AI papers — 2026-07-14

Problem: RLVR (reinforcement learning with verifiable rewards) becomes prohibitively expensive to re-run on every new strong reasoning model because rollout generation cost scales with model size, so the paper asks whether RL learned cheaply on a small "weak" model can be reused to improve a larger "strong" target model without running RL on the target.

Method: The authors propose Direct On-Policy Distillation (Direct-OPD): instead of distilling the post-RL weak teacher's final policy (vanilla on-policy distillation, OPD), they compute the teacher's RL-induced policy shift ∆T(y|x) = log πT(y|x) − log πTref(y|x) between the post-RL teacher πT and its pre-RL reference πTref. Under the KL-regularized RL objective this log-ratio is shown to be mathematically equivalent to the reward (β⁻¹rT) that trained the weak model, so it is used as a dense, per-token implicit reward applied on the student's own on-policy rollouts (top-k action-restricted, Rao-Blackwellized policy gradient with a stop-gradient weighting term), regularized by a KL penalty back to the student's own initialization πS. An adaptive KL-coefficient controller (updates α via a sign-based rule with ϵ=0.01, α∈[0.5,2.5]) compensates for the unknown, uncalibrated scale of the transferred reward across different teacher-student pairs.

Key results:

• Transferring the R1-Distill-1.5B → JustRL-1.5B policy shift raises Qwen3-1.7B from 48.3% to 58.3% (+10.0) on AIME 2024 and 36.8% to 43.2% (+6.4) on AIME 2025, in about 4 hours on 8 A100 GPUs.
• Same transfer improves Qwen3-4B (72.5%→77.6%, +5.1 on AIME24) and R1-Distill-7B (56.7%→63.1%, +6.4 on AIME24; 40.5%→48.8%, +8.3 on AIME25), even though both students already start above the post-RL teacher (JustRL-1.5B: 51.3% AIME24).
• A second, independent teacher pair (Nemotron-1.5B → QuestA-Nemotron-1.5B) also transfers: Qwen3-1.7B +10.7 AIME24/+6.3 AIME25; R1-Distill-7B +4.9 AIME24/+4.5 AIME25.
• Vanilla OPD (imitating the teacher's final policy) instead degrades R1-Distill-7B from ~56.7% to ~50%, illustrating the weak-to-strong failure mode Direct-OPD is designed to avoid.
• At matched RL steps, RL on R1-Distill-1.5B (160 hours/32 A100s for 1500 steps) plus a ~4-hour Direct-OPD transfer outperforms running RL directly on R1-Distill-7B (320 hours/32 A100s), both in accuracy and compute.
• Sequential composition: applying the JustRL shift then the QuestA shift to Qwen3-1.7B accumulates gains — 48.3%→58.3%→63.8% on AIME24 (+15.5 total) and 36.8%→43.2%→46.8% on AIME25 (+10.0 total).
• Analysis shows transfer succeeds even under low teacher-student top-k token overlap (unlike standard OPD) and that a short 2k-token training horizon generalizes to much longer evaluation rollouts, without actor entropy collapse.

Why it matters / caveats: The results suggest RL outcomes can be reused across model scales as implicit reward signals rather than only as models to imitate, potentially making post-training cheaper as models scale; the paper notes the transferred signal's usefulness depends on which small-teacher RL checkpoint is used (not monotonic with more RL steps) and that longer training horizons (6k tokens) can shift the actor further on a diagnostic metric yet validate worse than the 2k setting (45.6 vs 48.8), indicating some sensitivity to hyperparameter/horizon choices.

Problem: Existing generalist visual-language navigation (VLN) foundation models rely on monolithic policies that map observations directly to actions, causing coordinate drift, poor handling of long-tail semantics, and black-box decisions that lack interpretability across diverse embodied navigation tasks (point-goal, object-goal, POI-goal, instruction-following, person-following).

Method: ABot-N1 uses a slow-fast dual-system architecture that decouples cognition from control: a slow 4B-parameter vision-language reasoner (Qwen-3.5-4B) performs explicit Chain-of-Thought reasoning over tri-view (left/front/right) images and historical frames to produce an "affordance pixel" (next safe waypoint) and/or "target pixel" (goal location), while a fast 2B-parameter action expert (Qwen-3.5-2B backbone) fuses this CoT+pixel guidance with real-time observations via a QFormer-style cross-attention module and an MLP head to emit continuous SE(2) waypoints at native control frequency. Training is a two-stage recipe: supervised pretraining on 30M samples (13.3M slow-system, 16.4M fast-system) across five tasks using cross-entropy loss (CoT/pixel tokens) and smooth-L1 loss (waypoints), followed by GRPO-based reinforcement post-training on the slow system with a composite reward (format validity + target alignment + safety clearance) using a theoretically derived 5:3:2 Safe:Critical:Danger zone sampling ratio for training-set balance.

Key results:

• POI-goal (ABotN-POIBench, new benchmark): SR@2m raised from 42.3% (POINav baseline) to 77.3% (+35.0 points); SPL from 40.3% to 72.6%.
• Point-goal outdoor (ABotN-PointBench, new benchmark): 92.9% SR (relaxed collision criterion), +16.0 points over ABot-N0, with largest gains (+22.7 pts) on the hardest difficulty tier (88.0% vs 65.3%).
• Point-goal indoor (ABotN-PointBench): 95.4% SR (strict zero-collision criterion), +5.8 points over ABot-N0.
• Instruction-following (VLN-CE R2R-CE Val-Unseen): NE 3.32m, SR 70.9%, SPL 67.5%, best among all reported methods including depth/odometry-based baselines.
• Instruction-following (RxR-CE Val-Unseen): best NE (3.13m); SR 73.9%/SPL 63.9%, second only to Qwen-RobotNav-4B (trained on larger in-domain RxR data).
• Object-goal (short-horizon OVON): SR +11.7 and SPL +16.4 points over ABot-N0; DTG compressed from 1.44m to 0.82m (1.8x reduction).
• Person-following (EVT-Bench): STT 90.1% SR/89.8% TR (+3.2/+2.2 over ABot-N0); AT split 70.0% SR/87.8% TR (+2.7 SR/+8.3 TR); DT split TR +9.0 points.
• Trained on 30M samples from 800+ indoor and 22 outdoor 3D Gaussian Splatting simulation scenes; releases two new benchmarks (ABotN-PointBench: 31 real-world scenes; ABotN-POIBench: 11 commercial regions, 163 POIs).
• Deployed at 10Hz on a 12-DOF quadrupedal robot (AMap TuTu) with Jetson AGX Orin 64GB, running fully on-device.

Why it matters / caveats: The pixel-goal interface unifies five heterogeneous navigation tasks under one architecture with a single checkpoint achieving positive cross-task transfer (ABot-N1 matches or exceeds single-task specialist variants on most tasks) rather than the negative transfer typically seen in generalist navigators, while also providing human-interpretable CoT + visual anchor decision traces for safety auditing. Caveats noted in the paper: person-following shows a trade-off of higher collision rate on harder splits (DT/AT) from a more aggressive approach policy; POI-goal training only supervises the slow system since no simulator covers commercial-street settings, relying on the fast system to inherit execution capability from point-goal data.

Problem: Long-horizon embodied agents lack a general runtime layer that bridges high-level VLM/VLA reasoning and low-level physical execution while providing persistent, multi-modal, cross-embodiment memory, leaving gaps in reasoning-execution coupling, embodiment generalization, and durable memory.

Method: ABot-AgentOS is an Agent OS built on an edge-cloud dual-LLM core (a lightweight edge Tiny LLM handling every turn, escalating to a cloud Large LLM on demand) with an Agent Harness composed of a main LLM (scene-conditioned planner), a Skill Runner (isolated-context subagent for procedural execution), and a Verifier (runtime, skill-level, and finish-time verification) forming a verification-aware ReAct loop. Its memory system, Universal Multi-modal Graph Memory, encodes dialogue, vision, spatial/temporal context, and task traces as a typed graph G=(V,E) with provenance-tagged nodes, retrieved via a fixed hybrid graph retriever (semantic/lexical/metadata seed selection plus typed-edge expansion), and improved through a failure-driven lifelong self-evolution loop that diagnoses retrieval/writing/answering failures offline and promotes gated "evo-assets" only to later evaluation splits (preventing current-split leakage). A separate deployment-oriented pipeline distills long-horizon tool-use from teacher agents into a smaller model via text-based environments, SFT, RL with LLM-as-a-Judge rewards (matching the GiGPO reward interface), and a self-evolving reward engine (Meta-Judge validation + multi-agent Cluster/Analyzer/Refiner/Validator prompt refinement).

Key results:

• EmbodiedWorldBench (introduced benchmark; 16 scenes, 200+ tasks): with Qwen3.6-Plus, ABot-AgentOS improves TSR by 11.99 points (49.97%→61.96%) and GCR by 10.84 points (57.95%→68.79%) over a single-controller ReAct baseline; switching the main LLM to DeepSeek-V4-Pro further raises TSR to 68.18% and GCR to 74.62%.
• LoCoMo (conversational memory): ABot-AgentOS Static scores 87.5 overall, beating Mem0 (85.6) by 1.9 points and approaching human performance (87.9); self-evolution raises it to 88.7.
• OpenEQA EM-EQA: Static reaches 59.9 overall (24 frames, 62.8 on ScanNet), outperforming scene-graph/caption/3D-snapshot/3DGS memory baselines; self-evolution improves it to 60.4.
• Mem-Gallery: Static reaches 88.6 overall, above listed textual and multi-modal memory baselines; self-evolution raises it to 89.0.
• NExT-QA: Static reaches 76.5 Acc@All, +3.2 points over GraphVideoAgent (73.3).
• EgoLife (EgoLifeQA): ABot-AgentOS-Qwen3.5-Flash achieves 65.4% average accuracy using only 1 retrieved frame from 1FPS video, best among agentic baselines (best per-category on EntityLog, EventRecall, HabitInsight, RelationMap); improves to 66.2 with self-evolution.
• Reward-engine ablation: initial LLM-as-a-Judge achieves ~60% human alignment, improved to above 90% after Meta-Judge-driven self-evolution.

Why it matters / caveats: The results support the idea that a general Agent-OS layer (hierarchical control + verification + structured graph memory) can improve long-horizon embodied task success and yield near-human, auditable long-term memory performance across diverse benchmarks. Caveats stated by the authors: agent evaluation is only on an initial EmbodiedWorldBench subset (not a full leaderboard), remaining failures stem from coarse visual feedback and indoor/outdoor scene confusion, the self-evolution loop still relies on structured traces and post-hoc (ground-truth or human) correctness signals, large-scale real-world validation under noisy perception/actuation/latency is still needed, and the small-model training pipeline currently uses text-only sandboxes without visual observations.

Problem: Reconstructing high-fidelity 4D dynamic human-scene content (novel view synthesis of moving people plus background) from only a handful of sparse, low-overlap, uncalibrated cameras, where existing dense-camera or sparse-view 4D Gaussian Splatting methods and video diffusion models leave large unobserved regions and produce artifacts or geometrically inconsistent humans.

Method: StudioRecon decouples background and human reconstruction so each can use a different prior. It first runs a feed-forward 3D reconstruction model on first-frame multi-view images to get point clouds, depth, and camera poses, then uses SLERP-interpolated camera trajectories and a camera-controlled video diffusion model (Ren et al. 2025) to synthesize hundreds of novel background views for dense supervision of background 3D Gaussians (optimized with L1/SSIM/LPIPS losses plus density regularization, masked to exclude humans, with an iterative refinement pass for undersampled regions). For humans, it performs per-view SMPL pose estimation, cross-view identity association via a weighted spatial+pose affinity score (Hungarian matching), triangulates 2D keypoints with Huber-loss robust optimization to get 3D joints/vertices, and fits SMPL parameters via 3D-to-3D optimization; human Gaussians are then initialized on the SMPL mesh in canonical pose and deformed per-frame via Linear Blend Skinning plus a temporal MLP for appearance residuals. Finally, a recursive enhancement module applies a single-step diffusion model (Wu et al. 2025b, built on Difix3D+-style artifact removal) with a "motion-adaptive consistency injection" — blending each frame with RAFT-optical-flow-warped previous outputs via per-pixel confidence-weighted exponential moving average — to harmonize the composited background+human render and reduce temporal flicker.

Key results:

• Evaluated on 4 real-world datasets: EgoHumans and Harmony4D (360° camera coverage), Mobile Stage and SelfCap (180° coverage), using 4 sparse training cameras ~90° apart, tested on 4 held-out intermediate-angle cameras.
• Outperforms Dyn-3DGS, MonoFusion, and STG across all metrics; e.g., on 360° scenes averaged, StudioRecon reaches PSNR ~18.6–22.75 / LPIPS ~0.16–0.34 vs. best baselines around PSNR ~15–20 / LPIPS ~0.4–0.7 (e.g., Fencing: 22.75 PSNR / 0.164 LPIPS vs. MonoFusion's 20.09 / 0.472).
• On 180° scenes: Dance 21.74 PSNR / 0.145 LPIPS and Yoga 21.63 PSNR / 0.115 LPIPS vs. best baseline STG at 16.52–18.72 PSNR / 0.289–0.563 LPIPS.
• Ablations: dense view synthesis adds +2.4 PSNR and 36% LPIPS reduction over a 4-view baseline (18.13→20.54 PSNR); enhancement module adds a further 27% LPIPS reduction.
• Cross-view identity association: hybrid spatial+pose affinity achieves 97.8% accuracy with 100% precision (vs. 93.3% spatial-only, 81.4% pose-only), averaged over 8 scenes.
• Motion-adaptive consistency injection reduces frame-to-frame warp error (Warp-L2) by 23% (0.119 → 0.092).
• Seen vs. unseen region fidelity gap: 2.3 dB PSNR (20.88 seen vs. 18.55 unseen) and temporal L1 of 0.059 (seen) vs. 0.178 (unseen).
• Robustness tests show small degradation under simulated segmentation/SMPL/keypoint noise (0.07–0.20 dB PSNR drop).

Why it matters / caveats: The decoupled-prior approach (video diffusion for backgrounds, parametric body models for humans) enables practical 4D capture ("in-the-wild studio capture") with as few as 4 uncalibrated, widely-spaced cameras, and supports applications like novel camera trajectories and human replacement, moving volumetric capture closer to non-professional settings. Stated limitations: fine details like faces/hands remain hard to recover from sparse views; the SMPL-based pipeline cannot handle dynamic hand-held objects (e.g., a basketball disappears); and static-scene shadows baked in at t=0 do not follow subsequent human motion.

Problem: Personal AI assistants on smartphones and AI glasses continuously perceive users' lives through visual and audio streams, but lack a lightweight multimodal memory system that can continuously accumulate, organize, and retrieve long-term everyday experiences to answer queries spanning different temporal horizons.

Method: LightMem-Ego is a deployable streaming multimodal memory system that ingests egocentric visual-audio streams aligned on a shared timeline (using relative timestamp τ = t−t0), segments the stream into micro-events via temporal continuity and cross-frame change signals, and organizes them into a three-level hierarchical memory: current memory (rolling buffer for ongoing scene), short-term memory (recent event segments with visual/transcript evidence), and long-term memory (consolidated into episodic memory for past events and semantic memory for routines/preferences). Costly operations (ASR backfilling, event refinement, indexing, semantic extraction) run asynchronously on the backend, while a lightweight client only samples, compresses, timestamps, and uploads frames/audio. At query time, a memory router selects the cheapest sufficient memory source (current/short-term/long-term) based on temporal scope and intent, fuses retrieved heterogeneous evidence into a compact evidence view, and generates an answer via modular ASR, vision-language, and language model components.

Key results:

• Retrieval (across object finding, conversation recall, life summarization): overall R@1 51.9, R@3 74.1, R@5 77.8, MRR 0.627; life summarization highest (R@3/R@5 = 100.0, MRR 0.944), object finding lowest R@1 (22.2).
• QA accuracy: overall 51.9 LLM-judged / 55.6 human-judged accuracy; life summarization best (77.8/77.8), conversation recall weakest (33.3/33.3).
• Latency (P50/P90): short-term memory QA end-to-end 5.86s/10.95s on phone, 7.01s/9.96s on glasses; long-term memory QA end-to-end 14.87s/35.15s on phone, 19.96s/42.70s on glasses.
• Capability comparison: LightMem-Ego is the only system among 10 compared (ChatGPT Memory, Mem0-style, Memories.ai, Gemini Live, Ray-Ban Meta AI Glasses, Vinci, VisualClaw, VisionClaw, Egocentric Co-Pilot, EgoButler) marked as fully supporting all five capabilities: real-time visual-audio stream, current/short-term multimodal memory, long-term multimodal episodic memory, long-term semantic memory, and timestamped evidence retrieval.

Why it matters / caveats: The paper argues explicit long-horizon multimodal memory (vs. flat retrieval or single-context-window approaches) is a promising direction for personal AI assistants on wearables/phones, demonstrating feasible interactive latency for short-term queries. Caveats stated by the authors: the system relies on upstream API calls for ASR, vision-language, and generation, making accuracy/latency sensitive to external services; object finding and conversation recall remain more challenging than life summarization in the current prototype; and evaluation is conducted on a small set of manually annotated queries.

Problem: Existing mathematics benchmarks for LLMs are limited to competition-style or answer-centric evaluation and rely on final-answer or coarse solution matching, leaving the correctness of advanced natural-language mathematical proof reasoning inadequately and unreliably assessed.

Method: The authors build AdvancedMathBench, comprising ProverBench (245 proof problems curated from undergraduate exams, doctoral qualifying exams, the S.-T. Yau College Student Mathematics Contest, and textbooks, split into 200 UG and 45 QE problems) and VerifierBench (888 model-generated proof trajectories with full-chain expert annotations distinguishing Fatal vs. Recoverable errors). Problems and proofs are curated via a pipeline of PDF parsing/filtering, verifier-uncertainty-based difficulty pre-screening, and PhD-level expert quality control. To score ProverBench proofs at scale, they train an automatic verifier starting from Intern-S2-Preview-35B using GRPO reinforcement learning with a meta-verifier (gpt-oss-120b) reward, augmented with large-scale expert annotation (~2k examples), positive-sample augmentation via proof repair (~1.2k examples), and "pessimistic verification" (a proof is accepted only if all 8 independent parallel verification passes judge it correct).

Key results:

• On ProverBench, the best model, GPT-5.5-xhigh, scores only 64.5 (UG) and 48.9 (QE) pessimistic verification, far below scores on answer-centric HMMT (~98%) and competition-style USAMO (~87%) benchmarks.
• Clear UG→QE difficulty gradient: e.g., Claude-Opus-4.8 drops 59.0→40.0, Gemini-3.1-Pro-Preview drops 46.5→17.8, Kimi-K2.6 drops 48.0→20.0, gpt-oss-120b drops 20.5→2.2.
• On VerifierBench, the best model (DeepSeek-V4-Pro) achieves only 65.1 Meta-Verification Balanced F1 (GPT-5.5-xhigh: 64.9); moving from Rough to Meta-Verification scoring lowers average Balanced F1 by 9.0 points and average TNR by 12.3 points.
• Models tend to over-accept flawed proofs: e.g., gpt-oss-120b and Intern-S2-Preview-35B reach >95 TPR but only ~30-32 TNR.
• The trained automatic verifier reaches 82.1 Rough Balanced F1 / 73.9 Meta-Verification Balanced F1 on a 94-example held-out set, outperforming LLM-as-judge baselines GPT-5.5-xhigh (70.6/61.6) and DeepSeek-V4-Pro (69.9/63.0).
• Ablations show each pipeline component contributes: Meta-Ver-RL over Rough-RL raises Meta-Ver Balanced F1 from 59.5→62.6; positive-sample augmentation raises it to 72.2; pessimistic verification raises it to 73.9.

Why it matters / caveats: The benchmark reveals that frontier LLMs remain far from saturating advanced (undergraduate-to-doctoral level) mathematical proof generation and are especially weak at rejecting plausible-but-invalid proofs during verification, suggesting binary validity judgments overstate true verification capability. The evaluation methodology depends on a learned, model-based verifier trained via RL and expert annotation rather than formal (machine-checkable) proof verification, so its judgments — while shown to agree well with experts on held-out data — are not fully formally guaranteed.

Problem: It is unclear whether, when, and to what extent large language models exhibit genuine metacognitive abilities (monitoring and regulating their own cognition), and how such abilities—if present or induced—could be measured, improved, and leveraged to make LLM-based systems more capable, reliable, and self-driven.

Method: This is a survey/taxonomy paper (not an empirical study). The authors define metacognition as an internal monitoring–control loop (drawing on human cognitive-psychology concepts: metacognitive knowledge, regulation, experience, sensitivity, and calibration) and organize existing LLM research into a taxonomy covering: (1) methods to measure/elicit metacognition (psychologically-grounded measures like meta-d′/M-ratio from signal detection theory, neurofeedback-based measures, confidence-based measures, interpretability-based probes, task-specific paradigms, and benchmarks such as CogEdit, MetaMedQA, ObjexMT, and AwareXtend); (2) techniques to instantiate/give LLMs metacognitive abilities (frameworks, architectures, applications to reasoning models and LLM agents); and (3) metacognition-inspired methods to improve LLM capabilities (task performance, confidence calibration, hallucination reduction, knowledge boundary detection, interpretability, learning efficacy).

Key results:

• Metacognitive sensitivity and efficiency in LLMs are generally weak-to-moderate, though some proprietary models can exceed human sensitivity on narrow tasks (e.g., predicting football outcomes), often in small-sample studies (<50 examples).
• Findings on calibration are mixed and task-dependent: some studies find underconfidence (e.g., coaching tasks), others find widespread overconfidence and inability to retrospectively adjust confidence based on past performance.
• Metamemory (judgments of learning) appears largely absent: GPT models largely fail to predict their own future memory performance.
• Both open- and closed-source frontier models largely fail to faithfully express intrinsic uncertainty, though metacognitive prompting and using metacognitive accuracy as an RL training signal significantly improve faithful calibration.
• Reasoning does not reliably improve metacognition: enhanced reasoning capability can act in opposition to metacognitive sensitivity, and DeepSeek-R1 fails at simple tasks like detecting the length of its own reasoning trace, while some distilled variants (Qwen32B, LLaMA-8B distills) retain reliable internal position estimates.
• Metacognitive ability generally correlates with model size, but a proposed "capacity threshold" of roughly 7–9B parameters exists below which metacognitive prompting can be harmful.
• Post-training (RLHF) effects are model/domain-specific: comparing Llama 3 base vs. instruct models showed instruct-tuning slightly decreased M-ratio, with a more pronounced effect on STEM tasks than humanities.
• Introspective/self-report abilities exist but are fragile and limited to a low-dimensional subspace of a model's activation space (tested on models up to 70B and down to ~8B parameters).

Why it matters / caveats: The paper argues metacognition is central to LLM reliability, hallucination reduction, human-AI collaboration, and interpretability, but the field lacks consensus definitions, standardized benchmarks, and consistent experimental protocols, making cross-study comparison difficult; most cited findings derive from small-sample or narrow task-specific studies, limiting generalizability.

Problem: Existing LLM post-training methods (reward optimization like PPO/GRPO and distribution-matching methods like on-policy distillation/OPD) tightly couple expensive policy exploration to the specific model being trained, preventing update signals from being generated asynchronously, cached, reused, or transferred across different models.

Method: The paper proposes Proxy-guided Update Signal Transfer (PUST), a three-stage pipeline: (1) Proxy Exploration — a lightweight proxy model πϕ is optimized with standard GRPO reward optimization on a target dataset to obtain πϕ+; (2) Update-Signal Extraction — a token-level relative improvement signal Δϕ(a|st) = log πϕ+(a|st) − log πϕ(a|st) is computed, capturing the reward-induced direction rather than the proxy's absolute distribution; (3) Signal Transfer — this signal is transferred to a primary policy πθ via an anchor-calibrated objective that penalizes deviation from a frozen reference policy πref using a calibration coefficient λ, formulated as a KL-divergence-style loss so gradients only update πθ while πϕ, πϕ+, πref stay frozen.

Key results:

• Math domain (Qwen3 family, Mean@16 on AIME24/25, HMMT25 Feb/Nov): PUST (4B→8B) reaches 47.5 average (+30.2 over Qwen3-8B base 17.3), and PUST (1.7B→8B) reaches 37.2 average (+19.9), using proxies trained with GRPO for 500 steps on DeepMath-103K.
• Code domain (HumanEval+, MBPP+, LiveCodeBench): PUST (4B→8B) achieves 60.5 average (+4.6 over Qwen3-8B base 55.9), with LCB gain of +8.3 (16.7→25.0).
• Reusability: the same update signal extracted once from a Qwen3-4B proxy (500 GRPO steps) can be transferred to Qwen3-1.7B, Qwen3-4B, and Qwen3-8B primary models in as few as 50 training steps, each yielding consistent gains (e.g., 4B→8B: +30.2 average).
• Transitivity: signals can be chained across models (4B→1.7B→8B) and still improve the final model (+26.6 average vs base), though with some drift/degradation compared to direct 4B→8B transfer (+30.1).
• Sensitivity analysis: optimal calibration coefficients are λ=1.08 (4B proxy) and λ=1.57 (1.7B proxy) when transferring to Qwen3-8B, both fit with cubic polynomials (R²=0.993 and 0.999).
• Exploration-quality comparison: proxy-explored signals are slightly weaker than signals explored directly by the primary model itself, but larger proxies and longer proxy exploration (800 vs 500 GRPO steps) narrow the gap.

Why it matters / caveats: PUST decouples exploration from alignment so that costly reward optimization can be done once on a small, cheap proxy and its update signal reused/cached/transferred across many differently sized primary models (including weak-to-strong transfer), reducing post-training compute. Limitations noted by the authors: the calibration coefficient λ is currently fixed rather than adaptive, signal quality is bounded by the proxy's data/exploration strategy, and repeated (transitive) transfers introduce signal drift.

Problem: Dexterous-hand robots lack steerable, free-form language-following manipulation policies because large-scale, language-aligned, action-accurate demonstration data for dexterous embodiments is extremely difficult to collect directly on real robots.

Method: The authors build a full-stack system with three components: (1) EgoSmith, a four-stage data pipeline (pre-filtering with optical flow/YOLO hand detection, 4D motion estimation via DPVO + Any4D for camera/depth/world-space hand trajectories, multi-granularity language labeling with Qwen3.5-VL-Plus, and multi-level post-filtering) that curates in-the-wild egocentric videos into 9.6K hours of annotated data across 12 datasets; (2) a unified robot teleoperation/inference stack (PsiBot SynGlove-Air gloves + Vive Trackers, mink-based IK) with a relative-motion-mapping scheme enabling seamless human-in-the-loop DAgger correction; and (3) EgoSteer, a flow-matching VLA pairing a Qwen3-VL 2B backbone with a DiT-based action expert, augmented with a training-only world-model expert that regresses future DINOv3 features to improve action accuracy, plus training-time Real-Time Chunking (RTC) to avoid execution pauses, trained via Conditional Flow Matching.

Key results:

• EgoSmith achieves 9× higher throughput and better annotation accuracy than prior SOTA (HaWoR), yielding 9.6K hours / 2.09M episodes / 1.04B frames of curated data.
• Robot stack collected 187 hours of teleoperation data across 193 tasks; human-in-the-loop handover success rate exceeds 85%.
• EgoSteer achieves 75% average success across 40+ diverse tasks (80+% on 22 tasks), 65% on compositional generalization, and 62% on unseen tasks.
• DAgger refinement (3 iterations, 8.3 hours of correction data) raises success rate from 22.5% (EgoSteer-FT) to 62.5% (EgoSteer-DG) on failure-prone tasks.
• Scaling pre-training data (0→3K→6K→9.6K hours) improves real-robot success rate from 30% (scratch) to 73%, with lower converged pre-training loss.
• On 10 easier tasks, EgoSteer (74%) outperforms baselines π0.5 (22%) and Being-H0.5 (39%) post-trained on the same robot data.
• Ablations: removing the world-model objective drops average success to 31%, removing training-time RTC to 39%, and using unfiltered/noisy egocentric data to 33%, versus 44% for the full method (EgoSteer-1K setting).
• Few-shot adaptation to long-horizon tasks: 75% success on 18-step "box folding" (RealMan, 120 demos) and 83% on 9-step "cake unboxing" (AgiBot G1, 200 demos), while Diffusion Policy, IMLE, and a from-scratch variant all achieve 0%.

Why it matters / caveats: The work demonstrates that systematically curated egocentric human video at scale, combined with a world-model auxiliary objective and DAgger-based real-robot correction, can produce steerable, free-form-instruction-following dexterous manipulation policies that generalize across tasks and embodiments with data-efficient post-training. The authors note limitations: robot degrees-of-freedom restrict transfer of the most intricate human dexterity, lack of tactile feedback limits contact-rich task performance, and pre-training scale could still be expanded for broader generalization.

Problem: It is unclear whether the internal representations of LLMs form reproducible, system-level functional organizations (analogous to brain cognitive systems) that can explain model behavior, failure modes, and their relationship to human cognition, rather than being understood only at isolated local-mechanism or broad-representation scales.

Method: NeuroCogMap extracts sparse autoencoder (SAE) activations from LLMs across a curated capability space and clusters features by task-evoked response profiles into "parcels" (270 parcels selected via a combined clustering-quality/description-quality/non-redundancy score) that form a functional atlas; it estimates a directed cross-layer structural connectome among parcels, assigns interpretable functions via semantic matching and intervention/inhibition tests to build a cognitive atlas, links parcels to cognitive capabilities via correlation and perturbation analyses, and organizes capabilities into a four-level Bloom-inspired hierarchy (perception, representation, abstraction, application). The framework is applied to Gemma2-2B, Gemma2-9B-IT and Llama-3.1-8B (with Pythia models 70M–410M used for hierarchy/prerequisite experiments), and pathologies are analyzed via contrastive circuit/parcel/capability/hierarchy-level comparisons of normative vs. pathological responses, with detection via AUROC-based classifiers and mitigation via targeted parcel intervention/steering.

Key results:

• Optimal parcellation at k=270 parcels (combined score 0.771); intra-parcel semantic similarity (0.638) significantly exceeds inter-parcel similarity (0.433, P<0.0001); parcel functions and activation patterns are partially conserved across Gemma2-2B, Gemma2-9B-IT and Llama-3.1-8B.
• Hallucination detection AUROC across HaluEval, MedHallu, Dolly, NQ-Open, SciQ, TruthfulQA: mean 0.681 (Gemma2-2B) and 0.840 (Gemma2-9B-IT), exceeding SelfCheckGPT and hidden-probing baselines.
• Refusal-failure detection on AdvBench/JBB-Behaviors: near-ceiling AUROC of 0.992 (Gemma2-2B) and 0.990 (Gemma2-9B-IT).
• Sycophancy detection: AUROC 0.685 (Answer) / 0.646 (Feedback) on Gemma2-2B; parcel-guided steering reduced sycophancy rate from 28.4% to 26.4% (Answer) on Gemma2-2B.
• Cortical response prediction (LeBel story-listening fMRI, Schaefer 100-parcel atlas): NeuroCogMap mean r=0.407, outperforming WordRate (r=0.123), Word2Vec (r=0.158), BERT (r=0.219), Language-Context (r=0.361) and Language-Standard (r=0.367); largest gains localized to Default-mode network language regions (Δr up to 0.268).
• Two-step-task decision fMRI prediction: NeuroCogMap mean r=0.355 vs. Word2Vec (0.205), BERT (0.339), Language-Context (0.344), Language-Standard (0.341).
• Predicted LLM-vs-human behavioral fit (AIC) across five Psych-101 task families: r=0.598 (episodic memory), 0.556 (multi-attribute decision), 0.522 (Shepard categorization), 0.508 (intertemporal choice), 0.213 (bandit learning).
• NeuroCogMap-guided refinement of a classical Dual-systems decision model achieved lower (better) held-out AIC than the original and behavior-only discovered models: 262.45 vs. 268.77/268.73 (Two Step Task One), and 465.95 vs. 478.29/473.53 (Two Step Task Two).

Why it matters / caveats: The framework offers a mesoscopic, brain-inspired interpretability layer that connects mechanistic features to behavior, distinguishes qualitatively different failure modes, and enables mechanism-guided detection/mitigation as well as cross-species comparison with human cortex and decision models. The paper explicitly notes it does not claim to solve alignment/safety, and correspondence with human cortex/behavior is presented as functional alignment without asserting mechanistic equivalence between LLMs and brains.

Problem: Existing virtual try-on (VTO) systems realistically transfer garments onto a person but give users little control over how the garment is worn—its size, style (tucked/untucked, open/closed), and spatial placement on the body.

Method: The paper makes two contributions. First, VIP-SAM solves a new task, Visual-Instance-Prompt Segmentation (VIP-Seg), which segments the exact garment instance shown in a support (flatlay) image within a query (on-person) image; it injects support-image features into the SAM (ViT) or SAM2 (Hiera) query encoder via cross-attention adapters at intermediate stages, rather than only at the prompt level as in VRP-SAM. Second, CtrlVTON recasts VTO as an image-editing problem instead of inpainting: a base model (CtrlVTON-base) is a fine-tuned image-editing diffusion transformer (based on FLUX.1 Kontext) conditioned on reference/target image triplets plus discrete garment-class and task tokens (full_swap, partial_swap, add); CtrlVTON then adds pixel-level spatial control by training a LoRA adapter on top of the frozen base model that consumes three VIP-SAM-derived segmentation masks, injected via channel-wise concatenation of VAE latents, trained with a flow-matching objective. Training data is synthesized via a data pipeline and released as the VITON-HD-edit benchmark (2,032 images).

Key results:

• VIP-SAM outperforms PerSAM, Matcher, VRP-SAM, and ProSAM on Fashion-val/test, COCO-20i, and PASCAL-5i (e.g., mIoU 97.2 on Fashion-val, 96.6 on Fashion-test, 74.0 on COCO-20i, 79.1 on PASCAL-5i for the best variant, vs. VRP-SAM's 94.4/94.3/59.4/68.7).
• On single-garment VTO (VITON-HD, OmniTry Bench), CtrlVTON-base beats inpainting baselines (IDM-VTON, CatVTON, Leffa, Voost, CORAL) and editing baselines (Any2AnyTryon, OmniTry), e.g. M-DINO 0.8054 vs. OmniTry's 0.7421 on VITON-HD.
• On multi-garment VTO (DressCode-MR, Garments2Look), CtrlVTON-base surpasses FastFit, OmniTry, and BootComp on M-DINO/M-CLIP-I/GTC/PBC/PR metrics.
• On mask-controllable VTO (VITON-HD-edit), CtrlVTON achieves IoU 0.961 vs. 0.811–0.873 for proprietary editing models (Nano Banana Pro, GPT Image 1.5, Seedream 4.5, FLUX.2 [pro]), and lower Hu-moment distance (0.0022) and Hausdorff distance (26.05) while remaining competitive on garment fidelity and quality.

Why it matters / caveats: By combining an editing-based (mask-free) formulation with mask-based spatial control, CtrlVTON provides pixel-precise styling control (tucking, layering, sizing, per-garment control in multi-garment outfits) that even strong proprietary image editors lack, positioning VTO as an interactive styling tool rather than one-shot generation. The approach depends on accurate instance-level segmentation (VIP-SAM) trained on a ~10K human-labeled mask dataset, and no public benchmark yet exists for multi-garment mask control, so those results are only qualitative.

Problem: Existing text-to-image personalization methods faithfully reproduce a person's identity but offer little precise control for modifying it (e.g., adding a beard, changing eye openness, adjusting freckles) while keeping the edited identity consistent across diverse generated images.

Method: The authors analyze the frozen latent identity token space produced by a Q-Former-based personalization encoder (built on the IP-Adapter architecture, using Omni-ID and PuLID encoders with the Flux.dev text-to-image model), requiring no additional training. They introduce token selection methods to isolate the subset of tokens governing a given facial region — one using paired patched/unpatched images to measure token-wise embedding shifts, another using per-token linear SVMs trained on CelebA attribute labels — and then apply editing techniques within the selected token subspace: global/local identity interpolation, part-level identity transfer between subjects, supervised linear directions (mean-difference and SVM hyperplane normals) for specific attributes, and unsupervised directions via PCA (global, per-token, and per-group), all applied with a norm-preserving update rule.

Key results:

• PCA computed over FFHQ (70,000 face images); supervised attribute directions trained on CelebA (202,599 images, 40 binary attributes).
• Custom benchmark with basic and advanced edits, over 3,000 generated images per method; identity consistency measured via mean pairwise ArcFace cosine similarity, edit/prompt adherence via VQA-based judge.
• On basic edits: Ours (OmniID) achieves ID consistency 0.47, edit adherence 0.79, prompt adherence 0.91, versus PreciseControl (0.25/0.66/0.75) and Weights2Weights (0.31/0.64/0.27).
• On advanced edits: Ours (OmniID) 0.47/0.78/0.89 vs. Flux Kontext Direct 0.39/0.78/0.81 and Kontext Sequential 0.49/0.71/0.80; Ours (PuLID) reaches 0.49 ID consistency, 0.81 edit adherence, 0.88 prompt adherence on basic edits.
• User study (250 pairwise judgments): method preferred over W2W, Kontext Direct, and Kontext Sequential on all five criteria, with overall preference rates of 94%, 85%, and 83% respectively.

Why it matters / caveats: The work shows that identity latent spaces of Q-Former-based personalization encoders are structured and partially disentangled, enabling training-free, fine-grained, localized, and continuous facial identity edits that are difficult to express through text prompts alone, while maintaining identity consistency across diverse generated scenes better than baselines like PreciseControl, Weights2Weights, and Flux Kontext. The authors note that quantitative edit-adherence metrics do not fully capture artifacts introduced by baselines like Flux Kontext, and suggest future work should pursue perceptually aligned metrics and interactive user-guided workflows.

Problem: Existing video motion transfer methods rely on predefined human/character skeleton structures and skeleton-conditional training, which fails to generalize to cross-species or cross-morphology subjects (e.g., different animals, or animate-to-inanimate transfer) where a shared skeletal template does not exist and paired training data is scarce.

Method: Motion4Motion is a training-free framework built on top of WAN-T2V-14B (a Diffusion Transformer with Flow Matching and 3D RoPE), applied entirely at inference by manipulating self-attention. It first establishes cross-image correspondence between anchor points on the source subject (sampled within a Grounded SAM-2 mask) and the target subject via diffusion-feature semantic matching (DIFT), then extracts the source subject's dense pixel-level motion flow across frames using point tracking, and composes the two mappings to obtain the target motion flow. A new module, TransPE (Transferring Positional Encoding), injects this target motion flow into the DiT's self-attention by padding the Key and Value tensors with the target subject's appearance features, re-embedding the padded Keys with RoPE according to the target trajectory coordinates, and concatenating them into the attention computation, while leaving the Query unchanged; manipulation is applied across attention layers 0-40 up to denoising step 35 of 50. For real images, a WAN-I2V model plus deterministic inversion is used to obtain the starting latent.

Key results:

• On human (123 pairs) and animal (33 pairs) motion transfer benchmarks, Motion4Motion achieves best scores across TS, MF, TC, AC vs. MotionDirector, RoPECraft, MotionClone, and FlexiAct — e.g., overall Motion Fidelity 0.448 vs. FlexiAct's 0.384, and overall Appearance Consistency 0.967 vs. FlexiAct's 0.938.
• Introduces a new 50-pair image-video Pose Similarity (PS) benchmark across diverse animal categories; Motion4Motion scores PS 0.543 vs. 0.415 (FlexiAct), 0.408 (MotionClone), 0.355 (RoPECraft), 0.342 (MotionDirector).
• User study (10 raters, 50 test cases, blind pairwise vs. WAN-I2V-14B base model): Motion4Motion preferred 92.5% vs. 7.5% on Motion Consistency and 87.8% vs. 12.2% on Appearance Consistency — the largest margins among all compared methods (e.g., FlexiAct only reached 78.4%/62.1%).
• Ablation on manipulation step count shows step 35 (of 50) is the optimal balance; step 0 fails to relocate features, and full-step (50) manipulation causes semantic confusion/texture artifacts.
• Demonstrated cross-morphology application "Teaching a Table Walking" — animating a static table with a human walking gait using SAM2-based leg binding, which vanilla WAN-I2V/T2V baselines cannot produce.

Why it matters / caveats: The method removes the dependency on skeletal priors and per-case fine-tuning/training entirely, enabling motion transfer across arbitrary morphologies (including animate-to-inanimate objects) using only a pretrained video DiT and inference-time attention manipulation. Caveats: it relies on external point-tracking/matching components (DIFT, Grounded SAM-2, CoTracker) whose correspondence errors could propagate, and challenging cross-morphology cases still require manual "bone binding" of body parts to compensate for structural ambiguity.

Problem: Existing flow-matching mesh generators encode continuous vertex geometry and discrete connectivity in a single joint latent space, which entangles the two dissimilar signals and causes drifting vertices and broken surfaces.

Method: LATO.2 factorizes explicit mesh generation into two stages anchored to a shared coarse voxel scaffold: a Vertex VAE (V-VAE) that encodes a Vertex Displacement Field (via PointNet features rasterized into a 1024³ grid, compressed by sparse 3D convolution/attention into a 64³ sparse latent) and decodes vertices through coarse-to-fine sparse subdivision plus a per-vertex sub-voxel offset/drift head; and a Topology VAE (T-VAE) that encodes connectivity as per-vertex features using adjacency-masked self-attention and decodes pairwise edge probabilities via an MLP classifier, with faces recovered by loop detection. A sparse Flow-Matching Transformer (V-Flow, adapted from TRELLIS, conditioned on DINOv2 image features and a log2(vertex-count) control signal) generates the vertex latent with a rectified-flow objective, followed by a vertex-conditioned Topology Flow (T-Flow) that samples the topology latent conditioned on realized vertex positions and coarse geometric context. Training losses include asymmetric focal loss for occupancy/connectivity, MSE offset regression, and KL regularization; the model is trained on ~450K assets (TRELLIS500K, Objaverse, Objaverse-XL, 3D-FUTURE, Toys4K, ABO) plus 100K procedurally generated synthetic meshes, using 8 H100 GPUs.

Key results:

• V-VAE reconstruction: CD(L1) 0.0003 vs. LATO 0.0038, MeshGPT 0.0185, MeshCraft 0.0524, PivotMesh 0.0395; HD 0.0069 vs. LATO 0.0083.
• Geometry-conditioned mesh generation (vs. AR and flow baselines): CD(L2) 0.0407, CD(L1) 0.0596, HD 0.0657, |NC| 0.8341 — best across all metrics, beating LATO (0.0421/0.0617/0.0738/0.8262), MeshFlow, DeepMesh, MeshRipple, BPT, MeshSilkSong, FastMesh, MeshAnythingV2.
• Topology generation ablation: GT-Verts+T-VAE upper bound CD(L2) 0.0351; full pipeline (V-Flow vertices + T-Flow) achieves 0.0407, showing most reconstruction quality is preserved through the generative stages.
• V-VAE ablations: removing the offset head keeps occupancy metrics (F1 0.9735) unchanged but increases CD(L1) from 0.0003 to 0.0011 and HD from 0.0069 to 0.0082; removing synthetic-mesh training or the high-res VDF downsampling path substantially degrades occupancy accuracy (e.g., IOU drops to 0.8873–0.9034).
• Part-wise generation scales face count with number of parts (e.g., 6→9→34 parts yields 12,853→15,690→24,003 faces on one example), producing higher resolution than monolithic generation at the same latent capacity.
• Parameter counts: V-VAE ~320M, V-Flow ~160M, T-VAE ~180M, T-Flow ~240M (trainable parameters).

Why it matters / caveats: The vertex/topology factorization improves geometric fidelity and connectivity coherence over prior joint-latent (LATO, MeshFlow) and autoregressive (BPT, DeepMesh, MeshRipple) mesh generators, and uniquely enables controllable vertex count, part-wise high-resolution generation, and topology-adaptive editing (mesh stitching, part transformation) without re-optimization, since connectivity can be regenerated purely by re-running T-Flow on updated vertices. The paper notes a concurrent, independently developed work (Nexus) takes a similar vertex-then-topology approach but with different stage designs.