In briefLM-DSB: A lightweight multi-agent system that automatically writes daily construction site safety bulletins
SummaryThis paper presents LM-DSB, a lightweight multi-agent system that automatically generates daily safety bulletins for construction sites. It combines retrieval-augmented generation with a locally deployable large language model that runs on consumer-grade computers, grounded in 37 health, safety, and environment documents. Expert evaluation (N=40) rated the outputs as practically acceptable, with mean scores of 3.50 to 4.20 on a 5-point scale.
In briefPredicting rock mineral content from chemical data using simulated labels
SummaryThis paper trains a machine learning model on simulated petrological labels to predict mineral types and abundance distributions from whole-rock chemical compositions. The model combines self-supervised pretraining on about 788,000 compositions with prediction heads trained on nearly one million simulated labels, and a recalibration layer aligned to measured mineralogy. On 640 volcanic rocks, it zero-shot identifies the presence of most mineral categories and ranks rocks by mineral richness, but its predicted abundances are systematically offset. Recalibrating with just 25 labeled rocks recovers most of the likelihood improvement.
Matthew J. Liu, Aaron D. Kaplan, Michael L. Whittaker et al.
In briefFORESEE-MS: Forecasting Future MS Lesion Locations from Baseline FLAIR MRI
SummaryThis study examines whether baseline FLAIR MRI can identify areas at risk of developing new multiple sclerosis lesions, and introduces FORESEE-MS, an evaluation protocol that uses a fold-specific population-prior atlas as the baseline. Five deep learning models were compared with this atlas using 5-fold cross-validation on 39 subjects from the MSSEG-2 dataset. The atlas achieved the highest pooled voxel-level performance, while some deep learning models scored higher on within-patient lesion localization, though most confidence intervals included zero. The authors conclude that it remains uncertain whether baseline FLAIR improves forecasting beyond the population-prior atlas in this small cohort.
Maryam Akhavan Aghdam, Sudip Mittal, Noorbakhsh Amiri Golilarz et al.
In briefProperty-based tests for a more robust softmax and cross-entropy
SummaryA 2026 paper by Barron et al. identified two numerical issues in the softmax and cross-entropy functions used by most standard machine learning libraries. This work reproduces those issues with simple property-based tests written in Hypothesis, and finds an additional edge case where the existing implementation fails. The authors then propose a more robust softmax implementation that passes these tests and supports causal attention masking, and compare it with the original. The code and tests are publicly available on GitHub.
In briefRepulsive Population Policy Gradients: a reinforcement learning method for diverse, high-quality behaviors
SummaryThe paper proposes RePPO, a method that adapts PPO to jointly optimize a compact population of policies using a population-dependent, dynamically shaped reward. On standard continuous-control tasks, it achieves a strong quality-diversity tradeoff compared with state-of-the-art QD-RL methods. It also maintains PPO-level quality on pixel-based tasks while substantially increasing diversity, and it is applied to fine-tuning large language models.
Saeed Hedayatian, Harsh Shah, Celine Mneimneh et al.
In briefSO-101 Bench: A Simulation Benchmark for Spatial Robot Arm Manipulation
SummaryThis paper presents SO-101 Bench, a simulated benchmark for language-instructed spatial manipulation on the low-cost SO-101 robot arm, built on an Isaac Lab digital twin of a real setup. It includes four tasks, 138 objects, 916 evaluation episodes, and 3,629 simulated demonstrations. Policy rankings in simulation match those on the physical robot, though simulation overestimates success rates. The best learned policies reach about 61%, while an LLM agent that writes code and delegates grasping to a learned policy reaches 83.2%.
WhyProposes a simulated SO-101 benchmark whose policy rankings match the real robot exactly, with substantial evaluation; useful for a niche subfield.
In briefSpeeDiff: A Scalable End-to-End Latent Diffusion Model That Trains Its Autoencoder Jointly
SummaryThe paper presents SpeeDiff, a latent diffusion method that trains the VAE and diffusion model together from scratch rather than using a pretrained frozen VAE. To prevent latent space collapse caused by unrestricted gradient backpropagation, the authors introduce a Tweedie Pixel Reconstruction loss, which decodes a predicted clean latent from a noisy state to provide pixel-level feedback. The authors report that SpeeDiff-XL trains over 140x faster than Vanilla SiT and 61x faster than REPA, reaching an unguided FID of 1.50 on ImageNet 256x256. With a 32x compressed VAE, it reaches an unguided FID of 1.53 on ImageNet 512x512.
In briefVerifiable Limits of LLM Unaided Planning in Sliding Puzzles
SummaryThis paper examines how well large language models plan innately, generating complete solutions to 4×4 sliding puzzles in one response without tools or execution feedback. It proposes a verifiable framework that splits success-weighted plan efficiency into execution reliability, coordination among legal plans, and action economy among completed plans, checked by trajectory replay and certified optima. Evaluating 8,400 full-plan outputs from seven model conditions, the authors find that models completing every trial become less economical at longer horizons, and the framework distinguishes coordination failure from inefficient completion.
In briefPersistence, Exposure, and Credit Routing in Linear Temporal Predictive Coding
SummaryThis paper analyzes temporal credit assignment in linear temporal predictive coding (tPC) networks using tools from cellular sheaf theory. It describes credit assignment as a tug-of-war between spatial displacement and temporal errors, modulated by persistence and exposure operators P and Q. The authors identify a persistence–error tradeoff in spatial modes and show that recurrence can mitigate it by routing credit between persistent and exposed modes.
In briefExpert rock-type labels do not improve learning for predicting marine sediment properties
SummaryThe researchers trained a self-supervised embedding model on 236,963 observations from 296 boreholes in the LILY ocean drilling database, using six routinely logged measurements such as bulk density, magnetic susceptibility, and colour reflectance. The embeddings separated geologist-assigned lithologies, but training with those labels added nothing and higher label weighting degraded reconstruction. The embeddings also predicted physical properties absent from training and performed better than depth alone on unseen boreholes.
In briefkw-bench: A benchmark testing AI coding harnesses on real investigative accounting tasks
SummaryThis paper introduces kw-bench, a benchmark built from real audit, tax and similar engagements described by specialists. Its first task reconstructs five years of VAT records from raw extracts across three generations of SAP systems, graded deterministically without an LLM judge. On the same model, two independent harnesses scored about 0.53 to 0.54, while two lab harnesses scored 0. On a frontier model, two harnesses finished in roughly 22 to 25 minutes with near-perfect scores.
WhyIntroduces an investigative benchmark from real audits with deterministic grading, comparing four harnesses and revealing large performance gaps relevant to agent evaluation.
In briefWhen Can a Learned Design Surrogate Be Trusted in a Self-Driving Microfluidics Laboratory? An Offline Closed-Loop Framework
SummaryThe paper proposes a measurable framework for deciding when a learned design surrogate in a self-driving microfluidics lab can be trusted, covering competence ratios, target-measurement budgets, measurement ordering, and when to stop adapting. The authors test it offline on six public droplet-microfluidic and liposome tables without wet experiments. They report that warm-start residual boosting restores most chemistry groups to near their reference performance within 20 target rows, but no label-free statistic reliably certifies individual candidates.
In briefSpecLST: Spectral-enhanced long-short term spatio-temporal multimodal traffic forecasting
SummaryThis paper proposes SpecLST, a framework for forecasting traffic in urban road networks. It uses a Fourier-transform-based frequency representation to efficiently encode multi-week historical patterns, fuses weather and road-incident context via attention, and adds a graph-augmented sparse Mixture-of-Experts backbone with a frequency-enhanced residual head for abrupt anomalies. At the 60-minute horizon, the authors report mean relative error reductions of 14.10% in MAE, 10.35% in RMSE, and 15.78% in MAPE across eight neural baselines and four datasets.
In briefMAREval: A Multi-Agent Framework for Judging Recommendation Explanations
SummaryThis paper introduces MAREval, a framework that uses multiple large language model agents to evaluate natural language explanations produced by recommender systems. The agents include a planner using a Chain of Debate prompting strategy, a moderator, an arbitrator, and a Monte Carlo sampling method to improve robustness. The authors report experiments on the public TopicalChat dataset and a proprietary e-commerce dataset, showing better alignment with human judgments and lower variability. They also report statistically significant gains in engagement and revenue in online A/B testing.
Reza Yousefi Maragheh, Jayesh Uddhav Kudase, Aysenur Inan et al.
In briefLoRA under a budget: a method for compressing groups of adapters while preserving function
SummaryThis paper studies how to compress many LoRA adapters under a shared resource budget. The authors build a functional rank-distortion frontier for each adapter and allocate the shared rank budget across the fleet using minimax allocation. The proposed Functional Rank Allocation (FRA) method is unlabeled and measures module truncations end to end, combined with a spectral proposal. Across 78 adapters from three public fleets, FRA improves worst-case and tail utility over uniform spectral compression in constrained-budget settings.
In briefSOMA: A benchmark testing AI on fine-grained body motion from first-person video
SummaryThis paper introduces SOMA, a benchmark of 1,632 human-verified questions for understanding body motion from egocentric video, covering body configuration, motion dynamics, and temporal reasoning. Thirty-seven multimodal models were evaluated, with humans reaching 68.5% accuracy and the best model, Qwen3.6-27B, reaching 50.0% zero-shot. The authors also release a training dataset and a multimodal baseline aimed at closing this gap.
WhyA fine-grained egocentric body-motion benchmark evaluated on 37 models against humans, with clear comparisons and error analysis, useful for the subfield.
Matea Tashkovska, Na\"el Ouerghemi, Haozhe Qi et al.
In briefFrom Generation to Scoring: Scaling LLM Semantic Labeling at LinkedIn
SummaryThis paper addresses industrial LLM labelers that pack many candidate labels into one prompt and generate YES/NO answers together, which couples independent decisions. The authors reformulate the labeler as semantic scoring, fine-tuning it on single-candidate prompts to produce each decision as one answer-token score in a prefill-only pass. They also adapt Multiple-Item Scoring to score hundreds of candidates in one sequence with attention masked between them. Deployed for profile qualification labeling at LinkedIn, the approach changes at most 0.01% of decisions under candidate reordering and improves throughput by 1.24x offline and 1.9x in model serving.
Jian Sheng, Sriram Vasudevan, Sundara Raman Ramachandran et al.
In briefHIDRA: A Hierarchical Routing Method for Replay-Free Lifelong Imitation Learning
SummaryRobots learning from a continuous stream of demonstrations often cannot store past data, so their expert selection can degrade over time. HIDRA uses a hierarchical dual-attention mechanism that separates instruction-level expert retrieval from context-dependent refinement, plus key-level geometric regularization to keep routing stable under representation drift. The authors evaluate it on several LIBERO benchmark suites and report better AUC and forward transfer than replay-free and replay-based baselines, with low forgetting.
In briefBoxWorld: World Engines Control and Video Generators Render Interactive Video
SummaryBoxWorld builds real-time interactive video worlds with user-defined rules by separating interaction reasoning from visual rendering. A coding agent creates a Box World Engine that determines outcomes and outputs object states as layout boxes, attributes, and status variables. A Box-State Generative Renderer then autoregressively renders this state stream into video. The paper also introduces a three-stage training strategy and the PairedStateBench benchmark, and reports that BoxWorld substantially outperforms end-to-end world models.
In briefIOSage: Detecting HPC I/O Bottleneck Patterns to Guide LLM Recommendations
SummaryThis paper presents IOSage, a system that turns HPC application Darshan logs into I/O fix recommendations. A classifier trained on many labeled logs detects I/O bottleneck patterns, and an LLM then writes recommended changes using those patterns, configuration files, per-file evidence, and benchmark-derived cards. In manually applied, repeated paired runs, the suggested changes sped up an openPMD kernel 4.38x and hourly WRF restarts 1.42x.
WhyA trained classifier, not the LLM, detects I/O bottlenecks, a novel design, backed by repeated paired runs and multiple benchmark evaluations.
Mahdi Banisharif, Tabassum Mahmud, Ahmad Tarraf et al.
In briefCombining kidney cancer slide and CT scan images with cross-attention to predict patient prognosis
SummaryThe paper links histopathology slides and CT scans of renal cancer through bidirectional cross-attention to predict patient outcomes. In 210 TCGA-KIRC patients, 5-fold evaluation gave a C-index of 0.760 and a 3-year AUROC of 0.773, the highest averages among the unimodal and fusion baselines tested. External testing in 180 CPTAC-CCRCC patients, with CT withheld, gave a disease-specific survival C-index of 0.629.
In briefComparing Models for Short-Term Blood Sugar Risk Prediction with a Focus on Reliability
SummaryThis paper presents a benchmarking framework for predicting short-term low and high blood glucose risk in people with type 1 diabetes at 30, 60, and 120-minute horizons. Using the T1DiabetesGranada dataset, it compares three neural network models built from over 5.5 million prediction anchors. No statistically significant differences in predictive performance were found, but the models differed in calibration, explanation reliability, and robustness. The authors stress that reliability assessment and external validation are needed before clinical deployment.
Pir Bakhsh Khokhar, Carmine Gravino, Fabio Palomba
In briefHELICS: Generating Synthetic Genomes at Biobank Scale with Latent Flow Matching
SummaryThis paper presents HELICS, a framework for generating synthetic genome-wide SNP-level genotype data from biobank-scale datasets without requiring haplotype phasing. It compresses over 600K SNPs per genome into 4,096 latent tokens using convolutional autoencoders, then models them with a transformer-based conditional flow matching model. Trained on over 417K UK Biobank genomes, it better preserves real allele frequencies and SNP correlation structure than existing haplotype-based simulators, and can be conditioned on ancestry and genetic liability.
WhyApplies latent flow matching to phasing-free genome-wide SNP generation, validated on 417K UK Biobank genomes with correlation and allele-frequency fidelity.
In briefAK_QMUL's ArGuard entry: one fine-tuned JAIS model for detecting harmful Arabic prompts
SummaryThis paper describes AK_QMUL's submissions to the ArGuard-2026 shared tasks on Arabic harmful-prompt detection, a binary safe/unsafe task and a seven-way harm-domain task. Both submissions use a single JAIS-6.7B backbone fine-tuned with 4-bit QLoRA and a linear classification head. The team ranked first on both tracks, with macro-F1 scores of 0.984 on the binary task and 0.790 on the seven-way task.
In briefWAKMan: A Workload-Aware KV Cache Manager for LLM Serving
SummaryThis paper presents WAKMan, a KV cache manager that selects its eviction policy based on live traffic for prefix caching in LLM servers. The authors find that agentic requests return within seconds and favor recency, while popular prompts and long conversations favor frequency, so no single policy is best across workload mixes. WAKMan chooses policies without knowing request types or reading prompts. It improves hit rate over LRU by 5 percentage points on average.
Suhas Kumar Morisetty, Vaibhavi Lokegaonkar, Sachin Mathew Varghese et al.
In briefDetecting input ambiguity in LLMs before resolving it
SummaryLarge language models often answer ambiguous inputs as if they had one intended meaning. This paper proposes a label-preserving perturbation framework that detects input ambiguity using perturbations that stay close to the original input. It implements this with a black-box rule-based perturbation library and a white-box importance-guided method. Experiments on four open-domain, conversational, and visual QA datasets show it outperforms output-side uncertainty and LLM-paraphrase perturbation baselines.
In briefBayesian Replicator Dynamics and Strict Dominance Reversal in Bimatrix Games
SummaryThis paper studies replicator dynamics from evolutionary game theory, interpreting them as continuous-time Bayesian inference in which an agent accumulates evidence about an opponent's strategies. It compares the standard expected-utility payoff with the logarithmic payoff of Ay and Erb. Under expected utility, the dynamics match Bayesian updating only in games without a dominant strategy, while logarithmic payoffs allow this representation in every game. In games with a dominant strategy, the flows reverse between the two payoff definitions, so cooperation emerges in the Prisoner's Dilemma.
In briefDeep Representation Learning for Out-of-Distribution Detection in DNA Methylation Tumor Classifiers
SummaryThis paper addresses the problem that DNA methylation array-based tumor classifiers can make faulty predictions for tumor types absent from their training data. The authors propose an end-to-end out-of-distribution detection framework, training neural networks on public sarcoma and brain tumor cohorts and comparing five detection methods. The 1D-SSA approach performed best, reaching a mean AUROC of 0.976 on an independent clinical test set, suggesting it as a quality-control layer for clinical methylation-based tumor classification.
Md Zahangir Alom, Quynh T. Tran, Breuer Alexandar et al.
In briefBalanced Active Subspaces for Optimal Control via Trajectory Optimization
SummaryThis paper introduces BASOC, a model reduction framework for optimal control of high-dimensional nonlinear systems that addresses the curse of dimensionality in solving the Hamilton-Jacobi-Bellman equation. It uses trajectory optimization and the Pontryagin minimum principle to collect state and co-state data along locally optimal trajectories, then balances the resulting covariance matrices to obtain an oblique projection for approximating the value function. The method is demonstrated on a nonlinear three-dimensional system and the complex Ginzburg-Landau equation.
In briefRerunning Two Early Programs That Modeled How People Write Code
SummaryThis paper revives two historical programs by running their original code: Herbert Simon's 1961 Heuristic Compiler, which treated programming as problem-solving using means-ends analysis, and the 1984 KBEmacs prototype from MIT's Programmer's Apprentice project, which built programs by combining familiar patterns called cliches. The authors reproduce the compiled code from Simon's 1963 paper and the SQRT sequence from KBEmacs's December 1984 self-test. Both systems adapt existing routines, and the paper argues that their explicit theories of programming are largely absent from today's AI coding assistants.
In briefCrowdCC: A benchmark jointly testing crowd counting models for image corruption robustness and quantization
SummaryThis paper introduces CrowdCC, a benchmark that tests crowd-counting models against 23 deployment-motivated image corruptions at three severity levels, while also comparing FP32, FP16, and INT8 precision. It covers four crowd-counting datasets and 26 models, with throughput and energy measured on Jetson Orin NX hardware. The findings show that capture-chain corruptions cause the largest degradation, INT8 changes robustness more than FP16 without a consistent direction, and precision alters model rankings. The authors conclude that model selection should weigh corruption robustness together with measured deployment cost.
WhyJointly evaluates corruption robustness and quantization across 23 corruptions, 26 models, and edge measurements, with substantial empirical support.
In briefCRLoMTL: Resolving Task Conflicts in Multi-Task Learning via Low-Rank Matrices
SummaryThis paper proposes CRLoMTL, a parameter-efficient framework that reduces interference between tasks in multi-task vision learning. Each task learns its own LoRA low-rank update, and these are merged into one deployment-time update using an SVD-based shared-subspace resolver. Experiments on PASCAL-Context, NYUD-v2, and CelebA with Swin Transformer backbones show gains over single-task baselines while training far fewer parameters.
In briefOA-RAG: Ontology-based agentic retrieval for more faithful LLM reasoning
SummaryThis paper proposes OA-RAG, an agent that combines formal logical reasoning with LLM reasoning, using the logical structure and informal annotations of domain ontologies to answer questions. For each answer, it reports the ontology observations used as evidence along with a grounding level (entailed, supported, related, or unrelated) indicating evidence strength. On benchmarks built from three large-scale ontologies and six real-world medical question-answering tasks, OA-RAG achieves the best overall performance among the evaluated methods, with particularly strong results on reasoning-intensive tasks.
In brief"Plenty of Room in the Middle": Finding GPU Cluster Optimization Gaps Through Fine-Grained Monitoring
SummaryThis paper studies a university research GPU cluster by monitoring resource usage at 100 ms intervals using a lightweight tool built on Nvidia's DCGM library. Over a 10-day period covering 7,399 jobs on 319 GPUs, the authors find heavy underutilization, frequent mismatches between requested and actual job durations, and highly periodic communication patterns in multi-GPU jobs. They argue that substantial utilization gains are possible through better scheduling and GPU multitasking, and release their monitoring tool as open-source software.
In briefStability of LLM-generated proofs under behavior-preserving program refactoring
SummaryThe paper introduces RefactorProof, which creates 458 Lean-certified variants of VERINA tasks through behavior-preserving refactorings while keeping the specification and proof fixed. Structural refactorings reveal substantial brittleness, with specialized prover proofs surviving unchanged more often than the supplied reference proofs on matched tasks. The authors note that proof failure does not necessarily imply high repair cost, since simple one-identifier edits repair most failures.
In briefTraining-free speedups for video language models: methods, trade-offs, and evaluation
SummaryThis survey reviews training-free methods for speeding up inference in video large language models while keeping task-relevant spatiotemporal information. It groups existing approaches by the pipeline stage they intervene in, and treats post-training quantization separately. The paper also analyzes why local efficiency gains may not produce end-to-end speedups and evaluates accuracy-latency trade-offs experimentally.
In briefOntology-guided multi-agent reasoning for explainable differential diagnosis
SummaryThis paper proposes an ontology-guided multi-agent framework for differential diagnosis that mirrors the iterative reasoning of physicians. It grounds diagnostic inference in a structured, expert-curated medical knowledge base and breaks diagnosis into explicit, auditable steps. The system also identifies the most discriminative missing symptoms and generates targeted follow-up questions to reduce uncertainty.
In briefAHEAD: A neural accelerator that speeds up solving very large linear programs
SummaryLarge linear programs with millions of variables are slow to solve with first-order methods, which need many iterations. AHEAD is a neural accelerator that predicts step sizes for such solvers while keeping their update rules, and a learned policy sets how many warm-start iterations to run before the solver finishes on its own. On packing linear programs with up to 10^7 variables, it reports 6.5 to 11.6 times end-to-end speedups over Gurobi's first-order solver from a cold start. The paper states the gains extend to PageRank linear programs.
In briefConfidence Is Not a Promise: Auditing What Typed Decision Models Can Guarantee
SummaryThis paper audits what the probability distributions output by typed decision models actually license, comparing the open-weight Laya model across five checkpoints with the API-only Jev 1.13. It argues that probabilities support claims about groups of decisions rather than single ones, and tests threshold-selection methods across 190 exchangeable evaluations to see which keep accuracy promises. The authors also examine labeling needs, retraining under drift, and explanation faithfulness, and release frozen answers, a datasheet, a promise check, and code.
In briefLinking Exploration to Loss of Plasticity in On-Policy RL
SummaryThis paper studies Loss of Plasticity, where deep reinforcement learning agents lose the ability to adapt to new data. It analyzes Proximal Policy Optimization with isotropic Gaussian policies and examines how the policy standard deviation σ affects the extent of this loss. Experiments include benchmark evaluations against standard baselines and real-world deployment on a 7-DoF robotic arm.
Andrea Vasco Grieco, Evelyn D'Elia, Edoardo Caldarelli et al.
In briefRetrieval Alone Is Not Enough: Outdated Facts in Agent Memory
SummaryThe paper studies why memory agents fail on a single-hop fact-update task, separating cases where an update is missed from cases where a retrieved update is not used. Accuracy drops as the retrieved evidence budget grows, and most errors occur on questions that also contain an obsolete statement. Under equal evidence size, replacing an unrelated statement with an obsolete competing one causes larger accuracy losses for both Qwen models tested.
Oleeviya Babu Poikarayil, Tony Basil Babu, Abdulrahman Nahhas et al.
In briefSpiking Neural Networks Resist Model Stealing: A Benchmark Study
SummaryThis paper benchmarks four model stealing attacks against spiking neural networks (SNNs) and conventional artificial neural networks (ANNs) on MNIST, Fashion-MNIST, and CIFAR-10. It finds that surrogate models of SNNs reach 10% to 20% lower agreement with the target and 15% to 25% lower adversarial transferability. The authors attribute this to the time-dependent, spike-based dynamics of SNNs, which yield decision boundaries that ANN-based surrogates struggle to approximate.
In briefFASST: A Compact Stateless Atmosphere Model for Long-Term Ocean Coupling
SummaryThis paper presents FASST, a coupled climate model that keeps the POP2 ocean and replaces the atmosphere and sea ice with lightweight U-Net emulators. The atmosphere emulator diagnoses surface fluxes from the ocean-ice state every 6 simulated hours, and the sea-ice emulator advances the ice monthly. In 140-year unforced runs without SST restoring, the model keeps a stable global-mean climate and captures Niño-3.4 variability, though with weaker amplitude than CESM.
Pratham Aggarwal, Duncan Watson-Parris, William E. Chapman
In briefTesting whether independently built AI agents will work together safely, even those built later
SummaryThe paper studies how independently developed autonomous agents, such as self-driving cars, can be guaranteed to be jointly compatible when sharing an environment. A regulator publishes seed agents, and each developer tests its own agent for compatibility with every seed-compatible agent, a property the authors call future-proofness. The authors formalize compatibility for general multi-agent systems, characterize future-proof agents as the closure of the seed, and show this closure is monotone in the seed. They also examine how far executable tests that interact only with the seed can distinguish these future-proof agents.
In briefA dual-objective model that decides which rental listings get instant booking
SummaryThis paper proposes a smart instant book filter system for hospitality marketplaces, deciding which listings in search results receive instant booking. The system aims to increase the instant book rate while keeping the overall booking rate from decreasing, a dual-objective optimization problem. A/B tests with global users show it outperforms previous human-expert-tuned models on both instant book rate and overall booking rate. The system has been deployed to production and serves millions of users daily.
WhyProposes a dual-objective filter maximizing instant book rate without lowering booking rate, validated by global A/B tests and deployed to millions of users.
In briefEVA-Flow: Learning Shared Chemistry to Generate Molecular Shapes Across Different Environments
SummaryThis paper introduces EVA-Flow, a unified model that generates molecular 3D conformations while accounting for the environment, such as vacuum, protein-ligand docking, solvation, or crystal packing. It learns shared chemical structure and adapts it through environment-conditioned generative modeling. The authors report that pretraining and cross-environment finetuning yield the largest gains in three data-scarce settings.
Bing Yan, Anuroop Sriram, Zachary Ward Ulissi et al.
In briefCross-Modal Steering Vectors Work for Both Image and Text Generation
SummaryThis study examines whether concepts and mechanisms in unified multimodal models are shared across modalities or implemented separately. The authors find vectors that steer generation in one modality and test whether they transfer zero-shot to the other. Steering vectors for concepts such as emotion and object size transfer in both text-to-image and image-to-text directions. Vectors with strong transferability also point in similar directions for the same concept despite the modality gap between text and image representations.
In briefInbox as a Coordination Tool for LLM Agent Societies
SummaryThis paper studies how large language model agents coordinate messages. It proposes a "pull" approach, where messages wait in an inbox outside the agent's context and each recipient's own rules decide what is read and kept, compared with the conventional "push" approach. In simulated societies of up to 16,000 agents, the authors report that push fails when agents join many coordinations at once, that an inbox is necessary under heavy traffic, and that the best rules for each inbox decision depend on traffic and agent role.
Karthik Sreedhar, Leo Yuan, Pradyumna Chari et al.
In briefWhat an information bottleneck encoder should learn: separating representation learning from rate control
SummaryThis paper asks whether, in the information bottleneck (IB) framework, the learned representation can be separated from the compression-rate control parameter. For jointly Gaussian data, the authors show the optimal solution factorizes into compression-independent information directions and scalar gains, and they derive bounds on the separation gap governed by a sufficiency deficit. Experiments on MNIST, SVHN, CIFAR-10 and STL-10 use a frozen CNN representation with one compressor-decoder pair per budget, reducing training compute and wall-clock time versus end-to-end VIBs while achieving higher budget-averaged accuracy.
In briefcuArms: GPU-accelerated planning that coordinates multiple robot arms without collisions
SummaryThis paper presents cuArms, a conflict-based planner for coordinating multiple robot arms in a shared workspace, designed alongside GPU parallelism. It breaks down the search space using reusable multi-modal roadmaps for each arm and resolves conflicts and deadlocks through lazy search. A high-level scheduler proposes candidate skeletons that cuArms turns into coordinated manipulation plans. The authors report real-time initial solutions and anytime improvements in motion quality and scheduling, without the days of domain-specific training that learning approaches require.
In briefHierarchical multimodal reasoning for recognizing relationships between two people
SummaryThis paper addresses recognizing the social relationship between two individuals from their multimodal interactions, including facial, bodily, acoustic, and textual cues. It proposes HiRe-DRR, which organises relationship evidence by feature level, interaction type, and relationship hierarchy, using a Mixture-of-Interaction Transformer and a Hierarchical Conditional Loss. Experiments on three dyadic relationship benchmarks show it outperforms recent image, video, and multimodal methods. The routing analysis indicates distant relationships rely more on low-level homogeneous interactions, while closer ones focus on high-level heterogeneous interactions.
In briefFreeze the pretrained policy prior, rather than keep cloning it, when offline data is gone
SummaryThis paper studies offline-to-online reinforcement learning when the offline dataset is no longer available during fine-tuning. The authors find that continued online behavioral cloning erases the pretrained expressive prior, causing representational collapse and loss of action diversity. Across seven OGBench domains with IFQL, FQL, and action-chunked RL (QC), freezing the prior while stopping online behavioral fitting substantially improves fine-tuning.
Sami Nur Islam, Johan Obando-Ceron, Doina Precup et al.
In briefAlgorithm guaranteeing baseline protection and robust performance across adversarial and stochastic settings
SummaryThe paper proposes DUET, a horizon-free phased mixing algorithm for combinatorial semi-bandits with arbitrary finite action sets. Two observable tests decide when to double the learning weight and when to double the empirical-regret budget and restart. Assuming the baseline selects every arm with positive probability, expected regret relative to the baseline stays logarithmic in T. Regret to any fixed action is Õ(√T) under non-anticipating adversarial losses and polylogarithmic and gap-dependent under i.i.d. losses with a unique optimal action.
WhyA horizon-free phased mixing scheme gives baseline protection, adversarial robustness, and stochastic adaptivity with proven regret bounds, affecting a combinatorial semi-bandit subfield.
In briefBehaveRec: When Long-Term Behavior History Helps Sequential Recommendation
SummaryThis paper proposes BehaveRec, which combines a session transformer, a 90-day user history encoder, and a GraphSAGE module over a user graph built from behavioral embedding similarity, using a learned gate to blend the three signals. On an industrial dataset of 137M sequences from 12.6M users, it improves R@5 by 3.2% over SASRec. The gains grow with history length, becoming negligible for users with few interactions, and the authors derive a deployment rule of roughly 20 interactions per user.
Aysenur Inan, Reza Yousefi Maragheh, Jayesh Uddhav Kudase et al.
In briefMARQ: A modular, LLM-agnostic tool that scores requirement quality using systems engineering knowledge
SummaryThis paper presents MARQ, a modular prototype that equips large language models with a knowledge stack built on INCOSE requirement quality criteria to assess requirements. It compares MARQ against off-the-shelf models, including Claude Opus 5, GPT-5.6 Sol and Gemma 3 12B, on an expert-labeled NASA requirement set over 100 executions per model. Knowledge infusion raised median F1 scores for every model, though the gains and trade-offs differ by model and by criterion.
Jannatul Shefa, Alejandro Salado, Paul Wach et al.
In briefBeyond IceBERT: Open-data ModernBERT encoders for Icelandic, 12M to 1B parameters
SummaryThis paper trains Icelandic encoders from 12M to 1B parameters on 9B tokens of open data using the ModernBERT architecture. Under one evaluation protocol with confidence intervals, it reruns IceBERT's benchmarks and five newer tasks across 30 encoders. The 1B model trained from scratch significantly outperforms IceBERT on all eleven tasks, and a version warm-started from the English Ettin encoder scores best on every ranked task. Controlled tests show warm starts were never significantly worse than scratch training, and most of the transfer benefit comes from the source model's transformer layers.
In briefTelling AI coding agent commits from human ones without authorship signatures
SummaryThis paper examines whether commits made by AI coding agents can be distinguished from human commits after authorship signatures such as bot accounts and Co-authored-by trailers are masked. The authors analyze about 7.84 million commits from 1,226 repositories covering five coding agents and human commits. Detectors reach a ROC AUC of up to 0.859, with most signal coming from message verbosity and change size. Precision at the observed agent share is only 0.332, and detection transfers poorly to held-out Claude Code and Copilot commits.
In briefDirectional statistics on circles and spheres: why angles need special treatment
SummaryTwo compass bearings near north can average to a southward direction because angles lie on a circle rather than a line. The paper uses real wind observations, diffusion-MRI orientations, and geometric examples from machine learning. It introduces circular and axial averages, directional variability, and probability models on circles and spheres. Its guiding principle is to identify the geometry and symmetries of the data before choosing a statistical method.
In briefSparse autoencoders reveal how ESMFold's trunk folds proteins
SummaryThis paper uses sparse autoencoders to interpret the folding trunk of ESMFold, a protein structure prediction model. The authors analyze how the single and pair representations divide the work, interpret the learned features with a reasoning language model, and trace how concepts combine across blocks. They report that the sparse dictionaries capture the pairwise geometry the trunk computes, and that concept activations can forecast the model's own fold quality.
In briefEREBOTS 4.0: A Framework for Conversational Assistants Using A2A Agents and MCP Tools
SummaryThis paper presents EREBOTS 4.0, an open-source multi-agent framework for conversational assistants that combines language-model agents with deterministic tools. Language-model agents plan and route requests, while tools compute exact results over the Model Context Protocol, and an orchestrator calls specialist agents via the Agent2Agent protocol. New agents can be added by listing their URL in a configuration file. The framework's A2A server passed all 97 tracked requirements in the official A2A compatibility kit with no failures.
Elia Pacioni, Davide Calvaresi, Andrei C. Coman et al.
In briefDiagnosing and fixing bias in multimodal LLMs for long-video question answering
SummaryLong-video question answering depends on a few key moments, but memory limits force multimodal LLMs to sample frames sparsely, which can miss critical evidence. The paper introduces S-MME, a benchmark for systematically diagnosing shortcut-biased behavior in these models. It also proposes CLEAR, a training-free framework that estimates bias using augmented views and intervenes on hidden states at inference. Experiments across five benchmarks show it generalizes across different MLLMs with consistent gains in accuracy and robustness.
In briefCounterfactual profile perturbations explaining recommendations: a reproducibility study
SummaryThis reproducibility study re-examines LXR (Learning to eXplain Recommendations), a counterfactual method that finds changes to a user's ratings that explain why an item is recommended. The authors extend the original work to explanations for items beyond the top-ranked one and compare LXR with alternative baselines such as item popularity and latent factor similarity. They find that LXR produces shorter perturbations, meaning better explanations, when less popular items are explained.
In briefApproximation Families and Statistical Efficiency of Denoising Score Matching: a study of score-estimation variance and convergence under different approximation spaces in diffusion models.
SummaryThis paper studies score matching over finite-dimensional spaces built from gradients of eigenfunctions of a forward diffusion's generator. It compares polynomial approximation under Ornstein-Uhlenbeck flow with trigonometric-polynomial approximation under heat flow on the torus. In the Ornstein-Uhlenbeck setting, the asymptotic score-estimation variance is shown to grow exponentially along a subsequence of polynomial degrees for non-Gaussian initial distributions, and operator-informed score matching (OISM) reduces variance but does not remove this instability at small times. Under torus heat flow, OISM over trigonometric polynomials gives finite-sample score error control and attains the classical minimax rate for smooth densities under expected Kullback-Leibler loss.
In briefConcorde: a link prediction method that uses geometry-aware energy minimization
SummaryThis paper proposes Concorde, a framework for predicting links in sparse graphs by splitting link likelihood into three decoupled branches: a parameter-free spectral anchor, a mixed-curvature residual on a learnable hyperbolic-spherical product manifold, and a continuous vector field on the tangent bundle of hyperbolic space. The branches are fused via learned gating and trained with an N-pair ranking loss. The authors report state-of-the-art early retrieval precision across five benchmarks, with P@10% nearly doubling leading baselines in sparse settings.
In briefUT-HMR: Uncertainty-Guided Temporal Modeling for Robust 3D Human Mesh Recovery
SummaryThis paper presents UT-HMR, a one-stage framework that recovers 3D human meshes from a single image, targeting cases such as distant or occluded people and motion blur in sports. It predicts correlated pose uncertainty across joints for each person query while keeping inference deterministic, and fuses neighboring-frame evidence early in the pipeline. The method reports state-of-the-art results on ice hockey and soccer benchmarks.
In briefBase models can detect hidden intentions that prompted classifiers miss
SummaryThe paper introduces persona-header likelihood evaluation (PHLE), which scores outputs under a pre-trained base model with an appropriate persona header. PHLE distinguishes number sequences generated under different persona prompts, such as "You love owls" versus "You love eagles". It also tells apart outputs from a model with a hidden desire to harm users, filtered to remove noticeable misalignment, from those of a model with no system prompt. Prompted classifiers, including the generating model itself, perform near chance on these tasks, and the signal partly transfers across model families.
In briefcount-catcher: Turning model counting certificates into Lean theorems
SummaryThe paper presents count-catcher, a tool that converts a CPOG certificate from the exact model counter d4 into a Lean theorem stating that a CNF formula has exactly N models over V variables. With an encoding lemma, this becomes a theorem about the number of combinatorial objects. The authors combine these results with Mathlib and prove the number of Latin squares of order 8 as a weighted sum of hundreds of certified counts drawn from about eight terabytes of certificates.
WhyTurns model counting certificates into Lean theorems, proves Latin square counts of order 8, and finds bugs in CPOG's checker.
In briefTeaching Vision-Language Models to Read Braille from Images
SummaryThis paper introduces BrailleBench, a benchmark of 259 real-world Grade-1 and Grade-2 Braille images. It finds that standalone general-purpose vision-language models perform poorly at transcribing Braille, while tool-assisted systems perform much better. The authors build a synthetic Braille generation pipeline, and a 2B model trained on this data reaches 87% exact match on Grade-1 and 61% on Grade-2 BrailleBench. The paper also reports that high accuracy on synthetic validation data does not carry over to real photographs.
Kartik Jangra, Atharva Gawane, Siddharth Pulipaka et al.
In briefWiener-Informed Decoding to Ease Training of Pixel-Space Flow Models
SummaryThe paper addresses the difficulty of learning velocity fields directly in pixel space and finds that frozen Wiener estimates based on second-order image statistics improve velocity learning. It proposes Wiener-informed Decoding (WiD), which combines a trainable Wiener prediction path with coefficient interfaces initialized from data statistics. On ImageNet 256 x 256, PixelDiT-XL with WiD reaches FID 1.53 after 500 epochs, matching PixelDiT quality with 37.5% fewer training epochs.
In briefLOTUS: Detecting LLM Hallucinations via Low-rank Multilinear Decomposition
SummaryThis paper proposes LOTUS, a framework for detecting hallucinations, where large language models produce fluent but factually incorrect answers. It summarizes hidden-state token representations with statistical descriptors and applies low-rank multilinear (Tucker) decomposition to extract compact features. These features train a lightweight classifier. Experiments across several LLMs and question-answering benchmarks show better detection than recent state-of-the-art baselines.
Devansh Sharma, Faizul Rakib Sayem, Shahana Ibrahim
In briefBayesian evidence accumulation for graph learning in dynamic data-driven systems
SummaryThis paper introduces B-EAGLe, a probabilistic framework for graph classification in Dynamic Data Driven Application Systems, where observations arrive sequentially. It combines a graph neural network encoder with a Bayesian accumulator that updates class posterior beliefs as new observations arrive and commits to a decision only once sufficient confidence is reached. Experiments on CIFAR10 show competitive performance and principled confidence estimation under sequential observations.
Moises Ponce, Sarah Harkins Dayton, Ioannis Schizas et al.
In briefSkill Dreaming: Evolving repository-specific coding skills from pull request histories
SummaryThis paper introduces Skill Dreaming, a framework that lets LLM coding agents learn from a repository's pull request history. It builds a repository-specific skill library from validated PRs and keeps refining those skills as new PR outcomes appear. The skills guide agents in implementing features and maintaining tests, documentation, compatibility, and repository conventions. The framework combines PR analysis, skill management, code execution, and continuous integration in one workflow.
Xinyuan Song, Hasibul Haque, Jay Somasundaram et al.
In briefFrankly Speaking: Same Election Words, Different Company in Congressional Newsletters
SummaryThis paper analyzes 10,420 congressional e-newsletters sent from 2010 to 2025 to examine how Democrats and Republicans use election terms such as "integrity," "security," and "fraud." Using diachronic word embeddings, it measures how each party's usage differs and whether their nearest semantic neighbors match. The two parties use the same words but surround them with different language. By 2024, Democrats paired "voter," "election," and "integrity" with access and administration, while Republicans paired them with citizenship verification, noncitizen voting, and border security.
Michael A. Robinette, Christina Barsky, Amanda D. Clark et al.
In briefCan AI Guide Testing Without Writing Tests? A Comparative Study of AI-Assisted Unit Testing
SummaryThis paper introduces TestCompass, a generative AI tutor that gives targeted feedback on weaknesses in students' test suites without generating executable tests. In a classroom study with 18 students, it was compared with a general AI assistant on two Java unit-testing tasks. Students using the general assistant produced more tests, while TestCompass students spent more time developing and verifying tests themselves. Most test-suite quality measures were similar between groups, and in one task TestCompass students reached similar requirement coverage with significantly fewer test methods.
Zihan Fang, Ron Asoulin-Handelman, Helen Wu et al.
In briefLipi-Ghor: A Large-Scale Bengali Speech Dataset with Speaker Labels and a Reusable Build Method
SummaryThis paper presents Lipi-Ghor, a weakly supervised Bengali speech corpus of 885 hours drawn from 998 YouTube videos across 596 channels and 17 domains such as audiobooks, drama, sermons, news, and dialects. Recordings are labeled in SSTT format (speaker, speech, transcription, timestamp), combining automatic speaker diarization with Bengali YouTube captions. The authors also describe a replicable pipeline for other languages and benchmark four common ASR models, which perform worse on emotive, dialectal, and long-form content. A human-verified gold evaluation split, checked by 17 volunteers, is released alongside the corpus.
In briefLearning How Far Athletes Can Realistically Move Using Quantile Regression on Tracking Data
SummaryThis paper studies time-to-arrival (TTA) in sports, the shortest plausible time a player needs to reach any location on the pitch given their current movement. The authors frame TTA as a conditional quantile regression problem that estimates the lower quantile of locomotion time, using optical tracking data from all 64 matches of the PFF FC FIFA World Cup 2022 dataset. A lightweight multilayer perceptron outperforms six physics-based baselines and a learned reachable-hull baseline. The model is also tested on basketball and a single NFL game to examine how player motion dynamics transfer across sports.
In briefDynamical coarse-graining of neural-network training: a framework for macroscopic modeling of learning
SummaryThis paper develops a dynamical coarse-graining approach that describes neural-network training in a state space much smaller than the parameter space. The authors define a learning operator based on the training law, prove a descent theorem for it, and then present an exact closure theorem and a dynamic coarse-graining theorem as a framework for building and assessing such models. As an example, they derive an exact coarse-grained model for a finite-width nonlinear ReLU network with a fixed pattern, composed of five independent moments, which exactly predicts the training outputs and the evolution of the learning operator.
In briefHow instruction tuning shapes language models' random number choices
SummaryThe paper measures, without sampling, the probability language models assign to each integer from 1 to N when asked for a random number. It evaluates 18 open-weight model lines across several decoding configurations, comparing 15 base and instruction-tuned checkpoint pairs. Instruction tuning made the distributions less uniform in all pairs, installing round-number avoidance and sharpening endpoint avoidance.
In briefDon't Pick a Winner: Knowledge Conflicts Need a Prior, Not a Rule. A study on how retrieval-augmented language models should resolve conflicts between retrieved context and their own knowledge.
SummaryThe paper studies how retrieval-augmented language models should answer when retrieved context and the model's own knowledge disagree. It argues that fixed design-time rules are wrong in some settings and that resolving conflicts requires separate evidence from each source and a prior on how often the context is right. On the GRACE benchmark and ClashEval, the authors' method, which voices knowledge notes with calibrated confidence and commits or abstains under a prior, outperforms baselines.
WhyProposes a decision-theoretic prior-based framework replacing fixed conflict rules, with a two-source selector; novel, and supported by comparisons on GRACE and ClashEval.
In briefPrinGS: Privacy Preservation in Gaussian Splatting
SummaryThis paper addresses privacy risks in 3D Gaussian Splatting, whose learned representations can encode sensitive content such as faces, license plates, and documents. It proposes a framework that anonymizes a trained splat directly, using a user-provided reference view to locate and remove sensitive regions in the 3D representation. The authors also introduce a privacy-by-design training strategy that keeps sensitive content out of the representation during optimization. Reported results include a Face Similarity Measure of 9.79% and an OCR Character Error Rate of 1.00.
Gaurav Behera, Abhinav Raundhal, Archisha Panda et al.
In briefStratum: Persistent Repository Memory for Coding Agents
SummaryThis paper proposes Stratum, a tiered memory system that lives with a code repository so coding agents can retain knowledge across sessions instead of re-reading files from scratch. In a multi-session experiment on eight real bugs, a fresh agent that inherited memory from a prior session found the root-cause file in 21 of 24 trials, versus 6 of 24 for a retrieval-only baseline, at a 4.5x lower median token cost. Retrieval over a store of one million nodes takes 1.7 ms at the median.
In briefReplication vs. Corroboration: An Axiomatic Framework for Multi-Agent Debate
SummaryThis paper studies whether matching reports in multi-agent debates repeat the same evidence or come from separate evidential roots supporting the same claim. It formalizes this as a minimal judgment game and proposes axioms such as root replication invariance, along with a lineage quotient that separates replication from corroboration. The framework is extended to quantitative bipolar argumentation frameworks and checked through exhaustive profile audits and profiles derived from AVeriTeC.
Hoang-Loc Cao, Truong Thanh Hung Nguyen, Phuc Truong Loc Nguyen et al.
In briefAny3R: Feedforward 3D Reconstruction from Any Camera Type
SummaryThis paper presents Any3R, a feed-forward network that recovers camera poses and dense geometry from images captured by any central camera. It replaces the pinhole-camera assumption in positional encoding and training loss with a per-pixel ray direction field, so no lens model is needed. On a benchmark of thirteen central cameras across three datasets, Any3R was the most consistently accurate method when each approach used its own predicted poses.
In briefConfidence bands for distribution functions valid at any time
SummaryThis paper develops confidence bands for distribution functions of a data stream that hold simultaneously across all times and all points, even when observations depend arbitrarily on the past. It first derives such bands for the time-averaged conditional CDF and shows that their widths are set by the depth of a binary search tree. The result is then generalized into a framework for building bands uniform over time and an arbitrary index set, illustrated with importance-weighted, right-censored, and multivariate CDFs.
Bror Hultberg, Dave Zachariah, Ant\^onio H. Ribeiro
In briefLearning Prompt Editors That Keep Scene Structure Intact
SummaryThis paper introduces StruPO, a structure-aware prompt editor that uses a calibrated scene graph to guide rewriting of text-to-image prompts while preserving the requested scene. Edits are made through token-aligned deletion, reordering, and insertion, and training uses smooth reward shaping and averaging over multiple generated images. Experiments on three disjoint 2K-prompt test sets and four frozen generation backbones report better faithfulness-aesthetics trade-offs and consistent gains in relational fidelity.
In briefSpatialPC: Deep Computation via Sublinear Local Credit Propagation
SummaryThe paper argues that the number of rounds a bounded-range local learning rule needs before loss signals reach the earliest parameters depends on communication geometry rather than network depth. It constructs a unit-range, degree-2d directed acyclic graph and introduces SpatialPC, which propagates credit within predictive coding via adjoint prediction dynamics. On MNIST and Fashion-MNIST, finite-round gradients remain aligned with settled gradients, and test accuracies of 97.6% and 88.1% are reported after 20 epochs.
In briefSymmetry breaking under limited model capacity: a paper on why optimal models lose rotational symmetry
SummaryThe paper studies why the best model for a rotation-symmetric problem can break that symmetry when capacity is limited. For a Gaussian problem with a bilinear predictor, it derives an exact log-loss price of symmetry and extends the result to attention heads fitted to symmetric similarities. It reports that with peaked softmax attention, broken heads beat every whole-shell symmetric head at 16 of 20 head dimensions.
WhyDerives an exact log-loss price for symmetry under limited capacity and verifies predicted symmetry-breaking dimensions in attention heads with solid experiments, though impact stays within ML theory.
In briefContrastive Evidence Paths: Guiding Search Agents Past Misleading Connectivity
SummaryThe paper proposes Contrastive Evidence Paths (CEP), which scores candidate documents by their strongest individual paths in a query-conditioned document graph and calibrates these scores against the graph's background visitation distribution to reduce structural bias. Building on CEP, the authors introduce Structured Corpus Interaction (SCI), which organizes candidates and connecting paths into a structured Map that lets agents verify sources and expand from anchor documents. Experiments on BrowseComp-Plus and TREC-COVID report state-of-the-art answer accuracy and nDCG@10 with the lowest generation cost.
In briefConformalMOS: Uncertainty-Aware Speech Quality (MOS) Prediction with Conformal Intervals
SummaryThis paper proposes ConformalMOS, a framework for predicting Mean Opinion Scores (MOS), the human ratings used to judge synthetic speech quality. Alongside point estimates, it uses conformal prediction to produce prediction intervals with statistically guaranteed coverage under exchangeability. During training, MOS labels are converted into Gaussian-smoothed distributions to preserve score ordering. On BVCC datasets, the method reports competitive accuracy (MSE = 0.08) and empirically validated coverage.
Kehinde Elelu, Joshua E. Siegel, Mohammadali Saffary et al.
In briefSuperJEPA: Supervision-Augmented Predictive Pretraining for High-Energy Physics
SummaryThe paper introduces SuperJEPA, a joint-embedding predictive architecture that adds a supervised objective to pretraining for high-energy physics jet tagging. Compared with supervised pretraining alone, it improves downstream jet-tagging performance across multiple tasks, model sizes, and pretraining dataset sizes. The gains hold for both frozen and fine-tuned encoders, which the authors say motivates predictive self-supervision in future large-scale particle-physics foundation models.
In briefWhy and How to Do Science When AI Surpasses Humans: A Framework Based on Human Understanding and Direction
SummaryThe paper asks why humans should keep doing science as AI matches or exceeds researchers at core tasks, and how AI should be used in research. It argues that science's indispensable aim is durable, collective human understanding of the world, pursued for human ends and values. From this it derives a two-part test for delegating research tasks to AI: human understanding and human direction. The authors apply the test to each phase of the CRediT taxonomy and say the framework holds even if AI outperforms human researchers.
In briefContext as a Pipeline: Information Loss in LLM Agent Tool Outputs
SummaryThis paper studies how much information from tool outputs is lost before reaching the language model in an LLM agent that diagnoses injected faults from Linux kernel traces. Across 720 runs, about a third of 11,189 tool results were cut by fixed-length JSON slicing, producing broken JSON and uneven field visibility. The authors model context as a five-stage pipeline and report controlled experiments in which the cost of cutting appears first as extra tool calls and computation rather than outright wrong answers.
In briefDirichlet Process Mixtures for Conditional Density Estimation in Semiparametric Bayesian Networks
SummaryThis paper proposes using Dirichlet Process Mixtures (DPMs) to model non-parametric conditional probability distributions in semiparametric Bayesian networks, replacing kernel density estimators. DPMs adaptively cluster observations into a more compact representation, lowering inference costs. On synthetic data, the approach yields lower density estimation errors than state-of-the-art Bayesian network models for continuous variables, and it also outperforms kernel-based networks in structure learning and on real-world data.
Victor Alejandre, Pedro Larra\~naga, Concha Bielza
In briefBeyond Update Counts: A Control Map for Evaluating Test-Time Model Updates
SummaryThis paper introduces the Control Map, an evaluation framework that records the model updates made before each scored answer and assesses behavior change, learning, retention, locality, and performance in new settings separately. Experiments show that schedules with the same number of updates can produce very different accuracy, that learning an edit does not guarantee it persists or stays contained, and that no editing method performs best across all models, benchmarks, and tests. The authors recommend reporting update histories and evaluating acquisition, retention, and locality separately.
In briefCohortFlow: A Steerable Agent for Clinical Cohort Curation with Localized Corrections
SummaryThe paper introduces CohortFlow, a language model agent that applies natural language corrections to clinical cohort specifications as localized updates, reconciled with a persistent execution graph. Only computations affected by recorded semantic dependencies are invalidated and recomputed. On 100 clinical-trial-derived scenarios with 126 correction turns, localized steering achieved 77.0% requirement micro-F1 versus 76.4% for full regeneration, a difference that is not statistically significant, while reducing correction latency and model calls. A proof-of-concept over MIMIC showed all 13 cohort states executing successfully.
In briefMinimax-optimal small-loss bounds for online linear optimisation on ℓp balls
SummaryThe paper studies full-information online linear optimisation over the unit ℓp ball for fixed p between 2 and infinity. It presents a follow-the-regularised-leader algorithm whose regret depends on the cumulative excess loss and the dimension, and shows this rate is minimax-optimal for each fixed p. It also analyses a switching extension that keeps the sharp constant, and computes the exact log T coefficient of expected regret for i.i.d. losses.
WhyMatches minimax lower bounds with an optimal constant for small-loss regret, plus analysis of FTL failure paths; impact mainly theoretical.
In briefInstruction Capacity: how many rules a language model can follow under load
SummaryThe paper proposes a measurement instrument that treats rules replacing meaningless single-token words as a controlled instruction load, with each replacement counted as a Bernoulli trial with a known reference answer. A statistical model describes how the probability of a correct replacement changes with text length, word position, replacement density, and rule number, yielding an instruction capacity with uncertainty. Measurements on GPT-4o mini (2,958 requests, 642,430 trials) gave about 37–38 rules at 0.95 correct replacements and about 18–19 rules at full replacement density.
In briefHow Much the Hint Matters: Conditions for Hindsight Self-Distillation in Financial Reasoning
SummaryThis paper addresses financial institutions that must run small open-weight models locally due to data-residency rules. It proposes Fin-SD, in which a model identifies its first incorrect reasoning step, writes a short critique, and is then distilled, under reverse KL, from a hint-conditioned copy of itself on a correct prefix, without any external model. Replacing the critique with generic feedback yields the same accuracy gain, suggesting the benefit comes from where supervision is applied rather than hint content. Fin-SD matches full-reference self-distillation on accuracy while generating 9.5% fewer tokens, with no material regression on general reasoning or compliance audits.
In briefClosed-loop planning that scores candidate speeds without imitation, for end-to-end autonomous driving
SummaryThe paper proposes CLAP-AD, a planning head that plugs into existing end-to-end driving models. It learns the path through imitation but scores candidate speeds with a learned function supervised by physically grounded safety constraints, allowing explicit speed control at test time. On Bench2Drive-Speed it reaches a Speed-Adherence Score of 82.53, 12.9 points above the previous best baseline TFv6, while keeping strong closed-loop driving performance.
WhyProposes a plug-in planning head enabling test-time speed control without speed-conditioned training, with clear gains on Bench2Drive and NAVSIM; impact is confined to end-to-end driving.
Emad Aghajanzadeh, Thomas Schwarzfischer, Marc Kaufeld et al.
In briefLearning quasi-stationarity diagnostics from low-dimensional surrogates
SummaryThe paper addresses estimating decorrelation times needed by accelerated dynamics methods such as parallel replica dynamics. It replaces the MCMC-based diagnostic with one built on coarse-graining of non-linear probability flows and simulation-based inference, and proves uniform-in-time, non-asymptotic bounds comparing a surrogate process's flow to the true flow. A recurrent neural network is trained to estimate decorrelation posterior probabilities and is compared with the MCMC diagnostic on synthetic and molecular systems. When its prior assumptions hold, the learned diagnostic performs better in ROC analysis and is less conservative.
