In briefCounterfactual probing for safe parallel token unmasking in hidden-forest distributions
SummaryThe paper studies parallel sampling from masked generative models, where token dependencies are unknown and must be discovered at extra model-evaluation cost. It proposes a sampler for discrete distributions with hidden forest structure that shares hypothetical-reveal evaluations across dependence tests to find safe parallel batches without fully recovering the forest. Under stated regularity conditions, the authors show total evaluations and sequential depth are sublinear in sequence length, and that comparable irreversible product-commit samplers require more counterfactual submissions or commit rounds in the worst case.
In briefCan vision-language models stay helpful when facing hidden risks? A cheaper training method that teaches safe and useful answers
SummaryThe paper introduces Intent-Privilege On-Policy Self-Distillation (OPSD), a training method that uses evidence-grounded intent as extra supervision so vision-language models can spot implicit cross-modal risks and still give safe, useful answers instead of refusing. It uses only 1,447 safety examples, cutting training time by 5x compared with multi-rollout GRPO-style training and reducing average inference length by 7%. On pooled SIUO and HoliSafe evaluations, the joint safety-helpfulness success ratio rises from 43.9% to 53.5%.
In briefMatching behavior does not mean matching internal representations; neural networks keep traces of their training history.
SummaryResearchers trained paired convolutional networks on reversed task orders, then applied identical common training. In 20 MNIST pairs, 16 met a predeclared behavioral-matching criterion, yet their internal representations still retained a measurable history score and about 3.1% prediction disagreement. Extending common training to 50,000 updates did not remove this difference, even though the accuracy gap was only 0.18 percentage points.
In briefMixture of Self-Improving Branches for Agent Harness Optimization: parallel evolving search branches whose harnesses a router selects per input
SummaryThe paper studies recursive self-improvement of agent harnesses, where agent-generated modifications are refined using execution feedback. Instead of a single search trajectory, the search is organized into branches with evolving development subsets and proposal policies to reduce the risk of local optima. A router selects one branch head harness for each new input before execution. Reported relative gains over Meta-Harness are 34.8% on Olympiad-level mathematical reasoning, 11.6% on Terminal-Bench 2.0, and 3.8% on SWE-bench Lite, with selection based only on development data.
In briefCan a Cacheable Decision Model Follow Rules? A study of a small non-generative decision model
SummaryThe paper studies Certo, a small non-generative decision model built on Qwen3-4B, which scores candidate actions from their text and returns a probability rather than generating an answer. It compares a joint scorer that reads state, rules, and each candidate together against a cacheable independent-encoding design that encodes candidates once for reuse but separates state from candidates. Results show rule-sensitivity drops sharply under the cacheable design, and transfer to unseen real rules is not established, while the joint scorer keeps its edge.</summary>
In briefReCaVSR: One-step streaming diffusion video super-resolution using recycled latents and learned cache routing
SummaryThis paper presents ReCaVSR, a one-step streaming video super-resolution method built on Wan2.2 for real-time online use. It reuses super-resolution latents to keep local temporal context and lets each transformer layer learn its own key-value cache time span. The method also adds a multi-scope discriminator and an LR-conditioned decoder. At 1080×1920 output on a single A100-80GB, it reportedly reaches 21.20 FPS with 15.16 GB peak GPU memory.
In briefWhen Correct Memory Misfires: Fuzzing How LLM Agents Use Stored Memory
SummaryThis paper shows that persistent memory in LLM agents can be used incorrectly even when the stored content is correct, especially when queries change or memory states evolve. The authors frame this as a fuzzing problem and present U-Fuzz, which mutates queries or memory states starting from memory checkpoints to uncover memory-use failures. The reported experiments show U-Fuzz finds more confirmed failures across several memory systems and in output-only settings where retrieval is hidden.
In briefDIET: Deletion-response expert pruning for video diffusion transformers without retraining
SummaryThis paper proposes DIET, a training-free method for pruning experts in mixture-of-experts video diffusion transformers. It records how each expert's removal changes model behavior in one calibration pass and selects retained experts by minimizing Overall Diversity Loss, with a layer-wise budget search. On LingBot-Video 30B-A3B, pruning 50% of experts cuts the checkpoint from 57 GB to 30 GB, fitting a 48 GB GPU, and VBench Total rises from 0.7941 to 0.8115 under a fixed protocol.
In briefPrivy to the Foil: Recasting Value Estimation with a Self-Privileged Critic for RLVR
SummaryThis paper addresses the difficulty of assigning credit to intermediate steps when training language models on multi-step reasoning with sparse terminal rewards. It proposes πPPO, a self-privileged actor-critic framework that reuses verified rollouts from the same prompt as contrastive evidence to help the critic judge intermediate reasoning. The authors report that πPPO substantially improves value-estimation quality and outperforms representative baselines on challenging mathematical reasoning benchmarks.
In briefHow transformer residual states become specific to their final prediction, explained through geometry
SummaryThis paper studies how intermediate residual states in transformer language models become specific to their eventual outputs. Across six pretrained models, the authors compare each intermediate state with its own final state and with final states from other contexts. They find that competing endpoints generally shrink with depth, though directional alignment can improve while Euclidean distance changes little. The paper also offers a high-dimensional model and proofs about straight-path convergence.
In briefMeasuring which language models complement each other when forming an LLM team
SummaryThis paper studies how to select a team of large language models that work well together. The authors propose a framework that profiles each model offline and measures complementarity using two signals: decorrelated error patterns and divergent predictive behavior. Team selection is framed as a combined quality and complementarity objective, solved with an efficient greedy search. Experiments on multiple benchmarks show it outperforms quality-only baselines under controlled candidate pools and team sizes.
In briefDeep Thinking, Direct Speech: Recurrent Latent Reasoning for Paralinguistic Spoken Dialogue
SummaryThe paper introduces LoopSLM, a spoken dialogue model that uses looped Transformers for latent reasoning, reusing a decoder block to refine hidden states with acoustic cues at each pass. A two-stage training process separates learning to reason from learning to respond, allowing direct inference without generating chain-of-thought. On EchoMind, LoopSLM outperforms Qwen2.5-Omni-7B on paralinguistic understanding, reasoning, and reply quality, and against a CoT-SFT baseline it gains over 20 points in reasoning accuracy while generating 64.5% fewer tokens at half the latency.
In briefMinkowski Attractor Networks: closed-form hyperbolic flows for visual representations
SummaryThis paper proposes Minkowski Attractor Networks (MAN), which learn visual representations in pseudo-Riemannian Minkowski spacetime to model hyperbolic geometry with constant negative curvature. The method combines linear Lorentz group transport, non-linear cone lifting, and closed-form radial rescaling, computing in a single forward pass without numerical ODE solvers or iterative retractions. The authors present a two-dimensional version (MAN-2D) as the main visual backbone and a four-dimensional version (MAN-4D) that evaluates Lorentz isometries via two commuting planar maps.
In briefWINGS: Reference-free 3D Gaussian splatting inpainting using a native 3D generative prior
SummaryThis paper presents a method for filling in missing regions of 3D Gaussian Splatting scenes directly in 3D space. It uses the embedding space of a large pre-trained 3D prior together with a structure completion network to generate the missing geometry and appearance, without inpainted reference views. The authors report that the approach is faster than 2D-based methods and demonstrate its effectiveness through experiments and a user study.
In briefExplore, Execute, Evolve: A Skill Loop That Lets Robot Agents Learn and Reuse Skills
SummaryThe paper introduces RoboSkill, a framework that links skill acquisition and reuse through an Explore, Execute, Evolve loop, where the agent updates its skill library from execution records and reuses it in the next cycle. The approach adds tactile feedback and reusable code to improve efficiency. On LIBERO-10, it raises first-episode success rates by 12.5 to 25.0 percentage points and cuts average runtime by 7.6% to 72.4% across four agents. On real robots, it improves success rates by 8.3 percentage points.
In briefPixel-level Transformers for satellite canopy height mapping
SummaryThis paper studies Transformer models that use pixel-level attention to predict canopy height from medium-resolution satellite imagery, a method for assessing forest condition. The authors report that pixel-level attention generally outperforms models relying on larger patch sizes. They also test efficient attention variants to balance prediction quality against resource needs, and compare against established models, finding Transformers can perform better with suitable hyperparameters.
Sven Ligensa, Jan Pauls, Karsten Schr\"odter et al.
In briefProtein optimization that recalibrates predictions using lab feedback
SummaryThis paper addresses how to choose which protein variants to test next under a limited experimental budget. It proposes Batch-Aligned Tail Arbitration (BATA), which uses experimental feedback to adaptively combine prior-informed and task-specific rankings, focusing calibration on the batch-aligned high-fitness region. On measured GB1, PABP, and TrpB landscapes, BATA achieves the best mean task rank (1.67) in final best fitness after 480 measurements.
In briefCompOrca: Compliance Labels for a Large Instruction-Tuning Dataset
SummaryThis paper provides compliance labels for all 4,233,923 examples in the OpenOrca corpus, using an open-weight LLM judge (LongCat-2.0) with five passes per example. Results are released as 94.75% unanimously compliant, 1.28% unanimously noncompliant, and 3.97% split decisions, with raw vote counts. Against 450 human-annotated examples, unanimous compliance and noncompliance labels are 97.3% and 86.7% precise.
In briefByteTraX: a refined ByteTrack tracker with a unified matching threshold
SummaryThis paper proposes ByteTraX, an enhancement of the ByteTrack multi-object tracking architecture. It uses a single unified matching threshold and stricter track initiation criteria to reduce identity switches. Reported results across benchmarks including GMOT-40, SportsMOT and TeamTrack show gains in HOTA, IDF1 and FPS, with identity switches reduced by over 40%. The source code and YOLO integration instructions are made publicly available.
In briefBandits in the Wild: Warm-Started Recommendations for New User Cohorts
SummaryThis paper addresses recommender systems that repeatedly face new users with little interaction history. The authors propose CohortMix-TS, a warm-started mixture bandit that learns latent user groups from earlier cohorts and uses metadata to build priors for new users, combining Thompson sampling with diversity and inventory-depletion controls. The method was evaluated through simulation, semi-synthetic experiments, and a 25-day randomized deployment with 713 participants in a campus quiz application. Results report that cross-cohort transfer improves early recommendation quality and user-level regret.
Serafima Lebedeva, Sumantrak Mukherjee, Ali Arshad Sadal et al.
In briefGLaS-JEPA: Speech Self-Supervised Learning Without Hand-Designed Prediction Targets
SummaryThe paper introduces GLaS-JEPA, which predicts the current encoder's continuous representations at masked positions without contrastive learning, discrete targets, or separate EMA target encoders. It uses SIGReg representation-space regularization to prevent representation collapse. Pretrained on 960 hours of LibriSpeech, the 57M-parameter model reaches 6.89% WER on frozen-encoder SUPERB ASR and 25.87% CER on slot filling, outperforming the best non-distilled sub-90M baselines by 43.1% and 22.0%. The authors conclude that competitive speech representations can come from a much simpler training recipe.
Gaspard Bott\'e, S\'everin Baroudi, Samir Sadok et al.
In briefA neural network that keeps and recalls memories according to context
SummaryThe paper trains a recurrent neural network with an episodic memory buffer that infers context via Bayesian inference while predicting upcoming scenes in naturalistic movies. When inferred context modulates the recurrent connectivity in a low-rank manner, the model's activity best matches fMRI responses from humans who watched the same movies. Context also shapes memory retrieval, which uses both content and context similarity, and this model learns to retrieve memories much faster than one without context modulation.
In briefPredictive self-supervised learning can provably recover random signals despite nuisance noise
SummaryThis paper proves that common self-supervised learning methods that predict in latent space can separate a prediction-relevant stochastic latent signal from observation-specific nuisance information. The authors attribute this ability to two principles: predictive mutual information maximization and latent distribution matching. In simulations with Gaussian predictors, the true signal is recovered up to an affine transformation, even in dynamic, nuisance-laden environments.
WhyProves common latent-prediction SSL methods identify stochastic signals under nuisance, with simulation confirmation for Gaussian predictors, a notable theoretical contribution.
In briefA Proposed Rubric for Scoring Clinical Reasoning in LLM Answers
SummaryThe paper proposes a multidimensional rubric for scoring the clinical reasoning expressed in free-text model responses to gold-standard clinical vignettes. It draws on medical education assessment frameworks, clinical LLM benchmarks and general LLM reasoning evaluation research, and adapts the factuality category as "groundedness". The authors state that the rubric has not yet been tested for inter-rater reliability, construct validity or clinical utility, and its aim is to make evaluation decisions explicit before empirical testing.
In briefConvergence analysis of Adam on generalized smooth objectives using only second-moment gradient information
SummaryThis paper studies whether Adam converges on generalized smooth objectives when stochastic gradients are controlled only by second-moment information, without strong tail assumptions such as almost-sure boundedness or sub-Gaussianity. It extends the self-normalization framework of Jin et al. (2026) to an L0-Lp smoothness condition and a generalized second-moment ABC condition. The authors establish high-probability convergence rates for p<2 with confidence dependence of order δ^{-1/2} and show via a hard instance that this dependence is sharp. For p<1, they also obtain convergence guarantees in expectation.
In briefCHOQOLATE: Organizing Concept Bottleneck Spaces Using Choquet Integrals
SummaryThe paper proposes CHOQOLATE, an interpretable-by-design layer based on 2-additive Choquet integrals that merges correlated concepts into compact nodes in concept bottleneck models built on CLIP-style vision-language models. Across four datasets, it reports a favorable accuracy-interpretability trade-off with sparse, semantically coherent nodes. On standard bias-mitigation benchmarks, suppressing spurious concepts after training performs comparably to methods that require group annotations or retraining.
In briefPlanetary Feature Fields: a compact way to store many Earth datasets together
SummaryThis paper introduces Planetary Feature Fields (PFFs), which model satellite observations, precomputed embeddings, and map products jointly as continuous functions of space and time at planetary scale. Each field uses a hybrid explicit-implicit neural representation, sharing a factored feature volume across products with lightweight decoders for each product. The authors report that at 1800x compression, reconstructed features retain about 90% or more of the original performance on segmentation, change detection, and patch classification, while cutting feature access latency by an order of magnitude.
Arjun Rao, Sebastian Loeschcke, Anthony Fuller et al.
In briefRxnOptBench: Testing LLMs on choosing organic reaction conditions
SummaryThis paper introduces RxnOptBench, a benchmark that asks language models to pick the best catalyst, ligand, solvent, and other conditions from real wet-lab screening tables in organic methodology papers. Items come from 2025 publications and are scored by a continuous relative score based on yield combined with ee, dr, and rr. A paired design with and without literature precedents separates use of in-context evidence from memorized knowledge. The authors report substantial headroom, with chemistry-specialized models near random baseline on multi-axis selection, and release the reviewed test set and evaluation code.
In briefA benchmark and dataset for detecting AI-altered visual evidence in courts
SummaryThis paper introduces the CIFAR Synthetic Evidence Corpus, a benchmark for authenticating photographic evidence in court and justice-system settings. It contains 1,505 items, including 720 authentic controls and 785 manipulated or fabricated images from surveillance, dashcam, and consumer-photo sources. Baseline tests with publicly available manipulation detectors show error profiles that remain problematic for evidentiary use. The dataset, metadata, prompts, code, and evaluation scripts are released.
Kelly McConvey, Sajad Ebrahimi, Nima Jamali et al.
In briefSelecting What Matters: Semantic Compression for Long-Context Embeddings
SummaryThis paper proposes SCSP, a training-free method for building long-document text embeddings. It splits a document into sentence-aware chunks, adds a semantic compression prompt to each chunk, and uses the attention patterns from these prompts to estimate which tokens matter most. The selected tokens' intermediate-layer representations are then aggregated into the final embedding. Experiments on long-context embedding benchmarks show it can be plugged into zero-shot and fine-tuned models and consistently improves their performance.
In briefHiRAE: Hierarchical Autoencoder with Residual Budgets for Better Image Reconstruction and Generation
SummaryThis paper presents HiRAE (Hierarchical Representation Autoencoder), which groups encoder layers by depth and learns residual corrections to the deepest representation, with norm caps that give shallower groups tighter budgets. On ImageNet-256, HiRAE-24 lowers reconstruction FID from 0.299 to 0.209 compared with RAEv2 while keeping competitive guided generation quality. For text-to-image generation, it improves alignment over RAEv2 on GenEval, DPG-Bench, and GenAI-Bench, and post-fine-tuning GenEval rises from 84.86 to 87.70.
In briefOmniVCBench: A Benchmark for Evidence-Based Multimodal Reasoning in AI Virtual Cells
SummaryOmniVCBench is a benchmark that tests how well AI virtual cell models interpret experimental evidence from scientific figures, with answers traceable to their sources. It includes 6,077 question-answer pairs built from figures and experimental context in the scientific literature, covering three scientific reasoning tasks. The paper also introduces AIVC-Judge, an MLLM-based evaluator for open-ended answers, and a method that turns model errors into multiple-choice distractors.
In briefHow accurate must papyrus handwriting recognition be? Character-error tolerance of four papyrological tasks on Greek texts
SummaryThe authors used 63,846 published Greek texts from papyri.info as ground truth and simulated recognition output with character error rates from 1% to 50%. They tested four tasks: document type, dating, documentary-versus-literary classification, and keyword search. Tolerance varied by task, with classification tolerating the most error and dating the least. Retraining models on text containing errors largely removed the sharp performance drop seen above 15% error.
In briefCross-entropy guided routing for sparse mixture-of-experts language models
SummaryThis paper addresses how sparse mixture-of-experts language models choose which experts process each token, noting that current router training does not directly link routing scores to token-level errors. The authors propose two forms of token-error supervision, one predicting per-expert error scores and one aligning router affinities directly with the model objective. Tested on two sparse MoE backbones across four multiple-choice question-answering benchmarks, the method improves accuracy on Granite by about 2.3 points on average over a parameter-matched baseline.
Yury Nahshan, Nati Daniel, Jacob Goldberger et al.
In briefReal-World Performance of Commercial AI for Detecting Pulmonary Embolism Across Hospitals
SummaryThis retrospective study evaluated two FDA-cleared AI algorithms from one commercial platform, one for pulmonary embolism triage on CT pulmonary angiography and one for detecting incidental pulmonary embolism on routine contrast-enhanced CTs, across a 17-facility academic health system. Both models showed sensitivity below FDA-clearance benchmarks while exceeding cleared specificity. Performance declined for peripheral and non-acute emboli, and the authors call for standardized post-market surveillance of AI-enabled medical devices.
Aawez Mansuri, Mohammadreza Chavoshi, Theodorus Dapamede et al.
