Skip to main content

Daily report

AI and science frontiers · 2026-11-16

Only content delivered through the publication boundary on this date is included.

Sealed edition v1

arXiv

WeatherDreamer learns data assimilation itself: hourly global analyses in about 1.5 minutes, without reanalysis

WeatherDreamer folds data assimilation and forecasting into one learned cycle trained directly on 18 Earth-observing systems, without reanalysis. It corrects a latent state with new observations every hour; at held-out stations its ensembles beat NOAA's GEFS for temperature, dew point and wind out to 12 hours and are better calibrated, while its assimilation matches GSI near the surface and in moisture channels but shows departures up to about five times larger in temperature-sounding channels. One global hourly analysis takes about 1.5 minutes on a single GPU.

arXiv

MoE serving re-splits expert replicas every decode step: Relief's integer relaxation nears the per-step optimum, and an on-die engine lifts effective utilization to 99.5%

In expert-parallel MoE serving, each layer finishes only when its most loaded GPU does, and routing shifts at every decode step. Relief replaces the per-step linear program with an integer bidirectional pairwise relaxation that halves load gaps along replica edges and nears the per-step optimum in a few passes. On decode routing from three open MoE families, scored as effective utilization relative to that optimum, the deployed static scheme reaches 66% to 72%, the GPU kernel 93%, and the on-die engine 99.5%, with the largest recovery over static placement at 37.3 percentage points.

arXiv

Powder diffraction structure solving goes to AI agents: exact recovery on pXRD-Bench rises from 27% to 69%

Crystal structure determination from powder X-ray diffraction has long depended on manual trial and error, which is hard to plug into self-driving laboratories. The task is recast as a verifiable agentic search problem: candidate structures can be scored during inference by forward-simulated profile agreement and during evaluation by structural matching to a known reference. On the 200-problem simulated single-phase pXRD-Bench, an adapted agent raises selected-structure exact recovery from 27% to 69%, with further gains from repeated sampling and verifier-guided inference-time scaling.

arXiv

Edge LLMs stop paying fixed width per weight: LazyWeights shrinks codes and cuts Orin token time by up to 36%

Edge LLM weight formats spend equal bits on every integer code, although the codes are far from uniform. LazyWeights keeps llama.cpp's GGUF formats, tunes GPTQ so the codes concentrate, Huffman-codes them losslessly in 32 interleaved lanes, and decodes them on demand in a small unit between DRAM and L2, leaving GPU kernels unchanged. At 1% and 3.5% perplexity budgets, coded Phi-3-mini, Mistral-7B and SmolLM-1.7B are 5.7–31.4% smaller than the best llama.cpp format; on a Jetson Orin Nano Super, with the decoder unit, latency and energy per token fall by up to 36% and 38%.

arXiv

A few specimens can define a whole family of hyperelastic laws: hierarchical Bayesian PANNs separate specimen spread from noise

A hierarchical Bayesian physics-augmented neural network (PANN) learns the constitutive laws of several specimens of one material as a population: all specimens share polyconvex features learned by the network, each specimen is described only by its own amplitude and feature shares, and the population mean and covariance and the measurement noise are inferred in one posterior. On synthetic populations it recovers the population from few specimens and predicts new specimens more reliably than three published methods; on Treloar's rubber and on healthy and aneurysmal aortas it reaches the accuracy of a deterministic PANN and assigns no aorta specimen to the wrong cohort with confidence, although its bands for untested curves are too narrow.

arXiv

Debate as a training protocol gets its first learning analysis: cooperating debaters can cut mistakes to a constant in some cases

When debate is used as a training protocol, a human judges which of two debaters is right instead of supplying labels. Earlier work analyzed only the game between debaters under a known judge, leaving open how debaters learn the correct function. This work formulates debate as an online learning problem and shows there are cases where adversarial debate cannot learn but cooperative debate needs only a constant number of mistakes, with a new combinatorial dimension characterizing the optimal mistake bound for cooperative debaters in the realizable setting.

arXiv

SinkState stores recurrent states as adaptive low-rank factors: hybrid-model decoding peak memory falls to as little as 1/2.6 of baseline

Hybrid language models compress history into fixed-size recurrent states, yet serving systems still treat those states as dense matrices, reading and rewriting them whole at every decoding step, so memory traffic grows with batch size. SinkState exploits the concentration of queries, keys and values in linear attention into a stable low-dimensional subspace, keeping each state in an adaptive low-rank factorization throughout decoding, merging new tokens with a local SVD and consolidating the frozen part asynchronously on the CPU. Across four Mamba-2 and GDN hybrid models from 4B to 120B parameters, peak recurrent-state memory falls to 0.38–0.63 of the baseline and decoding speeds up by up to 1.58x, with accuracy maintained.

arXiv

Scientific agents stop leaning on the last round: Sage-Taiyi keeps what it learns mid-search and tops Taiyi-Bench at 54.0

Sage-Taiyi splits scientific discovery into two levels: an object level that searches over a target tree and an idea graph, and a meta level that keeps a revisable cognitive state recording which judgments have evidence and which premises remain untested, then uses it to decide what to investigate next and which method to apply. On Taiyi-Bench, 265 tasks drawn from 50 papers across ten disciplines, it scores 54.0 overall against 47.9 for the strongest baseline, AI Scientist-v2, ranking first on all eight metrics. Five expert-guided collaborations in chemistry, materials and engineering carried some proposals to experimental or computational validation, including a cobalt-catalysed carbonylation model reaction raised from trace amounts to 84% yield.

Qwen

Alibaba releases Qwen4-Max: a 2.4-trillion-parameter model with a million-token context, scoring 52.69 on Work360 versus 47.74

Alibaba's Qwen team released Qwen4-Max, built on the new Qwen4 architecture with a 2.4-trillion-parameter backbone, 71B parameters activated per token, native vision and a one-million-token context window. It scores 52.69 on the team's Work360 industry-task benchmark, up from 47.74 for Qwen3.8-Max, uses fewer tokens at matched effort levels on Work360 and Terminal-Bench 4.0, and carries lower API prices. The release also documents pre-training, long-running agent and safety evaluations.

arXiv

Symbolic execution stops stalling on environment setup: an AI agent supplies the database model, lifting CrossHair's executable functions from 4.5% to 81%

CrossHair, a symbolic execution tool, barely runs on real Python web applications because the code ships without contracts and its database calls cannot be executed symbolically. A pipeline pairing an AI coding agent with CrossHair automatically writes contracts for functions and synthesizes a symbolic model of the database, raising the share of executable functions from 4.5% to 81% across 11 open-source applications and 1,156,044 lines of code, and finding 221 bugs reproduced by tests. These bugs are mostly input-domain corner cases, and some are missed by direct inspection with the same model.

arXiv

GeneFlow reads 1,000 genes straight from H&E slides, cutting inference to 0.31 GFLOPs

Predicting spatial gene expression from H&E histology has relied on complex architectures and multi-stage pipelines, usually limited to fixed panels of about 200 genes. GeneFlow pairs flow-matching generative modeling with histology and gene tokenizers to predict each spot's expression profile directly, reaching the highest PCC across four public datasets, a 17.1% gain on PRAD PCC-1000, with a 0.31 GFLOPs trainable module and 0.61 s sampling per sample.

arXiv

Diffractive optical networks stop prescribing output patterns: weak supervision lets handwritten scripts self-organize, cutting training steps by about 82%

Diffractive optical neural networks usually assign each class to a fixed detector region and train the light field to reproduce that preset output. A team at the Singapore University of Technology and Design and the National University of Singapore instead used script-level coarse labels and a modified contrastive objective that constrains only relations among samples, letting optical representations of handwritten Devanagari, Arabic and English characters cluster into three groups on their own. At a shared training endpoint the three-layer network reached 93.4% mean test accuracy against 84.7% for supervised optical networks, and needed about 18% of their training steps to reach a stable 84.5% test accuracy.

arXiv

A decision model that acts badly can still teach: JEV scores whole trajectories and lifts a 1.5B policy from 72.8% to 82.9%

Decision models are usually judged by how well they act, yet the same models are pulled toward an easier subgoal when acting on ALFWorld. Used instead as a teacher that sees the whole trajectory after an episode and is asked whether each action helped, JEV raises a Qwen2.5-1.5B policy trained with GRPO from 72.8% to 82.9% success on ALFWorld and from 56.8% to 71.4% on WebShop, while JEV as an actor completes only 42.9% of ALFWorld tasks.

arXiv

WaveGen folds zero-shot TTS into one model: text goes straight to waveform, hitting WER 1.47 and SIM 0.75 on Seed-en

WaveGen pulls zero-shot speech synthesis back from multi-stage pipelines into a single diffusion model: text produces raw waveform directly, without a neural audio codec, Mel vocoder, VAE latent or duration predictor. Spectral Forcing makes the same model learn a low-resolution spectral trajectory alongside the target waveform trajectory, giving text-speech alignment a structured guide in high-resolution waveform space. Trained in one stage on open Emilia data, the 0.6B-parameter WaveGen reaches WER 1.47 and speaker similarity SIM 0.75 on the Seed-en benchmark.

arXiv

TraceKV shrinks reasoning-VLM KV cache about sixfold: on-policy calibration keeps reasoning accuracy near the dense model

Reasoning vision-language models must hold both multimodal context and a growing reasoning trace in their KV cache, and existing low-rank compression calibrates on static inputs, so its divergence from the dense model grows with generation length. TraceKV instead calibrates on reasoning traces the dense model generates itself, weights each KV state by its sensitivity to policy divergence before the SVD, and allocates ranks globally. Across four models including Qwen2.5-VL-7B, a 70% rank ratio with 4-bit quantization gives 73.41% overall accuracy versus 73.58% for the dense model, with the KV cache reduced 5.64 times.

arXiv

Picking the right image is not enough: Qwen2.5-VL gains 16 circular-accuracy points with five distractors from answer-consistency training

Vision-language models lose accuracy when irrelevant images sit beside the target, and the loss persists even when the prompt names the target image, because middle-layer distractor-to-text connections carry the interference into reasoning. Using the single-image prediction as a reference, answer consistency regularization aligns answer distributions with and without distractors on Qwen2.5-VL, raising circular accuracy by an average of 16 percentage points across five perception benchmarks with five distractors while preserving single-image performance.

arXiv

S1Pert stops gene-perturbation models from reciting response frequencies: macro-F1 up to 0.850 across four cell lines

In CRISPR perturbation screens, models can score well by memorizing whether a gene tends to rise or fall historically, without resolving which perturbed target caused it. S1Pert removes statistically inconclusive responses from supervision and caches frozen-encoder representations of targets, query genes and historical evidence offline, leaving only a lightweight decoder for online inference. Across HepG2, Jurkat, RPE1 and K562, its macro-F1 is 0.850, 0.698, 0.785 and 0.769, 5.8 to 6.4 percentage points above the strongest baseline.

arXiv

Open-source agent platform OpenLayout takes over VLSI physical design, lifting atomic-task Pass@1 well above a standalone LLM

OpenLayout turns VLSI physical design from a scripted flow that senior engineers tune command by command into an automated loop run by specialized agents on commercial EDA tools. Its companion OpenLayout-Bench evaluates agent capability at TSMC 28nm: standalone Gemini 3.1-pro reaches only 18.9% Pass@1 on Tool A atomic tasks, while OpenLayout reaches 80.0% on the same metric, and on Tool B the standalone model reaches 19.3% versus 84.1% with OpenLayout. On the 10-design holistic flow suite, small and medium designs close timing, while NVDLA-large and Open910 retain residual DRC violations after Routeopt.

arXiv

Landslide inventories record observers as well as hazards: splitting occurrence from recording pays off on one earthquake, not all

Landslide susceptibility maps learn from inventories that are systematically biased, because landslides near roads and towns are more likely to be reported, so a classifier absorbs the observation process into the hazard signal. This work adapts ecological occupancy models to landslide inventories, treating several independent inventories of one event as repeated surveys and modelling true occurrence and recording probability separately, with a small exhaustively mapped anchor area to set the scale. Under simulated reporting bias in three earthquakes it reduces road bias and estimates the overall landslide rate more stably than standard practice, but its ranking advantage over neutralisation appears on only one earthquake, Iburi.

arXiv

Sewer digital twins stop retraining per pipe: the FLOOD operator predicts across unseen pipes and networks, about 5,500-fold faster on one pipe

Sewer hydraulic digital twins have been limited by the cost of solving the Saint-Venant equations, while existing machine-learning surrogates need retraining for each new pipe or flow regime. FLOOD learns the solution operator of the Saint-Venant equations with a multi-branch neural operator, taking pipe geometry, hydraulic parameters and water levels at both manholes as inputs and returning depth and flow rate. On pipes excluded from training it reaches coefficients of determination of 1.000 for depth and 0.999 for flow rate, speeds up single-pipe online prediction by about 5,500-fold, and also serves roughness calibration and leakage localization.

arXiv

A neural controller takes over nonlinear time-optimal MPC: half the training data and 1 kHz instead of 1.4 Hz

Nonlinear time-optimal model predictive control (TOMPC) drives constrained systems to a target in minimum time, but online solving is too slow for fast dynamics. This work imitates a long-horizon, nonconvex TOMPC with hard terminal constraints using a neural network, and introduces self-informed sampling: the network also predicts time-to-go, and the expert solver is queried only where a rollout departs from time-optimal behavior. On a Furuta pendulum, this sampling reaches the performance of standard dataset aggregation in roughly one-third of the training time and less than half the data, while the resulting controller keeps the expert's region of attraction and GPU inference raises the control frequency from about 1.4 Hz to 1 kHz.

arXiv

Parallel reasoning stops writing every path to the end: PreG scores short prefixes first and saves tokens in 54 of 60 settings

Parallel test-time scaling for large language models normally generates every reasoning path to completion before internal confidence can score it, so compute on low-quality paths is already spent. PreG instead generates short prefixes in parallel, ranks them by internal confidence, and completes only the most promising candidates before aggregating as before. On Qwen3-8B and Llama-Nemotron-4B across math competition sets and GPQA-Diamond, it reaches the accuracy of the original method at a 1000k-token budget with fewer tokens, saving tokens in 54 of 60 configurations and up to 80%.

arXiv

Detection error in latent spaces also double-descends: high-dimensional theory predicts miss rates at a fixed false alarm

Once foundation models encode images, text and audio into high-dimensional latent vectors, how detection and classification error probabilities vary with the ratio of model size to training samples has lacked a computable characterization. The work derives explicit large-deviation rates for linear and quadratic decision statistics and finds the rate depends non-monotonically on the shape parameter, including double descent and, in some regimes, multiple descent; in latent-space experiments with DINOv3, Llama-Nemotron-Embed-1B-v2 and BEATs, predicted rates and detection probabilities at fixed false alarm track the measured curves closely.

Science Robotics

A 5-tonne skid-steer loader learns two-wheel acrobatics: reinforcement learning policies swing, balance and cross bumps on the real machine

A 5-tonne John Deere 332G skid-steer loader trained with reinforcement learning in simulation and transferred to the real machine can rock itself up onto its front wheels, balance on two wheels and track velocity commands. Against a model predictive control baseline, the two perform comparably at static balancing and velocity tracking, but the baseline loses balance beyond a target speed of 3.5 and cannot cross 75-millimetre undulations, while the learned policy can. The pipeline needs no prior machine model; an actuator network trained on measured data closes the sim-to-real gap.

arXiv

Crystal sequences stop following coordinates: topology-guided ordering lifts single-shot structure-prediction match from 56.10% to 61.92% at the same model size

Crystal generative models write three-dimensional periodic structures as one-dimensional sequences, but ordering atoms by real-space coordinates lets small geometric perturbations reshuffle the sequence and inject variability unrelated to connectivity. CGTransformer instead orders atoms by periodic quotient-graph topology, combining element, symmetry and graph-embedded coordinates. On MP-20, the same 86M model raises single-shot structure-prediction match from 56.10% under coordinate ordering to 61.92%, and in conditional generation of 3.0–3.5 eV wide-bandgap semiconductors it increases the share of stable, unique and novel candidates.

arXiv

Composed image retrieval stops squeezing queries into one vector: FactEmbed decomposes the edit first, then reranks candidates

Composed image retrieval takes a reference image plus a modification text and must return the target image, and the common recipe compresses that multimodal query straight into one retrieval vector, leaving the model no explicit chance to work out what the text changes. FactEmbed inserts a set of learnable factorized latents into Qwen2.5-VL-7B's input and trains them with a target-imagination objective that reconstructs target details, so each latent grounds a semantic region, then carries that decomposition into retrieval through a factor-level contrastive loss and attention distillation. On the refined CIRR and FashionIQ benchmarks it scores above both CSMCIR and ReCALL on Recall@1 and average Recall@10, though the improvement margins stated in the text do not match the table values.

arXiv

Cutting three-quarters of visual tokens, APS-VLA keeps 95.75% robot-policy success and drops inference to 68.75 ms

Vision-language-action policies process hundreds of visual tokens per step, which drives up inference latency. APS-VLA keeps the backbone frozen and uses a lightweight scorer conditioned on the current observation, instruction and proprioception to pick tokens, training the retained representation both to reproduce policy outputs and to predict the subsequent visual change. On LIBERO with 128 tokens retained, mean success is 95.75%, 1.25 points below the unpruned 97.00%, and inference latency falls to 68.75 ms.

arXiv

Pruning 1.6% of neurons cuts name inference in fine-tuned LLMs by up to 7.4 points, with no retraining

Large language models fine-tuned on proprietary records memorize names and other personal data, and existing privacy-neuron editing inspects only the leaking model, with no PII-free reference to separate identity neurons from useful context. JEPA-Prune trains a public model on surrogate and masked views of the same records, never on the original personal values, and uses a cross-model representation probe, integrated gradients and a scale-invariant specificity score to locate PII-responsive neurons, which are zeroed before deployment without retraining. Across three PII datasets and five models, pruning 1.6% of neurons lowers PII inference accuracy by up to 7.4 and extraction recall by up to 13.9, while unseen-text perplexity stays within 16% of the unpruned model for most LLMs.

arXiv

A frozen graph foundation model can flag multivariate time-series anomalies: Gfm4Mtsad reaches best or second-best on most metrics across three real datasets

Multivariate time-series anomaly detection has long relied on graph detectors trained from scratch on limited target-domain data, whose learned sensor topologies are prone to spurious correlations and cannot reuse relational priors from pretrained graph models. Gfm4Mtsad wraps a frozen graph foundation model in three peripheral modules: SICA aligns temporal features with normal-connectivity priors and restores projection-lost detail, MAPS synthesizes near-boundary pseudo-anomalies through residual-preserving local dependency breaks, and TGEC re-ranks multi-granular evidence only in the upper score tail. Across MSL P-15, SMAP T-3 and SMD 1-1 with AnomalyGFM and OwlEye backbones, the adaptation reaches best or second-best results on most metrics.

arXiv

ForceWAM splits contact learning in two: demos set the goal, play data teach force response, reaching 95% on one real-robot task

In contact-rich manipulation, expert demonstrations cover successful paths but say little about how force responds to different actions, while task-agnostic play records broader contact dynamics without a task goal. ForceWAM gives the task-compatible motion and six-axis force/torque future to a force world model trained on demonstrations, and gives the question of which actions realize that future to an inverse dynamics model trained on both demonstrations and play; at deployment, fresh wrench feedback corrects actions between world-model updates. Across force-regulated sliding, box flipping with and without disturbances, and ring placement, the full system reaches 85%, 95%, 85%, and 90% success, above the listed baselines.

arXiv

Reconstruction residuals miss a drifting sensor: CLOSURE adds a second, separately learned channel-consistency score and reaches 88.13 Aff-F1 on SMD

Multivariate time series anomaly detection has long leaned on reconstruction residuals, but a residual only certifies that a window can be reproduced, not that its channels still explain one another as they do under normal operation. CLOSURE splits recoverability and cross-channel consistency into two prediction tasks trained separately on normal data: a diffusion reconstructor conditioned on a discrete memory of normal temporal bases scores the first, and a relation operator that predicts each channel only from its peers scores the second, with the two combined after calibration on normal data. Across the five public benchmarks SMD, MSL, SMAP, SWaT and PSM, its Aff-F1 exceeds the listed methods, reaching 88.13 on SMD.

arXiv

Gaussian splats skip the rasterizer: GaussianRenderFormer renders directly, reaching 38.08 dB on unseen objects

GaussianRenderFormer replaces the hand-written rasterizer in 3D Gaussian splatting with a transformer: one token per Gaussian, and a view-dependent decoder that turns small patches of camera rays into pixels. Trained on 26,820 Objaverse objects, it reaches 38.08 dB PSNR on 300 unseen objects against renders of the full splat, while rasterizing the same pruned input reaches 44.02 dB; diagnostics point to a view-decoder bottleneck, though under-training is not ruled out.

arXiv

Near-OOD detection gets a complement, not a replacement: PLUS lifts all 126 detector-backbone pairs across 18 detectors

Near-out-of-distribution (near-OOD) inputs resemble the training classes closely, and existing post-hoc detectors that excel on far-OOD inputs lose that edge here. PLUS does not replace a detector: it uses a sparse autoencoder to pick each class's most active neurons, spans a low-dimensional subspace, and scores an input by that class's whitened Mahalanobis distance, adding the signal to any calibrated base score at one fixed weight. On the ImageNet-1K OpenOOD v1.5 benchmark, near-OOD AUROC improves in all 126 detector-backbone combinations across 18 detectors and seven backbones, while far-OOD performance is broadly maintained, with declines of at most one point where the base score is already near-saturated.

arXiv

Aggregation is itself a spectral valley: SpecValley learns where the valley sits, leading four of six heterophilic graphs

In node classification on heterophilic graphs, the step before propagation has stayed fixed by degree normalisation, which acts spectrally as a valley whose zero is set by the normalisation rather than by the data, exactly where dense graphs' eigenvalues accumulate. SpecValley learns the valley's centre, sharpness and flank gains per graph, then feeds an ordinary ChebNet spectral filter. Across six heterophilic benchmarks it matches or exceeds the best published result on four, sits within the confidence interval of the best on Tolokers, and trails on Actor; crossing five aggregators with four filters attributes the gain to aggregation rather than to the downstream filter.

Diffusion language models learn to count characters: no fine-tuning, summaries land within 3.1 characters of target

For summarization settings such as headlines and previews that must fit an exact character count, Delta estimates the expected character length of the current output inside a masked diffusion language model's denoising loop and dynamically inserts or removes masked positions to steer generation toward the target. It needs no fine-tuning and applies to off-the-shelf LLaDA and Dream. Across the XSum, CNN/DailyMail and arXiv summarization benchmarks, character-length error falls by an order of magnitude relative to the baselines while summary quality stays competitive under automatic and human evaluation.

arXiv

NEO-VLA drops the vision encoder: action signals peak before the final visual layer

NEO-VLA removes the fixed visual encoder from vision-language-action models: images are tokenized by lightweight patch convolutional layers and evolve with language through the backbone, and a flow-matching action expert draws on instruction-conditioned visual representations at different depths. In controlled comparisons the native architecture beats encoder-based counterparts; full fine-tuning narrows but does not erase the gap, and extra scale alone cannot recover it. Layer-wise probing shows action-critical signals often peak before the final visual representation.

arXiv

Delayed feedback meets a memory cap: CA-OCRC keeps online risk control on target with capacity growing only logarithmically in the maximum delay

Online conformal risk control has assumed each round's loss arrives immediately and intact, yet real systems delay feedback and can track only a limited number of pending predictions, so some feedback is lost for good. CA-OCRC designs the tracking policy and the prediction sets together: each round draws a random proxy delay, is tracked while capacity allows until its feedback arrives or the proxy delay expires, and a two-level inverse probability weighting compensates for lost feedback. Risk bounds for arbitrary and stochastic delays follow, and optimizing the proxy delay distribution shows that capacity growing only logarithmically in the maximum delay suffices to match the convergence rate of OCRC with unlimited feedback.

arXiv

ScanDreamer gives quadrotors an online map to imagine with: 96.3% success and 3.6% collisions in unseen simulated layouts

For quadrotor navigation through unknown clutter, ScanDreamer builds an online 2D scan-memory map from depth images and vehicle pose, and conditions its world model, actor and critic on that map, so imagined rollouts advance the pose over a frozen map and read geodesic progress and wall-contact signals from mapped geometry. Across 1560 test episodes in NVIDIA Isaac Sim spanning unseen layouts and parameter ranges, three seeds average 96.3% success and 3.6% collisions, against 59.6% success and 28.8% collisions for AirDreamer at the same online-interaction budget.

arXiv

EmoReward stops emotional video RL from trusting fixed rewards: last run's traces set the next run's weights

Emotional video generation has no single correct target, and fixed fusion of visual, motion and text-alignment rewards captures emotional intent poorly. EmoReward distills reward traces from one completed training run into per-emotion-regime reliability estimates, converts them into bounded weight multipliers, and combines them with current-video instance gates in the next run without updating any reward-model parameters. On HunyuanVideo 1.5 with DanceGRPO, FlowGRPO and SAGE-GRPO, VideoAlign Overall rises under all three algorithms, including SAGE-GRPO with EmoReward moving from -0.8269 to -0.2624.

arXiv

Video action explanations split into what the model sees and when it moves: lowest deletion AUC on most axes across three datasets, with masks about five times sparser

Explanations for video action recognition usually fold spatial appearance and temporal motion into a single saliency map, so a user cannot tell whether the model predicted from the object or from the movement. This work separates the explanation into a spatial appearance mask and a temporal motion mask, optimized jointly against a frozen classifier, with nearest-frame replacement used to remove temporal dynamics. On EPIC-Kitchens-55, EGTEA Gaze+, and Something-Something V2, the method attains the lowest deletion AUC on most datasets and axes, with masks roughly one-fifth the size of the strongest baseline.

arXiv

Multi-talker transcription stops decoding each speaker alone: Pandemonium lets speakers see each other and reaches 14.7 average cpWER on meeting benchmarks

Target-speaker recognition for multi-talker meetings previously decoded each speaker independently and was confined to 30-second windows. Pandemonium moves the diarization conditioning onto the longform Parakeet backbone, which decodes a whole conversation in one pass, and adds cross-speaker attention plus a mask refiner so that the speakers of one conversation become visible to each other inside the encoder. Trained on synthetic, pseudo-labeled and real data, it reaches state-of-the-art results across several meeting benchmarks, with an average cpWER of 19.2 under predicted diarization.

arXiv

CrossPressor sidesteps the fusion bottleneck in long-sequence 3D reconstruction: training-free routing cuts 256-view cross-view fusion from 125 ms to 4.8 ms

As feed-forward 3D Gaussian splatting scales to more views, anchor-based compression fixes the decoder-facing representation but lets support-to-anchor fusion grow with view count. CrossPressor leaves released checkpoints untouched and routes support features before K/V projection with coverage-cycled offsets, spreading retained spatial evidence across support views and fusion layers. At 256 views it keeps 1.85% of dense QK interactions, cutting cross-view latency from 125.28 ms to 4.80 ms while rendering quality stays close to dense fusion.

arXiv

Reasoning search stops voting only after the fact: VGSC resamples steps during decoding and reaches 63.1% average accuracy

Test-time search for multi-step LLM reasoning has long traded between two imperfect signals: post-hoc voting keeps diversity but cannot steer generation toward correct paths, while beam-style verifier guidance intervenes early but over-concentrates and amplifies verifier errors. VGSC multiplies step-level correctness with a future-outcome prediction into an acceptance weight, applies bounded rejection-resampling to candidate steps during decoding, and keeps multiple trajectories for final self-consistency voting. Across five reasoning benchmarks with the Qwen2.5 Math 1.5B Instruct generator and Qwen2.5 Math PRM 7B verifier, VGSC reaches 63.1% average accuracy at verifier-trust exponent 10, above the strongest non-VGSC baseline BS+SC-Future at 60.7%.

arXiv

A2WM refines frozen imitation policies without retraining, lifting mean success from 27.18% to 59.51%

Small execution errors in imitation policies accumulate into failure, and earlier fixes relied on additional on-policy learning or costly online planning. A2WM learns local action sensitivity offline from action-perturbed rollouts of a frozen policy, predicting task-relevant geometry as a nominal trajectory plus a linear response to corrections, which turns online refinement into one model evaluation and one convex quadratic program solve. Across four simulated manipulation tasks it reaches the highest mean success among the evaluated baselines, 59.51% versus 27.18% for Diffusion Policy, with 3.8 ms mean refinement overhead per control step excluding base-policy sampling.

arXiv

Byte order, not just byte count, sets boundary prediction: UBE cuts one-symbol balanced grapheme error from 8.76% to 0.06%

Byte-level language models group bytes into patches with a learned boundary predictor, and what that predictor can see at limited lookahead depends on the text encoding. In UTF-8 many scripts take three bytes per codepoint, and the first byte carries only 4 high codepoint bits, so after the same left context a mark extending a grapheme cluster and a letter starting the next can share their first byte. Training small causal predictors on matched FineWeb-2 text and evaluating on 17 FLORES-200 languages, balanced grapheme-boundary error at one-symbol lookahead is 8.76% with UTF-8 and 0.06% with UBE; the same-length UBE-BE variant gives 8.62%, so the gain follows byte order rather than stream length.

arXiv

Gating each ReLU feature with a partition-of-unity weight stops regression spectral bias from growing with window count, but on Poisson it makes the lowest frequency the slowest mode

In shallow ReLU models, the spectral bias that makes high frequencies learn slowly is tied directly to the condition number of the feature dictionary's Gram matrix. Multiplying each ReLU feature by a hat partition-of-unity weight makes that Gram matrix block-tridiagonal for any local bilinear form, and its block-Toeplitz symbol sets the learning rate of each frequency. Under a regression loss the symbol stays away from zero, and the measured spread of frequency-dependent learning rates falls from four orders of magnitude to within a factor of five at matched model size; under the stiffness matrix the symbol vanishes at zero frequency, so gating leaves conditioning intact and instead makes the lowest frequency the slowest mode.

arXiv

AskALCF assigns each HPC support failure to a component: unattributed turns fall from 41.4% to 14.2% across 2,174

AskALCF, an agentic system for high-performance computing user support, attributes each failed question-answer turn to one of five components—knowledge, retrieval, prompt, model or tools—instead of returning only a quality score. Across 2,174 production support turns it cut unattributable failures from 41.4% to 14.2%, and on a 150-turn expert reference it raised agreement from 108 to 123 turns. The object of improvement is the rules, prompts and review policy rather than model weights, and experts retain authority over every production-facing change.

arXiv

Multimodal rotary position encoding split apart: text-image spatial bias barely helps, and removing it sharpens grounding

Multimodal Rotary Position Embedding (MRoPE), widely used in vision-language models, assigns height and width coordinates to text tokens that have no inherent spatial structure, adding positional preferences to text-vision attention. The study separates intra-visual spatial encoding from cross-modal spatial effects, probing official pretrained checkpoints without training and running controlled training at 2B and 4B from identical initial parameters. Cross-modal spatial effects provide little extra benefit as spatial references and can even impair performance, while intra-visual spatial encoding remains useful for fine-grained grounding.

arXiv

Meeting transcription stops reading only who is talking: structured speaker conditioning steadies target recognition under overlap

Meeting transcription must output text for a requested speaker, but four-state activity conditioning collapses different interferer counts and hand-over histories into one target condition. Structured speaker conditioning encodes each speaker's activity track separately, combines it with session-level acoustic identity, and controls the acoustic encoder through condition-dependent low-rank cross-channel modulation. On Qwen3-ASR-0.6B with oracle diarization, the complete method uses 1.46M conditioning parameters and achieves a 6.23% mean relative cpER reduction over four-state diagonal conditioning across four meeting test sets.

arXiv

Conditional flow matching can transfer too: a correction on top of a frozen source representation cuts energy discrepancy 26.4% on five-dimensional responses

When target samples are limited, conditional flow matching can borrow a covariate representation learned from a related source population. TFM-Freeze keeps the source encoder fixed and trains only a target velocity head; TFM-Adapt jointly learns the head and a representation correction under a functional anchoring penalty. In simulations with related five-dimensional representations and 10000 source samples, unpenalized TFM-Adapt lowered energy discrepancy from 0.3158 for TFM-Freeze to 0.2325, a 26.4% reduction; target validation falls back to a target-only generator when transfer does not help.

arXiv

Search pipelines can run clean and still miss the point: WISE traces the bad page back to ranking and recall

Industrial search pipelines can return results on time and still show pages that fail the user's intent, while existing monitoring watches latency, availability and errors. WISE splits the diagnosis in two: it first judges whether a page has a quality problem using only user-facing multimodal evidence, then pulls internal pipeline evidence to attribute confirmed problems to supply, recall, relevance scoring, ranking or reranking. On WISE-Bench, 12,000 real search cases, Qwen3.5-35B-A3B rises from 35.25% to 75.20% end-to-end exact match after trajectory-supervised fine-tuning and AHPO.

arXiv

LoRA ranks no longer need to match: per-domain ranks cost averaging and utility selection nothing, while conflict-resolution merging loses up to 4 points

Model merging usually trains every expert at the same LoRA rank, yet each domain has its own best rank. Across 36 experts on mathematics, instruction following and multilingual data at four ranks and three model scales, six merge methods were compared at uniform ranks and at each domain's own best rank: simple averaging and loss-utility component selection are unaffected and gain up to 1.8 pp of multi-domain accuracy, while magnitude-based conflict-resolution methods such as TIES and KnOTS are inconsistent and can lose up to 4 pp. The cause is that selecting components by size discards the instruction adapter, whose singular values are smallest, almost entirely, and IFEval falls by 16 to 25 pp.

arXiv

MMFlash reweights visually dependent tokens in diffusion drafting, lifting VLM decoding to 3.00–3.05× on image benchmarks

MMFlash extends diffusion-style speculative decoding to vision-language models by measuring how much each token depends on the image and upweighting those tokens during training. The measurement shows DFlash drafter acceptance falling from 46% in the lowest visual-dependence quartile to 22% in the highest, a gap that persists after multimodal training. On the Qwen3.5 series, MMFlash reaches average greedy-decoding speedups of 3.00–3.05× across seven image benchmarks, 10.3–10.9% above DFlash, and 2.33–2.71× on long-video benchmarks with sparse attention and light fine-tuning.

arXiv

Diffusion LLM guidance drops the extra forward pass: GCG edits attention-head features and lifts LLaDA-8B's Ruler score from 26.2 to 29.5

Diffusion language models generate text in parallel through iterative denoising, but existing guidance methods build a weak model and pay an extra forward pass at every denoising step, doubling inference cost. GCG first shows these models lean on local context and underuse distant evidence during denoising, then amplifies the outputs of attention heads that carry global context directly in feature space, without a weak model. On LLaDA-8B, Dream-7B and LLaDA-1.5 it beats U-CFG and A-CFG across several benchmarks, while LLaDA-8B keeps the same FLOPs as the unguided baseline.

arXiv

Structure confidence picks the wrong proteins: 97% of high-confidence designs expressed, but only 16.7% bound terbium

In designing rare-earth-binding variants of the lanthanide-binding protein lanmodulin, candidates selected on structural confidence expressed at 97% yet only 16.7% bound terbium, against 83.3% for candidates from a function model trained on measured metal-binding data. No structural-confidence metric consistently ranked binders above non-binders within the three design panels, while the function model's prediction gave an ROC AUC of 0.79 to 0.92 in every panel. The comparison holds one protein family fixed, and the function model was trained largely on lanmodulin variants.

arXiv

D2ToF distills diffusion into one step: a single-frame single-photon waveform yields depth in about milliseconds, at roughly one-third the error of Marigold-DC

Low-cost single-photon LiDAR arrays deliver time-of-flight waveforms at very low spatial resolution, too coarse for direct depth imaging. D2ToF distills a multi-step conditional diffusion model into a one-step generator and adds a differentiable histogram data-consistency term to the training objective, so the reconstructed depth, re-rendered into low-resolution waveforms, matches the measurement; a high-resolution reflectivity image and a frozen monocular depth model supply geometric priors. On simulated Middlebury scenes under clean conditions the one-step model reaches an average RMSE of 11.7 mm, about one-third that of Marigold-DC at 32.9 mm, in a single forward pass of roughly milliseconds, far faster than its 250-step teacher.

arXiv

RiverScan stops averaging tributaries: area-weighted accumulation tops the evaluated baselines at 608 Danube gauges for daily runoff

Large-basin runoff prediction must carry heterogeneous subcatchment responses downstream along the river network, yet standard graph message passing does not explicitly account for the contributing area each tributary represents. RiverScan writes this accumulation as an area-fraction-weighted selective state-space recurrence, combining completed tributary states with local encodings in topological order and letting tributary state and reach attributes control retention. On 608 Danube gauges it gives the best overall predictive performance among the evaluated temporal and graph baselines without observed discharge inputs, and structural controls show that the benefit of correct topology grows along the gauge hierarchy.

arXiv

Ancient-terrace point clouds fail across regions: MELD corrects each feature level with unlabeled target data, lifting F1 from 0.397 to 0.459

Between the Teyuna and Miraflores lidar surveys in the Sierra Nevada de Santa Marta, Colombia, terrace detectors trained in one region degrade in the other because terrain, vegetation, acquisition settings and terrace morphology differ. MELD trains a separate diffusion denoiser on unlabeled target features at each of four PointNet++ levels, then uses these priors to refine the multi-scale features fed to the segmentation decoder during training and inference. F1 rises from 0.397 to 0.459 for Teyuna-to-Miraflores transfer and from 0.557 to 0.594 in reverse, the highest among the compared adaptation methods in both directions; adding KNN local-geometry descriptors raises these to 0.542 and 0.606.

arXiv

Latin text reuse gets a word-by-word account: copy, inflection or substitution, matched with 200 corrected pairs

Computational work on Latin intertextuality could previously only say whether two passages are related, or roughly split references into verbatim citation and allusion, without explaining what the later author changed word by word. A pointer network links each word of the reusing passage to a source word or to none and names the operation, such as copy, inflection or substitution, in the same step; synthetic pairs built by applying known operations to Latin text train a first model, and active learning selects the pairs an annotator corrects. On the 1,490 references of the Loci Similes benchmark, 200 pairs chosen by uncertainty bring the model level with one trained on every annotated pair (macro F1 84.58 against 84.44).

arXiv

Medical retrieval's dense vectors split into readable features: sparse codes beat original embeddings on CURE, and explanations still predict activations across corpora

Medical retrieval systems usually compress queries and documents into dense vectors whose matches are hard to explain. This work trains sparse autoencoders on those dense embeddings with an InfoNCE contrastive loss that preserves query-document relevance, raising NDCG@10 above the original embeddings for all three embedders on CURE. Neuron interpretations built only from CURE texts still positively predict whether a neuron fires on five out-of-domain corpora, and concept-level interpretations outperform raw example snippets.

FutureHouse

FutureHouse opens its scientific agent platform: Crow, Falcon, Owl and Phoenix reachable through a web interface and API

FutureHouse has launched a publicly available platform of scientific agents, with four agents covering general literature question answering, deep literature review, a check on whether anyone has done a given thing before, and chemistry experiment planning, reachable through a web interface and an API. The platform says Crow, Falcon and Owl outperform all major frontier search models on retrieval precision and accuracy, and beat PhD-level researchers in head-to-head literature search tasks on precision; Phoenix is labeled experimental, less thoroughly benchmarked than the other three, and may make more mistakes.

arXiv

Nonlinear system identification stops picking excitation signals: a reinforcement learning agent learns to fill the state space itself

Excitation design for nonlinear system identification has relied on predefined signal classes and repeated optimization, and classical design mainly minimized parameter variance for linear models. Here space-filling input design is recast as a reinforcement learning problem in which an agent issues amplitude-bounded inputs step by step from system observations, with no input class to choose and no frequency band to tune. On two nonlinear benchmarks, a mass-spring-damper with nonlinear stiffness and a Van der Pol oscillator, the learned policy gives the best overall space-filling performance under the same amplitude bound and distributes samples more uniformly than a state-of-the-art space-filling multisine design.

arXiv

Nano-FTIR stops discarding its noisy-but-clean harmonics: a cross-harmonic network pulls third-harmonic phase noise to about 0.02 rad

Hyperspectral nano-FTIR imaging has relied on the third harmonic, which is nearly background-free but noisy, while the higher-SNR first and second harmonics were discarded because of background contamination. A cross-harmonic neural network (CHNN), trained on experimental data, reconstructs third-harmonic spectra from first- and second-harmonic spectra, reducing noise on a PMMA/PS blend while preserving spectral features and nanoscale spatial contrast, and it suppresses noise better than a single-harmonic autoencoder.

arXiv

Quantum extreme learning skips the likelihood: on a transverse-field Ising chain, 15 training points bring maximum absolute error to 0.013

Quantum sensing usually infers an unknown parameter with Bayesian inference or maximum likelihood estimation, but both estimators require constructing a likelihood function, which needs full knowledge of the probe dynamics and is often impractical. A quantum extreme learning framework instead places the unknown parameter in the reservoir Hamiltonian, lets a fixed initial state encode it through parameter-dependent evolution, and trains a linear readout on local observable expectation values, avoiding explicit likelihood construction. Estimating the transverse field of a transverse-field Ising chain, 15 training points give a maximum absolute error of 0.013 over 200 test values, and combining measurements at two times into a larger feature vector improves accuracy further.

arXiv

High-entropy alloy strength prediction stops paying per test: ROSE-HEA matches a model trained on 185 records using 125

Tensile-strength data for high-entropy alloys are costly to obtain, and random sampling often spends the budget on similar compositions and processing routes. ROSE-HEA ranks candidate experiments by how much their possible strength outcomes would reduce worst-case predictive risk across a reference composition–processing domain, while weights on competing smoothness hypotheses update with each new response. Across 501 records and 30 matched runs, it reached a mean RMSE of 152.84 MPa with 125 observations, below EI-BO and UCB-BO at the same budget and below XGBoost, the strongest conventional regressor, trained on 185 observations.

arXiv

Nuclear-mass prediction swaps the gated recurrent candidate state: complex product units bring interpolation error to 0.2698 MeV

Machine-learning predictions of nuclear masses and charge radii usually feed neighboring nuclides into a gated recurrent network as an ordinary sequence. AM-PU-GRU embeds learnable product-unit mappings in the GRU candidate state and works in the complex domain to keep both amplitude and phase, reaching a mean mass interpolation RMSE of 0.2698 MeV and a temporal RMSE of 0.1905 MeV across three data partitions with three training seeds each. Without mass inputs, MI-PU-GRU predicts charge radii with an RMSE of 0.019947 fm, 14.40% below the fixed WS∗ prior.

arXiv

WebArena-Pro adds audio and video to web-agent testing: 20 self-hosted sites, 300 tasks, and a 62.3% ceiling for the best setup

WebArena-Pro is a reproducible web-agent benchmark of 300 human-validated tasks across 20 self-hosted web applications in six domain categories, treating audio and video as core observations alongside text and vision. Prior self-hosted benchmarks covered few application categories and limited modality to text and vision, while live benchmarks broaden site coverage but drift between runs. Codex with GPT 6.1 Sol and Claude Code with Claude Opus 5.5 tie for the best success rate at 62.3%, and cross-site tasks generally score below single-site tasks.

arXiv

No Detector Training Needed: Lens Finds Look-Alike Targets Across Satellite Tiles From a Single Click

For a satellite mosaic spanning thousands of tiles, an analyst previously had to collect examples and train a detector before asking where a given kind of target appears. Lens instead encodes each tile with a frozen self-supervised vision transformer, applies two linear corrections that remove positional bias and the area-wide common component, and matches the averaged features around a click against a byte-quantized index by cosine similarity. On the DFC2020 test set, a single click reaches 0.09 to 0.80 pixel F1 across eight land-cover classes, mean 0.39, and beats a pairwise INSID3 baseline on six of them at roughly four hundred times lower query cost.

Diffusion cache skips the full recompute: DACE+DTC cuts Qwen-Image-Edit latency 12.1% and LPIPS 18.9%

Diffusion inference uses learned feature caching to skip full denoising forwards, but cache prediction errors are written into the sampling state and accumulate. DACE runs cache decomposition and prediction only for target tokens and manages history by future read positions; DTC measures an endpoint counterfactual at scheduled full steps, estimates the interval defect integral with signed step sizes, and applies correction after the normal update. On Qwen-Image-Edit background replacement under the N7 schedule, the pair lowers GPU-resident latency by 12.1% and mean LPIPS distance by 18.9% relative to LinCa, while DTC adds 2.1% overhead to DACE and further lowers LPIPS by 18.6%.

arXiv

Fine-tuning a MACE foundation model fixes sodium clustering in hard carbon, cutting formation-energy error from 32.43 to 1.65 meV/atom

Pretrained MACE atomistic foundation models describe sodium intercalation in graphite-like domains but treat the metallic attraction between sodium atoms on graphene as repulsion, so they miss the clustering and pore filling that shape hard-carbon storage. Targeted fine-tuning on 1,985 sodium-cluster configurations plus 300 expanded graphite and sodium reference configurations lowers the formation-energy mean absolute error on the sodium-cluster data from 32.43 to 1.65 meV/atom and raises R² from -36.32 to 0.87. After fine-tuning on a broader sodium-carbon set, the model still gives a 2.03 meV/atom formation-energy error on sodium-filled carbon nanotubes excluded from training, and it links closed-pore radius to plateau voltage.

PLoS ONE

Battery voltage prediction stops restarting each cycle: SeqBattNet pairs discrete states with aging adaptation to reach RMSE 0.0257 and 0.0253 on two NASA cells

SeqBattNet turns terminal-voltage prediction into a discrete-state model advanced step by step, with an encoder that estimates aging-adaptation parameters such as ohmic resistance, RC time constants and state of health from a short opening segment of each discharge cycle, and a decoder that computes voltage from an equivalent circuit model. Earlier physics-informed models either needed many hard-to-identify physical parameters or advanced continuous differential equations whose errors accumulate and which struggle to stop at the cutoff voltage. On the public TRI, RT-Batt and NASA datasets, trained on a single cell and tested on unseen cells, it reaches RMSE of 0.0257 and 0.0253 on NASA RW4 and RW5, below PDB-Net's 0.0382 and 0.0406.

Fixed transforms no longer need guessing: the Adaptive Interpolating Transform beats fixed-transform baselines with nine added parameters in vision and three per encoder layer in language models

Fixed transforms such as the discrete Fourier transform and discrete cosine transform are widely used in machine learning to extract features, but choosing among them usually requires knowing the dataset's properties, and the approach weakens when that knowledge is missing. The Adaptive Interpolating Transform (AIT) turns the transform itself into learnable parameters, learning data-driven mappings tailored to the dataset and task. On ImageNet-2012 and the General Language Understanding Evaluation tasks, AIT-enhanced models outperform fixed-transform baselines under both from-scratch and fine-tuning workflows, adding only nine parameters to vision models and three per encoder layer to language models.

NVIDIA open-sources SkillEvaluator: agent skills pass a static quality gate first, then a with-and-without skill run comparison

NVIDIA released SkillEvaluator, an open-source command-line tool that checks AI agent skills and measures their effect on task results. A skill is a folder of instructions and supporting files defined by the Agent Skills specification. The tool runs a static quality check first, then compares runs with and without a skill; the example skill text-analyzer scores 86.2/100 (B), exits 0, and the default passing threshold is 70. That score covers static quality only, and publication evidence stays marked incomplete.

Plannotator moves agent review into a local browser: plans, diffs and terminal replies can be annotated and sent back

Plannotator is a local, browser-based review surface for AI coding agents including Claude Code, Codex, Copilot CLI and Gemini CLI. When an agent proposes a decision, plan or HTML artifact, or finishes work, the app opens in the browser so the user can mark it up and send feedback straight back to the agent to act on. The new Plannotator Inbox gathers messages, questions, files and guided reviews from multiple agents into one local window; answering does not hold a session open, and the reply returns to the session that asked.

Claude-BugHunter packs bug-hunting know-how into a skill bundle: 83 skills load by topic, no name-calling required

Sachin Sharma released Claude-BugHunter, a drop-in skill bundle for the Claude Code skills system that splits bug hunting and authorized red-team work into 83 skills and 15 slash commands. The skills are drawn from 681 disclosed reports, 433 of them individually cited and auditable, across 24 core vulnerability classes, with enterprise identity and infrastructure attack matrices. A user describes the target in plain English and the relevant skill loads by topic, with no invocation by name.

Stitch skills ship as a plugin marketplace: six coding agents can install design and build skills on demand

The google-labs-code repository has released stitch-skills, a collection of agent skills and plugins for Google Stitch that follows the Agent Skills open standard. It organizes design, build and utility plugins as an addable marketplace for Codex, Antigravity, Gemini CLI, Claude Code and Cursor, while OpenCode requires manually copying skill folders and configuring the Stitch MCP server. The skills depend on a configured and running Stitch MCP server, and selective installs must include the dependencies between skills.