In briefLiveAnon: a streaming voice system that hides speaker identity while keeping emotion and speech clear
SummaryThis paper presents LiveAnon, a voice-conversion-based streaming speaker anonymization system that hides who is speaking while preserving linguistic content and emotion. It conditions an autoregressive codec language model on hidden features from a pretrained streaming ASR encoder, and introduces capped RVQ decoding to reduce latency. Under the VPC24 and VPC26 evaluations, it improves intelligibility and emotion preservation over the prior Stream-Voice-Anon system and achieves a strong privacy-utility balance.
In briefAdapting a 30B language model to eight Indian languages by expanding its tokenizer in place and continuing pretraining
SummaryThe study adapts Nemotron-3-Nano-30B-A3B to eight Indian languages without changing its backbone architecture. The vocabulary grows from 131,072 to 163,072 entries, reducing tokens per word by 38.6%. After about 1.27 trillion tokens of continual pretraining, the annealed checkpoint gains 12.9 points on Indic MILU but loses 3.4 points on English MMLU. The paper also reports several negative results.
Avinash Benki, Prasanjith Pasunuti, Sutanshu Saksena Raj et al.
In briefFewer Bits, More Tries: Regaining Model Capability Under Memory Limits
SummaryThis paper asks whether a smaller or compressed language model can close its capability gap with larger models by generating multiple outputs and selecting among them using public or model-generated tests. Across four code-generation benchmarks, verified retries recovered much of the accuracy lost to smaller size or compression. The authors report that an 8-attempt 7B Qwen model essentially matches a 32B model on MBPP+ and exceeds it on the other three benchmarks, using about 4x less model memory and running up to 6x faster.
In briefA sovereign reasoning model built first for Indian languages
SummaryThis paper presents Gnani-Evon-v3.3-30B-A3B, a long-context hybrid Mamba-Transformer mixture-of-experts model for reasoning, instruction following, and tool use in English and eight Indian languages. It starts from NVIDIA's Nemotron-3-Nano-30B-A3B and expands the vocabulary, then applies continual pretraining and post-training. Reinforcement learning improves mathematics, graduate-level science, and instruction following but degrades Indic extractive question answering, with English answers rising from 8.7% to 91.7%.
Sutanshu Saksena Raj, Prasanjith Pasunuti, Parijat Chatterjee et al.
In briefANTE: Learning Coding Agents' Internal World Models by Verifying Their Predictions
SummaryThe paper proposes ANTE, a reinforcement learning framework where coding agents predict the outcome of each action before taking it, and these predictions are checked against real execution. This verification gives per-step credit to actions and supports trust-or-abstain decisions. On SWE-bench Verified, Qwen3-32B's pass@1 rises from 25.0% to 57.2%, and its world model accuracy grows from 44% to 69% during training.
In briefTRACE: Monitoring and Editing Language Model Concepts to Reduce Harmful Outputs
SummaryThis paper proposes TRACE, a framework that tracks the risk of harmful content as a language model generates a response and selectively edits internal token activations linked to harmful concepts. It learns a codebook of recurring activation patterns and uses a GRU to predict harmfulness, with an external LLM labeling the concepts afterward for auditability. Experiments on three generators and the WildGuard and S-Eval benchmarks report improved streaming F1 in most comparisons and reduced Harmbench attack success with little change in refusals on safe responses.
In briefCertified detection of adversarial inputs for pretrained classifiers via Lipschitz-margin analysis
SummaryThe paper addresses small adversarial perturbations that can fool deployed image classifiers. It proposes a detection-oriented certification: certifying that adversarial inputs within an ℓ2 ball are detected, rather than that predictions stay invariant. The authors prove a reduction from certified robustness to certified detection that depends on a single Lipschitz constant, and introduce two detectors that work on any pretrained classifier without retraining. Experiments on MNIST, CIFAR-10, and ImageNet-1k with four backbones show performance matching or exceeding five baselines.
In briefUniForge3D: Jointly Reconstructing and Generating 3D Objects in a Shared Anchor-Guided Space
SummaryThis paper addresses generating complete 3D objects from sparse, unposed images by combining reconstruction of observed geometry with learned shape generation. The proposed UniForge3D models both tasks in a shared anchor-guided space to resolve a coordinate mismatch, exchanging features between branches to predict per-view point maps and a full sparse voxel structure. It selects the most informative input view for each voxel to help preserve fine geometric details. The authors report state-of-the-art appearance and geometry results on Toys4K and GSO.
In briefJSIR: A Multi-Task Jazz Dataset Aligned with Audio Recordings
SummaryThis paper presents JSIR, a dataset of 267 recordings of 44 jazz standards totaling 25.6 hours. It provides beat and downbeat positions, chords, and sections, annotated by a jazz musician and revised over several passes. The authors report that models trained on JSIR improve results on other jazz datasets. The annotations, the annotation tool, and a web application that displays annotations in sync with audio are publicly released.
In briefVideo Transformers Fine-Tuned to Detect Sexual Harassment in Footage
SummaryThis paper extends the Sexual Harassment Video (SHV) dataset to 600 YouTube clips labeled as harassment or non-harassment. Three video transformer models pretrained on Kinetics-400, VideoMAE, ViViT, and TimeSformer, are fine-tuned for binary classification. TimeSformer achieves the highest test accuracy at 92.27%, compared with 91.71% for VideoMAE and 91.16% for ViViT.
Jubayer- Al- Mahmud, Md Shamimul Islam, Md Nur Hossain
In briefSiam-ALS: An Explainable Siamese Deep Learning Framework for Early ALS Detection Using EEG and Eye-Tracking Signals
SummaryThis paper proposes Siam-ALS, a Siamese deep learning framework that jointly models synchronized EEG and eye-tracking signals to detect amyotrophic lateral sclerosis (ALS). On the public EEGET-ALS dataset with strict subject-level cross-validation, it reports 0.988 accuracy and 0.998 AUROC. Cross-dataset testing on BigP3BCI gave an AUROC of about 0.98. Six explainability methods were compared, with SHAP showing the highest, though statistically comparable, fidelity.
Abdur Rahman, A. A. M. Muzahid, Istiak Ahmed et al.
In briefGugak-VocalSet: A Studio Dataset of Korean Traditional Singing with Sigimsae Labels
SummaryThis paper introduces Gugak-VocalSet, the first publicly available audio dataset of Korean traditional singing with dense annotations of sigimsae, a structured pitch contour system for vocal expression. It contains 305 unaccompanied solo tracks totaling 24.5 hours, performed by 20 professional singers in a studio. The dataset includes a unified sigimsae ontology and about 68,000 annotated segments labeled by 14 expert annotators. Baseline experiments show that domain-informed F0 representations are competitive with, and outperform on classification, a frozen general-purpose music foundation model.
In briefCodeWalk: Generating Coding Benchmarks by Combining Existing Problems
SummaryThis paper proposes a low-cost way to create coding problems by composing verified atomic problems into multi-level tasks. A task is kept only if all its atomic problems remain unsolved and a joint reference solution passes the evaluation check, with no language-model generation involved. The method is applied to four settings: multi-bug repair, proof completion, complexity analysis, and clone equivalence. The generated tasks are hard for frontier models, whose performance drops as compositional depth grows, and rejection-sampling fine-tuning on them improves Qwen3.5-9B at every composition level.
In briefPhysical Adversarial Textures That Fool World and Video Models Across Viewpoints
SummaryThis paper studies whether one printable physical adversarial texture can mislead world models and video understanding models across changing viewpoints and object-camera trajectories. It proposes TARA, an attack using trajectory-aware optimization and contrastive alignment, and builds a reproducible testbed on EPIC-KITCHENS-100 and Something-Something V2. Real-world print-and-recapture videos are also collected, and the results show a single texture can persistently disrupt these models under physical conditions.
In briefStimFlow: Generating Brain Activity from Images via Flow Matching
SummaryStimFlow generates fMRI brain responses directly in native voxel space using conditional flow matching with a linear path, rather than in compressed latent spaces. It combines DINOv2 structural features with CLIP semantic features to match the visual cortical hierarchy. A single sample achieves a per-voxel Pearson correlation of 0.371, matching baselines that average five draws. Using only one hour of scan data, it recovers over half of full-data performance.
In briefMultiFlow: A generative model that predicts brain fMRI responses to vision, audio, and language stimuli.
SummaryThis paper introduces MultiFlow, a generative framework for fMRI encoding that predicts a distribution of brain responses to a stimulus rather than a single point estimate. It builds on Conditional Flow Matching and proposes Stimulus-Informed Source Flow Matching, which conditions the source distribution on the stimulus, with a Diffusion Transformer decoder. On the Algonauts 2025 dataset, a single MultiFlow checkpoint approaches the performance of large deterministic ensembles and also provides stimulus-conditioned uncertainty estimates.
Bui Van Hung, Hoang-Son Vo, Mai Duc Minh Huy et al.
In briefCoLA: Coordinating Multiple Vision-Language-Action Robots via Learned Adapters
SummaryThe paper introduces CoLA, which keeps each agent's Vision-Language-Action backbone frozen and trains a lightweight message channel and per-arm diffusion action head for fully decentralised execution. Explicit messaging matters mainly when information is private: on an occluded tray handover, success is 76.0% with messages versus 1.3% with messages zeroed. With a shared overhead camera, removing messages has little effect on two-arm and three-arm handover tasks. The architecture also works with a frozen pi0.5 backbone, matching a LoRA-fine-tuned baseline with far fewer trainable parameters.
In briefZOIF: Zeroth-order influence estimation for selecting LLM training data using only forward passes
SummaryThis paper proposes ZOIF, a method for estimating how much individual training samples influence large language models, which can guide data selection. It avoids backpropagation entirely, relying only on forward passes to reduce memory use. Experiments report that the selected data subsets achieve downstream utility competitive with, and sometimes better than, the evaluated baselines, with substantially lower peak memory and comparable computation time.
In briefFull-waveform inversion using a neural network as a learned preconditioner
SummaryThis paper studies full-waveform inversion (FWI), a method for reconstructing subsurface velocity models, and whether a neural operator trained on reverse-time-migration images can serve as a learned preconditioner. It compares a standard Fourier neural operator with a Jacobian-informed version, finding the Jacobian-informed model gives the lowest model error while the standard one is unstable. The authors conclude that predicting RTM images alone is insufficient, and accurate learned derivatives and objective-consistent use of the operator are needed.
In briefBaT-SAM: Fully automatic brain tumor segmentation in MRI by adapting SAM with convolutional experts
SummaryThis paper presents BaT-SAM, a parameter-efficient method that adapts the Segment Anything Model to segment brain tumors in MRI automatically. It adds local convolutional features to SAM's frozen encoder via Conv-LoRA and uses a lightweight box regressor so user prompts are not needed. The method reports a weighted mean IoU of 0.832 on BRISC2025, and Dice scores of 0.702 on Figshare and 0.784 on the LGG MRI Segmentation dataset.
Sumiaya Jannat, Md Mahedi Hasan, Md Shamimul Islam et al.
In briefAnchor-and-Harvest Curriculum: LLM-free, unmasked curriculum for Craftax
SummaryThe paper proposes AHC, a rule-based curriculum training system for the Craftax-Full environment that uses no LLM inference and keeps all 43 actions available. After about 2.0 billion PPO rollout transitions, three primary AHC runs reach a native-task return of 46.70±0.57, versus 41.91±1.01 for a native PPO-GTrXL baseline. Most of the gain comes from Intermediate achievements, while the deepest-floor achievements remain unsolved.
In briefCIViC: Evidence gates for checking audio-conditioned visual predictions
SummaryThis paper introduces CIViC, a four-gate protocol for checking whether changes in a visual predictor's output truly reflect the intended audio information. The authors test it on actor-disjoint face-raster tasks, a synthetic audio-feature mixture, and source-video-disjoint SpeakerVid proxy tasks. They report that output activation, held-out error, and endpoint controls support different interpretations across these tasks.
In briefTempoKV: Predictive KV cache compression for long video generation
SummaryAutoregressive video diffusion models generate long videos in a streaming fashion, but their growing KV cache uses a lot of memory. The authors find that the video KV cache is largely predictable over time: keys follow from their temporal history, while values depend mainly on the current key and a short history of values. They propose TempoKV, a training-free framework that predicts keys and values and stores only low-precision residuals, with predictors calibrated once offline by linear regression. Experiments on Self-Forcing, Causal Forcing, and LongCat-Video-13B show higher generation quality than existing methods at comparable or higher compression, reaching up to 6.71x KV cache compression.
In briefDeep neural mixed model for nonlinear longitudinal prediction accounting for individual differences
SummaryThis paper proposes a Deep Generalized Neural Network Mixed Model (DGNMM) that extends the GLMM framework, using a deep neural network for population-level nonlinear relationships and trainable regularized subject-specific effects for individual heterogeneity. It is estimated with prediction-oriented stochastic optimization. In simulations and real depression and anxiety data, DGNMM achieved the lowest MSPE for continuous outcomes. The magnitude of subject-specific effects was not significantly associated with predictive gain.
Mahmud Hasan, Niloofar Ramezani, Hrishikesh Chakraborty
In briefSuperAI: Merging Abilities Across Different AI Models (a paper on combining complementary skills from several architecturally different foundation models into one host model)
SummaryThe paper proposes SuperAI, a framework that integrates complementary capabilities from multiple source foundation models into a single host model despite architectural differences. It extracts task-relevant representations with shared anchor samples, applies PCA and orthogonal Procrustes alignment, and trains a lightweight residual MLP that is injected into the host. According to the paper, experiments on three host models across 15 capability domains show higher average domain scores than evaluated cross-architecture fusion methods and better retention of general capabilities as source models increase.
In briefChannel Match: Calibrating shape latents so 3D models can generate assets in just a few steps
SummaryThis paper speeds up native 3D generators, which normally need many slow sampling steps, by distilling them into a few-step framework. The authors also find that few-step sampling shifts the latent distribution, which keeps geometry correct but causes texture artifacts, and they propose Channel Match, a training-free per-channel latent calibration method. On TRELLIS.2 at 1024³ resolution, the two-step model samples geometry 4.5 times faster with quality comparable to the many-step teacher, and Channel Match further improves rendering quality.
In briefFENAS: Few-shot ensemble method for faster neural architecture search
SummaryThe paper proposes FENAS, a few-shot framework that predicts neural architecture performance by combining multiple zero-cost proxies. Its core component, CAST, uses copula alignment and synergy transformation to map these proxies into a compact feature space for lightweight meta-learners. On NAS-Bench-201 with only 0.5% labeled samples, it reaches a Kendall's Tau of 0.70 and evaluates the space in about 1138 seconds with competitive test accuracy.
In briefWisdomNet: Trustable Machine Learning Aiming for Error-Free Classification
SummaryThis paper introduces trustable learning, in which a classifier refuses to decide or defers to a human expert when it cannot classify accurately, thereby avoiding misclassifications. The authors propose WisdomNet, an architecture that adds an extra "conjugate neuron" indicating whether the network can classify an input correctly. According to the paper, WisdomNet achieved a 0% misclassification error across several architectures and datasets, with roughly 10% or less of the data labeled as "reject."
In briefCHEAR: Cost-Efficient Human Activity Recognition via Confusion-Hierarchy Evidence-Aware Reasoning
SummaryThis paper proposes CHEAR, a method for recognizing human activities from wearable sensors that shifts part of the knowledge organization offline. It builds an activity hierarchy from recurring cross-subject retrieval errors and creates discriminative knowledge for activity pairs. At inference, the hierarchy orders candidate comparisons, and the LLM progressively retains candidates, accessing local discriminative knowledge only for remaining activities. Across eight datasets, the reported results include 86.79% mean cross-subject accuracy, 75.91% candidate reduction, and 96.44% true-activity retention, with per-sample token use 85.15% lower on average than a variant without the hierarchy.
In briefHGA-Quant: Heterogeneity-Aware Quantization for Autoregressive Image Generation Models
SummaryThis paper addresses the high computational and memory costs that make autoregressive visual generation models hard to run on edge devices. It proposes a training-free post-training quantization framework with two parts: Tail-Aware Bounded-init Grid Refinement for weight quantizer calibration, and Coarse-to-Fine Activation Precision Scheduling for mixed-precision allocation. On the Infinity model, the framework achieves near full-precision image quality at an effective W4A5.6 activation precision, where existing methods at W4A6 still show quality degradation.
In briefAn LLM retrieval framework for human activity recognition that uses an activity hierarchy and personal memory
SummaryThe paper proposes OntoAR-HAR, a retrieval-augmented LLM framework for recognizing human activities from sensor data. It organizes activities in a tree-structured ontology so that classification becomes local decisions among related activities, combining static population knowledge with a personalized memory of user observations. The authors report gains in cross-subject and heterogeneous-sensing experiments across several LLM backbones.
In briefCo-DSL: Human-LLM co-design of reusable spatial concepts for generating layouts
SummaryThis paper presents Co-DSL, a framework where users and an LLM jointly design reusable spatial concepts for layout generation. Users build a personalized library of operations, with the LLM turning descriptions into implementations that are refined through lightweight feedback. Given a new request, the LLM writes a layout program describing which concepts should hold, and a solver searches for a satisfying layout. Across five spatial layout domains, the authors report that Co-DSL captures layout patterns that prior fixed-operation systems struggle with.
In briefASM-6D: Fast Certified Pose-Shape Inference Using Complete Correspondence Maps
SummaryThe paper presents ASM-6D, a method that estimates an object's 6D pose and shape from RGB-D images without an exact instance model. It generates complete 3D correspondence hypotheses that share point identities with a category-level Active Shape Model, then fits rotation, translation, scale, and shape directly. The authors report state-of-the-art results on the ROPE and HouseCat6D benchmarks, with real-time speeds of 6 to 38 Hz depending on the inference budget.
In briefProtein crystal detection in fluorescence images using a secondary color channel
SummaryThis paper proposes detecting protein crystals in fluorescence microscopy images by using color intensity at wavelengths longer than the emission peak, rather than the dominant channel. The authors built a segmentation method based on the percentile intensity of this subordinate channel. Compared with the dominant channel, it reduced misclassification of crystals as non-crystals from 9.71% to 2.02%, depending on the classifier, and reached about 94% accuracy. New wet lab tests on two other proteins also kept misclassification below 1%.
In briefFragFlex: a fragment-based model that generates flexible-size molecules to explore large chemical space
SummaryThis paper introduces FragFlex, a flexible-length discrete diffusion model that inserts and deletes nodes at the fragment level to generate molecules. During denoising, it jointly learns the fragment graph's size, fragment identities, and connectivity, then converts it to an atom-level molecule. In experiments generating 10 million molecules, it reports near-perfect chemical validity while keeping high uniqueness.
In briefWhat data releases must preserve for off-policy evaluation
SummaryThis paper studies data release for off-policy evaluation, showing that merging decision-time histories that affect a policy's value differently can make the evaluation unidentified even when actions, outcomes, and policies are known. The authors characterize which distinctions must be preserved, identify the coarsest exact release, and design releases that refine each record into at most m categories, with a finite-data repair algorithm. On 24 KuaiRec tasks, the repair reduces the median mean pairwise sharp width by 93% at m=64.
In briefNorth or South? Testing how well LLMs understand North Korean language, society, and knowledge
SummaryThis paper introduces LENS, a human-validated benchmark of 1,765 questions that tests large language models on linguistic robustness, sociocultural perspective alignment, and domain-specific knowledge about North Korea. Models perform modestly worse on North Korean language inputs, show a larger gap in sociocultural perspective alignment, and have severe limitations on domain-specific knowledge. The authors conclude that linguistic similarity alone does not ensure sociocultural or domain understanding, and call for grounded evaluation resources for underrepresented contexts.
In briefImperfect Verifiers, Exact Targets: Inference-Time Scaling for Continuous Generation
SummaryThe paper proposes Continuous-State Verifier-Guided Backtracking (CS-VGB), an inference-time method that samples from reward-tilted distributions of continuous generative models through repeated denoising and renoising. The authors state that with correctly paired transitions, its stationary distribution at the clean checkpoint equals the target even when intermediate verifiers are inaccurate. Experiments cover analytically tractable settings, prompt-guided ImageNet and CelebA generation, and multi-robot planning.
WhyProposes an inference-time method recovering reward-tilted distributions despite imperfect verifiers, with convergence guarantees and experiments on images and robot planning.
In briefGeoBridge: Evolving Reusable Decoder Structures for Geometric Vision
SummaryThis paper presents GeoBridge, an evolutionary method that automatically designs the internal structure of decoders by combining configurable components in residual, sequential, and parallel ways. Candidate structures are scored without training using a gradient sensitivity-alignment proxy, and selected structures are trained for depth estimation and Gaussian novel-view synthesis. The authors report the best depth accuracy among four baselines on NYU Depth V2 and KITTI, 27.11 dB PSNR on RealEstate10K, and support for open-vocabulary segmentation on LERF.
In briefDo correct answers rest on stable latent knowledge?
SummaryThe paper argues that checking whether a language model's output is correct does not reveal whether the internal knowledge representations behind the answer are stable. It proposes two metrics, latent state stability (LSS) and knowledge topology stability (KTS), and builds EpiStableBench, with 4,411 items and 76,766 controlled queries. Tests on 16 LLMs show that representation stability varies widely even among items answered correctly under the base condition.
In briefLeHami: Physics-structured latent dynamics for world models
SummaryThis paper proposes LeHami, a joint-embedding predictive architecture that learns world models in latent space using an action-conditioned Hamiltonian flow integrated with a symplectic leapfrog step. The structure preserves phase-space volume at every rollout step. On visual pendulum environments, LeHami improves long-horizon velocity prediction over an unconstrained JEPA baseline and keeps latent volume constant where the baseline contracts.
Romina Yalovetzky, Randall Balestriero, and Yann LeCun
In briefRECAP: Calibrating low-bit reasoning models on the history they read
SummaryThis paper presents RECAP, a post-training quantization method for compressing reasoning models. It keeps the long historical context during calibration and weights historical cache errors asymmetrically by their effect on attention outputs. Under W4A4KV4, RECAP improves accuracy over the baseline by 13.3 points on MATH500 with DeepSeek-R1-Distill-Qwen-1.5B and by 15.8 points on AIME-120 with Qwen3-4B.
In briefSequent-Prover: Training Agents for Checkable SMT-Based Reasoning
SummaryThis paper introduces Sequent-Prover, language-model agents that reason by writing executable SMT-LIB programs and querying the Z3 solver across turns. The training method, verifier-conditioned trajectory distillation, keeps successful multi-turn trajectories that are correct and executable, then applies supervised fine-tuning and preference optimization. Across 37 benchmarks and 7,254 problems per model family, the 4B model's per-attempt accuracy rose from 85.6% to 97.0% and its four-sample all-correct reliability from 64.1% to 91.7%. Average SMT calls fell from 4.10 to 2.52.
Debargha Ganguly, Weicong Chen, Mohsen Hariri et al.
In briefCARMA: a reconstruction model that tracks channel correlations to detect multivariate time series anomalies
SummaryThe paper notes that reconstruction-based detectors for multivariate time series mix temporal dynamics and inter-channel structure into one latent space, so they miss anomalies where each sensor looks normal but their joint behavior deviates. The authors propose CARMA, a variational autoencoder with a temporal latent and a correlation latent that reconstructs pairwise Pearson correlations. On the TSB-AD-M benchmark (17 datasets, 180 series) against 29 baselines, it reports a VUS-PR of 0.4929 and removing the correlation branch lowers performance.
Md. Johir Raihan, Zahra Zamanzadeh Darban, Yueyang Liu et al.
In briefObsSpec: speeding up agent RL by predicting tool outputs early
SummaryTool-using agents alternate between generating text and waiting for tool calls to finish, which slows agentic reinforcement learning that needs large batches of multi-turn rollouts. ObsSpec uses a small world model to predict tool outputs while the environment runs, so the policy can start its next generation early. Speculative work is kept only when predicted and actual observations match exactly, keeping rollout collection lossless. On terminal-use tasks, a 0.6B world model matched actual observations exactly in 66.4% of attempts, and agent latency fell by 9.3% for an 8B policy and 14.2% for a 32B policy.
In briefAutomated carotid plaque inflammation segmentation in PET with MSR-Net
SummaryThis paper presents a two-stage deep learning method, MSR-Net, that segments carotid artery inflammation on 64Cu-CANF-comb PET images and quantifies uptake while accounting for partial-volume effects. The method was tested on synthetic PET data built from MR maps and GATE Monte Carlo simulation, where it outperformed U-Net, Transformer U-Net, and DUCK-Net under simulated conditions. The authors state that the patient analysis is preliminary and needs external validation before clinical use.
In briefEvolvAttack: Backdoor Attacks on Self-Evolving Vision-Language Model Agents
SummaryThe paper introduces the self-evolving backdoor, an attack whose malicious payload stays dormant in an initial vision-language model agent and activates as the agent's memory evolves. Activation requires both a visual trigger and a sufficiently evolved memory, so static scanning with clean memory detects little. Tested across three VLM backbones, four agent pipelines, and three memory evolution methods, the attack had a 0.5% success rate before evolution but exceeded 95% within 100 interactions.
WhyProposes a new dormant backdoor requiring visual trigger and evolved memory, evading static scans; tested across multiple backbones and memory schemes.
In briefExplicit bounds on Hoffman constants for sparse regularizers, a measure of solution stability
SummaryThis paper derives explicit bounds on Hoffman constants for structured polyhedral systems arising in sparse regularization. Hoffman constants relate a residual to the distance from the solution set and affect error bounds and convergence rates. The authors decompose these constants into spectral, geometric, and interaction components, with ℓ1-regularized least squares as the main application, and extend the analysis to the elastic net and adaptive ℓ1 formulations.
In briefClinEdge: an expert-graded benchmark of real patient cases that frontier multimodal AI models fail
SummaryThis paper introduces ClinEdge, a benchmark of 32 de-identified real patient cases, each with primary imaging and a short vignette. Seven frontier AI systems produced 224 responses, which were graded by a senior orthopaedic surgeon and then reviewed by a panel of nineteen physicians. All systems performed poorly, often failing in the same ways, and the authors catalogue 18 recurring failure modes in five families. Four cases are released publicly, and the remaining 28 are held out for external evaluation.
In briefA framework for an AI-ready shared database of ARPES spectra
SummaryThis Perspective proposes a framework for a community database of Angle-Resolved Photoemission Spectroscopy (ARPES) data, a technique that has accumulated over four decades of spectra, though publication-centered workflows preserve only a small selected fraction. It reports outcomes of a community workshop and outlines benefits on three timescales. Near term, standardized, searchable, citable datasets; intermediate, cross-lab tools and automated experiments; long term, a corpus for foundation models of quantum matter.
In briefMoMaTune: A two-stage training method for language models on numerical records
SummaryThis paper notes that decoder language models often fail on numerical data, since numbers serialized as text receive weak supervision for digit magnitude and for dependencies between fields in the same record. The authors propose MoMaTune, a two-stage post-training approach: Stage 1 applies next-token prediction on label-free records (Model) and optionally a same-record consistency objective (Match), followed by Stage 2, supervised fine-tuning for downstream target prediction (Tune). Evaluated on 169,285 records from eight public datasets, the method improves masked-value imputation results compared with baselines.
In briefUsing Evidence-Centered Design to Guide Prompt Engineering in Measurement for Automated Essay Scoring
SummaryThe paper argues that current prompt engineering for large language models in education is largely ad hoc, which can threaten validity. It proposes Evidence-Centered Design, a framework from educational measurement, as a principled basis for prompt design by mapping its student, evidence, and task models onto prompt components and deriving five design principles. An empirical study on automated essay scoring compares ECD-designed prompts with state-of-the-art prompting baselines. The authors report that treating prompting as assessment design improves the validity, reliability, and fairness of AI-generated educational outputs.
In briefWords to World: a benchmark for LLMs acting in real physical environments
SummaryThis paper presents a benchmark that evaluates large language models in closed-loop physical interaction, where they control a harness connected to real robotic arms, cameras, and sensors. Models must issue instructions, receive physical feedback, interpret the resulting state, and adjust their next action. Experiments identify failure modes including hallucinated state changes, poor recovery from unexpected outcomes, and unstable long-horizon plans.
In briefUnderstanding Unified VLMs: Attention Probing and Guidance Tuning for Multimodal Generation
SummaryThis paper studies how a single shared Transformer in unified vision-language models handles conditions for both text-to-image and image-to-text generation. The authors probe each attention head to measure activation differences between two classifier-free guidance branches, finding that some heads are sensitive to condition changes while others are not, a pattern that is largely stable across samples and both generation directions. Based on this, they propose HeadTuned-CFG, an inference-time method that tunes conditional guidance at the attention-head level without extra training, with notable gains in text-to-image generation.
In briefZeroth-order fine-tuning of LLMs using learned low-rank subspaces
SummaryZeroth-order optimization can fine-tune large language models with less memory, but it has high variance in high-dimensional parameter spaces. This paper proposes AdaSub-ZO, which learns layer-wise projection matrices from finite-difference feedback instead of using fixed random subspaces. It also adds a subspace momentum mechanism to smooth noisy gradient estimates. Experiments on natural language understanding tasks report that it outperforms existing zeroth-order baselines, with lower gradient estimation variance and faster convergence.
In briefLLM-based persistent orchestration of robot teams by partially delegating to multi-agent protocols
SummaryThe paper addresses using large language models to coordinate heterogeneous robot teams in missions that do not end and in environments that keep changing, such as disaster response. It replaces the ever-growing interaction history with a Markovian state summary, delegates part of the robots to an autonomous multi-agent protocol, and keeps the team alert by flagging observation staleness and re-prompting on task and event triggers. In a simulated urban search-and-rescue mission with spreading fire and moving evacuees, the framework rescued 98% of civilians versus 89% for the strongest baseline, using six times fewer tokens.
Maria G. Mendoza, Victoria M. Tuck, Zachary Ravichandran et al.
In briefNon-Euclidean gradient descent can have many stability edges
SummaryThe paper studies the stability of non-Euclidean gradient descent on quadratic objectives. For general norms, the intermediate step-size regime between 2/S and 2/μ can contain arbitrarily many stability edges tied to distinct attracting directions. The authors show that only Hilbertian norms make 2/S an almost-everywhere sharp threshold.
In briefHarnessNet: hierarchical, feedback-driven co-evolution of the software harnesses around AI agents in multi-agent systems
SummaryThis paper proposes HarnessNet, a framework in which multiple AI agents revise their harnesses, the software managing inputs, tools, memory, and execution, using environment feedback while keeping foundation models frozen. Each agent's local optimizer sees and changes only its own harness, and feedback propagates upward through a layered graph. On AgentBench with several foundation models, HarnessNet matches the performance gains of a single global optimizer while preserving local privacy and using fewer tokens and less time per optimizer call. Evolved harnesses remained effective after swapping the foundation models, suggesting transferability across models.
In briefUsing fewer loops makes recurrent-depth language models stop refusing harmful requests
SummaryThis paper studies recurrent-depth language models, which reuse one block of weights for a chosen number of loops at run time. Halving the loops on Ouro-1.4B-Thinking kept about 89% of its GSM8K maths accuracy but raised its help with harmful requests from 0.038 to 0.205. The authors find the harmfulness information remains in the model's internal state, while the refusal that should follow is missing. They report that changing the training, using randomly chosen depths, works better than run-time fixes.
WhyShows fewer loops systematically weaken refusal and proposes random-depth training as a fix; multiple models and instruments support claims with broad impact.
Seyedarmin Azizi, Erfan Baghaei Potraghloo, Arya Fayyazi et al.
In briefPANINI-DR: Controlled deep research using a shared workspace that tracks sub-questions, findings, and evidence.
SummaryThe paper introduces PANINI-DR, a multi-agent system that separates overall investigation coordination from the research and verification of individual sub-questions. A shared workspace records sub-questions, dependencies, findings, and evidence, and a separate verification agent checks answers against that evidence. On three benchmarks (BrowseComp-Plus, FRAMES, MoNaCo) it reportedly achieves the strongest average performance against seven other deep-research harnesses.
M. Yigit Turali, Shreyas Rajesh, M. Efe Lorasdagi et al.
In briefMoSAIC: learning scheduling heuristics for task graphs from optimal schedules
SummaryThis paper presents MoSAIC, a learning framework for scheduling tasks described as directed acyclic graphs (DAGs) on heterogeneous computing systems. It uses the OR-Tools CP-SAT solver to produce optimal schedules, then uses Bayesian Optimization to learn compact priority-based scheduling functions, with large language models assisting the reasoning. Experiments on synthetic graphs and the PolyBench and MachSuite benchmarks report that the learned heuristics consistently outperform HEFT while keeping runtime overhead low.
In briefSocraticRepair: A Guided Tutoring System That Helps Students Fix Code Bugs by Asking Questions
SummaryThis paper introduces SocraticRepair, a Socratic tutoring framework that helps learners debug programs and correct misconceptions, building on earlier planning-guided tutoring work. It is applied to Python by fine-tuning Qwen2.5-Coder-7B-Instruct on synthetic multi-turn tutoring dialogues built from four preprocessed programming datasets. Experiments reported higher scores for stepwise scaffolding and question quality, and substantially less premature disclosure of solutions than a general-purpose LLM. The authors note that reliable generation of pedagogical plans remains an open challenge.'
Mykhailo Kopytskyi, Jianhua Yang, Mir Seyedebrahimi
In briefDo LLMs Internally Tell Apart Linguistic Traits? A Layer-by-Layer Study
SummaryThe paper examines whether two equal-sized pre-trained models, Llama-3.2-1B and Gemma-3-1B-PT, internally distinguish linguistic nuances such as conventional versus idiomatic meaning, and at which layers. It combines layer-wise probing with geometric distance measures, using the top embedding dimensions ranked by principal component loadings for each layer. The results show that even with equal capacity, the two models organize linguistic information in fundamentally different ways. The authors argue that evaluating how closely model and human language interpretation correspond requires analyzing internal representational geometry, not just behavioral outputs.
Samhita Bollepally, Takashi Yamauchi, Manju Muralee et al.
In briefContinuous risk-based scheduling for food inspections using history and consumer reviews
SummaryThe paper presents the Continuous Inspection Risk (CIR) system, which re-ranks food facility inspection queues as new data arrives, combining time-decayed historical scores with language-model-classified consumer reviews. Evaluated on 9,505 routine inspections in Santa Clara County, California (2024–2026), history-based scoring alone shortened mean time-to-detection of major violations by 69 days but missed all vermin-related closures. Adding post-inspection review signals predicted 7 of 8 vermin closures with an average lead time of 117 days.
In briefSAFE-KV: Behavior-preserving causal suffix repair for non-prefix KV cache reuse in RAG
SummaryThis paper studies reusing cached document states across retrieval-augmented generation requests, where cached states depend on their original prefix and position. The authors propose SAFE-KV, which re-encodes the causal suffix from a selected start document to repair a composed cache. On Qwen2.5-7B-Instruct, the method restores answers close to full-prefill results on HotpotQA-500 and 2Wiki-500, while prefix-conditioned composition alone preserves far fewer.
In briefDeltaNet-PC: A Test-Time Memory Method That Avoids Backpropagation Through Time
SummaryThis paper proposes DeltaNet-PC, which gives predictive coding, an algorithm that optimizes prediction error with local updates, a test-time memory for sequential data. The authors treat each step's memory update as a latent variable rather than a slow weight, deriving a PC-native rule for verbatim sequence memory. The paper reports perfect recall accuracy for storing single sequences, multiple sequences, and key-value associative memory pairs.
In briefTopology-Free Game-Theoretic Reinforcement Learning Defense for Virtual Power Plant Cybersecurity
SummaryThis paper addresses cyberattack risks to distributed energy resources aggregated by virtual power plants. It proposes a hybrid defense that combines game-theoretic planning with model-free reinforcement learning under a Bayesian belief-update framework, without needing feeder topology or line parameters. Evaluations on IEEE 13- and 123-Bus feeders show effective detection under attacker-model uncertainty and feeder re-configuration.
Pooja Algikar, Stefano Fenu, Kumarsinh Jhala et al.
In briefWord-aligned prosody mainly reflects speaker identity, not depression or PTSD severity
SummaryThe study examines whether word-aligned prosody measures are stable and clinically relevant for detecting depression and PTSD. Word-level speaker-specific patterns showed little recurrence across contexts, while speaker-level averages were highly consistent across recordings. However, these speaker-level measures showed associations with depression or PTSD severity of no more than 0.15, and near-zero within-person links to daily PTSD symptoms.
Hsiang-Chen Yeh, Aurosweta Mahapatra, Shreeram Suresh Chandra et al.
In briefUDHRBench: A Benchmark Testing Whether Large Language Models Respect Human Rights
SummaryThis paper introduces UDHRBench, a large benchmark for checking whether large language models comply with the Universal Declaration of Human Rights. It includes 1,124 human rights questions covering all 30 articles and 4,496 expert-annotated responses from four models. The authors found that violations often come from seemingly rational reasoning, such as appeals to efficiency or risk management, rather than explicit rejection of rights. They also propose an LLM-as-a-Judge framework for automated evaluation, reaching a 89.8% weighted F1-score in detecting misalignment.
WhyIntroduces a benchmark covering all 30 UDHR articles, with expert annotation (kappa 0.801) and an LLM-as-judge framework reaching 89.8% weighted F1.
In briefPhysics Foundation Models for Efficient Control of Fluid Flows: Reusing a PDE-forecasting model as a flow controller
SummaryThe paper asks whether a physics foundation model pretrained for PDE forecasting can be reused for closed-loop fluid-flow control. The authors attach lightweight Gaussian actor and critic heads to frozen pretrained representations and train them with proximal policy optimization. Tested on HydroGym's two-dimensional Kolmogorov-flow environment at Reynolds numbers 100 and 200, the controller achieves higher reward than a baseline MLP policy trained the same way, while using smaller control action magnitudes.
Mahindra Rautela, Shaifalee Saxena, Alexander Scheinker
In briefLexiCan: Lexicographic Canonicalization of Point Clouds for Machine Learning
SummaryThis paper introduces LexiCan, a method that resolves rotational and permutation ambiguity in 3D point clouds by selecting a unique representative, so standard architectures such as Transformers can process them. The authors report it is provably invariant under O(3) or SO(3), with rotational error around 10^-6 angstroms on QM9 and QM-sym and fewer discontinuity jumps on QM-sym, while running much faster than an existing method. They also state that non-invariant models perform better on QM9 property prediction and ModelNet40 shape classification when using LexiCan than with the prior canonicalization.
Daniel Sanchez, Djordje Mihajlovic, Federica Valeau et al.
In briefExeWorldBench: A Benchmark for Code-Based Executable 3D World Modeling
SummaryThis paper introduces ExeWorldBench, a benchmark that evaluates how well models understand a world and turn that understanding into executable code-based 3D world construction. It includes three tasks of increasing difficulty: realizing individual mechanisms from visual observations, building multi-stage systems with spatial, mechanical, and causal dependencies, and synthesizing executable worlds from high-level goals and physical requirements. The authors report that experiments on frontier models reveal substantial limitations in executable world modeling.
In briefCan active learning recover rare subgroups missed by a biased starting set?
SummaryThis paper examines whether active learning can correct bias in the initial labelled set used to train a classifier for a rare class, such as hate speech, when that set was selected by methods like keyword search. Using AG News and Measuring Hate Speech at roughly 1% prevalence, the authors compare random sampling with eight active learning strategies, measuring recall on subgroups omitted from the starting set as well as F1. Uncertainty-based strategies recover fewer omitted subgroups than random sampling, while discriminative active learning recovers 55% of them without a significant F1 loss. The authors note that aggregate metrics hide these differences.
In briefTRIG: A Metric for Measuring Texture Realism in Generated Images
SummaryThe paper argues that common metrics for evaluating generated images rely on backbones trained for tasks like classification or image-text alignment, which makes them poor at capturing texture realism. The authors propose TRIG, which uses the discriminator of a GAN pre-trained on a texture dataset as its backbone, with a plausibility head fine-tuned to detect local texture degradations. A distributional distance is computed between real and generated image features. The authors report closer agreement with human realness ratings than baseline metrics.
In briefSanzaru: Benchmark for Team Protocol Formation Under Unequal Information and Communication
SummaryThe paper introduces SANZARU, a benchmark built from cooperative bomb-defusal puzzles adapted from BOMBANANA!, in which three agents each lack a different ability: hearing, speaking, or seeing. It contains 120 puzzle instances across seven families and reports results from 1512 episodes. Across seven models, the overall success rate is 49%, and trace analysis attributes about 60% of 771 team losses to communication.
In briefCompressing LLMs by fusing non-adjacent similar layers via graph clustering
SummaryThis paper addresses depth compression of large language models, noting that existing methods mostly merge only adjacent layers, while non-adjacent layers show many more highly similar pairs. The authors propose FCM, which builds a full layer similarity graph, uses spectral community detection to group functionally equivalent layers, keeps the highest-influence layer per cluster, and applies LoRA recovery. At 2x compression, FCM reports lower perplexity than ShortGPT on twelve of thirteen models.
WhyUses full pairwise CKA graphs and spectral clustering to exploit non-adjacent layer redundancy, validated across thirteen LLMs with statistical tests and strong baselines.
In briefLearning links between propositions from sentence-level labels for neuro-symbolic NLI
SummaryThis paper studies neuro-symbolic natural language inference, where the premise and hypothesis are each translated into propositional representations. The authors propose a trainable atom linker that jointly predicts whether links exist, which source to select, and the relation type. It is trained only from sentence-level NLI labels, without atom-pair annotations. On SNLI, SICK, and MNLI, the system improves accuracy over a baseline and over the prior fixed-linker AMR-to-SAT system.
In briefRootSLAM: Square-Root Bundle Adjustment for Visual-Inertial SLAM
SummaryThis paper presents RootSLAM, a visual-inertial SLAM system for multi-camera setups built on Basalt, which estimates a device's motion and map from its surroundings. The authors adapt square-root bundle adjustment from structure from motion using a pose-graph-based reformulation, allowing the global optimizer to run in single precision and roughly double its speed. The authors report accuracy competitive with state-of-the-art methods while running up to 20 times faster, and state that the implementation will be open-sourced.
WhyAdapts square-root bundle adjustment to multi-camera visual-inertial SLAM via pose-graph reformulation, backed by benchmarks and ablations showing doubled optimizer speed.
Tom\'as Santucci, Mateo de Mayo, Daniel Cremers et al.
In briefJoint Data Fusion and State Estimation for Sparse, Multi-Rate Distribution Grid Sensors
SummaryThis paper addresses state estimation in distribution grids where sensors are sparse, heterogeneous, and sampled at different rates. The authors propose a heterogeneous latent graph ODE in which measurement functions and sampling processes are typed and placed on specific buses, paired with a graph-transformer estimator. The method performs fusion and estimation jointly rather than in two separate stages. Its effectiveness is tested across several IEEE standard topologies as observability degrades.
Pooja Algikar, Stefano Fenu, Kumarsinh Jhala et al.
In briefWhere Did We Go Wrong? Locating Errors to Repair Multi-Turn Chats
SummaryThis paper studies how LLM assistants in multi-turn conversations can misread a user's unclear intent and corrupt their working state, causing the conversation to fail. It introduces Loci, a framework that probabilistically identifies which updates excluded the true intent and selectively retains, verifies, or removes them. In controlled tests Loci localizes errors better than baselines, and in free-form conversations it reduces mean recovery cost by 46-79% relative to the strongest baseline.
WhyProposes probabilistic localization and selective repair of misinterpretations in multi-turn dialogue, validated in controlled and free-form settings with 46-79% lower recovery cost.
In briefHarmony Harness: Multi-objective self-evolution of AI agent frameworks
SummaryThis paper introduces Harmony Harness, which treats the framework (harness) around an AI agent as something to optimize for several competing goals at once. It uses decomposition-based evolutionary search to keep solutions across different trade-off regions, and evolves general mechanisms and benchmark-specific adaptations separately before combining them at inference time. On SecRepoBench and BaxBench, it consistently outperforms strong baselines, with gains exceeding 10% on both.
Weichen Yu, Lesly Miculicich, Bhavana Dalvi Mishra et al.
In briefCompute-optimal scaling rules for spiking language models, which use sparse binary activations to save energy.
SummaryThe paper studies how spiking language models scale, using an RWKV-based SpikeGPT family trained on FineWeb-EDU across a range of compute budgets. It finds that the compute-optimal balance between parameters and data stays close to the dense Transformer result, so spiking does not change that tradeoff. Compared with non-spiking baselines, the loss gap splits into a fixed backbone cost and a spiking cost that shrinks as compute grows, giving a recipe and a quantified accuracy cost.
In briefSingle-step retrosynthesis trained with forward-consistency reinforcement learning (RetroRLFF)
SummaryThis paper proposes RetroRLFF, a method for single-step retrosynthesis that predicts reactants from a target product. It starts from a supervised Transformer and applies reinforcement learning using a reward that combines agreement with reference reactants and forward consistency, checked by a frozen forward reaction model. On USPTO-50K it reports higher candidate-level round-trip accuracy than all 14 baselines at k = 1, 3, 5, and 10, with 90.1% top-1 round-trip accuracy and 57.9% exact-match accuracy. In an anonymized expert assessment of 131 cases, organic synthesis experts preferred RetroRLFF in 67.9% of comparisons.
In briefWhen Dynamic Scaling Fails: How Scale-State Feedback Shapes Recurrent-State Quantization in Linear-Attention and SSM LLMs
SummaryThe paper studies quantizing the fixed-size recurrent state of linear-attention and state-space language models to reduce decoding cost. It finds that the quantizer scale and the state it is derived from form a feedback loop, so errors compound across steps, and that numerical instability and functional degradation are separable failures. Experiments compare outlier-aware schemes against plain persistent scaling and examine learned bit allocation across layers.<br>
In briefTrust Hinders Bidirectional Human-AI Alignment: A paper arguing that trust calibration alone is insufficient for AI-assisted decisions with ill-defined goals
SummaryThe paper examines trust calibration, the dominant framing of AI-assisted decision making that aims to match users' trust in AI with its actual trustworthiness. The authors argue this framing suits well-defined, clearly scoped problems but overly restricts AI-supported decisions. They propose three dimensions: trusting AI is only one way to manage uncertainty, trust calibration is often misapplied to ill-defined large-world problems, and alignment in such settings requires addressing higher-order decision processes.
In briefLumen-bench: Cross-lingual safety gaps and the say-do gap in tool-calling LLMs
SummaryThe paper introduces Lumen-bench, a benchmark testing whether tool-calling language models behave ethically, across 30 models, five languages, and 194,400 trials. It finds higher harmful-action rates in several non-English languages than in English under the ethical prompt, and that a neutral system prompt raises pooled violation rates from 2.76% to 13.64%. In 8.57% of substantive-call cases, models refused or objected verbally while still making a harmful tool call.
WhyCombines say-do gaps in tool calling with cross-lingual safety differences, supported by 194,400 trials and multi-judge calibration.
Giovana Kerche Bon\'as, Thiago Laitz, Thales Sales Almeida et al.
In briefGesBench: When and Where Multimodal Models Perceive Social Meaning
SummaryThis paper introduces GesBench, a benchmark for testing how multimodal language models perceive speaker intent and social meaning as an utterance unfolds. It uses cumulative sub-utterance clips to measure when models form stable interpretations, and human gaze-guided corruption to test whether regions humans attend to matter for model judgments. Across open and closed models, humans form predictive understandings earlier and revise more often, while models often hedge or repeat across clips.
In briefDeferring to Experts or Answering Directly: Online Learning with LLMs and Congested Human Experts
SummaryThe paper studies how an AI system should decide whether a large language model answers a query or defers it to a capacity-limited human expert when the expert queue is congested. The authors propose KL-PBD, an online learning algorithm that uses Kullback-Leibler lower confidence bounds to learn the reliability of the LLM's confidence signal from delayed, selectively observed expert feedback. They report O(K log T) gap-dependent regret and Õ(√(KT)) worst-case regret, with matching lower bounds up to logarithmic factors. The method is validated on synthetic data and on GSM8K, NuminaMath, and MedMCQA using Qwen LLMs..
Vasileios Saketos, William Reveillard, Alexandre Proutiere et al.
In briefRIECNN: A Real-Time Image-Enhanced CNN for Recognizing Traffic Signs
SummaryThis paper proposes RIECNN, a real-time traffic sign recognition approach that combines image enhancement with a convolutional neural network to handle environmental effects and slow deep-learning inference. Tested on the GTSRB, BTSC, and rMASTIF benchmarks, it reports recognition accuracies of 99.75%, 99.25%, and 99.55%. Pre-processing and inference together take no more than 1.3 ms per image, and the method is reported to be robust to brightness and contrast variations.
Reem Abdel-Salam, Rana Mostafa, Ahmed H. Abdel-Gawad
In briefHow Data-Topology Compatibility Enables Personalized Decentralized Learning
SummaryThis paper studies decentralized learning where agents with compatible local objectives and communication topology form groups that reach similar distributions without explicit personalization. Under exact compatibility, the authors give a stability criterion for distributional synchronization and describe the stationary distribution's orbit structure. Under approximate compatibility, they prove O(ε) error bounds in 2-Wasserstein distance, and experiments on synthetic data, MNIST, and CIFAR-10 agree with the theory.
WhyIntroduces orbit-wise personalization via data-topology compatibility, proving stability criteria and O(ε) error bounds, validated on synthetic data, MNIST, and CIFAR-10.
In briefShellElastic: Elastic pretraining that scales and transfers hyperparameters
SummaryThe paper presents ShellElastic, an elastic transformer that arranges hidden dimensions into a dense core and successive shells, so all sub-networks can be trained in one full-width forward pass with a fixed computation graph. It also introduces μE, an extension of maximal update parameterization that keeps all sub-networks maximally updated and lets hyperparameters tuned on a small proxy transfer to elastic and larger supernets. According to the authors, a single run yields sub-networks that outperform MatFormer and fully elastic baselines at matched parameter counts and approach separately trained models.
WhyExtends muP to elastic architectures with provable simultaneous maximal updates and a fixed-graph shell design; experiments versus MatFormer are present but of limited scale.
In briefscHFBiMamba: A Compact Bidirectional Mamba Autoencoder for Heart Disease Classification from Single-Nucleus RNA-seq
SummaryThis paper presents scHFBiMamba, which combines a bidirectional Mamba autoencoder with an MLP head to jointly classify species, cell type, and disease state from heart failure single-nucleus RNA-seq data. The model uses about 13.2M parameters, roughly 14.6 times fewer than the dense denoising autoencoder baseline. It reports higher aggregate F1 than the baseline and provides SHAP-based gene attributions for interpretation.
In briefDrive3R: Block-wise streaming 3D reconstruction model for long autonomous driving sequences
SummaryThis paper presents Drive3R, a block-wise streaming geometric foundation model for dense multi-view 3D reconstruction in long autonomous driving sequences. It uses block-wise causal attention to address memory and complexity limits, and confidence-aware re-anchoring to reduce long-range drift. On nuScenes, it improves over the DVGT-2 baseline, lowering AbsRel from 0.1806 to 0.1081 and ATE from 1.5812 to 0.9092.
In briefEnergy-guided learning for the heat equation with a noisy source, with optimal estimation and generalization bounds
SummaryThis paper studies how to learn solutions of the heat equation with a noisy source term from finite, randomly sampled data, and analyzes the statistical estimation and generalization properties of that learning. The authors derive a variational formulation based on an energy-regularized, ellipticity-debiased Sobolev regression, using auxiliary samples of the source function as a physics-informed regularizer. They show that parabolic PDEs exhibit anisotropy between spatial and temporal regularity, and derive non-asymptotic rate-optimal estimation and generalization error bounds. Numerical experiments on various parabolic and hyperbolic PDEs report substantially lower test errors than several physics-informed neural network methods.
In briefStructural Chain-of-Thought: Reasoning via Executable Computational Graphs
SummaryThis paper proposes a Structural Chain-of-Thought framework in which language models generate executable computational graphs instead of free-form reasoning text, and the final answer is produced by deterministically running the graph. Training uses graph-based reinforcement learning that rewards structural validity, answer correctness, and node overlap with reference graphs. Experiments on five benchmarks and four model families report consistent accuracy gains and generation length reductions of up to 5-6x.
Petr Senatorov, Karim Vafin, Elena Tutubalina et al.
In briefRELAGRAM: A Recursive Latent Graph Model for Vehicle Routing with Capacity Limits
SummaryThis paper introduces RELAGRAM, a method that applies latent recursion, repeatedly passing internal representations through the same network, to graph combinatorial optimization. It targets the capacitated vehicle routing problem (CVRP), encoding instances as edge tokens that a compact transformer updates recurrently, with a decoder building feasible routes. On CVRP-100, the 0.53M-parameter model reaches gaps of 13.4% to 8.3% relative to a PyVRP reference, depending on decoding strategy. Trained only on CVRP-100, it generalizes zero-shot to CVRP-500 and CVRP-1000.
In briefRegExtract: A Benchmark for Extracting PFAS Regulatory Information with LLMs
SummaryThis paper introduces RegExtract, a benchmark for pulling structured information from PFAS regulatory documents. It contains 488 human-annotated field instances from 20 US state documents and 46 regulations, each paired with a provenance entry. Experiments with open and closed LLMs show that document identification fields remain difficult, while field-specific few-shot examples improve extraction.
Defne Circi, Krystelle Lionti, Sarathkrishna Swaminathan et al.
In briefPruning random features for multi-class classification by maximizing successive empirical risk differences
SummaryWeighted sums of random kitchen sinks (W-RKS) is a common feature extraction method for supervised learning, but many of its random features are redundant or ineffective. The paper proposes the SEED-MAX algorithm, which recursively computes empirical risks to select the most parsimonious yet informative feature set, and derives a lower bound on the probability of correct selection. Adding a Fisher test further prunes features, achieving about 92 percent pruning on UCR time-series datasets with negligible loss in classification accuracy.
In briefTIDE: Distilling Search into a Reusable Skill for Efficient Text-to-Video Prompt Optimization
SummaryThis paper proposes TIDE, a framework that distills costly iterative prompt optimization for text-to-video models into a reusable, interpretable prompt-rewriting Skill. The Skill uses source-conditioned rule modules that produce positive and negative prompts in one LLM call, followed by a single video generation, with no test-time visual search or model fine-tuning. Evolved using only 100 prompts over 14 hours, it achieves performance comparable to search-based methods and outperforms approaches designed for unseen video-generation models when transferred to them.
In briefAnchors Boost Control and Reasoning in Diffusion Language Models
SummaryThis paper studies how to control diffusion language models (DLMs) so that outputs meet task constraints while keeping their original abilities. It proposes AnchorDLM, in which an anchor generator produces a short output fragment and its position as control tokens for a frozen DLM, trained with reinforcement learning using rewards for answer correctness and constraint satisfaction. Experiments on code generation and mathematical reasoning across five DLM backbones show improved controllability and gains in reasoning performance on diverse benchmarks.
