In briefType-Checking for Pattern-Based Tree Transformations: checking whether rewriting expression trees with pattern pairs preserves a given tree type
SummaryThe paper introduces pattern-based tree transformations, where each (source pattern, target pattern) pair rewrites matching expressions, such as turning (x·y)+(x·z) into x·(y+z). The model is expressive, but checking equivalence of transformations is undecidable. The authors nonetheless prove that type-checking, which asks whether a transformation preserves a regular tree property, is decidable. Their method reduces the problem to the emptiness problem for alternating tree automata.
In briefTokenBank: Financial Infrastructure for Paying for AI Services
SummaryThis paper presents TokenBank, a financial infrastructure that uses structured contracts to represent commitments among AI service operators. It covers service-consumption rights, forwards that settle API price differences in cash, financing backed by limited future service revenue, and protection claims for specified failures. In evaluation, replaying 899,441 API requests, forwards reduced mean expenditure by USD 304.88 but increased its standard deviation.
In briefReCal: Recovery-aware calibration for structured pruning before on-policy distillation
SummaryThis paper proposes RECAL, a plug-and-play method that improves how pruned reasoning language models recover through on-policy distillation (OPD). Before pruning, it uses forward KL divergence between an unpruned teacher and a pruned probe to find teacher-supported predictions that pruning disrupts, then reweights calibration statistics to preserve them. Across multiple models and pruning methods, RECAL consistently improves mathematical reasoning after OPD, with gains of up to 16.7 percentage points on AIME.
In briefPathLang: A Language-Centered Benchmark for Pathology Vision-Language Models
SummaryThis paper introduces PathLang, a clinically grounded zero-shot benchmark that tests how pathology vision-language models respond to changes in diagnostic wording. It keeps slides and labels fixed and varies only the diagnostic language, covering classification, retrieval, paraphrase robustness, and open-vocabulary diagnosis retrieval. Across nine models and five public datasets, performance proved highly sensitive to clinically equivalent paraphrases, and better image-text alignment did not necessarily yield better class separability.
WhyBuilds a language-centered pathology VLM benchmark varying clinically equivalent diagnostic phrasing, evaluated across nine models and five datasets.
In briefMine Odyssey: A Minecraft benchmark for agentic spatial intelligence in real-world-like places
SummaryThis paper introduces Mine Odyssey, a benchmark that tests whether AI agents can explore, update their spatial understanding, and adapt actions across Minecraft recreations of real locations. It contains 180 tasks over 30 sites in 20 countries and regions. The best evaluated model reaches an 85.6% success rate, while the strongest open-weight model reaches 23.9%, indicating substantial room for improvement.
In briefAlways 10:10: Reference dials fix a clock bias that text prompts can't
SummaryThis study examines a bias in text-to-image models, which tend to draw analog clocks showing 10:10. The authors generated images of three clocks across 52 models, comparing prompts with no time, time written as digits, hand positions described in words, and an added reference dial. Adding a reference dial got all three clocks correct in 75% of images, versus 34% with digits alone.
In briefFinsler Flow Matching: Dynamics-Aware Interpolation for Cell Trajectories from Single Snapshots
SummaryThis paper introduces Finsler Flow Matching (FFM), a framework that learns continuous stochastic dynamics from discrete Markov transition graphs to infer single-cell trajectories from snapshot data. It builds a Finsler structure from first and second local moments, motivated by the Freidlin–Wentzell action. FFM improves recovery of withheld intermediate cell populations, especially when transition dynamics are strongly directional or anisotropic.
In briefDivMoE: Fine-grained MoE upcycling using experts from different domains
SummaryThe paper proposes DivMoE, a method for converting a pre-trained dense language model into a fine-grained Mixture-of-Experts model. The authors find that when all fine-grained experts come from one source model, routing collapses and accuracy falls to near-random levels. DivMoE initializes experts from domain-adapted dense models and uses diversity-constrained routing so each token activates experts from distinct domain groups. The paper reports gains over six upcycling baselines across 15 benchmarks on two base models.
In briefMetaEncoder: Testing How Far Bi-Encoders Go for Multimodal Decisions in Natural Language
SummaryThis paper studies System One models, which output constrained decisions and probability distributions rather than free-form text, using a fully natural language interface for both user requests and candidate options, with image and video inputs. The authors introduce MetaEncoder, a bi-encoder built by fine-tuning a pre-trained Muse-Glimmer 30B decoder and trained with unidirectional contrastive learning to align requests with candidates. The method is designed to handle both small closed candidate sets and very large open candidate sets. It is evaluated on 11 benchmark suites covering 190 tasks, where it is compared against state-of-the-art multimodal encoders and shown to have limits on reasoning-intensive tasks.
In briefHSVDQuant: low-rank 4-bit quantization for diffusion transformers
SummaryThis paper presents HSVDQuant, a framework for 4-bit weight-activation (W4A4) post-training quantization of multimodal diffusion transformers. It treats low-rank compensation and residual quantization as a coupled calibration problem, using a deflated Hessian and an activation-noise surrogate. Across five diffusion backbones and the Qwen3-8B LLM, rank-4 HSVDQuant is reported to outperform rank-4 and rank-32 SVDQuant baselines.
In briefFrom Retrieval to Reconstruction: Building Evolvable Cognitive Memory for Long-Term Dialogue (CogMem)
SummaryThe paper argues that existing retrieval-augmented generation (RAG) frameworks treat memory as passive storage, making it hard to separate source-attributed beliefs from unattributed events and facts, or to link evidence scattered across sessions. The authors propose CogMem, a cognitive memory architecture built on a PEC²F (Person-Event-Concept-Claim-Fact) graph schema, with dedicated Claim nodes that record the source and target of subjective statements. Experiments on LoCoMo and LongMemEval report strong results, particularly on multi-hop, temporal, and knowledge-update tasks.
In briefWhy finding the smallest two-layer ReLU network is computationally intractable
SummaryThis paper studies whether the minimum number of hidden neurons needed for a two-hidden-layer ReLU network to approximate a target function within a given error tolerance can be computed efficiently. The authors prove that computing this optimum exactly is NP-hard for every fixed input dimension, output dimension, and L^p error with p at least 1. The result gives theoretical support for using heuristic methods when designing such networks.
In briefMedBenchAgent: Systematic Automation of Medical VLM Benchmark Construction
SummaryThis paper presents MedBenchAgent, a multi-agent framework that automatically derives the specification of medical vision-language model (VLM) benchmarks, not just the evaluation items. It formulates construction as constrained compilation using a Benchmark Intermediate Representation, separating planning from instantiation. The authors report a Task-Space F1 of 90.9%, with 994 of 1,000 sampled items passing human audit, and evaluate twelve VLMs.
In briefFloorSAV: Giving audio-visual AI models a 2D floor map to reason about space
SummaryThis paper proposes FloorSAV, a framework that renders a dynamic 2D floormap from 3D point clouds, camera trajectories, spatial audio, and object landmarks. The floormap is fed to an audio-visual large language model as a stream synchronized with egocentric video. The authors also introduce SAVED-Bench, a benchmark of dynamic relativity, regional, and path reasoning questions. The method improves spatial reasoning on SAVED-Bench and SAVVY-Bench.
In briefFrom Geometry to Generalization: Why Row Normalization Can Beat Adam and Muon
SummaryThis paper studies whether different optimizers, which fit the same training data but select classifiers with different geometries, differ in population performance. The authors prove that row-wise normalization can achieve strictly higher population accuracy than full-batch Adam, a proxy for random-reshuffling Adam, and exact-SVD Muon in high-dimensional multiclass classification. The advantage is analyzed under an isotropic Gaussian-cloud model and extended to more general covariance settings, and is supported by synthetic and last-layer language-model experiments.
In briefAdversarial distillation that also uses teacher-favorable examples to make small models more robust to attacks
SummaryThe paper proposes CGARD, an adversarial robust distillation method that transfers robustness from a large teacher model to a compact student model. Within the same perturbation neighborhood, it jointly optimizes student-adversarial examples and teacher-collaborative examples, combining the latter with adversarial teacher supervision. Experiments on CIFAR-10 and CIFAR-100, including white-box and black-box transfer evaluation, report consistent robustness gains over strong baselines.
In briefBeliefScope: Diagnosing Whether LLMs Change Answers Due to Evidence or User Pressure
SummaryThis paper introduces BeliefScope, a black-box framework that separates shifts in a language model's response to the same proposition caused by relevant evidence from those caused by directional user pressure with no added fact. The authors test when these contrasts can be reliably attributed using controlled synthetic conditions and a 36-family Qwen/Llama study, with targeted checks including Gemma3-12B. The results are summarized as a conditional belief-response profile with stated validity boundaries.
In briefAn efficient method for transferring multi-level medical knowledge to generate radiology reports
SummaryThis paper proposes EMGKT, a method that automatically generates clinical descriptions from X-ray images to support radiologists. It uses a medical vision-language model for global knowledge and a fine-grained knowledge distillation task so the model learns detailed knowledge of key areas. At inference, only the student embeddings are used, so computational cost stays negligible. A mixture of disease diagnosis expert classifiers, initialized with disease embeddings, is also added. Experiments on two public datasets are reported to show its effectiveness and transferability to existing methods.
In briefiCATS: Faster Video Generation with Interaction-Aware Sparse Attention and Adaptive Sparsity
SummaryThis paper proposes iCATS, a training-free method that speeds up video generation in Diffusion Transformers by using sparse attention. It improves importance estimation by clustering based on query-key dot-product interactions and adjusts sparsity over denoising timesteps using an SNR-guided schedule. The authors report 2.03× acceleration at 31.017 dB PSNR on HunyuanVideo-T2V-13B and 1.55× acceleration at 29.301 dB PSNR on Wan2.1-T2V-14B.
In briefWhy code-generating AI models can be confidently wrong
SummaryThis paper studies why large language models that generate code often produce incorrect programs with confidence similar to correct ones. The authors test four open-source code models on three execution-based benchmarks and examine whether uncertainty metrics and token-level confidence reveal execution failures. They find that existing uncertainty signals give only partial, model-dependent evidence of failure, and that common mitigation strategies do not reliably resolve the overconfidence problem.
In briefDuplexAgent-RSI: Self-Improving Coordination Framework for Full-Duplex Voice Agents
SummaryThis paper presents DuplexAgent, a full-duplex voice agent system that keeps conversation live while delegating complex reasoning and coding tasks, coordinated by a harness split into six editable modules. It also introduces Duplex-Harness-RSI, a closed loop in which a simulator generates timed test conversations and failure traces guide revisions to the relevant modules. Experiments on intelligence, agentic, and duplex benchmarks report stronger spoken-knowledge and executable-tool scores than compared delegated systems, with strong interruption response, and show the loop outperforms repeated editing without diagnosis and repair archives.
In briefCombining MLP and KAN branches with wavelet fusion to represent multidimensional signals more accurately
SummaryThis paper proposes an implicit neural representation that pairs an MLP branch, which models smooth low-frequency structures, with a Kolmogorov-Arnold network branch that captures localized details. The two outputs are fused in the wavelet domain using discrete wavelet transform coefficients, with a band-separation regularization encouraging complementary frequency behavior. Experiments on 1D signals, 2D images, 3D volumes and SDFs, videos, and 4D light-fields report improved reconstruction fidelity across these modalities.
In briefORDO: Learning the best order of steps for speeding up mixed-integer optimization solvers
SummaryThe paper argues that the order of presolve actions strongly affects mixed-integer programming (MIP) solve time, while existing learning methods only tune parameters and cannot capture dependencies among actions. The authors recast presolve planning as autoregressive sequence generation over a unified atomic action space, and call the framework ORDO. It reports end-to-end zero-shot speedups on multiple unseen domains, and deployment uses concurrent sequence racing with the winning sequence kept.
In briefWhen to Use On-Policy Distillation: Offline Student Rollouts Often Work Better
SummaryThis paper compares on-policy distillation (OPD) with Semi-OPD, which distills from offline rollouts generated by the initial student model. Across 17 teacher-student pairs from 1.5B to 235B parameters, Semi-OPD outperforms OPD in 14 cases, with up to 13.6% higher accuracy and 11.4x faster training. The authors find that OPD helps only when the initial teacher and student are highly aligned, as measured by an output-token overlap ratio.
In briefOpen-ended scientific discovery using possibilistic reasoning, a framework for LLM-driven hypothesis testing that stays statistically valid.
SummaryThis paper addresses autonomous scientific discovery with LLMs, which must generate and test hypotheses as evidence accumulates while keeping statistical validity. The authors formalize this as Abductive Autonomous Scientific Discovery (AASD) using possibility theory, and introduce a computable measure of progress called abductive utility. They also propose possibility frontier search, the first algorithm for AASD, which is anytime valid and asymptotically reaches ε-optimal abductive utility under suitable conditions. Experiments on synthetic and real-world discovery tasks show strong performance.
In briefMemorization and Malign Generalization in Conditional Diffusion Models
SummaryThis paper theoretically studies how conditional diffusion models memorize training samples and generalize to new ones, using a random-feature conditional score model in a high-dimensional limit. The authors derive asymptotic formulas for training and test losses, and show that in the overparameterized regime, wider models better predict the condition-dependent mean but reduce within-condition variance, a effect they call "malign generalization." They also find that more informative conditions cause memorization at smaller model widths. Experiments with U-Net architectures on realistic data support these findings.
In briefREMORY: Learning residual memory to keep agents' context history usable after compression
SummaryLong-horizon agents compress their history to fit a limited context window, but a text summary alone may not support every later decision. REMORY is a neural memory network that adds a bounded sequence of soft memory tokens to the summary, learned to help a frozen LLM approximate its continuation with full history. On SummHay, it improves source attribution with nearly unchanged insight coverage, approaching full-context joint scores using only 5.2% of input positions.
In briefHow text drawn in an image can override the rest of the picture, and a first fix
SummaryThis paper studies a phenomenon in image generation models where words rendered inside an image carry meaning that the model may treat as part of its instructions, affecting non-text areas. The authors separate the main visual prompt from the rendered text and measure leakage and rendering fidelity. They show that harmful semantics in scene text can persist through LLM-based prompt enhancement even when the main prompt is benign. They also propose a preliminary mitigation on FLUX-2-dev that reduces unsafe semantic transfer while preserving text rendering.
In briefDynaTE: Speeding Up Diffusion LLMs with Dynamic Token Execution
SummaryThis paper presents DynaTE, a hardware-software co-design accelerator for diffusion-based large language models (dLLMs), which adjusts execution to token states during decoding. It skips computation for low-utility tokens and refines locally dependent tokens to reduce the number of denoising iterations. Evaluated on two representative dLLMs, DynaTE reports 2.05–2.78× speedup and 2.99–3.93× higher energy efficiency over state-of-the-art dLLM accelerators, and 2.55× speedup and 6.07× higher energy efficiency over Jetson AGX Orin.
In briefBeing-M0.7: A latent world-action model for humanoid robot control
SummaryThis paper introduces Being-M0.7, a latent world-action model for humanoid loco-manipulation. It learns visual-motion priors from over 10,000 hours of human data, including video-only, motion-only, and paired streams, then adapts them to robots through mid-training and action post-training. The model achieves the highest aggregate success rate among compared baselines on SIMPLE and matches the strongest baseline on real Unitree G1 tasks.
In briefPost-training on a Stand-in Reward: Envelope Sampling Curbs Reward Hacking
SummaryLarge language models are often post-trained against cheap stand-in rewards, such as LLM judges, because true rewards like human preference are costly to query. A miscalibrated stand-in can lead to reward hacking, where optimization produces unwanted side effects. The authors propose envelope sampling, which recalibrates the judge using a small set of expert-labeled outputs by minimizing an upper bound on post-trained model regret. Experiments on clinical note generation and a controlled sycophancy task show that recalibrating on envelope samples reduces reward hacking better than recalibrating on base-model samples.
Sanjit Dandapanthula, Shuvom Sadhuka, Samir Khan et al.
In briefPhonological Interference in Multilingual Speech Models: Single-language sound rules leak into code-switched and unseen-language speech
SummaryThe paper shows that phoneme-level speech models, which assume input comes from one language, lose sounds that exist only in the other language when given code-switched speech or unseen languages. On code-switched input, three models lost 32% to 79% of language-specific phonemes. The authors propose windowed language estimation (WLE), an inference-time fix that re-estimates the language from a short window around each position, removing 34% to 69% of this interference while leaving monolingual performance essentially unchanged.
In briefGated Memory: admission-controlled memory formation for chatbots
SummaryPersonalized conversational AI stores facts extracted from user messages in long-term memory. The paper argues that the moment a fact is first written is under-studied, and that context lost during extraction cannot be recovered later. It proposes Gated Memory, which adds an admission gate checking facts against the full utterance and a conditional enrichment stage with entity scopes and privacy constraints. On the LoCoMo-10 benchmark it reports an overall 2.6% relative improvement in LLM-judge accuracy over a strong baseline.
In briefTaReD: Tool-aware recursive task breakdown for long, multi-step AI agent tasks
SummaryThe paper notes that AI agents running reasoning and actions along a single chain become unreliable on complex tasks, as long histories obscure dependencies and early errors spread. The authors propose TaReD, which organizes tools into a hierarchy of capabilities, discovers tools on demand during execution, and recursively splits a complex task into a subtask tree aligned with the capabilities needed at each stage. In experiments on complex real-world tasks, the method raised end-to-end task success rate by up to 40 percentage points over the compared baselines. Code is available on GitHub.
In briefLR-V2X: Loss-Resilient Vehicle Collaborative Perception Over Low-Bandwidth Links
SummaryThis paper proposes LR-V2X, a framework that lets connected vehicles share sensing information reliably over low-bandwidth links with packet loss. It turns corrupted received latent features into a spatial prior and reconstructs missing bird's-eye-view content using the ego vehicle's context. The model is trained only under complete communication and applied directly to lossy conditions. Experiments on DAIR-V2X and V2XREAL report the strongest robustness under severe packet loss and a 64× reduction in communication overhead versus dense BEV feature fusion baselines.
In briefSLICE-Mem: A lightweight memory framework for efficient LLM agents
SummaryThis paper presents SLICE-Mem, a lightweight framework for long-term memory in large language model agents. A compact classifier selects user turns worth keeping, and a semantic-dependency extractor turns them into concise memory entries, with no generative LLM calls in the online stage. Cross-entry reconciliation is handled in an offline consolidation step. Compared with the MemOS baseline on Memora Weekly and LongMemEval-S, the online pipeline reaches similar accuracy while cutting total LLM token use sharply.
In briefResidual spectral instabilities in representation learning: a spectral cascade explaining stepwise latent-dimension collapse in VAEs
SummaryThe paper treats dimension-wise posterior collapse in variational autoencoders as a fluctuation theory around partially collapsed states, using the negative evidence lower bound as an effective free energy. It derives the exact mass spectrum of collapsed directions and a local reactivation criterion based on the residual spectral upper edge. Numerical continuation experiments show latent dimensions being lost successively near these spectral marginalities.
In briefLLM-IDEA: An agent that checks whether a stalled scientific model can be identified from data
SummaryThe paper proposes LLM-IDEA, a closed-loop autonomous discovery agent with an identifiability engine that returns one of three verdicts when progress plateaus: a capability limit, resolvable within the same experiment design class, or certified exhausted. It is evaluated on ODEBench, the DiscoverPhysics benchmark, and a two-body testbed called Alien Universe proposed by the authors. The results report how often the engine's verdicts match known identifiability and how discovery depth changes with the identifiable protocol.
WhyProposes a three-way identifiability verdict for plateaus, validated across several benchmarks and a new testbed; novel and well-evidenced, with impact mainly within autonomous scientific discovery.
In briefA fuzzy system that models human-like car-following driving using five memory types
SummaryThis paper introduces NeMeFIS, a hierarchical machine learning model that separately models acceleration and deceleration in car following by combining five human memory types. Tested on Arterial, Collector, and Rural Highway corridors with 54 trained models, it reproduces realistic driving behavior better than Linear Regression, ANFIS, and LSTM. The authors suggest the framework can model driver cognition and support safer decision-making in connected and autonomous vehicles.
In briefV-CoLA: Compressing Visual Tokens for Linear Attention Models
SummaryVision-language models are computationally expensive because visual tokens dominate the input sequence. This paper introduces V-CoLA, a training-free token compression framework designed for linear attention, combining a uniqueness-aware importance criterion with adaptive token merging. Experiments report retaining 99.5% of original performance with 50.0% of vision tokens, and over 88.0% with 12.5%, along with a 1.86x to 6.15x prefill speedup. The code is publicly available.
WhyProposes training-free vision token compression tailored to linear attention, validated across multiple benchmarks with clear speedups; useful for a subfield.
In briefWhy On-Policy Distillation Sometimes Fails: Vanishing Learning Signals
SummaryThe paper studies why on-policy distillation (OPD) with larger teachers hits early loss plateaus in code and math tasks. After 200 updates, OPD reduced loss by an average of 25.1%, versus 96.2% for self-RL teachers. The authors analyze OPD as a continuous-time dynamical system and prove a conditional local recovery guarantee for teachers close to the initial student. They hypothesize that limited representation adaptation contributes to learning-signal collapse, which remains untested.
In briefQuery-adaptive quantization for KV caches: reading only the precision each query needs
SummaryThis paper studies how to store and read key-value (KV) caches in large language model decoding, where stored entries must be kept before future queries are known. The proposed method, ReadKV, stores each key and value as a progressive code and allocates read precision per query, computing attention from the reconstructed keys and then allocating value precision. Across six base models, reading four bits on average from an eight-bit cache raises C4 perplexity by at most 0.66%, using about one quarter of the logical reads of a 16-bit cache.
WhyProposes query-adaptive progressive reading of KV caches, with exact allocation proofs and experiments on six models and GPU latency measurements.
Siddharth Bhandari, Lucas Gretta, Krishna Balasubramanian et al.
In briefReTTT: Reflection-Driven Test-Time Training with Experience-Based Memory
SummaryThis paper introduces ReTTT, a method that lets LLM agents reflect on feedback during multi-turn tasks, store the resulting lessons in explicit memory, and learn from them via an episode-local LoRA adapter while the backbone stays frozen. The adapted model both acts and reflects. Evaluated on ALFWorld, ScienceWorld, and SWE-bench Lite with two model families, ReTTT outperforms trajectory-based test-time training in all six settings, improving task success by up to 12.14 percentage points.
In briefTAP3D: Reconstructing 3D Human Point Clouds from Body Heat
SummaryThis paper presents TAP3D, a system that reconstructs 3D human point clouds using low-cost thermal array sensors that detect body heat signatures. It aims to offer advantages over LiDAR, radar, and depth cameras in cost, density, and privacy. A physics-informed design handles depth estimation, thermal interference, and separating multiple people, and it was evaluated on a dataset of 160K samples. Reported results include 91.46% accuracy for fall detection, 21.86 cm MAE for indoor tracking, and 4.87 cm error for human mesh recovery.
WhyFirst system reconstructing 3D human point clouds from thermal arrays, with physics-informed modules validated on a 160K-sample dataset for fall detection, tracking, and mesh recovery.
In briefAttributing How Diffusion Outputs Are Shaped, Not Just Which Training Images Matter
SummaryThis paper proposes CADT, an attribution method for diffusion image models that traces generated outputs back to training examples. It builds matched counterfactual pairs at identical noisy states to isolate factor-specific representation shifts, and records how these shifts evolve across the denoising process as trajectory descriptors. According to the authors, experiments on multiple public datasets show consistent gains over existing diffusion attribution baselines across hierarchical, compositional, and style attribution.
In briefBreaking the Group Size Barrier: Parameter-Efficient Group Dance Generation with Chain-of-Dancers
SummaryThis paper addresses generating coordinated multi-dancer choreography from music. The authors propose ChainDance, which treats group dance as a sequence of per-dancer generations, allowing one model to handle variable group sizes without retraining while keeping each dancer's identity. It builds on a frozen single-dancer diffusion model with two lightweight modules and a training-free noise optimization step at inference for spatial coherence. On AIOZ-GDance, it reports state-of-the-art motion quality and group coordination with 3-4 times fewer parameters and 3-6 times less training time.
In briefCalibrating Low-Cost Indoor Air Sensors: A Dataset, Test Scenarios, and a Lightweight Model
SummaryLow-cost air quality sensors suffer from nonlinear distortion, noise, and drift, so they need calibration. The authors release a six-month dataset from five indoor locations, pairing low-cost sensors with reference sensors and contextual metadata. They define four evaluation scenarios covering spatial generalization and robustness to gradual and abrupt shifts, and propose a lightweight temporal model. Experiments report strong calibration performance across all four scenarios with low edge-inference cost.
In briefMiniVer-V: Finding the Smallest Set of Evidence That Verifies Short Videos
SummaryThis paper addresses how to decide which evidence is enough to verify a claim in short-video fact-checking, proposing evidential sufficiency as the selection criterion. The authors introduce MiniVer-V, a benchmark of 195 short videos with supported, refuted, or insufficient labels and 5,510 multimodal evidence units, along with a two-layer verification framework and a sufficiency-driven greedy search. With Claude Sonnet 4, the method uses about 4.5 evidence units on average (16% of the full set) and reaches a Macro-F1 of 0.510, statistically indistinguishable from the full-evidence baseline (0.518), while better recognizing insufficient cases than the same search without abstention.
In briefHarness Compilation: Which Decisions a Small Vision-Language Model Should Keep
SummaryThe paper introduces Harness Compilation (HC), an offline procedure that adjusts how work is divided between a frozen small vision-language model and its external harness. A large teacher revises reusable content and control using the student's execution traces, and a validation set selects the deployed harness, with no weight updates or teacher calls at deployment. Across seven visual question-answering settings, HC improves scores over bare students by 9.9 to 23.9 points.
In briefSteerCast: Retrieval-based latent steering to improve time series forecasting without retraining
SummarySteerCast is a method that improves a decoder-only time series forecaster at inference time without updating its parameters. It builds a database from training data storing history representations and steering vectors, defined as differences between ground-truth-induced and model-predicted representations. At test time, it retrieves similar histories and injects their aggregated steering vectors into hidden states during autoregressive generation. Experiments across multivariate benchmarks and horizons report improved accuracy over the fine-tuned backbone and retrieval baselines.
Van Dai Do, Huu Hiep Nguyen, Minh Hoang Nguyen et al.
In briefMultimodal graph retrieval for sequential recommendation using collaborative filtering paths
SummaryThis paper proposes MGRASRec, a framework that helps multimodal large language models make sequential recommendations. It retrieves collaborative filtering paths tied to each candidate item from a user-item interaction graph, expanding coverage with multimodal similarity, and places these signals directly into the prompt. Inference requires only a single forward pass per candidate, and the method achieves the best results on all metrics across three public datasets, with especially strong gains in ranking quality.
In briefWhen Lower Reconstruction Loss Hurts: Distributionally Robust Refinement for Low-Bit LLM Quantization
SummaryThis paper shows that weight-only post-training quantization, which relies on minimizing reconstruction loss, does not always improve performance on new tasks and can even degrade it on the calibration data. The authors propose Distributionally Robust Quantization (DRQ), a post-hoc method that minimizes worst-case reconstruction loss over a constrained set of input activation distributions. DRQ refines integer codes within the existing quantization grid while keeping quantization parameters and inference operators unchanged. Experiments report gains for six representative PTQ methods, including AWQ, GPTQ, and ParoQuant, with no added inference overhead.
In briefSafeInferCom: Verifier-guided mid-generation checks that keep robot task plans valid
SummaryLarge reasoning language models can plan multi-step robot tasks, but continued reasoning may overwrite valid intermediate plans or leave constraint violations unfixed. The paper proposes SafeInferCom, a formal verifier-guided framework that monitors intermediate plans during generation, preserves valid ones, and directs error correction. Experiments across several models and planning domains show improved planning success and faster correction compared with one-shot inference, and lower token usage when combined with iterative refinement. The authors also evaluate it in VirtualHome and in a real-world robotic-arm demonstration.]
In briefCross-species learning aligns mouse and human brain activity and tracks drug effects
SummaryThe paper introduces a dual-rule contrastive learning framework that organizes electrophysiology representations by biological state rather than species, aligning corresponding mouse and human states. In epilepsy, it distinguished relationships between three mouse models and heterogeneous human patient groups. Drug-induced movement of treated mice toward the human-aligned healthy state retrospectively matched known clinical efficacy across ten model-drug combinations. The framework also identified shared disease-associated neural dynamics between Fmr1-knockout mice and human 16p11.2 copy-number variant carriers.
Marko Tvrdic, Justin Richmond Domingo, Jae Ann Buenaluz et al.
In briefHow Firm Should a Robot Grasp Be? A Visuotactile Study on Estimating Materials to Set Minimum Grip Force
SummaryThis paper proposes a visuotactile method that estimates an object's mass, elasticity, and surface friction in real time during grasping. The estimates are used to determine the minimum grasp force needed to hold the object securely without damaging it. The authors built a new dataset of real fruits and vegetables with measured physical properties and trained a force estimation model using simulations. Experiments with a parallel-jaw gripper show the robot can gently grasp a wide variety of objects, adapting to each one's properties.
In briefSocial World Model Evaluation Gap: Do Tests Check Whether AI Predicts Others' Responses?
SummaryThe study audits 113 works from 1999 to 2026 to ask whether evaluations of social world models directly test whether a system predicts how a person will respond. Only 2 of 9 primary evaluations that measured real human responses also compared an explicit response forecast with the actual response. The authors propose a 1-step negotiation experiment to separate these questions. The findings describe only the audited sample.
Denglin Jiang (New York University), Annie Wang (New York University)
In briefThe Lattice of Transition Laws: A Unified Cost Framework for Decoding Schedules in Diffusion and Autoregressive Models
SummaryThe paper treats diffusion, autoregressive, and hybrid generative models as paths on a single corruption lattice, defining the cost of a decoding schedule by the dependence its parallel steps discard. It shows that for data Markov on a graph, the minimum number of steps for a zero-cost schedule equals the graph's treedepth. Using a pairwise dependence kernel estimated from pretrained weights, the authors predict schedule rankings before decoding, and verify most predictions across text, image, and video generation.
WhyUnifies diffusion and autoregression as paths on a corruption lattice, characterizes minimal steps by treedepth, and validates predicted decoding rankings across text, image, and video.
In briefGATOR: Generating and refining 3D objects with textures and pose from casual photos
SummaryThe paper presents GATOR, a framework that reconstructs textured 3D objects and their pose relative to a scene from one or more casual images. It combines local modality mixing, text-guided generation for structure, geometry, and appearance, and a multimodal agent that refines results through an edit-render-review loop. According to the authors, GATOR shows strong geometric and appearance fidelity and pose recovery across synthetic objects, cluttered tabletops, and indoor scenes.
In briefLinking KV-cache quantization and linear attention, with RAM-Net weight migration from pretrained Transformers
SummaryThe paper contrasts KV-cache quantization, which compresses each key-value entry but keeps all of them, with linear attention, which merges history into a fixed-size continuous state. It proposes RAM-Net, which uses soft assignments over a discrete address space to unify both ideas, and proves under a restricted construction that these soft assignments extend hard quantized matching and locally approximate full-attention similarity. A soft-quantized intermediate construction enables migrating weights from Transformers to RAM-Net. Across nine pretrained models from 0.3B to 7B parameters, using a 500M-token budget per model, RAM-Net recovers on average 87.1% of the teachers' accuracy gains over random guessing on six commonsense and knowledge tasks.
In briefAliO: Why Consistent Outputs Matter in Long-Term Time Series Forecasting
SummaryThis paper studies long-term time series forecasting, which predicts future data from current sequences for uses such as weather and electricity planning. The authors find that current models give inconsistent predictions for the same future timestamps when the input sequences are lagged, which reduces reliability. They propose AliO, a method that reduces these discrepancies in both the time and frequency domains, and a new metric, TAM, to measure output alignment. Experiments report up to 58.2% improvement in TAM while maintaining or improving forecasting performance by up to 27.5%.
In briefAnalyzing Representation Structure in Vision Transformers: A Rigorous Architectural Study
SummaryThis paper systematically analyzes feature information in Vision Transformers (ViTs) across different architectural scales and examines how representation structure relates to generalization. The authors identify feature collapse at initialization and propose a reduction scheme, and show that entropy and minimum eigenvalue can indicate generalization. They also find that features from linear submodules are important for predicting generalization, and their proposed proxy improves correlation ranking by 18-48% over prior baselines.
In briefDo Flatter Minima Drive Better Generalization? A Study of Sudden Learning in Neural Networks
SummaryThis paper examines whether flat loss landscapes actually cause neural networks to generalize, using grokking, where models memorize data for a long time before suddenly generalizing, as a testbed. Sharpness-aware minimization (SAM) alone does not reliably trigger this transition, but combined with weight decay it can speed it up by up to 4x at the epoch level. A theoretical analysis with a minimal two-layer ReLU model shows that flatness alone cannot tell memorizing and generalizing solutions apart, while weight decay favors generalizing ones.
In briefUMAS: System-level uncertainty measurement for multi-agent LLM systems
SummaryThis paper introduces UMAS, a parameter-free, cost-neutral framework that measures uncertainty at the level of an entire multi-agent LLM system. It estimates how credible influence between agents is, based on each generation's intrinsic confidence, and uses this to update node-level uncertainty before aggregating it into a system-level estimate. On Debate and DyLAN across several benchmarks, UMAS outperforms state-of-the-art baselines by an average of 10.54 AUROC points. The authors also report that it supports hallucination detection and uncertainty-aware answer selection, improving accuracy by up to 12.6 points on specific tasks.
WhyProposes system-level uncertainty quantification exploiting multi-agent structure, beating baselines by 10.54 AUROC on Debate and DyLAN; evidence is solid but impact is limited to these frameworks.
In briefPageWeaver: KV-Guided Query Unions for Sparse Attention
SummaryThe paper presents PageWeaver, an execution design for sparse attention that groups queries by their selected KV-page affinity so that page loads are shared and Tensor Core tiles are better populated. Using FP8 KV on H200, the Union8 implementation reports a 1.70x geometric-mean complete-call speedup over the measured FlashInfer path across six captures. The authors also study online regrouping, its latency and prefill throughput effects, and cases on B300 where preparation cost and a stronger native kernel erase the advantage.
In briefComparing Classical and Deep Learning for Sensor-Based Activity Recognition with Group-Aware Evaluation
SummaryThis study compares four classical machine learning models and four deep learning models for recognizing human activity from wearable inertial sensors. Experiments on the UCI-HAR and KU-HAR datasets test accelerometer, gyroscope, and combined signals, using subject-grouped cross-validation. Deep learning did not consistently outperform classical methods, and the combined sensor setting generally performed best. The authors also report that random splitting noticeably inflates measured performance.
In brief3DTexMOR: Removing Multiple Objects from 3D Gaussian Scenes via Texture-Space Inpainting
SummaryThis paper presents 3DTexMOR, a method for removing multiple target objects from reconstructed 3D scenes and completing the occluded geometry and appearance. It performs inpainting in a shared texture space that aggregates multi-view observations, then reprojects the completed texture maps to supervise 3D Gaussian scene completion. The method uses view-independent intrinsic attributes instead of RGB colors and adds geometric regularization to the completed regions. The authors report improvements over existing methods, including a 5.8 dB PSNR gain and at least a 22% LPIPS reduction.
In briefSelective Listening: Controlling How Audio Sways Audio-Language Models
SummaryLarge audio-language models can have their text-based answers changed by audio that is irrelevant to the task. The paper introduces ICAP-Gate, which controls each model's late audio pathway based on the task and the model's internal mechanism. Across four models, two reasoning benchmarks, and two kinds of interference, it shows lower point estimates for influence rate and answer flip than ungated inference.
In briefOmniDex: Scaling Dexterous Hand Grasping to Diverse Cluttered Scenes
SummaryThis paper addresses dexterous robotic grasping in cluttered scenes, where large-scale training data is scarce. The authors build a benchmark of over 2.6 million scenes with 0.4 billion scene-specific grasp ground truths, and propose the OmniDex model, which combines Soft Winner-Takes-All learning with physical constraints and physics-driven ranking. The reported experiments show state-of-the-art performance and strong generalization across scenes, views, and unseen objects.
In briefOne Shared Model for Multiple Compute Budgets Across Conversation Turns
SummaryThe paper introduces Multi-Ratio ConceptMoE, a single shared language model that supports token-to-concept compression ratios of 2, 4, and 8 so compute can be assigned flexibly across turns. Its three fixed modes outperform compute-matched MoE baselines in all six comparisons on PTBench and SFTBench. On CodeChat and CoSQL, mixed-ratio inference improves accuracy while reducing FLOPs compared with routing among independent models.
In briefIs the Future Predictable? Reliable Temporal Reasoning via Future Predictability Modeling
SummaryLarge language models often give confident answers about the future even when the queried facts are highly volatile. This paper introduces ChronoRegime, a task asking models to judge whether historical temporal evidence justifies committing to a future answer. It also proposes CTR, a framework that estimates future predictability from temporal dynamics and commits to answers only when the evidence supports them. Across multiple LLMs, CTR improves strict overall accuracy by 6.7 points and reduces hallucination rates by 18.4 points on average over a plain QA baseline.
In briefRecursive Semantic Divergence: An Unsupervised Metric for LLM Agent Consistency
SummaryThis paper observes that LLM agents with tool use can behave inconsistently across independent runs from the same starting state. The authors propose Recursive Semantic Divergence (RSD), a trajectory-level consistency metric that uses natural language inference to detect logical disagreement between rollouts and is approximated by a neural network. Used as a fine-tuning reward, RSD reduced variation in task progress from 132% to 91% on tau2-bench and from 90% to 44% on toolsandbox.
Harshavardhan Abichandani, Penny Chong, Atin Ghosh et al.
In briefDeciding What Memory Enters and How It Is Shown in LLM Agents
SummaryThis paper studies persistent memory in LLM agents, separating admission (which recalled information enters the working context) from presentation (how admitted information is expressed). The authors implement two retraining-free inference-time designs that reduce pooled judge-assessed failure rates on an external benchmark compared with verbatim injection. The results suggest selection quality yields task-specific safety gains, while preserving beneficial memory use remains unresolved.
In briefRGBD-to-3D Mesh Refinement Using Depth Matching and Symmetry Propagation
SummarySingle-view 3D reconstruction methods often produce meshes that do not match the input view, especially near depth edges and self-occluded regions. This paper presents a lightweight, plug-and-play refinement method that improves any RGB-to-3D reconstructor without retraining. It corrects the visible surface by matching to depth points, transfers corrections to the occluded side across a detected symmetry plane, and propagates them with a smoothness solver. All stages are closed-form, making it much faster than optimization-based test-time refinement.
In briefPredictive Multiplicity in Cell-Fate Assignment: Label-Free Rashomon Sets and Per-Cell Certification Limits
SummaryThis paper studies single-cell trajectory inference, where several models that fit the data equally well can assign conflicting fates to the same cell. The authors propose FateMultiplicity, a label-free framework that builds a statistically admissible set of models (a Rashomon set) without lineage labels. On simulation ground truth, the per-cell certified fate margin discriminates misassignment less well (AUC 0.682) than the baseline configuration's own decision margin (0.965), so per-cell certification does not yield more reliable fate calls, though measuring and reporting multiplicity is worthwhile.
In briefWorldFact-Bench: Checking Whether Images Agree with Real-World Facts
SummaryThis paper introduces WorldFact-Bench, a benchmark of 1,274 real-fake image pairs across four verification regimes and ten semantic domains, for testing whether a single image is consistent with real-world facts. The authors also propose PERSIST-Agent, which iteratively verifies images using a persistent state of candidate facts, observations, and evidence. On evaluated 8B backbones, PERSIST-Agent improves pair accuracy over direct judgment and retrieval-augmented baselines.
In briefRAG-Stress: Testing How Far Retrieved Evidence Can Mislead AI Answers
SummaryThis paper introduces RAG-Stress, a diagnostic test that edits one sentence of retrieved evidence to support a wrong answer while keeping the question fixed. It compares two source priority instructions and three answer positions across fifteen systems on TriviaQA-RC, HotpotQA, SearchQA, and English and Chinese MedQA. Instructions that prioritize documents produce a higher misleading rate than those allowing prior knowledge. A paired audit finds more harmful overrides but no matching gain in beneficial corrections.
In briefMATE4D: Matrix-guided editable 4D generation from one image
SummaryThis paper presents MATE4D, a framework that turns a single input image into editable dynamic 4D content. It builds a spatio-temporal multi-view image matrix with text-guided background editing to optimize 3D Gaussian primitives, which a lightweight deformation module then animates into a 4D representation. The authors report that experiments on Objaverse-XL and Diffusion4D show better visual quality, efficiency, and controllability than strong baselines.
In briefReview of multimodal remote sensing image registration: challenges and outlook
SummaryThis paper systematically reviews mainstream methods for aligning multimodal remote sensing images, grouping them into region-based, feature-based, and deep learning-based approaches. It explains the principles, representative algorithms, advantages, and limitations of each category. The authors also summarize publicly available multimodal remote sensing datasets and highlight current bottlenecks and future trends in high-precision registration.
In briefHigher-order action supervision to make a stronger policy class
SummaryThis paper addresses control instability in imitation learning and reinforcement learning, which limits their use in robotics and autonomous driving. The authors argue the problem arises because methods supervise only zeroth-order action labels and ignore higher-order dynamics and temporal consistency. They propose a loss scheme that supervises both zeroth- and first-order actions, applicable to any policy model without structural changes and usable as a plug-and-play module in offline RL frameworks. Experiments on OGBench and D4RL report improved performance and robustness, along with better out-of-distribution generalization in low-data settings.
In briefIntactWorld: Joint World Modeling Using Uncompressed Intact Features
SummaryThe paper argues that current video generation models produce realistic visuals but lack a real understanding of real-world logic, and that existing learning methods compress features and lose structural information. It proposes IntactWorld, a joint world modeling architecture that uses uncompressed intact features and predicts the clean feature x0 at intermediate layers, along with a Full-to-Compact training paradigm that replaces full features with refined CLS tokens. The authors report 11.4% lower spatial memory use and 43.8% lower inference latency, and a 2.46-point gain over established baselines on the VBench 2.0 benchmark.
In briefVAMR: Multi-Question Agentic Reasoning for Efficient Long-Form Video Understanding
SummaryLong videos often need answers to several questions, but existing video agents handle each question in a separate tool-use trajectory. VAMR answers all questions about one video through a shared trajectory, using layered multi-question memory to keep common context and per-question evidence. On LVBench it reaches 62.1% accuracy while cutting reasoning rounds and processed frames substantially.
In briefSACQ: Structured forecasting head for long-term time series prediction
SummaryThis paper proposes SACQ, a plug-in prediction head that replaces the common flatten readout used in long-term time series forecasting while leaving the encoder unchanged. It first produces a coarse patch-grid forecast, then refines each future position via cross-attention over historical memory and merges the result using a learned per-patch gate. The authors also introduce a batch-adaptive scaled log-cosh loss for robustness to noisy labels. Reported results show top-tier test MSE/MAE across PatchTST, DLinear, and patch-Mamba backbones, and better performance than flatten readouts under input corruption and label-noise stress tests.
In briefPMTRM: A Pseudo-Memory Module That Helps Robot Policies Tell Repeated Phases Apart
SummaryRobotic manipulation often repeats motions that look similar but require different actions, which can cause policies to repeat finished steps or switch phases at the wrong time. The authors propose PMTRM, a lightweight plug-in module with 7.61M parameters that encodes a bounded history of past states and actions into a latent sequence for existing policies. A temporal heterogeneity objective and anchor and reconstruction losses help distinguish phases while keeping action-relevant information. Experiments in simulation and on a real robot report improved task success on phase-ambiguous tasks across several policy backbones, with little added computation.
In briefPIVOT: Perplexity-based scheduling that switches small models from teacher-guided learning to reward-based refinement in few-shot niche domains
SummarySmall language models struggle to classify text in specialized fields when few labeled examples are available. This paper proposes PIVOT, which uses the teacher model's perplexity on each sequence to decide whether a sample moves from On-Policy Distillation to GRPO reinforcement learning. Low-perplexity samples go to GRPO for reward-driven refinement, while high-perplexity samples stay in distillation to keep acquiring domain knowledge. On Banking77 and HWU64, PIVOT reportedly outperforms continued distillation and fixed-schedule baselines under the same number of post-warm-up optimization steps.
In briefCARE: A lightweight add-on module that corrects long-term forecasts and flags uncertain predictions
SummaryCARE is a lightweight plug-in that can enhance any deterministic forecasting model without redesigning its architecture. It runs alongside the base model, learning residual correction patterns from historical context aligned to the forecast horizon, and applies bounded updates gated by per-coordinate risk signals. Across eight benchmarks and three backbones, it improves accuracy with marginal parameter and latency overhead. Its risk gates identify high-error regions, with the highest-gate tertile showing nearly four times the error of the lowest on the Weather dataset.
In briefRideBench: A Large-Scale Benchmark for Ride-Hailing Forecasting with External Factors
SummaryThis paper releases Ride-Hailing, a large ride-hailing time series dataset synthesized from DiDi's marketplace data across 200 areas, spanning four years at half-hourly resolution. It includes three exogenous scenarios: weather disturbance, holiday effects, and large-scale events. Using this dataset, the authors introduce RideBench and evaluate over 30 forecasting methods for week-ahead and 8-week-ahead predictions. Results show that known future external variables help short-term forecasts, but existing models struggle with disturbance-induced changes and long-horizon forecasting.
In briefAutoAdapt: Reliable Few-Shot Adaptation for Clinical Distribution Shifts
SummaryThis paper addresses how to adapt large pretrained clinical models to a target hospital that has only a small labeled patient cohort. It introduces AutoAdapt, with an Adapter that defines an extensible space of adaptation recipes and an Automator that forms weighted recipe combinations from evidence in the adaptation patients. A reliability rule selects only strategies that perform best on most available patients. The authors report state-of-the-art results across critical care, emergency care, and diagnostic datasets using only a few patients.
In briefVGGTWorld-VLA: Intent-Conditioned 3D World Prediction for Driving
SummaryThis paper proposes VGGTWorld-VLA, an extension of VGGT-World that conditions future 3D scene geometry on driving intentions for autonomous driving. It adds an action-semantic conditioning mechanism and a geometry-language-action bridge so that the same scene yields different future geometry under different ego actions. Evaluated on NAVSIM, the method shows competitive geometry prediction against the baseline, and ablations support the value of semantic and action conditioning.
In briefLadderEdit: Memory-Efficient Lifelong Editing of LLMs via Per-Edit Residual Compression
SummaryLifelong editing of large language models requires storing thousands of edits, and the common approach of one LoRA adapter per edit grows linearly in storage. This paper proposes LadderEdit, which compresses each edit's adapter after acquisition by first storing it at low rank and checking whether it meets rewrite, generalization, and locality requirements on probe prompts, promoting it to higher rank only when it fails. On ZsRE, CounterFact, and WikiBigEdit with LLaMA-3-8B, Mistral-7B, and Qwen2.5-7B, the method reports 5.2x less memory than exact LoRA storage and remains effective at 50,000 sequential edits.
In briefLocal Prototype Reconstruction for Text-Compatible Speech-to-LLM Bridge Pretraining
SummaryThis paper studies pretraining of a small trainable bridge that connects a frozen speech encoder to a frozen large language model. The authors propose a fixed diagnostic for lexical manifold compatibility and a training-only regularizer, Local Prototype Reconstruction, which requires bridge tokens to be reconstructable from neighbourhoods of LLM token embeddings. On multilingual speech recognition and speech translation, the method improves transfer, with the largest gains on translation and low-resource adaptation.
In briefSocial Pain Disrupts Emotion-Action Brain-State Links in Self-Injuring Adolescents
SummaryThis study recorded EEG from 106 adolescents with depression, 67 with non-suicidal self-injury and 39 without, under social pain, physical pain, and resting conditions. Using microstate analysis and an interpretable deep sequence model, it identified social pain as the condition with the strongest discrimination, at 68.55% accuracy. Adolescents with self-injury showed weakened transitions between the MS3 (emotion-related) and MS5 (action-related) brain states, suggesting disrupted emotion-action coupling.
In briefCARE: Teacher-student distillation to refine interpretable attention for fine-grained image classification
SummaryThis paper proposes CARE, a framework for interpretable fine-grained image recognition. A class-specific attention student produces the final prediction and explanation, while a training-only auxiliary query teacher reads intermediate DINOv2 layers and transfers class-discriminative knowledge to the student. Diversity and sparsity terms regularize the student's attention heads to reduce redundancy. Experiments on CUB, Oxford-IIIT Pet, Stanford Dogs, and Stanford Cars show strong results under a frozen-backbone setting, with 78.5% Top-1 accuracy on CUB, and insertion and deletion analysis indicates top-ranked attention regions retain class-relevant evidence.
In briefDo LLMs Learn from Rewards in Context? Rethinking Reward's Role in In-Context RL
SummaryThis paper tests whether LLM agents truly perform reinforcement learning at inference time by conditioning only on raw trajectory-reward pairs in context. Controlled experiments on four benchmarks and six models show that flipping, randomizing, or removing rewards barely changes the improvement curve, while shuffled or corrupted trajectories work as well as real ones. The authors conclude that direct in-context RL is better understood as a special case of in-context learning, suggesting that input distribution and demonstrations may matter more than reward shaping for agent memory design.
WhyControlled experiments on four benchmarks and six models challenge reward's role in in-context RL, yielding a counterintuitive, fairly well-supported conclusion.
In briefA unified score matching method for detecting and anticipating video anomalies
SummaryThe paper proposes Uni-DSM, a unified framework based on denoising score matching that models anomaly patterns through likelihood estimation and score functions over the learned data distribution. The same score-based formulation serves both video anomaly detection, via autoregressive denoising, and video anomaly anticipation, via a self-distilled denoising score matching mechanism that avoids unavailable future ground truth. The authors report state-of-the-art performance and high efficiency on multiple benchmark datasets.
In briefSP-DocReader: Difference-Aware Self-Play for Precise Document OCR
SummaryThis paper presents SP-DocReader, a self-play framework that improves optical character recognition by targeting errors remaining after supervised fine-tuning. It aligns reference and generated tokens with a longest common subsequence and adds direct negative log-likelihood supervision at unmatched ground-truth positions, training only the OCR module while the backbone stays frozen. On Qwen3-VL-4B, the authors report about a 54 percent reduction in character error rate and a 3.7-point gain in DocVQA ANLS.
In briefWhat Transfers from Training-Only Priors? Separating knowledge, ordinal targets, and optimization
SummaryThis paper studies distillation where a teacher is available only during training, and a compact student deploys using local features and exported weights. The authors separate label, query, capacity, and teacher-access budgets across three public credit datasets and a new bankruptcy dataset. They find limitations in the historical Ranking Prior Alignment method, with no single tested repair consistently fixing it. The paper reports mixed results and explicitly does not claim universal superiority or an LLM-specific advantage.
In briefTight limits on testing if two unknown distributions match, using non-adaptive conditional samples
SummaryThe paper studies distribution testing when non-adaptive conditional samples are available, focusing on whether two unknown distributions are identical or at least ε apart in total variation distance. Its algorithm and lower bound show that Θ̃(log n/ε²) queries are both necessary and sufficient. The authors conclude that uniformity, identity, and equivalence testing all have complexity Θ̃(log n) under this model.
In briefImproving Image-Based Nutrition Estimation via Multimodal Food Verification and Recovery
SummaryThis paper proposes a framework for estimating nutrition from a single food image. It uses multimodal large language models to list visible foods and separately verify each food's identity and whether its region supports portion estimation. A whole-image review triggers at most one targeted recovery pass for missed foods, and recovered items are re-verified and merged into a final set. The authors report that, without task-specific fine-tuning, the method improves mass and energy accuracy against an adapted retrieval baseline on matched valid-output samples.
In briefMachine-learned fast construction of inner sphere trees for collision detection
SummaryThis paper proposes ML-IST, a machine-learning method for building inner sphere trees, a representation used for collision detection of complex non-convex objects. It treats leaf-sphere generation as a single-shot set-prediction problem, with a network predicting inscribed sphere candidates from voxel geometry, refined by differentiable coverage feedback and geometric projection. The authors report a 2.0 to 2.6 times reduction in mean leaf-generation time on a single GPU versus single-core CPU construction, and 55.89 FPS in an alpha-puzzle benchmark within the Jolt Physics and Bullet engines.
