AI papers — 2026-07-07

Problem: Training a single GUI agent to work across heterogeneous platforms (desktop and mobile) is hard because naively combining cross-platform trajectories via mixed SFT, mixed RL, or model merging mixes distinct platform-specific interaction conventions, causing an averaged policy, platform-specific capability degradation, and catastrophic forgetting.

Method: The authors build Uni-GUI, a ~160K-step, ~11.5K-trajectory cross-platform dataset (desktop and mobile) collected via a unified data-collection harness plus cleaned OpenCUA and OpenMobile trajectories, and propose UI-MOPD, a two-stage training pipeline. Stage 1 performs supervised fine-tuning of Qwen3-VL-32B-Thinking on Uni-GUI to produce separate desktop and mobile expert teacher policies. Stage 2 trains a shared Qwen3-VL-8B-Thinking student with online reinforcement learning (clipped policy-gradient objective with a rule-based structured outcome reward) combined with multi-teacher on-policy distillation (MOPD): each rollout is routed to its platform-specific teacher via platform-conditioned teacher routing, and a reverse-KL penalty (estimated with the K3 estimator, plus an adaptive group-level KL mask that relaxes the penalty when task reward is already sufficient) aligns the student to that teacher's behavior only on states the student itself visits.

Key results:

• UI-MOPD reaches 38.2% task success on OSWorld and 12.0% on MobileWorld, i.e., relative improvements of 12.7% and 55.8% over the Qwen3-VL-8B-Thinking base model.
• It outperforms integration baselines: Mixed-SFT (35.0%/6.4%), Model Merge via weight averaging (36.5%/6.8%), and Model Merge via TIES merging (36.8%/0%) on OSWorld/MobileWorld respectively.
• The 32B platform-specific teachers achieve 46.3% (desktop teacher on OSWorld) and 16.2% (mobile teacher on MobileWorld), while single-platform SFT of the 8B student causes catastrophic forgetting (SFT on OSWorld drops MobileWorld to 0%; SFT on MobileWorld reaches 12.8% but with weaker desktop gains) — UI-MOPD instead improves both simultaneously (+4.3 points on each benchmark over the base 8B model).
• On general/static evaluations, UI-MOPD improves AndroidControl* accuracy from 78.73% to 80.05% (vs. 74.01% for TIES model merging) and largely preserves grounding performance (ScreenSpot-Pro 43.14% vs. base 43.71%; ScreenSpotV2 90.88% vs. 91.27%; OSWorld-G improves 52.13%→52.84%), whereas TIES merging degrades all three grounding benchmarks (e.g., ScreenSpot-Pro drops to 37.13%).

Why it matters / caveats: The results suggest platform-conditioned on-policy distillation is a more effective way than model merging or mixed SFT to build a single cross-platform GUI agent that retains platform-specific behaviors while adapting to new platforms, without inflating inference-time cost (only one student model is deployed). The evaluation is limited to two platforms (desktop via OSWorld, mobile via MobileWorld) and relies on teacher models trained with proprietary data-collection pipelines (Kimi-K2.6, Gemini-3.1-Pro), which may limit reproducibility.

Problem: 3D scene reconstruction and generation are typically handled by separate paradigms (pixel-based regression vs. latent diffusion), and recent attempts to unify them still define the diffusion objective on intermediate VAE/RAE latent features rather than the actual 3D representation, causing information loss and requiring a pretrained autoencoder.

Method: PixWorld is a ~1.04B-parameter, from-scratch two-stream diffusion transformer (DiT) that partitions posed multi-view inputs into clean and noisy view subsets, jointly processes them (with PRoPE camera-pose injection and cross-attention text conditioning) through shared transformer blocks, and decodes features into a pixel-aligned 3D Gaussian Splatting (3DGS) representation via depth unprojection. Training applies a flow-matching (JiT-style) loss directly on differentiably-rendered multi-view images (plus MSE/LPIPS on clean views and a Huber depth loss against DA3 pseudo-depth) rather than on VAE/RAE latents, and adds a geometry perception loss that matches rendered vs. ground-truth features from a frozen 3D foundation model (π3) to supervise 3D structure beyond photometric cues. It is trained on RealEstate10K + DL3DV-10K (~67K scenes) plus 10M BLIP-3o single images, for ~200K steps at 336×448 resolution on 32 A800-SXM4-80G GPUs with AdamW.

Key results:

• On RealEstate10K/DL3DV-10K 4-view and 8-view reconstruction, PixWorld achieves best PSNR/LPIPS in all settings (e.g., 4-view RealEstate10K LPIPS 0.138 vs. 0.143 for next-best YoNoSplat).
• Single-image (1-view) generation on RealEstate10K: PixWorld improves PSNR by +1.06 dB over LVSM (18.88 vs. 17.82) and raises pose accuracy AUC@5 from 0.546 to 0.614.
• Single-image generation on DL3DV-10K: PSNR +0.75 dB over the next-best baseline (16.50 vs. 15.75) and AUC@5 rises from 0.420 to 0.485.
• Two-view generation: PixWorld attains best LPIPS (0.210 on RealEstate10K, 0.340 on DL3DV-10K) and best AUC@5 (0.649 and 0.534 respectively).
• On the WorldScore benchmark, PixWorld achieves the best overall average score of 71.04, leading in camera control (91.08), 3D consistency (91.39), and photometric consistency (93.84).
• Ablation removing the geometry perception loss (1-view RealEstate10K, 30K-step controlled run) drops PSNR by 1.13 dB (19.12→17.99), SSIM by 0.105 (0.717→0.612), and AUC@5 by 0.080 (0.642→0.562, ~12.5% relative).

Why it matters / caveats: By supervising diffusion directly on rendered pixels instead of latent codes, PixWorld avoids the information loss and extra training cost of a VAE/RAE while unifying 3D generation and reconstruction in one model, matching state-of-the-art reconstruction methods and outperforming prior latent-space generation approaches across RealEstate10K, DL3DV-10K, and WorldScore.

Problem: Producing the "last mile" dissemination artifacts of a paper (poster, talk video, blog post) is still a manual process, and prior automation tools re-extract each paper from scratch per artifact, ship one-way renders authors cannot reopen in PowerPoint/Word, and gate quality on soft VLM-preference scores that plateau while sections still read as empty.

Method: ResearchStudio-Reel is a five-skill composition built on Claude Code and Codex skill runtimes: a shared extractor Paper2Assets (parses paper text, figures, metadata, and a nine-section summary once into a reusable bundle), three editable generators Paper2Poster, Paper2Video, and Paper2Blog, and a convergence layer Paper2Reel that binds the three into a navigable HTML viewer. Paper2Poster composes posters along four axes (layout, style, header, Scan-to-Read block) and runs a discrete "measured-fill loop" that classifies each section into five fullRatio bands (EMPTY/SPARSE/FULL/SPILLAGE/OVERFLOW) and edits one section per round until convergence, then exports native PowerPoint via DOM-to-EMU reconstruction (including OMML equations). Paper2Video delegates deck authoring to the ppt-master workflow, plans narration/duration before TTS (Edge TTS), adds narration-aligned visual highlight cues, and outputs video.pptx plus subtitled/non-subtitled MP4s tied together by a timeline.json sidecar. Paper2Blog builds one shared evidence map and produces two register-specific DOCX articles (Chinese WeChat-style and English research-blog style) gated by a layout-aware DOCX check for orphan tails, near-blank pages, and under-filled images. Each generator releases only through a hard pass/fail deliverables gate rather than a continuous VLM-judge score.

Key results:

• On the 100-paper Paper2Poster benchmark (scored by claude-opus-4.8 and gpt-5.5 judges), ResearchStudio-Reel (Claude Code, claude-opus-4.8) reaches an aesthetic mean of 3.52 vs. the author ground-truth's 2.94, and an information mean of 3.90 vs. 3.63.
• It wins the overall per-paper score on 84–93% of papers and loses outright on at most four.
• Single-shot frontier LLMs (Claude-4.8 Opus, GPT-5.5, Gemini-3.1 Pro) score ~3.4–3.5 overall vs. ~2.1–3.3 for prior poster pipelines (Paper2Poster Tool, PosterGen, P2P).
• Ablation: holding the model fixed, wrapping claude-opus-4.8 in the skill/fill-loop machinery raises aesthetics from 2.76 to 3.52 and Layout from 2.83 to 3.97 (a 0.76 aesthetic gain from the loop, not the model); the same substitution under gpt-5.5 lifts aesthetics from 3.03 to 3.36 and Layout from 3.22 to 3.82.
• Swapping the harness/model (Claude Code+claude-opus-4.8 vs. Codex+gpt-5.5) while holding the skill fixed retains most quality: aesthetic mean 3.52 vs. 3.36.
• ResearchStudio-Reel is the only system among compared baselines that emits all three editable artifacts (HTML, PPTX, video, blog) per the capability audits (Tables 2–3).
• A full four-artifact bundle from one PDF takes ~89.2 minutes and ~2.6M input / ~276K output tokens per paper (mean over 5 papers on claude-opus-4-8), with Paper2Poster's fill loop and Paper2Video's ppt-master/QA gate as the heaviest stages.

Why it matters / caveats: By sharing one extraction pass and gating on categorical fill verdicts rather than soft aesthetic scores, the system produces posters that beat human author ground-truth on aesthetics while remaining the only pipeline to ship all three artifacts as natively editable files. The authors note PaperQuiz reading-comprehension scores invert relative to aesthetics (denser, less-designed posters score higher on comprehension), the system cannot fabricate bespoke explanatory diagrams the way human designers do, and it is calibrated only on ML/CV/NLP venues, with transfer to other domains untested.

Problem: Existing LLM research-ideation tools can generate plausible candidate directions but lack a reusable skill layer that grounds a problem in current literature, diagnoses real bottlenecks, differentiates a proposal from prior art, and audits it against known failure modes before implementation begins.

Method: ResearchStudio-Idea is a three-skill suite: Paper-Search (multi-source literature retrieval across arXiv, DBLP, OpenAlex, OpenReview, Semantic Scholar, and Crossref), Scoop-Check (a prior-art collision checker that decomposes novelty claims into problem framing, core mechanism, key insight, and application domain), and IdeaSpark (the end-to-end ideation skill). IdeaSpark is built from a corpus of 1,947 ICLR/ICML/NeurIPS papers (2021-2025, labeled Oral/High-Cited/Reject); Claude Sonnet 4.6 extracts a two-stage 12-field "innovation signature" per paper (8 base fields, then 4 domain-agnostic rewrites), the four abstracted fields are embedded with OpenAI text-embedding-3-large and reduced/clustered via UMAP + HDBSCAN (min_cluster_size=10) into 31 ideation sub-pattern clusters, and Claude Opus 4.7 induces these into 15 higher-level ideation-pattern cards (each with success conditions, failure modes, and reviewer-expectation evidence). At runtime, IdeaSpark's two-tier, phase-based workflow performs evidence grounding, bottleneck identification, pattern-guided candidate generation (selecting 1-3 patterns), collision retrieval, outcome-informed audit, and idea-card rendering.

Key results:

• Corpus: 1,947 papers (1,014 Oral, 260 High-Cited, 722 Reject, 49 Oral∩HC overlap) across ICLR/ICML/NeurIPS 2021-2025.
• HDBSCAN with min_cluster_size=10 yields 31 sub-pattern clusters with the best silhouette score (0.584) in the sweep, though 47.7% (902/1,891) of papers remain "unclustered."
• Opus 4.7 induces 31 sub-patterns into 15 top-level ideation patterns in a single pass.
• Multi-label tagging shows k=2 ideation patterns is the modal composition for 59.2% of all 1,891 papers, with a 33.6% tail executing k≥3 patterns; class means are tightly clustered (Reject 2.26, Oral 2.28, HC 2.32).
• Reject-only re-clustering maps every rejected-paper cluster onto the existing 15-pattern vocabulary with no out-of-taxonomy bucket.
• The strongest Oral-enriched pattern combination is Algebraic Equivalence Unification + Generative Process Redesign at pO=85.0% (+26.6 percentage points over the 58.4% dataset baseline, n=20).
• In a blind automated-judge evaluation over 100 ICLR-2026-Oral problem seeds (3 blind rounds each against Opus-4.8 bare, Opus-4.8 self-generated, and GPT-5.5 bare), IdeaSpark achieves the highest mean idea-quality in every one of 21 ICLR primary-area domains while maintaining competitive novelty.

Why it matters / caveats: The authors argue that large-scale conference outcomes (including rejected papers) contain reusable, operational signals about how impactful research directions are formulated and audited, and that these can be packaged as inference-time skills rather than acceptance predictors. The reported quality gains rest only on blind automated-judge evaluation at the idea stage; the paper explicitly states this is not validated by human peer review or downstream implementation/execution success.

Problem: The rapidly expanding landscape of over one hundred deep-learning optimizers for large language model training lacks a unified framework for comparing methods across mechanism, mathematical form, and measurable training objectives, making optimizer selection an ad hoc, protocol-sensitive problem.

Method: The paper presents OmniOpt, built on four coupled components: (1) a universal meta-pipeline that decomposes every optimizer update into six stages (S0 gradient acquisition, S1 parameter scoping/routing, S2 gradient transformation, S3 state evolution, S4 update reconstruction, S5 update finalization), with an "identity-mapping principle" showing most optimizers only modify one or two stages (e.g., AdamW is primarily S3/S5, Muon is S1/S2, GaLore spans S1-S4, Lion is S2/S3, SAM is S0/S5); (2) a norm-constrained linear minimization oracle (LMO) framework that unifies update directions (sign, spectral/polar, adaptive-box) into a four-axis decomposition of update domain, state estimator, geometry/precondition operator, and finalization wrapper; (3) a dual-dimension taxonomy classifying 108 optimizers along a methodological dimension (five families T1-T5) and an objective-oriented dimension (six effect objectives O1-O6: convergence efficiency, computation cost, memory cost, stability, hyperparameter robustness, generalization); (4) a cross-domain benchmark spanning language model pretraining and CIFAR100 image classification, plus a mechanistic ablation of Muon.

Key results:

• The taxonomy covers 108 optimizers across five mechanism families: T1 adaptive moments/scalar control (43 members), T5 curvature-aware/geometric regularization (25 members), T2 matrix-structured methods (16 members), T4 state compression/structural aggregation (11 members), and T3 discretization/quantization (5 members).
• Under BF16 training with FP32 master weights, Adam's two FP32 optimizer-state tensors (m and v) add 8 bytes per parameter, so a 7B-parameter model requires about 56 GB for optimizer states alone, before activations or distributed-training overhead.
• The benchmark spans language model pretraining from 60M to 1B parameters, across four architectures, at context lengths from 256 to 32k tokens, together with image classification on CIFAR100.
• Optimizers and taxonomy families are evaluated against six effect objectives (O1-O6), and the paper reports that no single optimizer dominates this multi-objective frontier across convergence, cost, memory, stability, hyperparameter robustness, and generalization.

Why it matters / caveats: OmniOpt gives practitioners a mechanism-aware coordinate system for selecting or combining optimizers under explicit compute, memory, and tuning constraints, rather than relying on fragmented, protocol-sensitive comparisons. The paper reports qualitative trends—aggressive state compression excels under short context but degrades sharply as input complexity grows, structured-matrix methods transfer most stably across architectures and tasks but at substantial per-step cost, and optimizer rankings exhibit systematic crossings with scale, context length, and domain—but the detailed quantitative benchmark tables (Section 6) were not present in the excerpted text.

Problem: Evaluating robot foundation models requires slow, costly real-world rollouts, and it remains poorly understood which properties make a learned video world model a reliable surrogate for real-robot policy evaluation.

Method: The authors build WMBench, a benchmark of paired teleoperated and policy-rollout trajectories (2,989 trajectories across 8 manipulation tasks) with a four-level World Model as Evaluator Score (WMES), and use it to run a controlled study of 7 video world models, 4 action-representation schemes, and 324,000+ annotated rollouts (plus a fine-tuned Qwen3-VL-8B-Instruct LoRA judge for scalable outcome annotation). Guided by the resulting design map, they build GigaWorld-1, an autoregressive diffusion-transformer world model on Wan 1.3B/5B backbones trained on ~12,980 hours of curated data (physical/internet videos, open-source robot data, egocentric data, Giga-collected demonstrations), using channel-concatenated/pixel-aligned control (EE pose maps for the head camera, ray maps for wrist cameras), LoRA adapters, and a hierarchical memory with a persistent first-frame anchor plus short/mid/long-range history for long-horizon rollout stability.

Key results:

• Corpus/benchmark scale: 324,000+ analyzed rollouts, 7 world models, 4 action representations, 12,980 hours of training data, and WMBench dataset with 50,000+ Hugging Face downloads.
• Metric correlation with WMES: Visual Fidelity group ρ=0.78, Geometry ρ=0.71, Semantics ρ=0.59; individually Subject Consistency ρ=0.88 and Perspectivity ρ=0.86 are strongest predictors, while Background Consistency (ρ=-0.45) and Photometric Consistency (ρ=-0.42) are negative predictors.
• VLM evaluator vs. human annotators (5,000+ videos): 87.80% exact agreement, 99.16% adjacent agreement, only 0.84% large errors, MAE 0.1304, RMSE 0.3836, quadratic weighted kappa 0.7349, Spearman 0.7574, Kendall τb 0.7507.
• Pretraining backbone comparison (average evaluator score): Cosmos-Predict2.5 0.6123, Wan 2.2 5B 0.5948, LTX 2.3 (22B) 0.5775, CogVideoX 0.5620, SVD 0.5569 (worst, Trajectory Accuracy only 0.0926).
• Data composition ablation on Wan2.1 1.3B: GigaData-only average 0.5654; GigaData+PhysData raises it to 0.6144 (+0.0490, driven by Photometric Consistency +0.3074); GigaData+AgiBot raises it to 0.5940 (+0.0286) but drops JEPA Similarity by -0.2426 and Trajectory Accuracy by -0.1084.
• Action-control interface ablation (Trajectory Accuracy): no control 0.1576, cross-attention 0.1620, ControlNet-style 0.2566, channel-concatenated control 0.3528 (best across all metrics tested).
• Long-horizon rollout quality (0-40s, chunked): adding hierarchical memory to Wan 2.1 1.3B improves PSNR from 14.46 to 19.82 (0-8s) and keeps it near 17.4 at 32-40s, versus other backbones (e.g., SVD) dropping to PSNR ~6.8-7.2 and FID/FVD in the 400+ range at longer horizons.
• GigaWorld-1 improves the core evaluator-alignment metric by 14.9% over competitive state-of-the-art baselines.

Why it matters / caveats: The study reframes world-model evaluation around long-horizon, action-faithful consistency rather than short-term visual realism, showing that some commonly used appearance-stability and interaction-quality metrics are actually negatively correlated with true evaluator reliability and can reward degenerate (e.g., static) generations; the authors release code, model weights, datasets, and toolkits, but findings on data/architecture trade-offs (e.g., AgiBot vs. PhysData) are drawn from ablations on a single backbone (Wan2.1 1.3B) and a GigaData-centered benchmark, which may limit generality to other robot platforms.

Problem: Self-supervised visual foundation models optimize for semantic invariance and treat boundaries/shapes as downstream outputs rather than native learning signals, leaving them comparatively weak at the dense, fine-grained spatial understanding (depth, segmentation, geometry) that embodied/physical AI requires.

Method: The paper proposes masked boundary modeling, a self-distillation framework (extending DINO/iBOT-style teacher-student EMA training) in which the teacher's own online boundary-field predictions identify boundary-bearing tokens and force them into the student's masked set; masked tokens are then "routed by geometry," with boundary tokens receiving both the standard semantic iBOT objective and a categorical boundary loss, while other masked tokens get only the semantic objective. Boundary geometry is represented via a dense, over-parameterized attraction/boundary field (distance, orientation, and endpoint angles per pixel) that is reparameterized as a per-pixel categorical distribution over discretized bins (K=32 at scale), which is stabilized with DINO-style centering/sharpening and doubles as an a-contrario (LSD-style) statistical validation signal; targets are bootstrapped online from scratch using a frozen single-block ViT corner-point detector plus vote-aggregation decoding, with no human labels, external edge detectors, or pretrained backbones. The recipe is scaled to LingBot-Vision, a ViT-g/16 backbone (~1.1B parameters, SwiGLU, RoPE, 4 register tokens) trained on a 161M-image curated corpus (from a 2B-image pool) in three stages (300k pretraining, 100k Gram-anchoring, 100k high-res adaptation at 512px), then distilled into ViT-L (300M), ViT-B (86M), and ViT-S (21M) students, and further used as the encoder for LingBot-Depth 2.0, a depth-completion system trained on data scaled from 3M to 150M samples.

Key results:

• On ImageNet-1K proof-of-concept (ViT-L/16), the full recipe improves k-NN top-1 from 81.6% to 82.4% and NYU-Depth v2 linear-probe δ1 from 81.4% to 84.9% (RMSE 0.474 to 0.440) versus a matched DINO+iBOT baseline.
• LingBot-Vision (ViT-g/16, 1B params) attains the best NYUv2 depth RMSE of all compared models (0.296), beating the 7B DINOv3 (0.309) and 2B V-JEPA 2.1 (0.307), and is 13% lower RMSE than the best same-scale competitor AM-RADIOv2.5 (0.340).
• On KITTI depth, LingBot-Vision reaches RMSE 2.552, best among sub-2B models, trailing only 7B DINOv3 (2.346) and 2B V-JEPA 2.1 (2.461).
• On semantic segmentation, LingBot-Vision scores 53.5/79.6/87.5 mIoU on ADE20K/Cityscapes/VOC12, within 1.3 mIoU of distilled DINOv3 ViT-H+ on ADE20K and beating DINOv2 by ≥4 mIoU on all three benchmarks.
• Training-free video object segmentation: 70.0 J&F on DAVIS-2017 and 73.5 on YouTube-VOS, on par with DINOv3 ViT-H+ (71.1/74.0) and 6.1-7.9 points above equal-size DINOv2.
• Uses less than one-third the training samples and roughly one order of magnitude smaller curated corpus (161M vs. DINOv3's 1,689M images) than DINOv3.
• The distilled 0.3B (ViT-L) student matches the 7B DINOv3's NYUv2 RMSE (0.310 vs. 0.309) with about 23x fewer parameters; the ViT-B student gets the best linear ImageNet accuracy (85.05) and best depth (0.339 NYUv2, 2.793 KITTI) in its size class.
• In an encoder-initialization ablation for masked depth modeling, LingBot-Vision initialization is the strongest at ViT-L across all benchmarks tested (e.g., DIODE-Indoor RMSE 0.094 vs. 0.114 for DINOv3 and 0.152 for DINOv2).
• Scaling depth-completion training data from 3M to 150M samples, the LingBot-Vision-initialized model's D10^2 accuracy keeps improving (0.777 to 0.795) while the DINOv2-initialized model saturates (0.752 to 0.755), and LingBot-Depth 2.0 sets leading performance on 14 depth completion benchmarks.

Why it matters / caveats: Treating boundaries as native, self-bootstrapped pretraining signals (rather than post-hoc outputs) lets a single self-supervised objective produce representations that are simultaneously semantic and spatially precise, closing much of the gap to larger or multi-teacher-distilled models on dense tasks while using far less data and compute; the main caveat is that LingBot-Vision still trails on pure image-level classification (e.g., 86.32 vs. DINOv3-7B's 87.87 linear ImageNet accuracy), reflecting a trade-off toward localized geometric structure over global semantic invariance.

Problem: Real-robot deployment of trained manipulation policies (VLAs, VAMs, WAMs) is fragmented into per-model, per-robot scripts, making it hard to reproduce, debug, compare, and evaluate policies across robots, and evaluation data is typically lost rather than fed back into training.

Method: EVA-Client is an open-source, component-decoupled client that sits between signal sources (openpi, StarVLA, GR00T, Dream-Zero, or human teleoperation) and robot execution (ROS1/ROS2/ZMQ transports), organized into five layers (transport, robot description, policy client, inference strategy, CLI/web) with narrow dataclass interfaces so robots, transports, and strategies can each be swapped independently. It unifies real-time inference/scheduling strategies behind one configuration surface: synchronous execution, asynchronous prefetch with latency-aware linear-overlap blending (Yu et al., 2026), ACT-style exponentially-weighted temporal ensembling (Zhao et al., 2023a), a naive-async chunk-replace baseline, and Real-Time Chunking (Physical Intelligence, 2025) with a final linear-overlap smoothing pass; a shared PyRoki-based (JAX) Levenberg–Marquardt inverse-kinematics solver converts between joint/end-effector action spaces. It exposes Debug, Collect, and Eval workflows through a single Viser-based web console, records teleoperated demonstrations directly into LeRobot-format datasets with automated quality control, and logs three parallel action streams (raw policy predictions, smoothed actions, executed commands) per evaluation trial for reproducible, checkpoint-comparable scoring.

Key results:

• Supports six robot embodiments out of the box (AgileX Piper, ARX R5, Franka, UR5e, AgiBot G2, Galaxea R1-lite) through one robot-agnostic backend, with a reference dual-arm Piper action vector of 14 dimensions (2 × (6 joints + 1 gripper)).
• Implements 5 distinct inference/smoothing strategies (Synchronous, Async Prefetch, Temporal Ensemble, Naive Asynchronous, RTC) behind a single configuration surface (Table 2).
• Defines 5 execution/debugging granularities (open-loop simulation, real single-chunk stepping, segmented sim-to-real, continuous execution, data collection) as shown in Table 1.
• ACT-style temporal ensembling uses a default decay coefficient m = 0.01, producing near-uniform averaging across overlapping chunk predictions.
• Qualitative real-robot demonstrations (not a controlled benchmark) show synchronous execution stalls a Piper arm during table-tennis (rally does not get going), while asynchronous scheduling and RTC/overlap smoothing sustain both the high-dynamics table-tennis rally and a long-horizon cloth-folding task using the same codebase with only the inference strategy and robot description changed.

Why it matters / caveats: By consolidating deployment, debugging, teleoperated data collection, and reproducible/logged evaluation into one open-source client (rather than a training method or policy), EVA-Client aims to close a deployment-tooling gap alongside mature training frameworks like openpi and LeRobot; the authors note the reported task outcomes are illustrative deployment observations rather than a controlled study, non-ROS robots still lack a unified camera interface, and IK support is currently limited to serial-arm manipulators.

Problem: Wan-Streamer v0.1's native-streaming audio-visual interaction model was limited to 192×336 output, which compresses body posture, nearby objects, and scene context too much for scene-grounded, wide-composition real-time interaction.

Method: Wan-Streamer v0.2 keeps the v0.1 native-streaming formulation (user/agent text, audio, and video on one causal timeline modeled by a single Transformer with block-causal attention) but raises output resolution from 192×336 to 640×368 at 25 FPS. To preserve latency, the model is split into a single-GPU "thinker" (streaming perception, causal audio/video encoders, language/state Transformer pass, KV-cache construction, and causal decoding) and a multi-GPU "performer" using Ulysses-style context-parallelism (DeepSpeed Ulysses) for the expensive flow-matching latent generation path. Each performer rank writes incoming K/V into a pre-sharded local cache; the long high-resolution latent video sequence is split across ranks and gathered via Ulysses all-to-all/gather communication, while the shorter audio latent sequence is generated without sequence sharding.

Key results:

• Output resolution increased from 192×336 to 640×368 while maintaining approximately 200 ms model-side signal-to-signal latency at 25 FPS.
• Total remote interaction latency remains approximately 550 ms when including the same 350 ms bidirectional network budget used in v0.1.
• Streaming unit cadence is 160 ms per response boundary, matching the v0.1 measurement convention.

Why it matters / caveats: The upgrade enables scene-grounded mid-shot agents with legible posture, gaze, hands, and local scene layout during real-time conversation, expanding beyond close-up video-call framing, without increasing model-side latency. The paper reports only qualitative visual observations (no quantitative visual-quality metrics), and the network latency term is treated as an external deployment assumption rather than a measured quantity.

Problem: Unified robot manipulation models that combine vision-language understanding with future-state prediction and action tend to erode the pretrained VLM's semantics, suffer interference among heterogeneous training objectives, and learn future prediction from scratch in pixel space rather than exploiting pretrained video generators' dynamics priors.

Method: InternVLA-A1.5 uses a Mixture-of-Transformers architecture pairing a pretrained Qwen-3.5 2B VLM backbone (hybrid attention: 3 Gated DeltaNet linear-attention layers interleaved with 1 full-attention layer) with a lightweight 460M-parameter unified expert of the same hybrid design, connected only through shared full-attention layers. Training is a multi-stage recipe: Stage 1 co-trains the VLM on VQA and robot data with next-token prediction over subtask descriptions and FAST-tokenized discrete actions (in the spirit of pi-0.5); Stage 2 adds the unified expert, which uses learnable "foresight tokens" that query the shared multimodal context to produce a compact latent code conditioning a frozen, pretrained WAN2.2-5B video generation model (replacing its T5 text encoder) via a flow-matching video prediction loss, while separate flow-matching action-query tokens generate continuous action chunks; the video branch is discarded at inference. Pretraining uses AdamW (300K steps Stage 1, 600K steps Stage 2, batch size 1024) on 1.2M robot episodes/861M frames from six sources (InternData-A1, AgiBotWorld, UMI, DROID, Galaxea, RoboMind 1.0) mixed with ~3M multimodal samples (InternVLA-M1 corpus: General QA, Box QA, Point QA, Trajectory QA) at a 0.15:0.85 robot:multimodal ratio using two-level grouped sampling (Re-Mix-derived weights).

Key results:

• Best overall results on all six simulation benchmarks: LIBERO 98.9% average, RoboTwin 2.0 93.2% average, SimplerEnv 80.8% average (vs. pi-0.5's 57.1%, a 23.7-point gain), LIBERO-Plus zero-shot 84.8% total, DOMINO zero-shot 27.7% SR (29.3% after fine-tuning), EBench Test SR 35.2%.
• Real-world long-horizon MOF chemistry task: 76.4% success vs. pi-0.5's 29.3% and Motus's 0.0%.
• Real-world Insert Tubes and Move Tubes tasks: InternVLA-A1.5 leads pi-0.5 by 20.8 and 7.8 points respectively; on Sort Tubes it trails pi-0.5 slightly (75.9% vs. 77.8%).
• On held-out (OOD) instruction bindings across all three instruction-following tasks, InternVLA-A1.5 achieves the best success rate, showing gains are not from replaying memorized bindings.
• Ablations: removing the video supervision loss and removing foresight tokens both degrade performance across LIBERO, LIBERO-Plus, RoboTwin, and DOMINO, with the largest drops on zero-shot LIBERO-Plus and DOMINO.
• Inference runs in real time on a single NVIDIA RTX 5090 GPU, about 0.1s per inference step, since the video generation branch is discarded at deployment.

Why it matters / caveats: The design shows that a unified VLA model can inherit world-model dynamics priors from a frozen pretrained video generator through a handful of latent foresight tokens, avoiding costly pixel-level generation while preserving VLM semantics for strong compositional and zero-shot generalization plus long-horizon execution. The authors note two limitations: foresight supervision only spans the short horizon of a single action chunk (no long-horizon imagination/planning), and the frozen, generic video generator bounds the inherited priors by its own pretraining coverage of embodied scenes.

Problem: Existing KV cache eviction methods for autoregressive LLM decoding rely on static heuristics or proxy attention scores that poorly predict which tokens will actually matter for future queries, causing brittle eviction as token relevance shifts over long reasoning traces.

Method: KVpop is a sparse-attention retrofit that gives each KV head a fixed budget B = s + w + k: sink tokens, a protected recent window, and a learned long-range top-k cache populated by lightweight per-head scoring modules (a stateless two-layer MLP or a stateful mLSTM-based scorer). Scorers are trained via distillation against a novel future-attention supervision target — the attention mass a token receives after leaving the protected window — computed cheaply during training with a transposed-attention pass that reuses log-sum-exp (LSE) normalizers instead of materializing a dense S×S attention map, combined with a boundary-aware pairwise logistic loss evaluated only at the eviction cutoff. The stateful mLSTM variant additionally delays scoring until a token reaches the eviction boundary, letting it accumulate near-future context (unlike DMS-style gates that fix retention at insertion time), and a Fenwick-tree-based running top-k algorithm implemented in Triton/FlexAttention computes per-query cutoffs without ever storing a dense mask. KVpop is applied to Qwen3-4B-Instruct-2507 and Qwen3-8B, trained for 2,000 steps on the Nemotron-Math v2 dataset (sequence length 16,384) with a combined KL-distillation and scoring loss.

Key results:

• On Qwen3-4B, KVpop retains 95% of dense-attention (teacher) performance at 75% KV cache compression and 94% at 88% compression (98% and 97% per the abstract's rounding); on Qwen3-8B it retains 95% and 99% at the same compression ratios.
• At 88% compression, DMS trails KVpop by 8–16 points in relative performance on AIME/HMMT pass@1, while training-free StreamingLLM and TOVA collapse (e.g., StreamingLLM relative score of 0.19 on Qwen3-8B).
• On out-of-domain GPQA Diamond and LiveCodeBench v6 (Qwen3-4B), KVpop stays close to the dense teacher (teacher: .59 GPQA-D / .35 LCB; KVpop: .57/.33 at 75% CR and .56/.34 at 88% CR) despite being trained only on math data.
• Peak allocated VRAM for dense attention grows from 18GB at 16k tokens to 36GB at 131k tokens, whereas both DMS and KVpop grow only 19% to 19GB at 75% compression; KVpop also sustains higher decoding throughput than DMS at long generation lengths.
• Delayed mLSTM scoring improves token accuracy by 0.2 points over immediate (non-delayed) scoring after 2,000 training steps.
• KVpop's learned eviction policy achieves a global mean recall of 81% agreement with the full-attention teacher's top-k retention decisions across heads and layers of Qwen3-4B at 75% compression.

Why it matters / caveats: KVpop shows that supervising KV eviction directly with a future-attention target (rather than proxy heuristics or differentiable relaxations like DMS) can nearly close the gap to dense-attention quality under aggressive cache compression while adding no inference-time overhead, and its uniform per-head budget makes it more GPU-efficient to execute than DMS's ragged, dynamically-gated caches. The authors note it is designed as a post-training retrofit for standard dense-attention Transformers rather than a from-scratch compressed-cache architecture, and leave broader exploration of alternative stateful scorers and hybrid dense-sparse layers to future work.

Problem: Multi-vector vision-language retrievers like ColPali and ColQwen2 store hundreds to over a thousand image-side patch token embeddings per image for MaxSim late interaction, making storage and scoring expensive, and existing token compression methods can remove or collapse object- and region-level evidence that future query tokens need to select.

Method: The paper proposes SaMer (Semantic-aware Merging), which compresses post-projector image-side tokens into K representative centroids via feature-spatial soft assignment (combining feature similarity and spatial coherence between token and centroid) while keeping the original MaxSim late-interaction interface unchanged. During training only, ground-truth object bounding-box annotations (from Flickr30K-Entities) are used as a merge prior—via a stop-gradient hard assignment that estimates each centroid's dominant bbox label and penalizes cross-instance mixing—to shape the soft assignment weights, without requiring bboxes or detectors at inference. SaMer uses projection-only adaptation, fine-tuning only the shared image-text projection layer (frozen vision encoder and language backbone) for 3 epochs with a multi-positive InfoNCE retrieval loss over the compressed MaxSim scores, and is applied on top of ColPali and ColQwen2 backbones with K=64.

Key results:

• At K=64, SaMer removes over 93% of image-side tokens (93.79% compression ratio) and reduces ColPali image-side storage by 16.09x.
• On Flickr30K, SaMer improves ColPali R@1 from 77.0 to 82.4 and ColQwen2 R@1 from 73.6 to 79.3.
• On MSCOCO, SaMer improves ColPali R@1 from 47.4 to 51.6.
• On ImageCoDe, SaMer improves ColPali R@1 from 5.4 to 5.9 and nDCG@10 from 13.2 to 14.4.
• Under matched K=64 adaptation, SaMer outperforms compression baselines H-Pool (Flickr30K R@1 73.7→79.2), SAP (80.2), and HPC (67.7) — SaMer reaches 82.4 R@1 on ColPali/Flickr30K.
• Grounding metrics improve substantially: training-free SaMer raises RegionHit from 10.5 (full ColPali) to 62.8 and CoverageIoU from 2.1 to 11.7; with object-aware adaptation, BoxMass rises from 41.3 to 54.2 (surpassing the full-token baseline of 51.8) and CoverageIoU reaches 16.4.
• Efficiency: for ColPali, SaMer reduces MaxSim operations by 16.1x and increases query throughput (QPS) by 4.3x on Flickr30K and 9.1x on MSCOCO; for ColQwen2, storage drops 3.73x (Flickr30K) and 1.99x (MSCOCO), with QPS gains of 1.9x and 3.1x respectively.
• Ablation: adding spatial coherence alone barely changes Flickr30K R@1 (80.7→80.4), while adding the object-aware merge prior raises it to 82.4 (and MSCOCO R@1 to 51.6), showing the object-aware component drives the main gain.

Why it matters / caveats: SaMer shows that efficient multi-vector retrieval compression should preserve query-selectable object evidence rather than just minimizing token count, achieving large storage/latency savings while improving accuracy on natural-image and compositional retrieval; however, it is not optimized for document-image retrieval (DocVQA), where sparse OCR/layout evidence rather than object-centric evidence matters, and it only remains "competitive" (not superior) there.

Problem: Post-training diffusion language models (dLLMs) for reasoning is hard because supervised fine-tuning is off-policy with exposure bias, reinforcement learning gives only sparse sequence-level rewards without tractable sequence likelihoods, and porting on-policy self-distillation (OPSD) from autoregressive models fails because it needs an external, instance-specific privileged-information (PI) reference that collapses the student onto a weak PI-free consensus policy.

Method: The paper introduces dOPSD (diffusion On-Policy Self-Distillation), which sources the teacher's privileged information directly from the student's own denoising trajectory instead of an external reference solution: the student is scored at a genuine intermediate, still-masked decoding step s_k, while the same model acts as teacher by averaging its masked-position predictions over all later, more-decoded trajectory states s_t (t>k) where that position is still masked, producing a teacher target that is distilled into the student via a token-level generalized Jensen-Shannon divergence (forward KL as default, β→0). Only rollouts whose final answer matches the gold answer are kept (rollout verification), and the method is applied as a LoRA (rank 32, α=32) fine-tuning stage on Dream-7B-Instruct and LLaDA-8B-Instruct, trained on the MixChain-Z-PRM12K math corpus with AdamW at peak LR 2e-5 for 3 epochs, using a trajectory mask threshold τ=0.5.

Key results:

• On Dream-7B-Instruct, dOPSD raises GSM8K accuracy from 81.41 to 83.04 and MATH500 from 38.97 to 42.20, and is the only method to improve every benchmark.
• On LLaDA-8B-Instruct, dOPSD improves MATH500 by 4.76 points (31.24→36.00) and GSM8K by 1.64 points (71.23→72.87).
• Out-of-distribution code generation also improves despite training only on math: HumanEval rises 4.17 points on Dream (52.54→56.71) and 3.51 points on LLaDA (36.12→39.63).
• Naive OPSD ports fail: OPSD (full-solution) collapses GSM8K by 14.2 points on Dream (81.41→67.25) and 13.4 points on LLaDA (71.23→57.86), while OPSD (answer-only) stays near the base model.
• SFT degrades all four tasks on both backbones, and GRPO yields at most a marginal +1.53 gain (LLaDA MATH500).
• Ablations: forward KL beats reverse KL on every task-backbone cell (e.g., LLaDA HumanEval 37.42→39.63); averaging the teacher target over the full remaining trajectory beats capping the look-ahead at 200 steps (GSM8K 83.04 vs. 78.14, MBPP 58.49 vs. 48.92); τ=0.5 gives the best four-task average (60.11) versus τ=0.25 (53.63) and τ=0.75 (59.49); even without rollout verification, dOPSD still improves the four-task average from 57.59 to 58.64 (verification pushes it further to 60.11).
• A uniformly random mask over a finished rollout lands on already-decoded "easy" tokens for ~90% of masked positions at a 10% mask fraction, versus 0% for genuine on-policy decode-step masks, explaining why naive random-mask supervision (which steadily degrades GSM8K accuracy during training) fails while genuine decode-step scoring stays high and stable.

Why it matters / caveats: dOPSD shows that diffusion decoding intrinsically supplies a self-generated "peek-ahead" privileged signal, letting self-distillation work without external reference solutions or reward models and even without answer labels (unverified variant), which the paper argues is an advantage unique to diffusion models since autoregressive decoding lacks a comparable ladder of increasingly-informed views of one generation. Results are demonstrated only on two backbones (Dream-7B-Instruct, LLaDA-8B-Instruct) and one training corpus (MixChain-Z-PRM12K), with LoRA-based fine-tuning rather than full fine-tuning.

Problem: Existing world models are almost all single-agent, treating other agents as part of the environment, so there is no world model that conditions on and stays coherent under the simultaneous actions of multiple players in a highly dynamic, physically-interactive environment.

Method: MIRA is a 5-billion-parameter multiplayer world model for 2v2 Rocket League that runs as a flow-matching diffusion transformer with factorized space-time attention, predicting causally in the latent space of a 600M-parameter video representation codec built on a frozen DINOv3-L feature extractor plus a trained linear bottleneck (2×2 spatial and 2× temporal downsampling, 1024→32 channels) and a causal space-time ViT decoder. The four players' latent views are tiled into a single grid with spatial attention spanning all views, and each player's actions are embedded and injected via AdaLN conditioning, with action dropout letting the model act as a policy for uncontrolled players. Training uses diffusion forcing (independent per-frame flow-matching noise levels) for long-horizon stability, followed by progressive self-distillation for few-step sampling, and a two-stage recipe that warm-starts the multiplayer model from a single-player checkpoint trained on the same codec.

Key results:

• The 5B-parameter model generates all four players' views in real time at 20 frames per second on a single Nvidia B200 GPU, with one full world-model-update-plus-decode step taking ~70ms (~35ms per video frame), inside the 50ms budget for 20fps.
• Trained on ≈10,000 match-hours of bot self-play data (82,983 matches, 331,932 per-view recordings) across 3 arenas.
• Although trained on 4-second clips, rollouts hold steady distributional quality out to 5 minutes (the longest horizon measured), with rollouts observed continuing for hours in practice without collapse.
• Latent-space modeling beats pixel-space world models by roughly an order of magnitude at matched training budget: gFID 10.7 (latent) vs. 104.9 (plain pixel) and 81.0 (pixel with JiT recipe); action recoverability ratio (ARR) 0.91 (latent) vs. 0.61 and 0.49 (pixel variants).
• A frozen pretrained DINOv3-L feature extractor yields far more stable long rollouts than training the extractor from scratch: over a 5-minute rollout, gFID drift is 1.3× (from-scratch + DINO distillation) and 1.7× (from-scratch, no distillation) relative to the pretrained-extractor baseline.
• The action probe used to compute ARR reaches 0.84 mean average precision (mAP) across nine controls on real held-out video, validating it as a controllability metric.
• The codec compresses video by about 192× in value count via its 2×2×2 spatial-temporal bottleneck while emitting latents at 10Hz from 20fps video.

Why it matters / caveats: This is presented as the first interactive multiplayer world model for a highly dynamic, physically-grounded environment, enabling directions like self-play reinforcement learning, multi-agent training, counterfactual policy evaluation, and human-in-the-loop interaction that single-agent world models cannot support. A key caveat is that all training data comes from four independent instances of the same bot policy (Nexto) controlling the cars rather than diverse human or multi-agent play, and the environment is restricted to three fixed arenas with largely deterministic dynamics, limiting behavioral and environmental diversity.

Problem: Flow-matching generative models produce high-quality images, video, and edits only after dozens of sampling steps (35–50), and existing acceleration methods (distillation, consistency training) require teacher models or auxiliary networks and are prone to mode collapse and artifacts.

Method: Perceptual Flow Matching (PFM) keeps the standard flow-matching training pipeline but changes only the supervision space: instead of regressing the predicted velocity against the ground-truth velocity in VAE latent space, PFM recovers the clean sample estimate x̂0 from the predicted velocity, decodes both x̂0 and x0 to pixel space with the VAE decoder, and applies a perceptual distance loss (via pretrained perceptual models such as VGG, DINOv2, SigLIP, or ConvNeXt) between their decoded feature representations. It requires no teacher model or distillation, uses standard consistency sampling at inference, and optionally bakes classifier-free guidance into training via prediction-side (blending conditional/unconditional velocities before the loss) or target-side (constructing a CFG-enhanced x0 target) strategies, avoiding the extra unconditional forward pass at inference.

Key results:

• On COCO 2014 val with SD3-Medium at 8-step generation, PFM achieves FID 33.93, CLIP score 31.70, and HPSv3 11.42, outperforming LCM (FID 34.18, CLIP 30.42, HPSv3 9.19) and DMD2 (FID 36.16, CLIP 31.54, HPSv3 10.19).
• At 4 steps, PFM reaches CLIP 31.89 and HPSv3 8.50, versus LCM's 30.62/6.68 and DMD2's 31.24/8.28.
• On MagicBrush image editing with Qwen-Image-Edit, PFM at 8 steps (no CFG) surpasses the 40-step original model on CLIP-I (0.9402 vs 0.8957), DINO (0.9187 vs 0.8312), L1 (0.0535 vs 0.1060), and L2 (0.0156 vs 0.0428), with only a 0.01 drop in CLIP-T.
• On VBench text-to-video with Wan2.1-1.3B, PFM at 8 steps achieves a total score of 0.792, exceeding both the 8-step baseline (0.735) and the 35-step baseline (0.774).
• The measured off-manifold penalty Rϕ is 1.594 for VAE latent space versus 1.9–2.03 for pretrained perceptual spaces (VGG, DINO, SigLIP, ConvNeXt, CLIP), and this ranking correlates with downstream 8-step generation performance (Table 6).
• A randomly initialized ViT (RandViT) perceptual model yields degraded results (CLIP 20.94, HPSv3 -9.49), confirming that learned (not merely nonlinear) representations are necessary.

Why it matters / caveats: PFM shows that standard flow-matching training, without any teacher model or distillation procedure, can become a strong few-step generator simply by changing the supervision space, reducing sampling steps from 35–50 to 4–8 while producing fewer artifacts than distillation-based baselines. Limitations include degraded performance below two sampling steps, and the paper does not identify which perceptual space is theoretically optimal, leaving stronger one-step generation to future work.

Problem: Scientific literature search agents typically use fixed retrieval pipelines or implicit language-only reasoning, making it hard for them to translate multi-turn, evolving, underspecified user feedback into concrete, controllable retrieval behavior.

Method: PAPERPILOT frames literature search as workflow induction: given an anchor paper and user query, the agent composes an executable DAG (PAPERPILOT-TOOLSET) of typed paper-search operators (keyword search, citation expansion, filtering, scoring, reranking, NLI filtering, evidence extraction) and refines the DAG via edits triggered by user feedback across turns. PAPERPILOT-9B is built from Qwen3.5-9B using a two-phase pipeline: supervised fine-tuning on 5,540 workflow trajectories (distilled from a teacher model over 2,723 anchor-query cases across five search directions—predecessor, successor, sibling, benchmark, survey), followed by preference optimization (IPO-style DPO) on 1,733 hard chosen-vs-rejected pairs built by corrupting successful workflows with structural/semantic errors (invalid references, missing inputs, incorrect operators, dropped nodes, shifted filters, vague NLI axes).

Key results:

• Multi-turn: PAPERPILOT-9B improves over base Qwen3.5-9B toolset agent, raising Hit@5 from 58.0 to 77.0, MRR from 47.5 to 59.4, nDCG@10 from 26.8 to 32.5, and Recall@50 from 34.8 to 40.0.
• Workflow execution errors drop from 9.5% (base Qwen3.5-9B multi-turn) to 0% for PAPERPILOT-9B.
• Adaptive toolset vs. fixed workflow (GPT-5.4, single-turn): Hit@5 improves from 52.0 to 79.5 and MRR from 36.4 to 62.8.
• Best overall system, GPT-5.4 + Web Search + PAPERPILOT-TOOLSET (multi-turn), reaches 84.0 Hit@5, 87.0 Hit@10, 89.5 Hit@15, 56.8 Recall@50, 71.8 MRR, 41.6 nDCG@10, at $0.1508/case, vs. OpenAI DeepResearch's 72.0 Hit@5 at $6.0903/case.
• Workflow-generation ablation: PAPERPILOT-9B improves whole-workflow TF-IDF cosine similarity from 0.0470 (base Qwen3.5-9B) to 0.3619 and signature similarity from 0.0281 to 0.2238, at a relative efficiency score of 3.26 vs. GPT-5.4's 1.00 (cost $0.00108 vs. $0.12565 per case).
• Human study (5 sessions/participant): PAPERPILOT achieves 74.7% success rate, 2.4 top-1 distance, and 4.2 question satisfaction score, vs. GPT-5.4 (32.0% SR) and OpenAI DeepResearch (8.0% SR).
• Sensitivity analysis: retrieval quality peaks near candidate pool size K=8–10; Recall@50 drops from 0.411 (K=8) to 0.374 (K=20), and Hit@5 drops from 0.705 (K=10) to 0.660 (K=20).

Why it matters / caveats: Explicit, editable DAG-structured workflows give a controllable and inspectable interface for turning multi-turn user feedback into concrete retrieval changes, and a 9B model trained this way can match or exceed much larger/proprietary systems at far lower cost. Limitations noted by the authors: the operator library is predefined and may not cover all domain-specific search behaviors, training data inherits teacher-model biases, evaluation focuses mainly on computer science literature with hidden-gold benchmarks, and multi-turn evaluation relies on an LLM-based user simulator rather than real users.

Problem: Whether an AI agent's performance improvement from interacting with real-world environments after deployment follows a predictable scaling law, analogous to pretraining scaling laws, remains unknown because existing benchmarks lack the long horizons and rich feedback needed to measure such learning.

Method: The authors build EdgeBench, a suite of 134 real-world tasks across six capability families (Scientific Problems & ML, Systems & Software Engineering, Combinatorial Optimization, Professional Knowledge Work, Formal Math & Theorem Proving, and Interactive Games & Simulators), each supporting at least 12 hours of continuous agent operation via a dual-loop evaluation harness (a fast local agent-driven feedback loop plus a slower judge-mediated submission loop with host-side auto-evaluation snapshots). They evaluate five frontier agents (Claude Opus 4.8, GPT-5.5, GPT-5.4, GLM-5.1, DeepSeek-V4-Pro) over roughly 38,000 hours of environment interaction, fit a three-parameter log-sigmoid curve S(t) = Smax / (1 + (tmid/t)^β) to best-so-far performance versus elapsed time, and derive this functional form theoretically as a frontier-expansion process on latent task graphs (dx/d log t = βx(1-x)).

Key results:

• Averaged over all 134 tasks, the log-sigmoid curve fits each of the five models' 12-hour learning trajectories with R² ≥ 0.997 (mean R² = 0.998).
• The same log-sigmoid form holds at the task-family level (R² between 0.972 and 0.998 across six families) and at longer horizons: R² ≥ 0.993 for 28-hour fits (80 tasks, 4 models) and 72-hour fits (18 tasks, 2 models).
• Fitting only the first 6.5 hours and forecasting the remaining 6.5–12 hours achieves R² ≥ 0.997 and held-out RMSE below 1.0 performance point for all five models.
• The log-sigmoid family achieves the lowest RMSE (0.390) among compared S-curve families (log-probit 0.398, log-Gompertz 0.402, Weibull CDF 0.404), versus 0.717 for a log-linear baseline.
• Agent learning speed (2-hour performance gain on a fixed 18-task slice) increased roughly 8x from GPT-5-Codex (Sept 2025) to GPT-5.5 (April 2026) over 221 days, corresponding to an approximate doubling every three months.
• On the 12-hour leaderboard, Claude Opus 4.8 leads with an overall score of 51.3, followed by GPT-5.5 (48.4), GPT-5.4 (39.3), GLM-5.1 (37.4), and DeepSeek-V4-Pro (31.0).
• Accumulated experience beats independent restarts: on 17 tasks with Opus 4.8, a continuous 12-hour run reaches 43.0 versus 36.1 for six independent 2-hour restarts (a gain of +6.9).
• A 1M-token context window gives Opus 4.8 a consistent advantage over a 200k window on a 42-task subset, +5.8 points at 2 hours narrowing to +4.4 points at 12 hours.
• Recorded human expert effort on EdgeBench tasks averages 57.2 hours per task, reaching up to 320 hours.

Why it matters / caveats: The results provide the first evidence that in-deployment environment learning, like pretraining, obeys a precise and predictable scaling law, suggesting learning speed itself is a measurable and rapidly improving frontier-model capability; the authors note the log-sigmoid law is an emergent, population-level regularity that only appears when averaging over many diverse tasks (individual task trajectories are noisy and idiosyncratic) and may fail for environments with strong bottlenecks, dispersed midpoints, or non-scale-free task-graph structures.

Problem: Standard LLM-as-a-judge verifiers collapse the scoring distribution into a single discrete token, producing coarse, tie-prone judgments that fail to reliably distinguish correct from incorrect agent trajectories and thus cannot unlock the large oracle-verifier headroom (e.g., 98.9% Pass@K on Terminal-Bench V2) available in generation.

Method: LLM-as-a-Verifier computes a continuous reward by taking the expectation over the full probability distribution of scoring-token logits (extracted from <score_A>/<score_B> tags on a 1–20 letter-based scale) rather than decoding a single discrete score, and converts pairwise reward differences into preference probabilities via a Bradley–Terry model. It scales verification along three axes—score-token granularity (G), repeated evaluations (K), and criteria decomposition into sub-criteria such as Specification/Output/Errors (C)—and introduces a Probabilistic Pivot Tournament (PPT) algorithm that selects a best-of-N candidate using ring-pass scoring plus pivot-based tournament comparisons, reducing verification cost from O(N²) to O(Nk²). The framework is training-free, using Gemini 2.5 Flash (or Qwen 3.6 35B for video) as the verifier backbone, and is deployed as "TurboAgent," a proxy extension for Claude Code/Codex.

Key results:

• Verification accuracy on Terminal-Bench V2 rises from 73.1% (G=1) to 77.5% (G=20) with granularity scaling, from 74.7% (K=1) to 77.5% (K=16) with repeated evaluation, and from ~75.2–76.4% (single criterion) to 78.3% with 3-criteria ensembling.
• Discrete LM judges produce a 27% tie rate on Terminal-Bench comparisons vs. 0% ties for the continuous verifier at K=16 (tie rate drops from 26.7% at K=1 to 5.5% at K=16 for the judge).
• Case study (query-optimize task): discrete 1–5 judge ties in 88/100 runs; continuous G=5 verifier ties 0/100 and ranks correctly in 69/100; continuous G=20 verifier ranks correctly in 77/100.
• State-of-the-art results: Terminal-Bench V2 86.5% (vs. Pass@1 83.1%, oracle 92.1%), SWE-Bench Verified 78.2% (vs. Pass@1 76.1%, oracle 84.4%), MedAgentBench 73.3% (vs. Pass@1 70.2%, oracle 75.0%).
• RoboRewardBench preference accuracy: 87.4% for LLM-as-a-Verifier vs. RoboReward-8B 81.4%, Robometer-4B 78.8%, TOPReward 74.7%, and discrete LLM-as-a-Judge 70.8%; MAE vs. human annotations drops from 1.11 to 0.72.
• Value-Order Correlation (task-progress proxy): 0.966 on RoboRewardBench (vs. RoboReward-8B 0.877, Robometer-4B 0.780, TOPReward 0.565); on Terminal-Bench V2, VOC is 0.848 for successful trajectories vs. 0.769 for failed ones (0.079 gap).
• As a dense RL reward, it gives ≈1.8× higher sample efficiency (and higher final success rate, 0.76 vs. 0.69) fine-tuning a π0 policy with DSRL-SAC on LIBERO, and ≈1.1× higher sample efficiency fine-tuning Qwen3-8B with GRPO on MATH.

Why it matters / caveats: By reframing verification as a probabilistic, scalable axis rather than a fixed judging step, the framework provides training-free, general-purpose gains across coding, robotics, and medical domains and can double as a progress monitor and dense RL reward signal; however, it relies on access to token-level logprobs (with only an appendix workaround for logprob-free frontier models) and its benefits (e.g., criteria decomposition) were demonstrated primarily on coding tasks.

Problem: Existing video diffusion models can generate either long single-view videos via temporal autoregression or short multi-view videos via bidirectional attention, but no method can generate long, temporally unbounded, 3D-consistent videos across an arbitrary number of viewpoints.

Method: MV-Forcing distills the bidirectional SynCamMaster (built on Wan2.1-T2V-1.3B) into a causal few-step student that composes temporal autoregression (causal attention over blockwise-masked latent frames, block size K=3) with view-sequential autoregression via SynCamMaster's multi-view synchronization (MVS) module. A recurrent 4D reconstruction model, CUT3R, maintains a persistent scene state updated from each decoded generated view and is queried via raymaps to render a geometric prior (RGB + confidence map) for the next viewpoint, injected into the DiT through a zero-initialized Conv3d layer (ControlNet-style). Training uses a joint denoising regime (both view slots start from noise with probability p=0.3) plus Distribution Matching Distillation (DMD) with "Spatio-Temporal Self-Forcing" — unrolling autoregressive generation along both the temporal and view axes and applying the DMD loss to self-generated outputs — to close train/inference exposure-bias gaps; for real-world data the teacher is swapped to a finetuned ReCamMaster.

Key results:

• Short-sequence setting (2 views, 81 frames): MV-Forcing beats the bidirectional SynCamMaster teacher on camera accuracy (RotErr 3.64 vs 3.83, TransErr 8.26 vs 8.83) and view synchronization (Mat. Pix. 251.13K vs 236.72K, CLIP-V 91.81 vs 90.13), though FID/FVD are slightly worse (167.90 vs 166.57 FID).
• Long-sequence real-world setting (3 views, 162 frames): MV-Forcing achieves FID 153.27 and FVD 1309.68 vs. 156.78/1363.23 for the best baseline (SF+ReCamMaster+SF), with lower RotErr (3.88 vs 4.89) and higher Mat. Pix. (239.37K vs 127.48K).
• Long-sequence synthetic setting: MV-Forcing attains FID 186.73 vs. 191.26 (SF(ft)+ReCamMaster+SF) and TransErr 8.41 vs. 10.73.
• Ablations (3 views, 162 frames): removing view-sequential unrolling causes the largest degradation (FID rises from 186.73 to 212.89, Mat. Pix. drops from 243.63K to 159.73K); removing CUT3R is the second-largest drop (FID 199.10, Mat. Pix. 173.75K).
• View scaling (81 frames): metrics stay nearly stable from 2 to 5 views (Mat. Pix. only drops from 251.13K to 250.95K, CLIP-V from 91.81 to 91.73).
• Temporal scaling (2 views): across an 8× increase from 81 to 648 frames, Mat. Pix. stays nearly constant (251.13K to 249.14K) while CLIP-F declines most (99.21 to 96.23).
• Training used 16 NVIDIA A100 65GB GPUs; ODE initialization stage took ~16 hours (3,000 iterations) and DMD distillation took ~26 hours (1,600 iterations); inference for one 81-frame view (4-step denoising) takes ~20 seconds plus ~30 seconds per view transition for the CUT3R update/query, versus ~6 minutes for SynCamMaster to generate 2 views at 81 frames.

Why it matters / caveats: MV-Forcing is presented as the first framework for long, geometrically consistent multi-view video generation at arbitrary viewpoint counts and unbounded temporal horizons using a single few-step model, avoiding the quadratic scaling of bidirectional attention. Limitations include reliance mainly on synthetic training data (SynCamVideo) with only minimal real-world finetuning, residual quality degradation over very long temporal horizons despite Self-Forcing, and DMD supervision from a teacher that only ever sees two simultaneous views, so consistency across more than two views at once is never directly supervised.

Problem: Existing safety testing for LLM agents relies on expert-curated scenarios and hard-coded verification rules that conflate stated intent with realized harm, making evaluation costly to extend as agents and their tool ecosystems rapidly evolve.

Method: VERA is a three-stage, self-reinforcing pipeline: (1) a Summary Agent performs literature-driven continuous risk exploration over arXiv/OpenReview papers, using create/update/merge/delete operations to build three hierarchical taxonomies (risks, attack methods, tool execution environments); (2) an LLM-based goal composer combines taxonomy leaf triples into concrete safety goals, which are compiled into executable safety cases σ=⟨g, s0, Vg⟩ with an LLM-initialized environment state and a deterministic Python verifier, then expanded into benign, single-channel, and multi-channel variants; (3) adaptive execution runs each case in an isolated Docker Compose sandbox (12 containers, including Mailpit, Gitea, Blnk, Databag, SearXNG behind an MCP-based configurable tool gateway) where a Control Agent conducts an observe-adapt-act loop over up to 10 turns, and an evidence-grounded verifier judges outcomes using a priority order of environment state > tool-call records > agent response (Equation 7). The released VERA-Bench dataset and a fine-tuned Qwen3Guard classifier are downstream artifacts of this pipeline.

Key results:

• Risk exploration processed ~800 papers to build taxonomies of 124 risk categories, 77 attack methods, and 30 environment categories, generating 39,078 candidate safety goals and retaining 1,600 executable base scenarios (VERA-Bench).
• Across four agent frameworks (OpenClaw, Hermes, Codex, Claude Code), overall execution success rate (ESR) was 90.6% under single-channel attacks and 93.9% under multi-channel attacks, versus 70.5% in the benign setting.
• The adaptive Control Agent raised ESR by 20.1 percentage points on average versus benign (largest gain +24.8 points for OpenClaw, smallest +15.1 for Claude Code), isolating the value of runtime-adaptive multi-turn steering.
• The multi-channel tool-gateway injection added a further +3.3 points on average over single-channel (ranging from -2.1 for Claude Code to +6.3 for OpenClaw), with Hermes reaching the highest multi-channel ASR at 97.8%.
• Per-agent overall ESR: Claude Code 88.6%, Hermes 86.6%, Codex 84.1%, OpenClaw 70.3%.
• All 80 cells in the risk×environment table and all 110 cells in the environment×attack-method table exceeded 40% ESR, validating that the combinatorial case-construction pipeline produces broadly executable, non-degenerate scenarios.
• Off-the-shelf guard models transferred poorly to VERA-Bench (e.g., LlamaGuard3: 0.438 accuracy/0.310 F1; base Qwen3Guard: 0.670/0.637); fine-tuning Qwen3Guard on VERA-derived data raised performance to 0.930 accuracy, 0.903 recall, 0.941 F1 (+26.0/+43.5/+30.4 points respectively), and this fine-tuned model also improved out-of-distribution accuracy on R-Judge to 61.7% versus 53.7-59.4% for baselines.
• Median execution used 155k input tokens, 3k output tokens, and 11 tool calls (95th percentile: 789k input tokens, 11k output tokens, 38 tool calls).

Why it matters / caveats: The results reveal a "capability-vulnerability alignment": agents with stronger tool orchestration and longer context (Claude Code, Hermes) are also more susceptible to adaptive multi-turn manipulation, suggesting safety and capability are structurally entangled rather than independent axes to optimize. The authors also note that poor off-the-shelf guard-model transfer indicates current moderation classifiers underestimate risks in richly environment-grounded, tool-mediated agent settings.

Problem: LLM agents for repository-scale vulnerability reproduction can execute low-level actions (navigating code, running commands, submitting proof-of-concepts) but frequently fail because they choose the wrong investigation strategy, and existing planner-training or memory-based agent frameworks address only part of this strategy-selection bottleneck.

Method: Mastermind is a dual-loop Curator–Planner–Executor–Verifier framework that separates transferable strategy learning from task-specific experience: a trainable Planner (Qwen3.6-35B-A3B, trained/served via Tinker) emits a compact natural-language "Strategy" (capped at 2,000 tokens) that guides a frozen Executor (GPT-5.4 mini, GPT-5.5, or GLM 5.1) to produce a PoC rollout, which CyberGym's Verifier scores against an 8-point milestone schedule (m0–m7). The Planner is warm-started with supervised fine-tuning on stepwise Claude Code/Codex trajectories, then optimized with milestone-based Group Relative Policy Optimization (GRPO) using 16 candidate strategies per task split into two 8-way slot-conditioned groups (e.g., minimal-reproducer, parser-format, bounds-allocation) for diversity, with a reward combining a log-compressed milestone term, a bounded strategy-length term, and a status penalty. A separate Curator maintains task-local experience (file locations, parser behavior, failed hypotheses) across up to N=8 sequential attempts per task, keeping volatile facts out of Planner weights.

Key results:

• On CyberGym (260 training tasks, 200 held-out evaluation tasks), with GPT-5.5 as frozen executor Mastermind reaches 84.5% (169/200) milestone-7 pass rate, versus 63.0% for Best-of-8, 70.5% for PAGENT-guided Best-of-8, 77.0% for iterative task-local improvement, and 60.0% for Level-3 open-book (ground-truth) context.
• The same Planner (trained only on GPT-5.4 mini trajectories) transfers without retraining: GPT-5.4 mini improves from 45.0% (Base Planner) to 60.0%, and GLM 5.1 improves from 58.5% to 71.0%.
• Mastermind also uses fewer rollouts: 560 for GPT-5.5 vs. 753 (iterative improvement) and 1,600 (Best-of-8), and reduces estimated cost per pass to ~$8 vs. ~$29 (Best-of-8) and ~$11 (iterative) for GPT-5.5.
• Strategy-sensitivity ablation under a fixed GPT-5.4 mini executor: Soft Oracle strategy reaches 39.5% pass rate vs. 23.5% for Null Strategy, 24.0% for Zero-shot Planner, and only 32.0% for Hard Oracle (ground-truth solution given directly), showing strategy quality causally affects outcomes.
• Dual-loop ablation on GPT-5.5: removing the policy loop (keeping only experience) drops pass rate from 84.5% to 77.0% (154/200); removing the entire dual loop (plain Best-of-8) drops it to 63.0% (126/200); for GPT-5.4 mini, removing the experience loop drops the trained-planner result to 45.5% (91/200).
• Independent Best-of-N sampling saturates: GPT-5.5 rises from 23.5% (attempt 1) to 63.0% (attempt 8), with most gains front-loaded in early attempts.

Why it matters / caveats: The results support treating compact, natural-language "strategy" rather than full action trajectories as the learning unit for long-horizon software-engineering agents, enabling GRPO training that is far cheaper than optimizing over full executor trajectories and that transfers across different frozen executor backbones without retraining. The authors note safety mitigations (sandboxed CyberGym execution, high training cost, no new execution capability created) but acknowledge the framework amplifies vulnerability-discovery capability and that three tasks may reflect CyberGym verifier granularity issues rather than true failures.

Problem: LLMs used for long-horizon behavior prediction (inferring a user's next action from a lengthy historical sequence) struggle to reliably induce latent behavioral patterns from thousands of scattered records and suffer from model-intrinsic cognitive biases, both of which prior context-compression memory methods fail to address.

Method: PraMem is a training-free framework that, instead of compressing the historical sequence, treats it as practice data: it iteratively samples segments as practice samples with ground-truth labels, prompts the LLM to make an explicit deep-thinking prediction under the current experiential memory (existing experience trial), then reflects on the prediction-label comparison to generate revise/prune/supplement proposals for two memory types — pattern experience (user behavioral patterns) and bias-alert experience (LLM cognitive biases). A self-review mechanism filters proposals via a groundedness review (perturbing the sample's history to see if the proposal still holds) and a generalizability review (rewriting the scene into a multiple-choice task to check the proposal isn't overly specific); only proposals passing both enter a proposal pool, and every T=5 rounds a consensus-driven adjustment step applies only operations supported by multiple proposals, over a total of 80 practice rounds using GPT-OSS-120B as the memory-construction LLM.

Key results:

• On OmniBehavior (GPT-OSS-120B backbone), PraMem improves overall ACC from 73.5 (Truncation baseline) to 84.7 and F1 from 24.7 to 31.6.
• PraMem outperforms the best prior advanced-memory baseline ProEx (77.7 ACC / 26.9 F1) by 7.0 ACC / 4.7 F1 points overall.
• On Qwen3.5-35B-A3B backbone, PraMem reaches 75.6 ACC / 25.3 F1 overall, versus 63.8 ACC / 20.9 F1 for Truncation and 67.6 ACC / 21.5 F1 for ProEx, showing memory built with GPT-OSS-120B transfers to another backbone.
• On MovieLens-1M, PraMem achieves F1=49.6 on the confusion-matrix diagonal-dominance metric, versus 38.9 (Long-context), 28.8 (Truncation), 28.1 (Mem0), and 27.3 (MemOS).
• Ablations (F1, OmniBehavior): removing deep-thinking trial drops overall F1 from 31.6 to 29.3, removing reflective proposal drops it to 30.9, and removing consensual adjustment drops it to 28.8.
• Experiential memory improves monotonically with practice: 79.0 ACC/28.3 F1 at round 10, 83.5 ACC/30.7 F1 at round 40, 84.7 ACC/31.6 F1 at round 80, while memory token length grows then stabilizes after round 60.

Why it matters / caveats: PraMem shows that reframing long historical sequences as a resource for offline "practice" rather than a burden to compress yields large accuracy gains and memory that transfers across LLM backbones without reconstruction. The main stated limitation is added memory-construction cost: the iterative practice process takes more time than conventional extraction/retrieval-based memory methods, though this cost is offline and does not affect prediction-time efficiency.

Problem: Existing cross-region LLM serving load-balancers only account for a subset of KV-cache locality, replica load, and network latency, causing uneven load/KV-cache concentration and suboptimal TTFT.

Method: GORGO is a proxy architecture that computes a per-replica TTFT cost combining network RTT (measured via EWMA ping), queueing delay (proxy-tracked in-flight token count), and prefill cost (uncached tokens), weighted by tunable parameters W_rtt and W_queue (W_prefill fixed to 1). A (1+1) evolutionary strategy with multiplicative log-space perturbation and Rechenberg's 1/5 success rule tunes these weights online to minimize rolling p95 TTFT. The authors also release ART-Chat-2.5M, a synthetic long-context, high-prefix-reuse dataset built in Mooncake's FAST'25 trace format from a week of production metadata, and evaluate against baselines (least-load, least-request, prefix-cache, simple-session-affinity) using SGLang engines serving Qwen3.5-35B-A3B (FP8) across us-ashburn, eu-frankfurt, and ap-seoul regions.

Key results:

• ART-Chat-2.5M has 89.4% intra-user and 89.7% global prefix reuse vs. WildChat-4.8M's 4.7%/32.5% and LMSYS-Chat-1M's —/3.4%, with 17,964 avg input tokens (6x WildChat's 2,925).
• On held-out evaluation windows, GORGO improved p95 TTFT by 6.9–15.5% and p95 E2E latency by 14.3–30.9% over baselines such as simple session affinity.
• Tuning converged after 672 samples (18th evolutionary step), learning w_rtt=0.276, w_queue=0.5 from initial values w_rtt=0.5, w_queue=0.1, reaching a best rolling-window p95 TTFT of 1.276s.
• In a load sweep, metrics improved over 3x for unsaturated policies once time_scale=3.0 was reached, and GORGO broke its saturation point earlier than other policies (e.g., at time_scale=2.0, GORGO was unsaturated with 1,822ms TTFT p95/5.65s E2E p95, while simple-session-affinity and least-request remained saturated).
• A reward-hacking case (Appendix D) showed GORGO could learn w_queue=0 and route ~100% of traffic to the nearest replica, achieving TTFT 17% better than the next-best policy but inflating E2E p95 to 12.49–12.6s.
• On the low-reuse WildChat-4.8M replay, GORGO was non-competitive against simpler baselines (e.g., gorgo-fixed TTFT p95 of 1.932s vs. simple-session-affinity's 1.025s).

Why it matters / caveats: GORGO shows that jointly modeling and online-tuning network latency, prefill cost, and queueing delay can substantially cut tail latency in long-context, high-prefix-reuse cross-region serving, but its advantage is workload-dependent (ineffective on short-prompt, low-reuse traffic) and the p95-TTFT-only objective can be gamed by continuous batching into concentrating load on one replica, trading away E2E and ITL tail latency; evaluation is also limited to a single production trace and a homogeneous three-region, two-GPU-per-replica fleet.

Problem: Best-of-N sampling pipelines for chest X-ray report generation score candidate reports in isolation against single-image reference reports, ignoring that radiologists actually write reports around the clinical change from a patient's prior exam to the current one.

Method: The paper introduces transition-aware best-of-N sampling: each report's Findings and Impression sections are split into sentences and embedded with a frozen all-mpnet-base-v2 sentence transformer (d=768) into unordered embedding sets; a directional set-to-set distance then reduces each (prior, current) report pair into a single fixed-dimensional transition vector. Four directional set distances are instantiated—Mean-Shift, Novelty (forward-only Chamfer residual), Dir-Hausdorff (directed-Hausdorff anchor displacement), and Cost-OT (cost-weighted optimal-transport displacement)—and a bank of ground-truth training transition vectors is cached once per corpus. At test time, each of K candidate reports is converted into its own transition vector and scored by cosine distance to the bank, aggregated via either min or kNN, with Findings and Impression distances summed to select the best candidate; the whole scheme is training-free and requires no changes to the underlying generator.

Key results:

• Evaluated on a multi-visit AP/PA cohort derived from ReXGradient-160K with 22,745 training transitions and 1,727 test transitions, across three VLMs (Gemini-2.5-Flash-Lite, Gemini-3.1-Flash-Lite-preview, Mistral-Small-2603) and three prompts (K=5 candidates).
• On Impressions, best configurations gave +8.5% ROUGE-L (Novelty/kNN), +13.6% RadGraph averaged-F1 (Mean-Shift/min), and +4.4% METEOR (Dir-Hausdorff/kNN) relative to random selection.
• On Findings, every method/aggregation combination beat random on every headline metric, with the strongest gains being +3.1% BERTScore-F1 and +5.1% RadGraph averaged-F1 (Cost-OT/min) and +2.2% CheXbert F-14 (Mean-Shift/kNN).
• Across 9 (model, prompt) configurations, transition-aware best-of-N beat the random baseline on every configuration for ROUGE-L and RadGraph averaged-F1 on Impressions, with the largest absolute gains under the weaker prompts P1 (zero-shot, image-only) and P2 (few-shot).

Why it matters / caveats: This is presented as the first training-free best-of-N scheme that explicitly conditions candidate scoring on the patient's prior exam, complementary to prior work that injects longitudinal context only at generation time; gains compress (but do not invert) under prompt P3, where the prior report is already given to the generator, and the authors note extensions like learned directional encoders and longer visit histories are still needed.

Problem: There is no diverse, large-scale real-world visuotactile dataset that lets researchers systematically compare 2D pixel-space and 3D geometric-space world models for predicting the dynamics of deformable objects.

Method: The authors build Deform360, captured with a rig of 41 surround-view calibrated cameras (720x1280, 30 FPS) and bimanual tactile-equipped UMI grippers, covering 198 daily-life objects (1D ropes/cables, 2D fabrics/garments, 3D volumetric plush/foam objects) across 1,980 interaction sequences. A markerless annotation pipeline reconstructs per-frame dynamic geometry with 3D Gaussian Splatting (segmented via SAM-style object masks), then tracks up to 1,600 grid points per view with CoTracker3, lifts 2D tracks into 3D via depth-based back-projection and RANSAC fusion, and refines the resulting particle trajectories by minimizing a physics-informed objective (Chamfer shape loss, ARAP local-rigidity loss, Laplacian smoothness loss, and a tactile-consistency loss from contact taxels) via gradient descent. Using the resulting particle trajectories, they benchmark 3D particle dynamics models (ParticleFormer, PGND, and the differentiable-simulation-based PhysTwin) against the action-conditioned 2D video model Cosmos-Predict2 (2B parameters, post-trained with 6D wrist-pose/gripper-openness conditioning) under per-episode, multi-episode, and multi-object generalization settings, and also train a transformer-based visual-to-tactile contact predictor and deploy PhysTwin-based dynamics in an MPC robot-planning pipeline.

Key results:

• Deform360 contains 198 objects, 1,980 interaction sequences, 23.3 million frames, and 215.7 hours of multi-view footage from 41 cameras (614,490 frames per viewpoint on average, ~10.34s per episode).
• Reconstruction quality (3DGS) averages PSNR 27.66 dB, SSIM 0.96, LPIPS 0.0708 across categories, with 3D volumetric objects highest at 30.00 dB PSNR.
• Visuotactile tracking cuts warped-point-cloud Chamfer error to 2.71x10^-5 m^2, a 5x reduction versus 1.41x10^-4 m^2 for vision-only tracking.
• Visual-to-tactile contact prediction reaches 88.67% mean accuracy (F1 = 0.8909) across 36 views, versus 50.31% random-guessing baseline.
• Per-episode setting: PhysTwin achieves the lowest reconstruction/prediction error (e.g., future-prediction CD 0.014, track error 0.025), outperforming learning-based PGND (CD 0.073) and ParticleFormer (CD 0.044).
• Multi-episode setting: Cosmos gets the best reconstruction PSNR (27.748 dB) but ParticleFormer wins future prediction (CD 0.051 vs. PGND's 0.130).
• Multi-object (zero-shot) setting: Cosmos generalizes best in image quality (PSNR 25.042 dB) versus ParticleFormer (23.312 dB) and PGND (22.049 dB, CD 0.429).

Why it matters / caveats: Deform360 is presented as an order-of-magnitude larger and more sensor-rich benchmark than prior deformable-object datasets, revealing a consistent trade-off where 3D particle models with structural priors excel in low-data regimes while 2D video models scale better with more data and generalize better zero-shot to novel objects, though Cosmos sometimes fails to follow commanded actions and is harder to use for reward design in planning. Limitations include reduced tracking quality under heavy self-occlusion, violated rigidity assumptions for highly plastic materials, and tactile sensors that only measure normal-axis pressure and cannot detect micro-slip.

Problem: Concept erasure methods must remove a target concept (e.g., gender, safety, sycophancy) from neural representations without damaging other, often correlated, information encoded in them (surgicality).

Method: The authors propose the Manifold Constraint Hypothesis (MCH): natural representations concentrate on a structured, lower-dimensional manifold, so erasure updates should be constrained to that manifold rather than moved arbitrarily in the full space. They instantiate this as MANCE (MANifold aware Concept Erasure), an iterative algorithm that trains a nonlinear MLP probe on the current representations (refit every τ=8 rounds), estimates a local tangent space at each representation via local PCA/SVD over its k nearest neighbors in the natural (unedited) representations X(0) — following LTSA (Zhang and Zha, 2004) with tangent rank r set by the TwoNN intrinsic-dimension estimator — projects the probe's gradient onto this tangent basis with spectral reweighting (exponent α=1), and takes the largest per-sample step allowed by a local-neighborhood radius cap (parameter ε, hard cap λmax). Two variants, MANCE+ and MANCE++, prepend closed-form linear preprocessing before the MANCE loop: MANCE+ prepends LEACE (removes first-moment/class-mean linear signal), and MANCE++ additionally prepends CovMatch, a rank-2 projection removing the leading second-moment class-conditional covariance asymmetry (a specialization of k-LEACE).

Key results:

• Applying MANCE on top of prior erasers on 39 NLP settings (13 LLMs × 3 concepts) reduces mean residual leakage |ΔS| at DY≤1pp from 19.1→1.5pp (LEACE), 15.2→1.8pp (INLP), and 11.5→1.6pp (IGBP), while maintaining or improving coverage.
• MANCE++ achieves state-of-the-art nonlinear erasure on NLP: leakage of +1.6pp at DY≤1pp down to 0.0pp at DY≤10pp, reaching chance on 19–35/39 settings, versus the strongest baseline Obliviator's +4.3→+2.7pp and chance on only 13–17/39 settings.
• On gender erasure specifically, MANCE++ reaches chance on 3/13 models at DY≤1pp (vs. 0/13 for Obliviator), widening to 12/13 vs. 0/13 at DY≤5pp.
• On CelebA vision settings (40 attributes), in the least-correlated control regime MANCE++ covers 39/40 and 40/40 attributes at DY≤1pp and DY≤3pp respectively, versus 15/40 and 29/40 for Obliviator; in the most-correlated regime it covers 19/40 and 34/40 versus 2/40 and 15/40 for Obliviator.
• The unconstrained ablation AmbCE++ (same probe loop and preprocessing, but full-space gradient steps of matched magnitude λ=29.31) leaves 6–10pp leakage versus MANCE++'s +1.6→0.0pp, isolating the manifold constraint as the source of the gains.
• MANCE adds roughly eight minutes of runtime (458.8–474.9s across variants on one NVIDIA B200) for the Gender concept, compared to seconds for LEACE/CovMatch and a few minutes for INLP, with ~50% of runtime attributed to per-round local SVDs.

Why it matters / caveats: The results provide empirical support for MCH as a general principle for representation interventions (potentially extending beyond erasure to activation steering), but MANCE relies on a local, first-order manifold estimate that degrades where natural representations are sparse or the manifold is strongly curved, and it is computationally heavier than closed-form one-shot linear erasers since it requires per-round k-NN queries and local SVDs plus periodic probe refitting.

Problem: Controllable generative models for 3D chest CT need to be simultaneously high-fidelity, natively 3D, and faithful to requested clinical conditioning, but standard likelihood-based training gives no guarantee that any single generated volume actually expresses its requested attributes.

Method: CONFLUX is a three-stage latent diffusion pipeline: a 3D convolutional VAE (downsampling factor f=8, latent channels C=16) compresses each CT volume into a normalized latent space using ℓ1 reconstruction loss, a KL regularizer, and a tri-planar LPIPS perceptual loss; a single-stream rectified-flow transformer (depth L=12, hidden width d=768, patch size p=2, 3D axial RoPE) is trained from scratch via flow matching to generate in that latent space, conditioned on an 18-dim finding vector plus sex, age, and reconstruction kernel through adaLN-zero modulation with classifier-free guidance dropout of 0.1. A third post-training stage applies Group-Relative Policy Optimization (adapted from Flow-GRPO) using a stochastic SDE sampler (to make ODE sampling tractable for policy gradients), a clipped PPO surrogate with KL penalty to a frozen reference policy, and a reward from a frozen findings classifier that scores decoded/un-normalized latents against requested conditioning.

Key results:

• Tri-planar 2D-FID of 32.3 for CONFLUX vs. 74.6 for MAISI and 145.4 for GenerateCT (VAE reconstruction ceiling is 22.6).
• Axial 2D-FID of 24.7 (CONFLUX) vs. 55.6 (MAISI) and 70.2 (GenerateCT).
• CONFLUX density/coverage of 0.056/0.105 vs. MAISI's 0.016/0.017 — described as 3.5x denser and 6x more manifold coverage than MAISI.
• Diversity (pairwise MS-SSIM) of 0.507 for CONFLUX, close to the real-data floor of 0.41.
• Reward/latent findings classifier reaches macro AUROC 0.793 over 18 labels on real validation data, outperforming CT-CLIP zero-shot on 17/18 findings (CT-CLIP zero-shot macro AUROC 0.731, CT-CLIP finetuned 0.756, CT-Net 0.629).
• RL post-training (GRPO) improves independent-judge faithfulness metrics: macro AP 0.330 → 0.344 and macro AUROC 0.684 → 0.699 (real-data ceiling AP 0.360, AUROC 0.746).
• Post-training recovers 47% of the AP headroom (24% of AUROC headroom) between pre-RL and real-scan reliability; gains are statistically significant via paired bootstrap (p=0.042 and p=0.014).
• Released dataset contains ~200,000 synthetic chest-CT volumes with paired conditioning metadata.

Why it matters / caveats: The work shows that RL post-training (GRPO) can directly improve conditioning faithfulness in a 3D medical latent diffusion model beyond what flow-matching alone achieves, and the authors state it is, to their knowledge, the first GRPO post-training of a 3D medical flow model; the faithfulness gains, while statistically significant, are numerically small, and the reward/faithfulness labels rely on NLP-extracted (not expert-annotated) findings, with baseline comparisons on synthesis quality not fully resolution-matched.

Problem: Crossmodal correspondences between sound and taste are well documented in psychology but have not been formalized as a content-based music information retrieval task, leaving no benchmark for predicting taste from audio or using predicted taste for music retrieval.

Method: The authors formalize taste-from-audio prediction as an MIR benchmark on a unified multi-source music-taste corpus (269 train / 68 val / 40 test clips, combining real-music, MusicGen-generated, and soundtrack subsets), predicting five normalised taste intensities (sweet, bitter, salty, sour, spicy) with a masked MSE loss. Ten frozen audio encoders spanning the four HEAR families (VGGish, PANNs, AST, HuBERT, MERT, Omar-RQ, CLAP, EnCodec, MULE, plus an MFCC DSP baseline) feed a shared two-layer MLP head (hidden size 256, ReLU, dropout 0.2, sigmoid output) trained with AdamW, and a gated late-fusion variant learns a per-encoder sigmoid gate over concatenated embeddings before the same MLP. Interpretability is probed via 5-fold ridge regressions from embeddings to nine psychoacoustic descriptors and a cross-encoder audio-bandstop (zero-phase 4th-order Butterworth) knockout analysis, and the predicted 5-D taste space is evaluated as a retrieval index against a CLAP-text baseline.

Key results:

• Best systems reach macro RMSE of 0.134 and MAE of 0.109 (single VGGish), versus the previous SOTA (five fine-tuned AST regressors) RMSE of 0.219.
• On held-out real music, model error (RMSE 0.13, MAE ~0.10) is less than half a single human rater's leave-one-out deviation from consensus (RMSE 0.28, MAE 0.227).
• Gated late-fusion (VGGish+MULE) raises macro Pearson r to 0.724 ± 0.020 versus 0.666 (VGGish alone) and 0.556 (SOTA), at the same RMSE (0.134).
• Ablation shows the frozen-encoder + masked-MSE + sigmoid redesign closes most of the RMSE gap alone (0.219 → 0.143), with fusion contributing the remainder (→0.134).
• On the 309-item retrieval pool, learned taste-space predictors saturate precision@k (P@5/P@10/P@20 = 1.0 for several systems) while a CLAP-text baseline collapses to chance (Spearman ρ = 0.122, AUC = 0.484 vs. 0.5 chance).
• AST+VGGish fusion is the best retrieval-ranking system (test Spearman ρ = 0.693), edging VGGish+MULE and the triple fusion (both 0.663) and SOTA (0.645).
• Ridge probes show CLAP best preserves spectral structure (R² between 0.78 and 0.93 across nine descriptors), though spectral decodability does not track taste-prediction accuracy (e.g., MERT and EnCodec decode spectra well but rank low on taste prediction).
• Bandstop knockout localizes taste-relevant frequency bands per encoder (e.g., AST bitter at 0–260 Hz, Δr −0.129; VGGish sweet at 260–610 Hz, Δr −0.060; MULE salty at 1.1–1.8 kHz, Δr −0.294), consistent with documented sound–taste correspondences.

Why it matters / caveats: The results show that simple loss/output redesign (masked MSE, sigmoid, single shared head) rather than architecture drives most of the improvement over prior work, and that a frozen-encoder taste predictor can already approximate group consensus better than an individual human rater, enabling a new content-based retrieval axis (e.g., "similar track but sweeter"). The evaluation is limited by a small corpus (269 train / 40 test clips, n=20 per source), favoring simple heads, and the underlying crossmodal literature is Western and sweet/bitter-centric, leaving cross-cultural generalization and the trigeminal spicy axis open questions.

Problem: Detect and characterize sexism in multimodal memes (binary sexism identification, source intention detection, and multi-label sexism categorization) while predicting the full distribution of annotator disagreement rather than a single majority label, as required by the EXIST 2026 Soft-Soft evaluation track.

Method: The system uses frozen Gemini Embedding 2 vision-language embeddings (768-dim) fed into a lightweight Gated MLP backbone built from stacked SwiGLU blocks (expansion factor 2, dropout 0.2), totaling only 3.5M trainable parameters, with three independent linear heads for the three EXIST 2026 subtasks. Each head is trained against empirical soft-label targets using KL divergence (binary KL per category for Task 2.3), combined via learned homoscedastic uncertainty weighting to dynamically balance the three task losses, plus conditional loss masking so Tasks 2.2/2.3 gradients only propagate for instances with downstream annotations. Hierarchical dependencies are enforced through probabilistic joint decoding (P(Subtask) = p̂(1)·p̂(sub)) and, as an ablation (Run 3), an explicit detached soft-gating mechanism that scales the shared representation by the stop-gradient sexism probability. Training used AdamW (lr 1e-4, weight decay 1e-2, batch size 8) with OneCycleLR (30% warm-up, cosine annealing), bf16-mixed precision, and early stopping on an 80/10/10 stratified split; three runs were submitted (two base-model seeds plus the gating variant).

Key results:

• Official Soft-Soft leaderboard: ranked 1st on Task 2.3 (ICM-Soft -2.8881, ICM-Soft Norm 0.3469), 4th on Task 2.1 (ICM-Soft 0.2323, Norm 0.5373), and 4th on Task 2.2 (ICM-Soft -0.6720, Norm 0.4285).
• Official Hard-Hard leaderboard (no threshold tuning): ranked 14th on Task 2.1 (ICM-Hard 0.2673), 12th on Task 2.2 (ICM-Hard 0.0020), and 8th on Task 2.3 (ICM-Hard -0.3953).
• Local test split: Run 2 achieved the best Task 2.3 soft result (ICM-S -3.2130, ICM-S Nr 0.3356) and Run 1 the best Task 2.2 result (ICM-S -3.6935, ICM-S Nr 0.3111); Run 3 (gating) performed comparably (e.g., Task 2.1 ICM-S 0.1900 vs. 0.1648–0.2025 for base runs).
• EEG separability analysis (PCA to 21 components, ~95% variance, MANOVA with 1,000-label permutation test): Task 2.2 reached statistical significance (p_perm = 0.0010) but with negligible effect size (Wilks' λ = 0.9946, partial η² < 0.01); Tasks 2.1 and 2.3 showed no significant linear separability, and no individual EEG channel survived Bonferroni correction (α = 6.25×10⁻⁴) on any task.
• Dataset comprised 3,984 training and 1,053 test memes across English and Spanish (5,037 total).

Why it matters / caveats: The approach shows that a compact (3.5M-parameter) head on frozen proprietary embeddings can achieve top soft-label performance on the hardest, most fine-grained subtask without fine-tuning large generative models, but the authors note reliance on the proprietary Gemini Embedding 2 limits reproducibility, the hard-label decoding used a fixed 0.5 threshold without calibration (so Hard-Hard results are not an upper bound), and the physiological signal analysis only tested EEG linearly, leaving gaze, heart-rate, and non-linear fusion unexplored.

Problem: Existing self-distillation approaches for unsupervised syllabic tokenization of speech, such as SD-HuBERT, use an utterance-level cross-entropy objective that causes the model to predict speaker identity rather than linguistic content, contaminating the purity of the resulting syllabic tokens.

Method: The paper proposes SylReg, a speaker-disentangled syllabic tokenizer built on a BYOL-style student-teacher framework initialized from pretrained HuBERT Base (with the last three Transformer layers pruned to nine layers). The student is trained to regress ℓ2-normalized, chunk-average-pooled (chunk size C=100 frames) projected hidden states of a speaker-perturbed input toward those of a moving-average teacher's clean input, using an MSE loss, with speaker perturbation restricted to gender-consistent formant/pitch shifts (male-to-female or female-to-male only, based on estimated mean pitch). The emergent block-diagonal syllabic structure (layer 8) is then distilled into pretrained data2vec 2.0 via multi-stage self-segmentation distillation (SylBoost) to produce SylReg-Distill, and syllable boundaries/tokens are extracted via minimum-cut segmentation, segment-average pooling, and two-step (K-Means + agglomerative) clustering. Downstream, a speech LM (SylReg-LM) is trained on interleaved syllabic-text sequences following the SpiRit-LM recipe (continually training Qwen2.5 7B or an 85M model), and a token-to-speech synthesizer uses a conditional flow-matching Diffusion Transformer with a BigVGAN-v2 vocoder.

Key results:

• SylReg outperforms Sylber on syllable purity (SP) by 10% while matching its segmentation F1 (72.2%), without using self-segmentation distillation.
• SylReg-Distill matches or exceeds SylBoost across all syllable segmentation and token purity metrics (e.g., F1 74.7% vs. 74.4%, token edit distance 7.66% vs. 14.4%).
• SD-HuBERT shows a speaker-normalized mutual information of 0.61 between speaker ID and predicted category, indicating strong speaker-identity dominance.
• Ablations show replacing BYOL+MSE with a DINO-style cross-entropy objective degrades all metrics and causes prototype collapse (only 32% of softmax categories become active); removing speaker perturbation increases token edit distance from 9.91% to 13.9%.
• Chunk size ablation: segmentation F1 improves with chunk size up to saturation at C=100 (F1 72.5%), while token purity (SP/SNMI) peaks at a shorter chunk size of C=20, showing frame-wise (C=1) and global (C=T) pooling are both suboptimal.
• SylReg-LM 7B achieves a 7% average relative improvement over phone-level SpiRit-LM 7B on sBLIMP and StoryCloze tasks (syntactic/semantic understanding).
• SylReg-LM 85M matches or exceeds SpiRit-LM 7B's generation quality (perplexity 59.6 vs. 62.0) while using only 0.2% of its training compute (8.2e18 vs. 4.0e21 FLOPs).
• The token-to-speech synthesizer built on SylReg-Distill matches the TWIST synthesizer's CER/WER (2.53%/5.46% vs. 2.59%/5.65%) at a 2.3x lower bitrate (76.8 vs. 174.8), and achieves the highest UTMOS score (4.31) among compared systems.

Why it matters / caveats: By disentangling speaker identity from syllabic representations, SylReg produces purer, more content-faithful syllabic tokens that narrow the granularity mismatch between speech and text tokens without changing downstream LM architecture, enabling more compute-efficient and linguistically capable speech language models and more coding-efficient speech synthesis; however, coarser tokenization (via lower merge threshold) improves boundary R-value but degrades downstream tasks requiring fine-grained lexical discrimination (e.g., sWUGGY, CER, WER).

Problem: In semi-supervised semantic segmentation with foundation-model teachers (e.g., DINOv2), pseudo-label confidence filtering already retains mostly-correct pixels, so existing pixel-contrastive memory banks (ReCo, U2PL) still admit a small fraction of confidently-wrong pseudo-labeled entries that contaminate the embedding space as false positives.

Method: PixCon adds a clean-positive pixel-contrastive branch on top of a DINOv2-Base ViT-B/14 encoder with a DPT-style decoder, trained jointly with UniMatch V2's weak-to-strong consistency branch (dual strong-augmented CutMix views with complementary channel dropout, teacher pseudo-labels filtered at τ=0.95). The branch projects fused decoder features through a two-layer 1x1-conv head to a 256-dim normalized embedding, and maintains a per-class FIFO bank (256 entries/class) that enqueues only labeled pixels the student already classifies correctly, guaranteeing contamination rate ρ_F=0 by construction. A supervised InfoNCE (SupCon) loss (temperature η=0.1) is computed over these clean anchors (capped at 1024/iteration) and added to the total objective with weight λ_pix=0.1 alongside supervised cross-entropy and the consistency loss.

Key results:

• Measured bank contamination rate ρ_F of confidence-filtered banks: 0.018 on Pascal VOC 1/8, 0.022 on 1/16, and 0.106 on ADE20K.
• Pascal VOC 1/8: 3-seed mean mIoU reaches 87.90, matching the published UniMatch V2-B figure, up from a 87.01±0.73 reproduced baseline (+0.89); improves all 3 seeds and 41/41 all-live training epochs.
• Controlled one-switch ablation (seed 0, Pascal 1/8): per-seed gain of +0.20 mIoU (87.40 to 87.60).
• Pascal VOC 1/16: +0.39 mIoU gain (85.05±0.52 vs. 84.66±0.20), improving 2/3 seeds.
• Pascal VOC 1/4: ties baseline (88.68 vs. 88.59, single seed).
• Cityscapes: ties baseline within single-seed noise (83.20 vs. 83.16 at 1/16; 83.88 vs. 83.96 at 1/8).
• ADE20K 1/8 (150 classes, most imbalanced benchmark): 49.23 vs. 49.10 baseline (+0.13 mIoU, a tie).
• Bank-admission-rule ablation: clean-positive and confidence-filtered banks tie within 0.02 mIoU at both Pascal splits (87.60 vs. 87.58 at 1/8; 85.53 vs. 85.54 at 1/16), both beating the no-bank baseline (+0.2 at 1/8, +1.1 at 1/16).
• Adds no inference-time parameters; the per-class bank overhead is ~1.4MB (21 classes x 256 entries x 256 dims for Pascal).

Why it matters / caveats: PixCon offers a low-cost, tuning-free default for foundation-model SSSS that removes the bank-confidence threshold ReCo/U2PL require, but the authors' own analysis shows the measured accuracy gain comes mainly from the "correctness" admission condition sharpening true-positive gradients rather than from the ρ_F=0 guarantee itself, since contamination is already under 11% (and often under 2%) with strong teachers; gains are concentrated at Pascal 1/8 (partly a variance-reduction effect, sign test p=0.125) while Cityscapes and ADE20K results are single-seed ties, and the paper notes the ρ_F=0 guarantee should matter more as teacher quality weakens, a regime not directly tested.