Method: The paper constructs CLOC (Cross-domain Large-scale Object Counting), a dataset of approximately 220K images, 619 categories, and 15M object instances spanning six domains: General Scene, Remote Sensing, Histopathology, Cellular Microscopy, Agriculture, and Microbiology. On CLOC, the authors train Count Anything, a generalist model using discrete instance points rather than density maps. It employs dual-granularity instance enumeration: a Region-level Sparse Counter for large and sparse targets, and a Pixel-level Dense Counter for small and crowded targets. A point-centric supervision strategy enables learning from heterogeneous annotations, and Complementary Count Fusion combines both counters without additional parameters.

Key results:

• Count Anything substantially outperforms existing open-world counting methods across multiple domains
• Specific numeric accuracy figures (MAE, NAE, etc.) not stated in the extracted text
• CLOC is described as the largest text-guided counting dataset in terms of image scale and instance annotations

Why it matters / caveats: Count Anything addresses a real gap by unifying fragmented counting research into a single generalist framework, analogous to what foundation models did for other vision tasks. The reliance on heterogeneous public data sources, which required unified preprocessing to resolve annotation inconsistencies, may introduce residual label noise.

Problem: Vision-language models fail to recognize when visual observations are insufficient or misleading for spatial reasoning, instead producing overconfident answers even under occlusion or perspective ambiguity.

Method: The paper introduces SPATIALUNCERTAIN, a controlled evaluation framework built on 3D simulated environments. It introduces two types of observation challenges: occlusion (inserting objects to hide target information) and perspective ambiguity (shifting the camera to create misleading geometric cues). Two evaluation tasks are defined: ViewSel (viewpoint selection ability) and AbstainViewSel (joint abstention recognition and viewpoint selection). Eight VLMs are evaluated spanning open-source (Qwen2.5-VL-7B, Qwen2.5-VL-32B, InternVL3-8B) and closed-source (GPT-4o, GPT-5-mini, GPT-5.4, Gemini-2.5-Flash, Gemini-3.0-Flash) families, with comparisons of structured prompting and fine-tuning mitigation strategies.

Key results:

• Average accuracy under occlusion: approximately 30%
• Average accuracy under perspective ambiguity: below 10%
• Some models perform near random chance on viewpoint selection under ambiguity
• Structured prompting partially improves abstention but degrades answerable accuracy
• Fine-tuning on diverse ambiguity conditions yields more robust uncertainty handling

Why it matters / caveats: The findings expose a systematic overconfidence failure mode in frontier VLMs for spatial reasoning under realistic observation limitations, which is directly relevant to embodied AI safety. The framework uses simulated 3D environments, so the difficulty of generating real-world equivalents with controlled occlusion and perspective conditions represents a limitation in ecological validity.

Problem: Existing machine unlearning benchmarks contain fewer than 0.06%–1.3% Why-type (causal/relational) questions, making it impossible to detect methods that fail specifically on causal knowledge unlearning.

Method: The paper introduces 5WBENCH, a balanced 5,000-sample benchmark with 1,000 examples per 5W category (Who, What, When, Where, Why), and MAAT (Multi-phase Adapter-Aware Targeted Unlearning), a three-phase framework operating on LoRA adapter weights without merging them into the base model. Phase 1 applies gradient-projected ascent that orthogonalizes forget updates against the retain gradient only when they conflict. Phase 2 combines SVD-based pruning of MLP adapter dimensions targeting rank components activated by forget-set inputs, with task vector negation on the top-scored forget rank dimensions. Phase 3 applies hybrid KL–hidden-state retain repair with an entropy term to prevent re-learning of forgotten content. Evaluation is performed on Qwen 2.5-7B using LLM-as-a-Judge, with additional tests on Llama 3.2-3B and Gemma 3-4B.

Key results:

• MAAT is described as the first method to simultaneously achieve high forgetting and high retention on Why-type causal knowledge
• Why-type answers average 40.1 tokens vs. 4.2–10.5 tokens for other categories
• 44% of Why entries involve multi-hop reasoning chains vs. ≤2% for other categories
• MAAT outperforms all baselines on the aggregate forget–retain tradeoff across Llama 3.2-3B and Gemma 3-4B
• Specific numeric forget/retain scores not stated in the extracted text

Why it matters / caveats: The benchmark reveals a structural blind spot in how machine unlearning has been evaluated, and MAAT advances the state of the art specifically for causally grounded knowledge removal. The evaluation relies on LLM-as-a-Judge rather than deterministic metrics, which may introduce its own biases.

Problem: Autonomous AI research agents lack a pre-execution filter capable of reliably rejecting methodologically flawed research proposals before incurring the computational cost of implementation.

Method: SoundnessBench is constructed from ICLR submissions: 35,209 initial submissions and 137,940 expert reviews are processed to extract 1,099 high-signal hypothesis–experiment pairings labeled with reviewer soundness sub-scores. The benchmark covers 16 ML subfields and uses reviewer-agreement filtering, near-verbatim proposal extraction, and retrieval-backed atomic-claim auditing to maintain traceability. Twelve frontier LLMs are evaluated under standard and aggressive prompting strategies, with additional controls for public-corpus contamination, paper-identifying phrases, and surface features.

Key results:

• Under standard prompting, mean false-positive rate (classifying low-soundness proposals as sound) is 74.0% across 12 frontier LLMs
• Under aggressive prompting, false-positive rate drops to 19.9%, but high-soundness recall collapses to 36.1%
• Benchmark contains 1,099 proposals across 16 ML subfields

Why it matters / caveats: The pervasive optimism bias and prompt-sensitivity of frontier LLMs indicate they are not yet reliable as standalone first-gate scientific evaluators, which has direct implications for the reliability of autonomous AI research pipelines. The benchmark is scoped to ML/CS proposal-stage soundness and does not generalize claims to other scientific domains or to full peer-review quality assessment.

Problem: Reasoning-augmented VLA policies that use textual chain-of-thought incur multi-second latency from autoregressive text decoding and suffer from weak visual grounding, making them impractical for real-time closed-loop robot control.

Method: VisualThink-VLA is a sparse visual intermediate-reasoning framework that uses a compact six-channel evidence layer (bounding boxes, edges, motion, spatial relations, depth, etc.) and a task-adaptive orchestration layer to route only task-relevant evidence channels at each decision step. Selected evidence is injected as learned soft states before action decoding in a frozen VLA backbone. A dense-to-sparse optimization uses a FULLSOFT teacher (consulting all four screened effective evidence channels) to train a sparse student via soft-hard collaborative masks and distillation. The VisualEvidence-Kit provides a VisualEvidence-Agent that constructs a 754.7K-instruction VisualEvidence-Set (from BridgeData V2, Fractal, RoboTurk, LIBERO, and UT Austin MUTEX) for route supervision and counterfactual faithfulness testing.

Key results:

• Highest success rate of 92.63% across evaluated benchmarks; best on 7 of 8 datasets
• Step latency reduced from 8.377 s (ECoT) to 0.367 s, a 22.8x speedup on BridgeData V2
• Multi-second latency of reasoning-augmented baselines reduced to sub-second regime

Why it matters / caveats: VisualThink-VLA demonstrates that visual-space intermediate reasoning can outperform both text-chain-of-thought and dense visual side information while achieving sub-second inference, resolving the accuracy-efficiency tradeoff in embodied control. The approach requires constructing a channel-grounded supervision dataset (VisualEvidence-Set) through a VisualEvidence-Agent, adding non-trivial data pipeline complexity.

Problem: This paper investigates whether the Markov boundary — the theoretically minimal sufficient feature set — is practically useful for tabular prediction given that causal discovery methods needed to recover it are computationally intractable at scale.

Method: The study introduces SCM3K, a 3,450-task synthetic benchmark based on structural causal models (SCMs) using Erdős–Rényi DAGs with six SCM families and feature counts from 40 to 1,000. Six regressors are evaluated — including LASSO, XGBoost, a neural network, TabPFN, and TabICL — comparing training on the oracle Markov boundary versus the full feature set, and also testing off-the-shelf Markov-boundary and causal discovery estimators as a pipeline.

Key results:

• Using the oracle Markov boundary improves prediction for most regressors, with gains growing as feature space becomes larger and sparser.
• Recovered (estimated) Markov boundary masks rarely beat training on the full feature set.
• Existing causal discovery estimators exhaust the compute budget before reaching the high-dimensional regime where the boundary advantage is largest.
• Specific accuracy numbers are not stated in the extracted text.

Why it matters / caveats: The paper reveals a fundamental gap between the theoretical promise of Markov boundaries and their practical utility: discovery algorithms optimize structural recovery, not predictive performance, and false negatives carry higher predictive cost than false positives. The finding motivates developing prediction-aligned feature selection methods rather than relying on causal discovery as a preprocessing step.

Problem: Existing learned image codecs cannot simultaneously achieve on-the-fly variable-rate adaptation, low encoding cost competitive with standard codecs, and real-time encoding on commodity hardware without GPUs.

Method: FRAPPE is a residual autoencoding framework that uses the full input to predict the residual output via a Projection Pursuit Encoder, formulating the encoding objective so that latent channels are sorted by importance. Each channel is an independent strided convolution projection, forming a directed acyclic graph (DAG) at encode time, eliminating the sequential dependency of RNN-based codecs (e.g., Toderici et al.) and RVQ-style codecs. A concrete image codec, FRAPPE-Image, is instantiated and trained under this framework, inheriting the asymmetric design philosophy of WaLLoC and LiVeAction.

Key results:

• At ~0.1 bpp (high compression ratio), FRAPPE-Image provides higher perceptual quality than AVIF with 47x faster encoding.
• Achieves real-time 1080p, 30 fps CPU-only encoding.

Why it matters / caveats: FRAPPE enables cloud-assisted robotics and wearables by making high-compression encoding feasible on resource-constrained sensors without GPUs. The evaluation focuses on a single image codec instantiation; audio and video applications are described as future directions.

Problem: Fully non-autoregressive decoding in diffusion language models is degraded when high model confidence is mistakenly assigned to end-of-text tokens or anchor-adjacent tokens too early, causing incomplete or inaccurate generation.

Method: The paper proposes Suffix-Anchored Confidence Modulation, a training-free plug-and-play method combining two components: (1) suffix anchoring, which inserts a short suffix (e.g., "The answer is" for reasoning, or return for code) near the end of the response region to discourage premature EOT generation, and (2) anchor-proximity confidence modulation, which down-weights confidence scores near the anchor early in decoding and gradually restores them as decoding progresses. The method is evaluated on LLaDA and Dream diffusion language models.

Key results:

• Consistent improvements on GSM8K, MATH-500, StrategyQA, and MMLU-Pro for text-only reasoning.
• Gains on MathVista and ChartQA for vision-language reasoning using LaViDa.
• Outperforms explicit EOT suppression without directly prohibiting EOT tokens.
• Larger gains than semi-autoregressive decoding under limited step budgets.
• Specific percentage numbers are not stated in the extracted text.

Why it matters / caveats: The method requires no training or architectural modification, making it broadly applicable to confidence-based fully non-AR decoding. It preserves the parallel decoding speed advantage of diffusion LMs while fixing a systematic failure mode of naive suffix anchoring.

Problem: It is unclear how modern composed image retrieval methods transfer to Earth observation (EO) imagery or whether they remain effective when retrieval relevance depends on scene identity and post-event state rather than attribute-level class matching.

Method: The study systematically benchmarks representative composed image retrieval methods — Pic2Word, SEARLE, FreeDom, MagicLens, and BASIC — across six vision-language backbones including CLIP, SigLIP, RemoteCLIP, and SkyCLIP on the PatternCom dataset under a standardized evaluation protocol. Additionally, the paper introduces xView2-CIR, a new change-centric dataset derived from xView2 for disaster and damage monitoring, where queries consist of a pre-event image and a textual modifier (e.g., "post-fire") targeting retrieval of the same scene post-event.

Key results:

• Training-free composition methods provide strong and scalable baselines for EO retrieval.
• Change-centric retrieval (xView2-CIR) presents harder challenges than attribute-based retrieval (PatternCom), particularly for preserving scene identity.
• Specific Recall@K or mAP numbers are not stated in the extracted text.

Why it matters / caveats: The work establishes the first unified benchmark for remote sensing composed image retrieval, with practical relevance to disaster response, infrastructure monitoring, and urban change analysis. A caveat is that current methods struggle with change-centric retrieval where scene identity must be preserved alongside semantic modification.

Problem: Current human motion generation methods are constrained to fixed modality configurations and task-specific architectures, limiting cross-modal generalization, and existing datasets lack the scale and multimodal alignment needed for arbitrary-modality-conditioned synthesis.

Method: The paper introduces OmniHuMo, a dataset of 3.2 million sequences (5,000+ hours) with text, speech, music, and trajectory annotations extracted from web-scale videos via monocular motion capture and automatic labeling. Built on OmniHuMo, AnyMo combines a Residual Finite Scalar Quantization (R-FSQ) motion tokenizer — using a base stream for coarse structure and residual streams for detail — with a LLaMA-based masked modeling transformer that uses bidirectional attention and modality-specific encoders. Parallel Mask Modeling predicts all residual streams simultaneously to improve efficiency.

Key results:

• OmniHuMo is 1.6x larger in sequences and ~2.5x longer in duration than the next largest dataset (MotionMillion at 2M sequences, 2000H).
• AnyMo achieves competitive performance across diverse motion generation tasks.
• Specific FID or other quantitative benchmarks are not stated in the extracted text.

Why it matters / caveats: OmniHuMo is claimed to be the largest human motion dataset integrating text, audio, and visual modalities, addressing a key bottleneck for multimodal motion research. The "any-modality" claim depends on the modalities covered (text, speech, music, trajectory), and quantitative comparisons to single-modality baselines are not detailed in the extracted text.

Problem: Per-instance prompting of foundation models like SAM for cell instance segmentation in histopathology images is prohibitively expensive (requiring hundreds of clicks per image), while cell-specific trained models fail on out-of-distribution cell types.

Method: The paper proposes Chain-of-Prompts (CoP), a training-free framework operating on the frozen features of SAM (specifically SAM3). CoP uses two components: (1) Hierarchical Similarity Gating (HSG), which combines SAM's high-resolution and low-resolution feature maps non-parametrically to identify reliable same-type cell locations given a single click, and (2) Farthest Prompt Recursion (FPR), which selects the spatially farthest reliable point as the next prompt to maximize spatial coverage. These steps iterate recursively to propagate a single per-type click across all instances.

Key results:

• CoP retains over 90% of per-instance SAM3 performance on three cell-type-annotated benchmarks using one click per cell type.
• On four morphologically homogeneous benchmarks, a single click retains over 99% of per-instance performance.
• HSG achieves precision above 96% without any learnable parameters.
• CoP achieves 92.7% of the per-instance upper bound (SAM3 with 245 clicks) using only 3 clicks — an 81.7x reduction in annotation cost.
• Outperforms fully-supervised methods on seven benchmarks.

Why it matters / caveats: CoP makes interactive segmentation practical for large-scale histopathology analysis by reducing annotation cost from O(N) to O(T) (number of cell types). The method relies on SAM's frozen encoder exhibiting cell-type clustering, which may not hold for all tissue types or image modalities.

Problem: Existing RL methods for diffusion large language models (dLLMs) that use ELBO as a likelihood surrogate introduce training-inference mismatch bias, degrading training stability and final performance.

Method: GDSD (Guided Denoiser Self-Distillation) reframes RL for masked diffusion models as direct denoiser self-distillation. Under reverse-KL regularized RL, the closed-form optimal policy induces an advantage-guided denoiser as a self-teacher. GDSD then matches the dLLM's denoiser logits to this teacher via a squared-logit distillation loss with a normalization-free reformulation (logit centralization to eliminate the partition function). This is off-policy by design and requires no importance sampling or ELBO surrogate. Experiments are conducted on LLaDA-8B and Dream-7B across planning, math, and coding benchmarks.

Key results:

• On planning tasks with Dream-7B, GDSD achieves test-accuracy gains of up to +19.6% over prior ELBO-based RL methods.
• With LLaDA-8B, GDSD improves performance across planning, math, and coding benchmarks with gains ranging from +0.6% to +5%.
• GDSD shows more stable training reward dynamics compared to prior methods (SPG, UniGRPO, ESPO, wd1, DMPO).

Why it matters / caveats: GDSD provides a principled, bias-free RL training procedure for dLLMs by eliminating the ELBO surrogate, and unifies prior ELBO-based methods as special cases of different distillation divergences. The gains are asymmetric — large for planning tasks, more modest for math and coding — suggesting room for further improvement.

Problem: Diffusion-based visuomotor policies trained via behavior cloning overfit to high-frequency noise in human demonstrations (jitter, pauses, micro-adjustments), causing erratic robot actions during deployment.

Method: The paper proposes the Frequency Guidance Operator (FGO), a diffusion guidance mechanism based on Discrete Cosine Transform (DCT). During forward diffusion training, FGO applies a bank of DCT low-pass filters at various cut-off frequencies to train the model on multi-band sub-frequency manifold mappings. During reverse denoising at inference, instead of targeting the full-frequency data manifold directly, FGO progressively routes the action trajectory through a hierarchy of sub-frequency manifolds with expanding spectral bands, thereby suppressing high-frequency noise implicitly. FGO is evaluated on 15 robotic manipulation tasks across 5 benchmarks in simulated and real-world environments.

Key results:

• FGO achieves superior performance in action smoothness and temporal consistency across 15 tasks on 5 benchmarks.
• Specific success rate percentages or smoothness metric numbers are not stated in the extracted text.

Why it matters / caveats: FGO is compatible with existing diffusion policy frameworks as a guidance mechanism and addresses a fundamental noise issue in behavior cloning without modifying the base architecture. Quantitative results are not detailed in the extracted portion of the paper.

Problem: Existing multivariate time-series forecasting benchmarks evaluate models as indivisible black boxes, obscuring which internal components (normalization, patching, attention variants, etc.) actually drive performance gains.

Method: The paper introduces TSCOMP, a benchmark that deconstructs deep multivariate time-series forecasting (MTSF) models into a hierarchical component pool covering series preprocessing, encoding strategies, network architectures (including specific deep models and large time-series foundation models), and optimization methods. Evaluation uses constrained orthogonal experimental design to isolate individual component contributions. A performance corpus of over 20,000 model-dataset evaluations is constructed, which trains a meta-predictor for zero-shot automated component selection on new datasets.

Key results:

• The corpus-driven automated component selection approach consistently outperforms state-of-the-art MTSF methods.
• Over 20,000 model-dataset evaluations are included in the performance corpus.
• Specific MSE or MAE improvements are not stated in the extracted text.

Why it matters / caveats: TSCOMP is the first large-scale component-level benchmark for deep MTSF, enabling automated model construction that surpasses manually designed architectures. The result that systematic component selection beats holistic SOTA models challenges the dominant paradigm of end-to-end architecture design in time-series forecasting.