In briefeSharp: Sharp Monocular Metric Depth from Event Cameras Using Diffusion Models
SummaryThis paper introduces eSharp, a monocular depth estimation framework that uses only event camera data to handle fast motion and high-dynamic-range scenes. It adapts pretrained diffusion priors for event-based depth prediction and trains with dense pseudo-labels distilled from a spatially co-registered frame camera and calibrated with LiDAR. Experiments report that eSharp outperforms prior event-to-depth baselines in metric depth accuracy and spatial fidelity.
In briefOne-step diffusion model predicting future brain MRI changes in aging and neurodegeneration
SummaryThis paper proposes a covariate-conditioned one-step diffusion framework that predicts longitudinal changes in 3D gray-matter density maps over prediction horizons of 2 to 10 years. The model predicts residual changes in a single denoising step and adds them to the baseline map to reconstruct future anatomy, conditioned on baseline anatomy, age, diagnosis, APOE4 allele count, and follow-up time. On the held-out ADNI test set it reports a Δ-Pearson of 0.478, and without retraining it reports 0.450 on the external AIBL dataset.
Natenaile Asmamaw Shiferaw, Ali Farki, Deepika Koundal et al.
In briefStructuring Survey Histories for LLM Digital Twins: Background, Decision Procedure, and Evaluation Study
SummaryThis paper studies a fixed way of organizing a respondent's survey history into Background, Decision procedure, and Evaluation (BDE), so that LLM-based digital twins can predict individuals' answers in new settings. On the Twin-2K-500 dataset, predicting held-out answers across 17 tasks, the full BDE profile improved overall accuracy over raw survey transcripts on both tested models. The authors find the advantage comes mainly from profile composition, especially adding Evaluation, rather than from compression or deeper extraction.
In briefDivide, Diagnose, and Compose: a method for optimizing AI agent pipelines
SummaryThis paper introduces DeCAF, a framework for optimizing agentic AI pipelines, covering prompts, topology, model choice, and tool policies. It optimizes each design dimension separately and records reasoning for every decision, then synthesizes compatible findings across dimensions. On a suite of agentic benchmarks, it improves average task performance by 2.58 percentage points, reduces inference cost by 4.30%, and cuts end-to-end latency by 42.45% compared to the best overall baseline.
Divij Handa, Christopher Tensmeyer, Jihyung Kil et al.
In briefRemembering Is Not Enough: Timing and Omissions in Long-Term AI Mental Health Support
SummaryThis paper examines whether AI chatbots with memory give appropriate mental health support as users' needs change over time. The authors analyzed 14,256 longitudinal simulations based on chatbot histories donated by 22 users, covering suicidal ideation, self-harm, and anxiety crises. They introduce the CARE framework, developed with mental health professionals, to assess the timing, quality, and omissions of advice. Findings show that post-crisis advice was less aligned with users' readiness, mainly due to premature recommendations, and that expected follow-up and next-step planning were frequently omitted.
In briefA memory-aware multi-agent design in which LLMs orchestrate conversational recommendation
SummaryThis paper proposes a multi-agent architecture for conversational recommendation in which large language models coordinate intent refinement, query construction, and interaction flow, with long-term memory as a core component. Experiments on the Amazon Reviews dataset show that memory-enabled interactions recover up to 63% of the gap toward explicit-query scenarios, and supervisory orchestration raises user acceptance by 28.9 percentage points. The authors also introduce the Semantic Consistency Score (SCS) as a complementary evaluation measure.
Emilly Pitman de Castro, Vanessa Silva, Gustavo Kanno et al.
In briefPatching Blind Spots: A Reversible LLM Rule Layer for Frozen Fraud Detectors
SummaryThis paper addresses blind spots in a deployed gradient-boosted tree fraud detector, where retraining may be unsafe. Large language models propose rules on features the model underweights according to SHAP, and these rules are screened and calibrated into a reversible additive overlay that leaves the original model unchanged. Averaging three LLM backbones improves worst-case stability. On a real e-commerce fraud detector, the overlay raised recall at equal precision by 2.26% and precision at equal recall by 1.60% on average.
In briefAn Educated Guess: Statistically Aligned Gradient Estimates for Zeroth-Order Optimisation
SummaryThe paper addresses training machine learning models on novel hardware such as neuromorphic chips, where backpropagation's global synchrony and exact differentiability requirements are limiting. It proposes Spectral Zeroth-Order methods (SZO-κ), which derive search directions from activation geometry in the forward pass to better align with the true gradient. The methods are evaluated on multilayer and convolutional networks against standard perturbation baselines and backpropagation, examining gradient alignment, bias-variance structure, and training stability.
Justus Huebotter, Serge Thill, Marcel van Gerven et al.
In briefReward design for fraud detection that also learns from blocked payments
SummaryFraud-detection models are trained only on transactions that were approved, since blocked ones never receive fraud labels. This paper uses a large language model to write a dual-stream reward for a reinforcement learning policy: one stream covers labeled transactions, and the other scores blocked, unlabeled ones. On a held-out four-month out-of-time test, the approach outperforms the prior labeled-only agent at each target recall and the production baseline on average.
In briefContextuality and Contradictions: How People Identify AI-Generated Writing
SummaryThis paper studies how English Wikipedia editors and students identify text as AI-generated, combining qualitative coding with computational analysis of lexical change. It finds that these identifications rely on varied, contradictory, and context-dependent evidence, often triggered by a perceived mismatch with the text's context rather than any specific feature. The authors argue such identifications can serve to define sociolinguistic norms, so tools that treat AI-likeness as context-free may enforce the norms of their training contexts at scale.
Michelle Morgenstern, Julia Witte Zimmerman, Tabia Tanzin Prama et al.
In briefModel errors as signals of irregular urban bus service
SummaryPretrained trajectory models that transfer to a new city produce errors, which are usually treated as performance limits. This paper instead uses these residuals to examine systematic local conditions, analyzing spatial and travel-time residuals of UniTraj and TrajFM with NetMob26 bus, ticketing, weather, network, and socioeconomic data from Niteroi. Explainable AI is used to relate high residuals to urban context variables. The authors treat high residuals as candidate proxy signals of transit irregularity, not direct evidence of service disruption.
Csaba I. Sidl\'o, Jamil J. Ghazal, Domonkos R\'ozsay et al.
In briefFoleyForge: a controllable model for editing sound effects from free-form instructions
SummaryThe paper presents FoleyForge, a generative model that edits sound effects from natural-language instructions while keeping the identity of the input clip. One model covers semantic, stylistic, and DSP-like changes. The authors build a data pipeline in which an audio language model authors both DSP-based acoustic edits and similarity-preserving semantic edits, then fine-tune a rectified-flow audio generation model on this data. The model also exposes controls for loudness change, edit strength, and an input-derived contour envelope. The authors report state-of-the-art instruction adherence with competitive audio quality and source faithfulness.
Giovana Morais, Hugo Flores Garc\'ia, Prem Seetharaman et al.
In briefRetraining-free accurate inference on analog in-memory computing accelerators
SummaryAnalog compute-in-memory accelerators perform matrix-vector products inside memory arrays to improve inference efficiency, but converter and device errors make pretrained models unusable without costly hardware-aware retraining. The paper proposes an offline method that changes how each layer is represented on the array without changing its computation, centering inputs and applying Hadamard rotations designed for analog arrays with low-cost shared calibration. Experiments on PCM and ReRAM across CNNs, vision transformers and LLMs up to 14B parameters show accuracy near floating-point levels. On ReRAM with 6-bit ADCs, ViT-B/16 improves from 37.4% to 82.9% and Phi-4 from 27.0% to 72.2%.
WhyThe authors report retraining-free inference via offline representation changes on analog arrays, tested on CNNs, ViTs, and LLMs up to 14B on PCM and ReRAM.
In briefCompressed per-player local methods for multiplayer federated learning with exact convergence
SummaryThis paper studies multiplayer federated learning (MpFL), where clients act as game players with individual objectives seeking an equilibrium. The authors propose C-PEARL-SGD, in which players share local strategy increments instead of strategies, which preserves exact convergence under compressed communication. They also introduce PEARL-SVRG and its compressed variant C-PEARL-SVRG, and numerical experiments, including distributed spectrum sensing, show practical gains over vanilla PEARL-SGD.
In briefCancellation alone is not enough: what client retries leave behind in an overloaded LLM engine
SummaryThe paper studies how clients retrying timed-out calls can keep an overloaded LLM engine from recovering even below capacity. Experiments on vLLM with three open models and scheduler traces show that propagating cancellation alone did not restore service in any run. Many retries started prefill shortly before timing out, and aborted executions consumed most of the prompt tokens computed.
In briefCompression of Emotion in Language Model Representations: how large language models internally organize fine-grained emotions
SummaryThis paper examines the internal representations of seven large language models on GoEmotions, an English Reddit dataset labeled with 27 fine-grained emotion categories plus neutral. The authors compare model predictions and hidden representations against the 28 categories and coarser emotion or sentiment groupings. They find a consistent overcompression relative to the fine-grained taxonomy: unsupervised clustering most clearly separates positive from negative emotion, and coarser groupings are recovered with fewer clusters. Linear probes, however, still retain substantial fine-grained predictive information.
In briefTransporting the Past: Optimal transport view of backtracking counterfactuals
SummaryThis paper studies backtracking counterfactual inference in Structural Causal Models (SCMs) using optimal transport. Because a factual observation typically induces a posterior distribution over exogenous variables, counterfactual inference is inherently distributional. The authors propose two complementary frameworks that transport this posterior onto exogenous configurations satisfying a desired counterfactual query, along with marginal, penalty-based, and hard-constrained solvers for finite posteriors. Synthetic experiments on non-bijective SCMs reveal trade-offs among transport cost, constraint satisfaction, and preservation of exogenous independence.
In briefGreedy Byte-Pair Encoding Reaches At Least 0.361 of Optimal Compression
SummaryByte-Pair Encoding (BPE) builds a tokeniser by repeatedly merging the most frequent adjacent pair of symbols, measuring its usefulness by the total string-length reduction from k merges. It was previously known that this greedy rule achieves at least one third of the optimal reduction. The paper proves that it always achieves at least 0.361 of the optimal reduction, for every string, budget, and tie-breaking rule, improving the one-third bound.
WhyThe authors improve the greedy BPE compression lower bound from one third to 0.361, give a proof, and explain why one third was a barrier.
In briefChainFlow: Few-Step Latent Flow Matching for Speculative Decoding
SummaryThe paper introduces ChainFlow, a new speculative decoding method and the first flow-matching drafter conditioned on the target model's hidden states. It needs only one or two flow steps to draft tokens, compresses hidden states with a variational autoencoder, and uses lightweight feedback heads to restore dependencies between tokens. On Qwen3.5 models, with matched drafter size and training data, it outperforms EAGLE-3 and DFlash in vLLM by 39% and 11%.
WhyThe authors introduce the first flow-matching drafter conditioned on target hidden states, reporting 39% and 11% gains over EAGLE-3 and DFlash on Qwen3.5 in vLLM.
In briefPdb or Not to Be: Comparing Debugging Interfaces with pass_X@k
SummaryThis study compares a debugger-only coding agent with a bash-only agent on all 500 SWE-bench Verified instances, keeping the rest of the scaffold fixed, across two models with three independent runs each. The authors introduce pass_X@k, a cross-agent pass@k metric that estimates the probability that at least one of k agents succeeds. They find that the two agents solve partially disjoint subsets of tasks, but this largely reflects run-to-run variance rather than tool complementarity. Two bash rollouts outperform one bash and one debugger rollout across both models.
In briefDynamic GPU voxel grids and sequential voxel processing in NanoVDB
SummaryThis paper adds two capabilities to the NanoVDB data structure: GPU-native dynamic topology updates and a voxel-sequential access method for sparse grids. The new topology operators include morphological dilation, grid merging, pruning, and cross-resolution coarsening and refinement, which the authors say can rebuild grids of hundreds of millions of voxels in a few milliseconds. The access method streams active voxels as a flat sequence, which the paper reports lets stencil operators approach peak memory bandwidth. The operators are benchmarked on applications including 3D reconstruction, fluid simulation, and large-scale mesh voxelization.
Eftychios Sifakis, Francis Williams, Christopher Horvath et al.
In briefJEPAhead: From Latent Inpainting to Causal Forecasting for Urban Driving
SummaryThe paper adapts a pre-trained V-JEPA1 predictor from bidirectional latent inpainting to action-free causal forecasting while keeping the encoder frozen. On CARLA Town12, the model predicts eight future token groups, equal to sixteen frames, in one forward pass and retrieves the correct future among 591 candidates in 69.0% of cases, versus 55.3% for persistence. The authors also report semantic and pixel-level evaluations, and note that downstream utility remains unverified.
Daniel G. P. Petrini, Davi Pincinato, Fabio Grassiotto et al.
In briefAttributing model components to jet physics via learned surrogates
SummaryThis paper examines the Particle Transformer (ParT), a transformer model used to identify jets in high energy physics. It proposes component attribution methods based on Kolmogorov-Arnold Networks (KANs) that link directly to the physical descriptors of jets. The goal is to test whether mechanisms such as Multi-Head Attention learn these physical quantities as discriminating features.
In briefScoring, Not Learning: How Catalog Consolidation Affects Retail Product Recognition Accuracy
SummaryThis paper studies a cheese product catalog built from web image search, containing 165,769 images across 1,824 query-named entries. After filtering and deduplication, the authors consolidate entries into 844 base types. Scoring the unchanged model against the merged label space gains about 4.27 points without retraining, while retraining on the consolidated catalog adds essentially nothing. The gain mainly comes from category coarsening rather than alias correction.
In briefA 3D segmentation pipeline for CT scans of prostate surgery practice models
SummaryThis paper presents an automated pipeline that segments 3D CT scans of prostate surgical phantoms, which are laboratory models used to test robotic surgery systems. It uses classical medical image-processing methods that encode known anatomy as geometric and intensity priors, so it needs no training data and transfers to new phantom revisions without re-annotation. Across 57 scans, it achieved a median Dice of 0.83 on the capsule interior in 11 contrast-suppressed scans and detected all 32 annotated lesions without false positives, taking 2 to 8 minutes per scan on a CPU versus 1 to 2 hours by hand.
In briefTurkish Video Fact-Checking: A New Dataset and Comparison of Two AI Systems
SummaryMisinformation in videos spreads quickly on social media, yet most fact-checking tools only handle English text, leaving low-resource languages like Turkish poorly served. The authors introduce TeViD, which they describe as the first multimodal video fact-checking dataset for Turkish, built from Teyit.org articles with translated Snopes articles and synthetic examples added. They build two verification pipelines on the same open-source base model: 3MFact-TR, an adapted multi-role framework with added modules for inconsistency checking, question management, and evidence retrieval that excludes fact-checking sites, and a fine-tuned model using LoRA. The paper reports accuracy, macro F1, and related metrics for both systems along with ablation studies.
In briefVoluCompose: Compact Visual Representations of 3D Medical Images
SummaryMedical vision-language models can interpret volumetric (3D) medical images, but processing hundreds of slices produces long, redundant visual-token sequences that are costly to handle. VoluCompose is a training-free, model-agnostic framework that separates selecting which visual content to keep from deciding how many tokens to allocate. It groups neighboring slices using hyperbolic token composition with Lorentz distance, and adjusts token budgets using visual-encoder attention entropy. Across nine calibrated settings on two tasks, three datasets and four backbones, it retains on average 22.78% of visual tokens and reports a mean relative quality of 121.68% against uncompressed references.
In briefIWIBench: a benchmark for gradient-free, model-driven editing of model weights
SummaryThis paper introduces IWIBench, a benchmark measuring how well a large editor model can modify the weights of a smaller target model (typically Gemma 3 4B IT) to achieve one of ten text-editing objectives without gradients, while preserving other capabilities. Evaluation of frontier models shows sharply varying abilities across both models and tasks, with a few models performing far ahead of the rest. Success on one task family correlated poorly with success on others.
Chris von Csefalvay, Pranav Doma, Achyuthan Unni Krishnan et al.
In briefFrom Geometry to Function: A Multi-Task LiDAR Benchmark for Parsing Urban Building Facades
SummaryThis paper introduces HFX3D, a large-scale LiDAR benchmark of about 492 million points from 26 buildings in Halifax, Canada, captured across four seasons, day and night, and varied weather. Each facade is annotated for semantics, instances, and functional attributes per instance, using a semi-automatic pipeline that combines geometry-guided heuristics with expert validation. The authors benchmark point-based, graph-based, convolutional, and transformer models. They find that semantic segmentation is already challenging under severe class imbalance, while instance separation of repetitive structures and functional reasoning from geometry are substantially harder.
In briefDecision Context: a system that makes AI agents remember behavioral corrections across sessions
SummaryLLM agents repeat the same mistakes because they do not learn from user corrections across sessions. The paper presents Forge Decision Context (forge_dc), which captures behavioral signals from human corrections and self-assessed tool-call failures, compiles them into imperative RULE directives, and injects them at the relevant decision points. On DC-Bench, a 71-scenario benchmark across nine behavioral dimensions, forge_dc reaches 95% compliance versus 55% with naive retrieval, while injecting about ten times fewer tokens per query.
In briefFinEvolveKG: Controlled, Traceable Prompt Evolution for Financial Knowledge Graphs
SummaryThis paper presents FinEvolveKG, a framework for building financial knowledge graphs from SEC 10-K filings. It separates shared general rules from persona-specific policy prompts and restricts the prompt optimizer to edit only the persona block, so every change is traceable. In experiments across three personas and three seeds, the constrained two-stage workflow outperformed both a frozen prompt baseline and an unconstrained full-prompt optimization workflow, and it also led on out-of-distribution shifts.
Abhinav Arun, Jayita Bhattacharyya, Fabrizio Dimino et al.
In briefTest-time method for picking the best AI answer through aspect-by-aspect pairwise contests
SummaryThis paper introduces Aspect-wise Pairwise Tournaments (APT), a method for selecting the best of several candidate solutions generated by a language model. Candidates face off in a single-elimination tournament, with each matchup decided by votes from multiple off-the-shelf aspect verifiers. Across five domains and three generator models, APT outperforms binary multi-aspect approval, self-consistency, and reward model verification, with the largest gains on open-ended tasks.
In briefForcing a Letter Suppresses Refusal: Mechanistic Evidence from WMDP
SummaryThis paper explains why language models refuse less often when safety questions are posed as multiple-choice, using the WMDP dataset and the Llama-3.1-8B and Qwen3-8B models. The models commit to an answer-letter token very early in multiple-choice format, and those activations sit about 3x further from the refusal direction than in open-ended format. Ablating the attention heads that carry the answer raises refusal, suggesting answer selection competes with refusal. The authors conclude that multiple-choice safety scores mainly measure knowledge rather than safety mechanisms.
Stepan Epifantsev, Anna Antipova, Oleg Chernikov et al.
In briefBenchmarking machine learning for predicting chaotic three-body system stability and trajectories
SummaryThe paper simulates three-body systems near the boundary between stable and unstable initial conditions and designs two learning tasks: stability prediction and trajectory prediction. It benchmarks machine learning methods, from simple regression to foundational physics models, against astrophysical approaches such as the Mardling-Aarseth stability criterion. Results show early dynamics support strong stability prediction, while precise long-horizon trajectory forecasting remains difficult. Code and dataset are publicly available.
In briefBEVCast: Path-adaptive bird's-eye-view maps for efficient self-driving
SummaryCamera-based self-driving models usually spend the same computation on every cell of a dense bird's-eye-view grid, even though a car only passes through a few cells within its planning horizon. BEVCast first encodes the scene on a coarse grid, then uses a heat head trained on the recorded path to refine the hottest cells to a finer resolution under a fixed budget. On nuScenes it reports a mean displacement error of 1.078 m versus 1.149 m for a full fine grid, and on NAVSIM navtest 0.577 m versus 0.598 m.
In briefSPRING-5: Benchmark Measuring Efficiency of Automatic Tool-Documentation Repair for LLM Agents
SummaryThis paper introduces SPRING-5, a benchmark that scores methods for automatically repairing tool documentation on call cost rather than only task accuracy. It covers 20 tool families and 344 tasks, each with a degraded specification, an honest reference specification, and an answer key. Against this benchmark, a schema-only method closes 1% of the baseline-to-ideal cost gap, an online self-evolution loop 13%, a single batch rewrite over all training transcripts 33%, and a variant given correctness labels 58%.
In briefLLM-based feature extraction to add customer review signals to multi-criteria inventory classification
SummaryThis paper proposes a method that uses large language models to turn unstructured data, such as customer reviews, into quantitative features via vector embeddings and LLM-generated evaluation criteria. These features are added to a multi-criteria inventory classification framework based on the Hadi-Vencheh model. Experiments show measurable reclassification when customer sentiment signals are included, suggesting that unstructured data can improve inventory prioritization.
In briefs³: Simple Scaling of Swarms via Cooperative Multi-agent RL (a method for letting agents organize themselves)
SummaryThe paper introduces s³, a multi-agent reinforcement learning method in which agents decide for themselves how long to work alone and when to share findings or read teammates' messages, without predefined organization or roles. With matched training and inference budgets, swarms evaluated with twice as many agents as in training solved over 25 times more coordination problems than the strongest baseline. In parallel exploration, communicating agents reduced repeated work by nearly 50% and outperformed best-of-N by up to 9.4 percentage points.
Kunal Jha, Gon\c{c}alo Faria, Shuyue Stella Li et al.
In briefNormGym: Testing AI Agents' Ability to Learn Social Rules Through Interaction
SummaryThis paper introduces NormGym, an interactive benchmark where a newcomer agent must infer a hidden, arbitrary norm from observed behavior and sanctions while pursuing its own task in a multi-agent community. The strongest evaluated large language model reaches a joint task-and-compliance success rate of 55.3%. The authors find that models can sometimes revise their behavior after being sanctioned for their own actions but struggle to infer the underlying norm from other agents' behavior and sanctions, and that fine-tuning may not be sufficient to improve this capability.
In briefInstaGen: Fast video generation using shared importance sampling
SummaryThe paper addresses the high computational cost of attention in diffusion Transformers used for video generation. It proposes InstaGen, which treats attention block selection as importance sampling: top-scoring keys are computed exactly, while the remaining keys are sampled in proportion to estimated attention and reweighted. Each query cluster shares one sample so the computation runs efficiently on GPUs. On Wan2.1-T2V models, InstaGen built on SVG2 achieves higher PSNR than SVG2 at matched attention density and at matched wall-clock time.
Kislay Aditya Oj, Kumar Krishna Agrawal, Sri Surya Varma Datla et al.
In briefJust CaloCHILL: fast simulation of hadronic showers in a learned latent space
SummaryThis paper presents CaloCHILL, a generative model for calorimeter simulation that produces a compact summary of a particle shower in a learned latent space and expands it to full resolution, giving each hit its position, energy and time. Tested on positive-pion showers in the International Large Detector calorimeter, a proposed detector for FCC-ee, it reproduces the main shower observables, including time structure. The authors report that it is 600 times faster than its predecessor while improving on all accuracy metrics considered.
Henry Day-Hall, Frank Gaede, Gregor Kasieczka et al.
In briefUser Embeddings as Combinations of Interpretable Behavior Patterns
SummaryThis paper argues that user embeddings can be understood as sparse combinations of interpretable behavioral modes. Sparse autoencoders recover these modes better than classical sparse baselines on synthetic data, including transformer-encoded embeddings. The method is then applied to classical and transformer-based user embeddings from two real-world datasets. The resulting natural-language labels enable zero-shot audience retrieval and characterization on embeddings never trained on text.
Jonathan Kozaczuk, Zheng Dong, Noyan Evirgen et al.
In briefStudy Compares AI-Revised Research Figures Against Originals for Clarity and Design
SummaryThe authors collected 371 motivation and framework figures from 200 papers and had a multimodal model rate six dimensions of communication and design. They then attempted source-constrained AI revisions of all figures, and 369 passed recorded fidelity checks. Revisions raised mean clarity ratings by 0.428 points and lowered aesthetic coherence by 0.111 points on a five-point scale. The authors conclude that researchers should lead figure design, with AI assisting exploration.
In briefWhere and Why Grounded Reasoning Helps Visual Reasoning Tasks
SummaryThis paper studies when grounded reasoning, where models output coordinates of relevant image locations during chain-of-thought, is actually useful. The authors propose an analysis method to measure how much models rely on their own grounding and run detailed ablations of a grounded reasoning training recipe. They find gains mainly in counting, with no significant improvement over baselines without coordinates on most spatial reasoning and perception benchmarks. The authors show grounding helps most when tasks need references that are hard to describe in language, and introduce BrickVision, a benchmark requiring such undescribable visual references.
In briefSwarmEvolve: Latent-Space Agent Swarms for Evolutionary Program Search
SummaryThis paper presents SwarmEvolve, a framework in which hundreds of low-cost agents work in swarms, reasoning in latent space and sharing information by passing KV caches instead of decoding text. A novelty filter keeps the most distinct ideas by measuring distances between agents' hidden state vectors, and idea crossover happens through a summarizer agent that receives the combined KV caches of the swarm. On 21 problems in mathematical discovery and scientific computing, the authors report the highest average score, with a median of 3 to 8 minutes to reach scores other frameworks reach in 60 minutes.
In briefHaNDF: Geometric neural distance fields for object-conditioned hand poses
SummaryThis paper proposes HaNDF, an object-conditioned neural distance field that serves as a data-driven prior for plausible hand poses during hand-object interaction. It uses Riemannian geometry to represent plausible hand configurations as the zero-level set of a conditional field and to project a given pose onto that field. The authors report that HaNDF supports pose refinement, reconstruction under occlusion, and physically plausible manipulation synthesis.
In briefScientist World: A Runnable Research Environment Built from Papers for AI Scientists
SummaryThe paper introduces Scientist World, which turns each research paper into an executable environment. An autonomous pipeline extracts each paper's open-ended research question and reproduces the authors' proposed method as a baseline, against which AI scientists are evaluated. The authors release the first 1,000 verified environments and report that frontier AI scientists surpass the human researchers 74% of the time. The environment also contains 6,400 hours of research traces and serves as a training ground, raising a small model's research ideation (+158%) to match frontier models.
WhyTurns each paper into an executable environment for evaluating and training AI-Scientists; the authors report releasing 1,000 environments and frontier models beating humans 74% of the time.
In briefEgoGeneralist: One Video Model for Egocentric Action and Video Tasks
SummaryEgoGeneralist adapts pretrained video generators into a single framework that handles action-conditioned egocentric video generation, pose estimation, and joint action-video generation. Its role-adaptive action renderer and cross-stream action-video coupling make metric 3D motion usable both as a control signal and as a recoverable generation target. The framework covers five action-video tasks and is adapted only on controlled laboratory recordings. It is reported to generalize to diverse unseen scenes and interactions.
In briefXStack: A Benchmark for Migrating ML Repositories Across Frameworks
SummaryThis paper introduces XStack, a benchmark that tests agents on migrating machine learning repositories between PyTorch on NVIDIA GPUs and JAX on Google TPUs. Each agent receives a PyTorch repository and a clean TPU machine and must deliver a working JAX repository, graded against the source repository's published results. The authors report that agent migration remains unreliable, with Claude Opus 5 passing 75% of tasks, GPT-5.6-sol 45%, and all other agents at most 40%.
In briefMeituan-Robotics-0: A Vision-Language-Action Model for Desktop Robot Manipulation
SummaryThe paper presents Meituan-Robotics-0 (MR0), a vision-language-action foundation model and embodied agent system for stationary-base desktop manipulation. It is trained on 147 million vision-language samples and 15,060 hours of robot trajectories across 12 embodiments, using a three-stage recipe and a Rollback method that pairs policy segments with human corrections. MR0 achieves 93.6% success on RoboTwin 2.0 and outperforms π0.5 on average success and progress score across three real-robot tasks.
Meituan Robotics Team: Caicheng Wang, Chen Zhao, Chengpeng Jin et al.
In briefSame Behavior, Different Mechanism: What Finite Experiments Can Certify About Integrated Information
SummaryThis paper examines whether mechanism-level quantities used in interpretability and machine consciousness assessment, specifically the integrated information φ from integrated information theory, can be identified and certified from finite, partial and noisy experiments. Using stochastic networks of four and five units with known ground truth, the authors show that identical input-output behavior can hide different mechanisms and that partial recordings can impose precision floors. They also prove that φ can be certified as a lower bound over a region of mechanisms from finite controlled transition data.
WhyThe authors propose identifiability, continuity, and certifiability criteria, and report exact characterizations and finite-data lower-bound certificates for integrated information in small networks.
In briefInterpreting any residual stream direction by scaling activation inversion
SummaryThe paper introduces Maximally-Activating Example Metamodels (MAEMs), which map a residual stream direction in a language model to text that strongly activates it. The training objective maximises the cosine similarity between the supplied direction and the activations of the generated text, requiring no labels. Experiments report that MAEMs invert held-out activations better than corpus search or natural-language autoencoders (NLAs), and activate 89% of held-out SAE features.
WhyThe authors train an RL-based metamodel mapping residual stream directions to maximally activating text, label-free, reporting gains over corpus search and NLAs across several benchmarks.
Celeste De Schamphelaere, Juan J. Vazquez, Ari Spiesberger et al.
In briefAdversarial training with large step sizes: how gradient descent behaves
SummaryThis paper studies deterministic gradient descent in adversarial training using large constant step sizes. It first analyzes adversarial linear logistic regression, examining where the parameters converge (implicit bias) and how fast the loss decreases (optimization). It then studies how loss sharpness evolves when training neural networks. The results show three differences from standard training: the implicit bias gap persists for arbitrarily large step sizes, optimization error has no uniform convergence rate, and sharpness follows a two-stage dynamic.
In briefCompute efficiency of strong-to-weak on-policy distillation versus GRPO
SummaryThis paper compares on-policy distillation (OPD), which trains a student on its own rollouts under a teacher's per-token supervision, with the reinforcement learning method GRPO by accuracy per training FLOP. With an external teacher, OPD generally underperforms GRPO. When the teacher is the student's own GRPO checkpoint, OPD recovers that checkpoint's math accuracy with about 50 times fewer distillation FLOPs than GRPO needs. The paper also presents multi-teacher OPD, which consolidates per-domain GRPO experts into one student and reaches joint GRPO's two-domain performance with 42.0% fewer total FLOPs.
In briefReproBridge: Making Machine Learning Research Artifacts Reproducible
SummaryThe paper introduces ReproBridge, an agent workflow that builds refined reproducible artifacts, including cleaned code, environment setup, experiment launchers, and a reproduction guide. It also presents ReproBridge-Bench, with 1,029 tasks from 387 experiments across 36 papers. Across four models, refined artifacts raised mean reproduction success from 76.4% to 92.6% and reduced downstream API spending by 19.7% on average.
In briefQ-OIL: Optimistic exploration when robots learn from human corrections
SummaryThis paper studies how robots learn from human interventions in the real world. Analysis of intervention data from human participants shows that people tend to step in when the agent's progress stalls, and their corrections are often suboptimal. Based on this, the authors introduce a progress-based human intervention model that better matches real user algorithm rankings in simulation. They also propose Q-OIL, which adds optimistic bonuses to intervention transitions in reinforcement learning, reporting 10-50% higher success in simulation and on real robots.
Pranav Teegavarapu, Sriyash Poddar, Rico Qi et al.
In briefVFlash: Fast, Accurate Visual Planning via Flash Primitive Decoding
SummaryThis paper presents VFlash, a method for visual planning in physical AI tasks such as robotics and autonomous driving, where plans take the form of points, bounding boxes, and trajectories. A lightweight drafter predicts all coordinates of a visual primitive jointly in one forward pass, instead of generating them one token at a time as autoregressive vision-language models do. On evaluated 2D and 3D robot trajectory tasks, VFlash consistently reduces prediction error compared with autoregressive decoding, and preliminary measurements show about 25x faster generation for 3D point motion forecasting.
In briefA hidden Markov model that predicts an attacker's next move online
SummaryThe paper proposes a hidden Markov model that gives a probability distribution over an attacker's next action before that action occurs. The latent state is the attacker's behavioral regime, which can switch within an episode, and action feasibility changes over time. On the CAGE 2 cyber defense benchmark, the loss begins to fall within three steps after a controller switch, compared with ten steps for a recurrent baseline.
Angel Daruna, Michael Cogswell, Tixiao Shan et al.
In briefEdge-assisted crowdsensing system for explaining how safe cyclists feel on routes
SummaryThis paper presents a system in which handlebar cameras capture street frames, a vision-language model turns them into structured traffic attributes, and an explainable boosting machine combines these with OpenStreetMap infrastructure and bike motion data to predict a crowdsourced 1-9 safety rating. On 1,362 annotated frames from two London routes, the model reaches 72.2 percent binary accuracy and a mean absolute error of 1.011. Per-feature effects are quantified, such as physical lane separation adding about 0.97 rating points.
In briefLabel-wording augmentation in a typed-decision encoder: helps unfamiliar words, not abbreviations or codes
SummaryThe paper compares two releases of Ru-Jev, a 307M-parameter multilingual cross-encoder, where the later release adds augmentation that presents labels in several surface forms. Holding questions fixed and varying only routing label wording, the authors evaluate ten label vocabularies. The augmented release is more accurate on held-out vocabularies of ordinary or borrowed words, but gives no usable gain on abbreviations and opaque codes.
In briefUnderstanding AI from the Inside: A Survey of Mechanistic Interpretability for AI Safety
SummaryThis survey reviews mechanistic interpretability, which studies AI models' internal representations and computations, as a route to AI safety. It introduces core concepts such as features, polysemanticity, transformer computation, and circuits, and covers methods including feature extraction, circuit discovery, and training-data attribution. The paper proposes a unified framework for evaluating mechanistic interpretability claims and examines safety uses such as knowledge editing, steering, and monitoring, along with their limitations.
In briefSM4RT: A Cascaded Model for Reconstructing Dynamic 3D Scenes from Video
SummarySM4RT is a multi-task model that reconstructs dynamic scenes from RGB videos by first estimating 3D geometry and then estimating motion from that structure. Each pixel's 3D motion is expressed in its own camera frame, which reduces ambiguity caused by camera movement. A weakly supervised loss uses labeled point positions at another time step to regularize motion for unlabeled points. The model performs best on most motion estimation metrics across Kubric, Waymo, and Point Odyssey, and is competitive with state-of-the-art methods on geometry reconstruction across five dynamic scene datasets.
In briefMedGraph: A Graph Memory That Keeps Patient State Accurate Over Time
SummaryThis paper presents MedGraph, a graph memory designed for long-term medical dialogue so that LLM-based assistants keep a patient's clinical state correct across visits. The authors identify three recurring error causes in five memory systems: outdated state, other people's health facts, and missing safety facts. MedGraph extracts dated facts, attributes them to individuals, and links records into chains that preserve earlier states. On MedMemoryBench it achieves the highest overall accuracy with both GPT and Qwen backbones.
In briefImproving academic advising with case-based reasoning, a method that solves new problems using past cases.
SummaryThis thesis applies Case-Based Reasoning (CBR) to academic advising, a time-consuming task in higher education. The author designs, develops, and validates a prototype case-based recommender system that supplements the conventional advising process. The system could eventually evolve to replace human academic advisors once its full potential is realised.
In briefMeta Access: Letting Chat Assistants Understand and Operate the Whole Interface
SummaryThe paper proposes meta access, giving an LLM assistant a structured, permissioned view of the chat interface's own state and capabilities. It describes a ladder of levels, from feature tutoring to full orchestration, and illustrates the idea with the Meta Chat prototype in the Chat AI client. The authors also note risks around grounding, privacy, prompt injection, and trust.
In briefFailure Prediction and Classification in Vision-Language-Action Models via Calibrated Latent Probing
SummaryVision-language-action models (VLAs) control robots but can fail unpredictably under distribution shift, which may damage equipment or endanger people. This paper presents VLA-Tug, a lightweight probe on the latent layers of a VLA backbone that detects failures early, bounds missed-detection and false-alarm rates, and predicts failure modes to enable corrective actions. The authors also release a dataset of 3,180 annotated manipulation episodes.
Michal P. Podolinsky, Yunhao Yang, Daniel Little et al.
In briefData-driven discovery of intrinsic Lie point symmetries in differential equations
SummaryThis paper presents a method for finding continuous symmetry generators of differential equation systems directly from sampled data, without symbolic determining equations or a closed-form governing system. It combines prolongation, sequential subspace restrictions, and a nonlinear dictionary with trainable coefficients, and adds a rank-stabilization test to flag systems whose symmetry algebra does not stabilize. On a controlled benchmark with an analytically known 11-dimensional symmetry algebra, the method recovers the complete generator set and reproduces its Lie-bracket structure to numerical precision.
In briefA paper on the limits of molecular representations
SummaryThis paper systematically evaluates the quality of molecular representations such as fingerprints and learned embeddings. It finds that state-of-the-art molecular embedding models do not capture graph structure, 3D shape, or ADMET endpoint similarity better than fingerprints or physicochemical descriptors. It also shows that classical fingerprints have their own limitations, with between 225M and about 4B molecules in the 10.8B-molecule Enamine REAL database having non-unique representations depending on the method, which affects virtual screening.
In briefBenchmarks as Software: Why Coding-Agent Tests Need Ongoing Upkeep
SummaryThe paper argues that benchmarks for coding agents degrade after release as models improve and evaluation infrastructure changes, a problem it calls benchmark maintenance debt. The authors propose a continuous-validation workflow combining community reports, automated oracle checks, and human-approved repairs, and apply it to Terminal-Bench 2.0. They find issues in about 30% of tasks and release the repaired Terminal-Bench 2.1, which raises accuracy by over 10 points and shifts rankings, while overall rank correlation stays high. The paper also shows that time limits and other resource settings substantially change agent performance.
WhyThe authors propose a continuous validation workflow, audit roughly 30% of Terminal-Bench 2.0, and report repairs raising accuracy over 10 points and time budgets materially changing agent performance
E. Kelly Buchanan, Crystal Zhou, Scott W. Linderman et al.
In briefSpectral and subspace dynamics of LoRA in supervised fine-tuning and reinforcement learning
SummaryThis paper examines how the singular spectrum and subspaces of LoRA updates change during supervised fine-tuning (SFT) and reinforcement learning (RL) on mathematical reasoning tasks. It finds that RL concentrates the spectrum at every learning rate, while SFT expands it at moderate learning rates, and RL rotates singular subspaces faster at small learning rates. By swapping ingredients between SFT and RL, the authors attribute these contrasts to group-mean baseline subtraction, which they explain with a quadratic multi-index model. Based on these findings, they tailor LoRA updates separately for each stage and report the best results among compared methods in most evaluated settings.
In briefT3R: A Training-Free Two-Stage Method for Efficient, Robust Robot VLA Models
SummaryVision-Language-Action (VLA) models control robots end to end, but their inference is costly. The paper proposes T3R, a training-free framework with two stages: segmentation-guided pruning removes spatially irrelevant visual patches, and saliency-guided attention biasing shifts attention toward action-critical regions and language instructions. On LIBERO-Spatial, T3R removes 85% of scene-camera visual tokens with essentially unchanged success rate and improves success by up to 7 points under out-of-distribution object-placement perturbations.
In briefAI-Assisted Math Proofs: Detecting Signs of AI in Proof Content
SummaryThis paper asks whether AI involvement in mathematical proofs can be detected separately from AI help with writing. The authors build five cohorts of 8,281 papers, including matched pairs on the same research problems, and introduce ProofTrace, which represents each paper as a graph of immediate proof support and extracts interpretable features. They report a gradient of AI involvement, recurring signatures of AI-generated proofs such as fewer externally cited results, and detection performance that stays stable under AI rewording while text-based detectors decline sharply.
In briefConformal prediction methods for uncertainty in time-dependent physics simulations
SummaryThis paper studies conformal prediction for surrogate models of time-dependent physical systems described by partial differential equations. The authors prove that in function spaces, distributions at arbitrarily close times can be mutually singular, making exact coverage guarantees impossible. In discretized settings, they show that total variation distance often saturates exponentially fast with grid resolution, requiring restrictions on spatial discretization to ensure coverage. For certain settings, they show how weighted conformal prediction yields finite-sample coverage over growing time horizons.
WhyThe authors show conformal coverage can fail for time-dependent PDE surrogates, prove mutual singularity and total variation saturation results, and give weighted conformal finite-sample guarantees.
In briefSocial Humanoid: a humanoid robot system that moves in coordination with human partners
SummaryHumanoid robots are usually trained as solitary agents, so they struggle to coordinate physically with human partners. This paper presents Social Humanoid, a hierarchical system that generates and executes socially responsive whole-body motion on a 29-DoF Unitree G1 robot based on a human partner's motion. The authors also build a new Human-Humanoid Interaction benchmark containing 260 hours of interactions.
Vongani H. Maluleke, James Ni, Kie Horiuchi et al.
In briefStepBack: A method that re-samples only the faulty reasoning step to make AI reasoning cheaper
SummaryThis paper proposes StepBack, a method for making AI chain-of-thought reasoning more efficient at test time. It uses a small set of draft reasoning trajectories and lightweight signals to locate the likely first erroneous step, then resamples only that step and what follows before taking a majority vote on answers. Across seven mathematical reasoning and multi-hop QA benchmarks, it matches self-consistency's average accuracy while cutting token usage by up to 41.26%. It can also be combined with adaptive inference methods such as Adaptive Consistency.
In briefNot All Misalignment Is Alike: A Decomposed Benchmark for Emergent Misalignment
SummaryThis paper examines emergent misalignment, where fine-tuning on a narrow task causes misaligned behaviour on unrelated prompts. It argues that common evaluations are sensitive to question sets, judges, and thresholds, and can confuse misalignment with capability loss. The authors introduce EMERGE, a benchmark that tracks eight distinct misaligned behaviours against matched benign controls while monitoring capability, coherence, and refusal. They find that models with similar overall misalignment rates can show very different behavioural profiles.
Afonso Domingues, Mubashara Akhtar, Bhada Yun et al.
In briefSynthetic Ovarian Cancer Multi-Omics Data: A Biology-Based Evaluation Framework
SummaryThis paper addresses synthetic multi-omic cancer cohorts, focusing on defining evaluation criteria rather than proposing a new generative architecture. The authors introduce a biology-informed, patient-disjoint framework for assessing paired RNA expression and copy-number variation (CNV) profiles, and apply it to the longitudinal DECIDER high-grade serous ovarian cancer cohort. The framework uses agreement between real training and held-out data as a reference, supplemented by training-data proximity diagnostics. Three multi-omic integration architectures based on variational autoencoders are adapted to generate paired profiles.
D\'eborah Boyenval, Alina Malyutina, Jaana Oikkonen et al.
In briefAdaptive low-light enhancement for detecting road hazards, with a benchmark and edge-device tests.
SummaryThis paper examines how autonomous vehicles can detect road anomalies in darkness, rain, and dust haze. It compares four detection models (YOLOv8n, YOLOv12n, RT-DETR, Faster R-CNN) and finds that applying the SCI low-light enhancement network uniformly hurts accuracy at dusk-level lighting, lowering mAP@50 by 16.1% to 30.7%. The authors propose a lightweight, training-free router that chooses among no enhancement, moderate SCI, or denoising plus strong SCI based on per-frame luminance, and validate the pipeline on NVIDIA Jetson Nano hardware.
Obed Honour Eje, Oluwatosin Kolade, Glory Bagai et al.
In briefLearning a patient world model for intensive care system simulation
SummaryThe paper proposes learning a shared patient world model from longitudinal electronic health records with a neural network, providing one clinical representation of patient trajectories for multiple simulation questions. Using ICU operations as the example, the model is trained on MIMIC-IV, frozen, and linked to shared ICU calendars, resources, and queues through scenario-specific mappings. Experiments show it matches Markov and semi-Markov comparators in aggregate occupancy and directional responses to capacity and dispatch policies, and reproduces policy gains from replayed observed trajectories.
In briefSCALE: A real-time method for agents to share internal model data over limited networks
SummaryThis paper introduces SCALE, a real-time latent communication method for cyber-physical agents that exchange intermediate model representations instead of raw sensor data or text. SCALE matches tokens across consecutive latent messages to find reusable information and avoid redundant transmission, and profiles the semantic gain and communication cost of sending different amounts of latent information. A real-time scheduler uses these profiles to decide which pending messages to send and how many latent tokens to include, under bandwidth and deadline limits. Evaluations on three cyber-physical benchmarks and a physical spider robot case study show improved task performance and on-time delivery in bandwidth-constrained networks.
In briefInterpretable hybrid machine learning for predicting rare landing-phase safety events from flight data
SummaryThis paper proposes an interpretable hybrid feature-fusion framework for predicting multiclass safety events during the landing phase, using strongly imbalanced multivariate flight data. An LSTM autoencoder trained on normal approach windows produces a reconstruction-error descriptor, which is fused with statistical flight-history features and fed to a class-weighted XGBoost classifier. The framework is evaluated on 53,750 NASA DASHlink approach windows from three arrival airports, achieving mean macro recall above 0.90 at all airports under repeated partitioning and retraining.
Taha Asghar, Bilal Mufti, Syed Qasim Zaheer et al.
In briefOptimizer tailored to graph neural networks, using graph-aware updates
SummaryThe paper argues that a graph neural network's effective weight geometry depends on the graph-propagated features it processes, so vanilla Muon can amplify updates. It introduces GraphMuon, which orthogonalizes updates in a metric defined by propagated layer-input covariance and scales step sizes by their effect on layer representations. On three large-scale OGB benchmarks it reaches target validation loss 1.32× to 2.42× faster than AdamW.
WhyThe authors extend Muon's orthogonalized updates to a graph-propagated covariance metric, analyze graph-induced update amplification, and report speedups over AdamW on three OGB benchmarks.
In briefMuir: A Linear Attention Model with Sensitivity-Based Gating and Matched-Filter Readout
SummaryThe paper introduces Muir, a linear attention model that merges its gates into a sensitivity-allocated low-rank projection and re-weights state rows with a matched-filter readout gate. It reinvests saved parameters in a wider feed-forward network and adds a parallel update designed for stable, efficient training at large chunk sizes. Across models from 166M to 1.34B parameters, Muir achieves lower perplexity than GDN-2 and Transformer and performs better on recall and long-context retrieval tasks.
Tushaar Gangavarapu, Gil Tabak, Clemens JS Schaefer
In briefPersonalized health coaching using a coachable health age and counterfactual wearable predictions
SummaryThis paper introduces Coachable Health Age (CHA), an interpretable health-age metric built from modifiable signals passively captured by consumer wearables. It also presents SensorWM, a wearable world model that forecasts multi-week physiological responses to candidate lifestyle changes, and integrates both into an agentic system that generates 12-week personalized coaching plans. In blinded paired reviews, certified lifestyle coaches preferred counterfactual-grounded plans over those based on wearable history alone.
Simon A. Lee, Zeinab Esmaeilpour, Menglian Zhou et al.
In briefMOMASPLAT: Novel-view demonstration augmentation for mobile robot manipulation
SummaryThis paper presents MOMASPLAT, a framework that augments robot demonstration data for mobile manipulation by synthesizing full manipulation trajectories from unseen robot base poses. It combines 3D Gaussian Splatting, feed-forward 3D foundation models, and kinematics-based action retargeting to reconstruct the scene and manipulated object separately and render them from new viewpoints, including dynamic object interactions. Evaluated on five real-world tasks involving rigid and articulated objects, the method more than doubles the average success rate of the strongest baseline and reduces the need for large-scale real-world data collection.
In briefNPCBench: A Clinical Apprenticeship Benchmark for Multimodal Care-Pathway Reasoning in Nasopharyngeal Carcinoma
SummaryThis paper introduces NPCBench, a benchmark for nasopharyngeal carcinoma that follows specialist training, covering guideline knowledge, case planning, longitudinal review, care-pathway integration, and open-ended consultation. It includes 4,777 component and consultation items plus a 100-patient CarePath cohort with 417 scored diagnostic and care tasks. Evaluation checks whether a model answers diagnostic and care questions correctly together for the same patient; among 21 multimodal models, strict CarePath success reached at most 23%.
In briefVisionSIM: A Multi-Modal Robot Perception Simulator with Realistic Sensor Models
SummaryThis paper presents VisionSIM, a unified simulation platform for robot perception that models the raw physical measurement process of conventional and emerging sensors. It reproduces failure modes that general-purpose simulators miss, such as LiDAR pileup and glare, iToF phase wrapping, and event camera refractory drops. The platform enables cross-modality comparison on identical scenes and trajectories. The authors also release the platform and a large indoor dataset with synchronized raw measurements and ground-truth annotations.
Sacha Jungerman, Kaustubh Sadekar, Avery Gump et al.
In briefSHORE: A stability-aware reinforcement learning method for online box packing
SummaryThis paper addresses online 3D bin packing, where a robot must choose which box to pack, its orientation, and placement while keeping the stack from collapsing. The proposed SHORE method combines a learned stability probability for each placement with an on-policy return estimate conditioned on stability, using a 3D U-Net trained on executed and simulated outcomes. In Isaac Gym, SHORE reaches 73% space utilization in about 0.4M steps and achieves 0.734 utilization on a 60-box set versus 0.588 for the GOPT baseline, with more placement failures. The authors also demonstrate the method on a real robot.
In briefStrataFlux: A Low-Bit LLM Inference Framework Built Around the Memory Hierarchy
SummaryThe paper presents StrataFlux, a framework for low-bit large language model inference that recovers accuracy lost to weight and key-value cache quantization while keeping the memory and speed benefits of quantization. It uses host RAM and solid-state drives as sources of additional precision, with selective precision streaming that reduces transfer volume, overlaps delivery with computation, and reuses fetched precision across decoding steps. The authors report that 2-bit Llama-3.1-8B-Instruct weights reduce WikiText-2 perplexity from 9.46 to 8.40, and that 2-bit weights and KV cache raise Qwen3-8B's AIME 2024/2025 accuracy from 11.0% to 47.1% at unchanged GPU memory use.
In briefSpecCTMC: Exact Speculative Sampling for Diffusion Language Model Continuous-Time Markov Chains
SummaryThis paper introduces SpecCTMC, a training-free speculative sampler for diffusion language models that needs no separate draft model. It proposes successive token updates from previously computed predictions, evaluates intermediate states in parallel, and applies an exact acceptance-and-correction procedure. The authors prove the method preserves the joint distribution of the underlying continuous-time Markov chain. On LLaDA-8B-Instruct and Dream-v0-Instruct-7B across five benchmarks, it requires 1.73 to 4.38 times fewer sequential model evaluations than direct continuous-time sampling.
In briefRadio-based information for correcting shared position drift in cooperative robot localization
SummaryCooperative localization for robot teams uses odometry and inter-robot measurements, which describe only relative geometry, so the team's pose can drift together over time. This paper proposes Radio-to-Pose Information (RPI), an interface that converts power residuals from a radio attenuation map, registered to the robot world frame, and their pose derivatives into a Gaussian information matrix and vector. Across 32 paired trials with a registered map in a 50×50 m arena, RPI reduces common-position RMSE from 0.542 m to 0.140 m compared with a baseline that uses radio only to weight relative measurements.
In briefOraRefer: patient-level oral cancer screening from multiple smartphone photos with a guaranteed sensitivity threshold
SummaryThis paper presents OraRefer, a model that screens for oral cancer and oral potentially malignant disorders using smartphone photographs of the oral mucosa. A segmentation-guided foundation model classifies each photo by its predicted lesion, then fuses all of a patient's photos with risk factors to make a referral decision via a threshold with a finite-sample sensitivity guarantee. On 3000 photographs of 714 patients, patient-level AUROC was 0.919, outperforming attention-based multiple-instance learning. At a 90% sensitivity target, the model detected 95% of patients with OPMD or cancer, at a specificity of 0.610.
Rupam Mukherjee, Mehakdeep Singh, Shirin Dasgupta et al.
In briefTheoretical analysis of a bilevel optimization method for tuning combined local-global optical flow models
SummaryThis paper studies a bilevel optimization approach for learning the regularization parameter of combined local-global (CLG) optical flow models. The lower level computes flow from a Lucas-Kanade data term and a quadratic or total variation (TV) regularizer, while the upper level chooses the parameter minimizing squared distance to ground-truth flow. For the quadratic model, the authors give conditions for existence, positivity, and strict local optimality of optimal parameters. For TV regularization, an optimal parameter always exists, and a sufficient positivity condition is provided that accounts for coupling between flow components via local structure tensors. Synthetic examples illustrate the conditions, including cases where they fail.
In briefUnifying central flow, rod flow, and edge coupling: a common basis for continuous-time models of gradient descent at the edge of stability
SummaryAt the edge of stability, gradient descent iterates oscillate along sharp Hessian directions, which standard gradient flow cannot capture. The paper introduces the edge coupling, a functional on consecutive iterates whose critical points are gradient descent's fixed points and period-two orbits, and derives central flow and rod flow from it. The two flows coarse-grain over the fastest and intermediate time scales respectively, and the framework extends to momentum methods such as heavy ball and Adam. Experiments on neural networks show the flows track discrete iterates across all training phases.
In briefUSB Gym: A Gymnasium Testbed for Closed-Loop Ultrasound View Matching
SummaryThis paper presents USB Gym, a testbed that simulates ultrasound from public patient CT scans to evaluate methods for moving a probe until its image matches a target view. The authors tested 20 published methods in the segmenter, controller and reward parts of a guidance system. The combined guide only moved the probe from an average of 16.0 mm to 13.7 mm, far from the pose oracle's 1.5 mm. The authors conclude that the task remains under-explored.
Thanapong Intharah, Zhichao Gao, Yiliang Liu et al.
In briefCascadeRec: Deployed Adaptive LLM Compute Allocation for Multi-Tenant Content Recommendation
SummaryThis paper presents CascadeRec, a two-tier LLM reranking cascade running in production at a large financial planning SaaS provider. A lightweight first-tier model reranks every request, and a router sends only likely-beneficial requests to a more expensive second tier. Over a month of production A/B testing, the system reduced inference cost by about 10% with minimal latency impact and no observed quality regression. On public benchmarks, uniform LLM reranking lowered NDCG@10 by up to 30% for sparse-history users.
In brief4D Point Splatting for Differentiable Rendering of mmWave Radar Micro-Doppler
SummaryThis paper introduces 4D Point Splatting (4DPS), a physics-based differentiable renderer that fits an explicit moving scene to real millimeter-wave radar captures. The scene is represented as oriented points with material, normal, and velocity, which are splatted into range and Doppler bins to produce raw complex radar returns. Reconstructions on the HuPR and RT-Pose datasets reach state-of-the-art correlation scores, and a pose model trained on the renders performs better on real data than the RF-Genesis baseline.
WhyThe authors propose a differentiable point-splatting renderer fitting moving scenes to real mmWave radar captures, reporting reconstruction and sim2real results on HuPR and RT-Pose.
In briefPreserving Global Distances in Compressed AI Agent Memory
SummaryThis paper studies how to compress the embedding vectors that AI agents use to store text without losing the relationships between distant items. Existing compression keeps nearest neighbors but can move many passages to different themes, which hurts tasks like memory consolidation. The authors propose locality-aware quantization (LAQ), which turns a proven bound on far-item ranking errors into a training objective, and report improved global fidelity across many tests, including scaling to a million passages.
WhyThe authors derive a provable far-end bound, turn it into the LAQ quantizer objective, and report gains across 52 held-out runs and million-passage scale.
In briefLearning rules that shape other agents' learning
SummaryIn general-sum games, self-interested learners can reach outcomes that are poor for everyone. This paper introduces DUOS (Discovered Update Rule for Opponent Shaping), which uses meta-learning to discover an update rule that shapes an opponent's learning in general-sum games. The rule reads only the agent's own experience, without access to opponent parameters or policies, and is meta-trained to maximise lifetime expected return. The authors report that it ranks first in normalised mean score and in self-play among tournament entrants.
In briefCultural Adaptation in Repeated Social Interactions: Testing whether LLMs adapt to social norms through repeated interaction
SummaryThe study places large language models in a sandbox environment with repeated social interactions to test whether they can learn cultural norms through interaction. It reframes cultural adaptation as a continuous process of absorbing norms through social exchange, rather than answering questions about culture. The findings show that memory mechanisms support longitudinal learning across repeated interactions and transfer to different social settings, and enable norm assimilation across five languages.
