In briefSkeleton Coding: Using Program Structure as a Specification for AI Coding Agents
SummaryThe paper argues that the program structure an AI coding agent chooses determines which tests are needed, so behavioral tests alone can leave a developer's intended program underdetermined. It introduces Skeleton Coding, where developers specify a high-level code-like structure alongside behavioral specifications. The authors show theoretically that skeletons can reduce the number of tests needed exponentially, and report that across four models and three benchmarks the best skeletons cut mean test error by 48.4% relative to the best end-to-end specifications.
Aditya Golatkar, Alessandro Achille, Renos Zabounidis et al.
In briefRainCheck: Using Historical Weather Data and Machine Learning to Forecast Rainfall in Bangladesh
SummaryThis study uses 618,650 station-day records from 35 Bangladesh Meteorological Department stations, covering 1948 to 2014, to compare machine-learning rainfall forecasts. It finds that a station-month climatology alone classifies high-rainfall months with 88.4% accuracy, and most learners do not significantly beat climatology on future years. The authors also release a leakage-aware evaluation protocol and code.
Shahran Rahman Alve, Muhammad Zawad Mahmud, Samiha Islam et al.
In briefGAAC: Lossless compression of the LLM memory cache for faster on-device inference
SummaryThis paper presents GAAC, a lossless key-value (KV) cache compression and serving system for large language models that exploits redundancy in BF16 exponents without approximating the cached values. Across five tested models, it achieves 1.405 to 1.407 times lossless BF16 KV compression. On NVIDIA DGX Spark and Apple M4, it speeds up attention computation, with reported gains of 1.19 to 1.27 times at a 64K-token context and 1.02 to 1.25 times respectively.
In briefProgressive coupling diffusion sampling for reducing scan overlap in ptychography
SummaryPtychography reconstructs a sample from overlapping diffraction measurements, but dense scanning makes acquisition slow. This paper proposes progressive coupling posterior sampling (PCPS), which explores the solution space broadly under high uncertainty and strengthens diffusion-guided consistency as reconstruction refines. Experiments on real X-ray ptychographic data at various scan overlaps show improved accuracy over traditional iterative methods and existing diffusion posterior samplers, especially at low overlaps. The results suggest diffusion priors can reduce the measurement redundancy needed for accurate reconstruction.
In briefAnnealed, Averaged Markov Chains Reach the Algorithmic Threshold for Tensor PCA (Part II: Langevin Dynamics)
SummaryThis paper studies tensor principal components analysis on the sphere, where an unknown vector must be recovered from a noisy observation of its k-th tensor power. The authors show that spherical Langevin dynamics, combined with annealing and ergodic averaging, achieves polynomial-time strong recovery above the algorithmic threshold from a uniformly random initialization. Together with Part I, the results show that local Markov chains can succeed down to the algorithmic threshold from an uninformative cold start.
WhyThe authors show annealed, averaged spherical Langevin dynamics recovers tensor PCA signal in polynomial time above the algorithmic threshold from random initialization, continuing Part I.
In briefTwo Birds with One Stone: Unifying Skill Retrieval and Use in LLMs via Reinforcement Learning
SummaryThe paper proposes Skill-Uni, a reinforcement learning framework in which a single language model learns both to retrieve reusable skills and to use them for solving tasks. Training uses a warm-up instruction tuning stage followed by GRPO. Across five skill-related benchmarks, it reports average gains of 66.59% over Qwen3-4B and 34.82% over LLaMA3.1-8B, and it works with various skill retrieval engines.
In briefFrom Natural Language to Real Schematics: Automated Power Electronics Circuit Design
SummaryThe paper presents a bounded dual-loop framework that converts natural language specifications into verified, editable KiCad schematics and manufacturing outputs. It covers Buck, Boost, Flyback, DAB, and half-bridge LLC topologies, using distributor APIs, datasheets, ngspice transient verification, and KiCad netlist round-trip checks. In two end-to-end case studies and comparisons with a Vanilla LLM baseline, component verification coverage, nominal value consistency, and SPICE simulation pass rates improved substantially.
In briefEmbodied Underwater Intelligence: Imitation Learning for Multi-Task Underwater Manipulation
SummaryThe paper trains 14 generalist Vision-Language-Action (VLA) models, called DynamicVLA, on 1616 underwater manipulation episodes collected in calm and turbulent waters to address visual distortion and currents underwater. Across three tasks, 10 models outperformed the baseline, and the best reached a 93.3% overall success rate. A marine archaeology case study compared a single-task Action Chunking Transformer with SmolVLA on grasping and mooring, with ACT substantially outperforming SmolVLA (47.9% vs. 7.9% pooled success). The BLUEMBODIED kit used for data collection, training, and deployment has been open sourced.
Greta Perez-Haiek (1), Luca Macesanu (1), Kevin Angers (1) et al.
In briefConvex-SKM-Net: A low-latency method for enforcing convex constraints in neural networks via block halfspace corrections
SummaryThe paper introduces Convex-SKM-Net, a convex constraint-satisfaction layer that builds linearization halfspaces from residual and subgradient oracles and corrects predictions using fixed-depth dual projected-gradient descent on minimum-displacement block problems. The predictor and a positive-definite mobility matrix are learned jointly to improve objective quality. Reported tests on second-order-cone model-predictive control and power-system convex relaxations keep outputs within a 10^-3 feasibility tolerance, with inference speedups of up to 144.3 times over compared methods.
In briefGearUp: DAG-aware serverless inference acceleration via slack-guided pre-loading
SummaryGearUp is a serverless system for running machine learning inference workflows structured as directed acyclic graphs (DAGs). It schedules each node's library, model weight, and CUDA kernel loading within the slack time the DAG exposes, hiding initialization behind upstream execution. The system also reuses initialized instances and uses a traffic-adaptive keep-alive policy. On the AMBER Alert and FINRA-ML workflows, it reduces median end-to-end latency by 64.5% and 75.9% compared with SMIless.
In briefA benchmark and prototype-guided method for detecting surface defects on shampoo bottles across many product variants
SummaryThis paper addresses shampoo bottle surface defect detection, where the same defect can look different across product variants (SKUs). The authors introduce Bottle-SKU, a benchmark with 13,119 images, 20,419 bounding boxes, seven defect categories, and 48 bottle SKUs. They propose ProDMM, which builds a visual prototype for each SKU and uses it to dynamically modulate a Grounding DINO detector while keeping its pretrained detection path. Experiments on Bottle-SKU and the public GC10-DET-SKU dataset are reported to show the approach's effectiveness.
In briefAdding Multimodal Grounding to Text Moderation for Detecting Fraudulent Inducement in Short Videos
SummaryThis paper addresses fraudulent inducement behaviors on short video platforms, such as prompts to like, follow, or join private chat groups. It compares three approaches: OCR+LLM, OCR+MLLM, and a large zero-shot multimodal LLM, using a dataset of 2,419 videos with 3,194 annotated events. The OCR+MLLM method localizes offending segments better than the keyframe-based baseline (mIoU 26.52 vs. 14.49) and detects non-textual inducements that text-only pipelines miss. It maintains 67.0 percent macro recall.
In briefIn-context learning for multi-fidelity multiscale model closure and cross-operator generalization
SummaryThis paper proposes an in-context multi-fidelity learning framework that uses one shared model to learn multiple multi-fidelity operators. The model takes paired low- and high-fidelity examples from multiple PDE operators as demonstrations, plus a new coarse-scale solution query, and predicts the corresponding fine-scale solution. After training, it predicts fine-scale solutions for a new operator from fewer than five coarse-fine demonstration pairs and a low-cost coarse query, without fine-tuning.
In briefDecompGym: A Benchmark for LLM Agents Designing Decomposition Algorithms
SummaryThe paper introduces DecompGym, a benchmark of 81 optimization instances from nine problem families, including five Benders and four column generation tasks, where agents must write solvers for prescribed decomposition methods. Each solver is checked for running within 60 seconds, numerical feasibility, an LLM judge's audit of the method used, and an objective within 1% of the stored optimum. Evaluating 12 GPT, Claude, and Gemini models over up to ten attempts, the authors find that reflection, which diagnoses failures first, succeeds on 66% of retries versus 39% for independent regeneration. The study concludes that decomposition agents need checks on how a solver reaches its answer, not only output checks.
WhyThe authors build a 81-instance benchmark testing LLM agents on Benders and column generation, comparing regeneration with reflection across 12 models, and separating output checks from method audits.
In briefScaling Lean Theorem Proving for Real-World Code Verification
SummaryThe paper presents EasyProver, a Lean proving system built from recursive agents, asynchronous search, and parallel tactic probing, which uses an off-the-shelf model without additional training to solve most problems on benchmarks such as PutnamBench and IMO 2026 formalizations. Because these benchmarks are near saturation, the authors introduce Aegis, which turns competitive programming problems into harder verification tasks over real Rust implementations translated to Lean. The pipeline yields 2,045 differentially validated Rust-Lean program pairs and 11,533 candidate specifications. On a stratified sample of 2,000 statements, EasyProver proves 1,054 (52.7%) within a 30-minute budget per statement.
Debargha Ganguly, Divya Raghunathan, Dejan Jovanovi\'c et al.
In briefGradient-Based Adversarial Augmentation for Countering Unlearnable Datasets
SummaryThis paper proposes a gradient-based adversarial augmentation framework for stress-testing unlearnable data methods (ULD), which add imperceptible perturbations to public data to block unauthorized model training. The framework learns an augmentation policy via stochastic gradient ascent over a differentiable pipeline of up to 18 augmentation techniques, training to maximize a surrogate model's classification loss. A perceptual-plus-pixel regularization preserves image semantics. Under a controlled ablation, it outperforms an equivalent random sampling baseline.
In briefUneven data sizes across sites improve federated learning for brain imaging
SummaryThis paper studies federated learning for neuroimaging, specifically classifying Alzheimer's disease from structural MRI scans. It compares federations where sites have unequal data amounts (one large site plus small sites) against federations where every site holds the same amount of data. The heterogeneous federations perform better, especially in low-data settings, and come close to centralized training. A large site often leads to faster convergence and higher final accuracy.
Paul Kurian, Tamoghna Chattopadhyay, Abhijith Shaji et al.
In briefB2P-Corr: Batch-to-population gradient estimators for correlation losses that cannot be split across mini-batches
SummaryThe paper observes that correlation metrics such as Pearson and concordance correlation depend on dataset-wide statistics, so naive mini-batch training yields biased, high-variance gradients. It proposes B2P-Corr, which converts these objectives into decomposable pseudo-losses using lightweight global moment sketches, with two variants, B2P-EMA and B2P-U. Experiments on two synthetic and three real-world datasets show more stable optimization and higher correlation than naive mini-batch training, with negligible overhead.
In briefReplaying Gradients at a Different ε: A Pre-Registered Test of Adam's ε Explanation for a Rare Embedding Row
SummaryThe paper examines how Adam's epsilon term affects updates to a rare token's output-embedding row. During training, the authors recompute that row's updates with a different epsilon while keeping the gradients fixed, and test this explanation against a rule written before the runs. When epsilon rises by four orders of magnitude, the replay predictions are clearly wrong, with the row keeping far more of its displacement than predicted. At three orders of magnitude the gap is smaller and less clear.
In briefNeighbourhood shortcuts in satellite building damage assessment, with a building-evidence plug-in to reduce them
SummaryThis paper examines deep models that assess building damage from very-high-resolution satellite imagery and finds they often rely on neighbouring buildings' damage levels as a shortcut, missing isolated damaged buildings. The authors propose a plug-in that adds two experts to an existing model without retraining it, one looking only at the building and one trained to ignore swapped neighbourhoods. The plug-in is reported to reduce neighbourhood-driven errors across nine models.
In briefATI-OPD: Adaptive Teacher Intervention for Stable Long-Sequence On-Policy Distillation
SummaryThis paper proposes ATI-OPD, a training method that improves on-policy distillation, where a student learns from its own responses using token-level feedback from a teacher. For long responses, the student's reasoning prefix can drift from the teacher's, weakening later feedback and destabilizing training. ATI-OPD replaces the earliest high-disagreement segment with a short teacher segment and lets the student resume reasoning from the revised text. Across five mathematical reasoning benchmarks and two teacher-student pairs, it improves average student-only accuracy over standard on-policy distillation.
In briefREEL: Same frames, different answers and costs
SummaryThe paper introduces REEL, an evaluation framework that compares three saved memories built from the same video frames: a text record, visual encoder outputs, and language-model prefix state. Each is compared against reprocessing the full frames using paired correctness changes, cumulative GPU time, and storage size. Across six vision-language models and two benchmarks, Text differs from Frames on 20% to 29% of questions despite similar mean accuracy, while Prefix State stays closest to Frames and has the lowest total GPU time, at the cost of up to about 21 GiB per video.
In briefSimulating creator ecosystems to test fair video recommendation allocation at billion-user scale
SummaryThis paper presents a user-simulation framework that models about 42,000 creators and their viewer interactions over 12 weeks, grounded in real platform distributions. It evaluates equitable view-allocation policies formulated as a budgeted multi-armed bandit, varying view targets, reward functions, and a small-creator reserve. A floor-protected approach meets 90% of view targets weekly and raises small-creator organic reach by about 33% per week.
Sohini Roychowdhury, Long Sha, Sunand Agarwal et al.
In briefMASS: A multi-mask selective self-distillation method for leakage-resilient reasoning (paper)
SummaryThis paper proposes MASS, a multi-mask selective self-distillation method for improving reasoning in language models. It creates multiple independently masked views of one verified solution so the teacher model can combine diverse reference signals, then selectively transfers reliable token-level corrections while focusing masking on influential reference spans. Experiments on mathematical reasoning benchmarks and several model families show that MASS consistently outperforms strong self-distillation baselines.
In briefStepControl: Generating Long Human Motion from Foot-Contact Timing Schedules
SummaryThis paper introduces StepControl, an autoregressive diffusion model that generates full-body human motion from a schedule specifying when each foot touches the ground. It needs no target root or body poses and no semantic prompts, representing contact as explicit temporal events injected via cross-attention with classifier-free guidance. Experiments show it follows prescribed contact timing more accurately than adapted spatially guided generators, and keeps control over long rollouts and edited schedules such as cadence changes, pauses, limping, and jumping.
In briefBridgeRouter: Per-instance routing among cross-modal bridge pathways for biosignal models
SummaryThis paper addresses transferring pretrained EEG foundation models to other biosignals without retraining their backbones. Instead of one global bridge pathway, BridgeRouter selects one of several frozen cross-modal pathways per instance, falling back to a validation-selected default route. It is evaluated on EMG-based freezing-of-gait detection and ECG-based sleep classification, where it improves balanced accuracy over the best single route.
In briefEstimating Implicit Expected Log Density Ratios: a study introducing the expected log density ratio (ELDR) and its finite-sample estimation limits
SummaryThe paper introduces the expected log density ratio (ELDR), which is the integral of log density ratios under a third measure. It shows that many implicit-density problems that naively use density ratio estimation (DRE) are actually ELDR estimation tasks. The authors prove fundamental limits on finite-sample ELDR estimation and argue that typical DRE methods are unsuitable for it. They also propose a simple modification to some DRE methods, which benchmarks show can reduce ELDR estimation error.
In briefSkillCT: A multi-dimensional benchmark and execution-free baseline for rating AI agent skills
SummaryThe paper introduces SkillCT, a benchmark that assesses agent skills across nine dimensions grouped into static, dynamic, and human-assisted assessment, going beyond task execution success. It also proposes an execution-free baseline that predicts a skill's quality profile by retrieving similar reference skills. According to the authors, this approach lowers token use and evaluation time, and SkillCT-guided optimization improves downstream task success and execution efficiency.
In briefEl Agente Quntur: AI agents choosing active spaces for CASSCF chemistry calculations
SummaryThis paper tests whether a multi-agent AI system can reliably choose the active orbital space for CASSCF multireference quantum chemistry calculations. The agents selected orbitals, checked target excited states, and judged convergence for 32 singlet excitations, and their results were compared with six fixed Active Space Finder protocols using published CASPT2 reference values. The agentic approach reduced mean absolute error by 48% relative to the best deterministic protocol. The authors also note that computational traces and costs varied substantially across runs, even for the same molecule.
Juan B. P\'erez-S\'anchez, Viviana Roman-Ventura, Olalla Nieto-Faza et al.
In briefDecompose, Don't Expand: Designing AI systems that work regardless of model size
SummaryThis paper examines a design choice in AI systems: when a task is too hard for a model, expand the model or decompose the task so a deterministic host handles the solvable parts and the model handles only the fuzzy remainder. The authors report measurements across nine models, finding that on the extracted sub-tasks performance no longer varies with model size. They argue the approach is a general design discipline applicable to physics setup, motion retargeting, and object-model editing.
In briefTheoretical Limits and Guarantees for Graph-Based Nearest-Neighbor Search
SummaryThis paper studies the theoretical guarantees of graph-based nearest neighbor search methods such as HNSW, DiskANN, and NSG, which perform well in practice. The authors prove that standard graph-based methods running greedy search on flat or hierarchical graphs cannot achieve sublinear query time in Euclidean space, even with arbitrarily large approximation factors. They show that allowing Steiner points in the search graph makes sublinear query time possible, with bounds comparable to LSH, and include initial experiments on incorporating Steiner points into practical methods.
WhyThe authors prove standard graph-based search cannot achieve sublinear query time, and that Steiner points yield LSH-comparable bounds, with initial experiments.
Majid Daliri, Mohammadreza Daneshvaramoli, Purna Dutta et al.
In briefA Separation Principle for Attacks and Defenses on Graph Neural Networks
SummaryThis paper treats adversarial attacks on graph neural networks as a problem shaped by graph structure, not only by gradients or learned policies. It derives feature attacks from the spectral properties of the normalized adjacency operator and states a separation principle: the optimal perturbation is rank one, splitting into a graph-dependent degree-based allocation and a weight-dependent top singular vector. Based on this, the authors propose SepAttack, a gradient-free attack, and SepAT, an adversarial training procedure.
In briefMiniMMD: Minimizing MMD for Sampling, Inference and Integration
SummaryThis thesis presents a unified approach to many problems in statistics and machine learning, which approximate a target distribution by minimizing the maximum mean discrepancy (MMD), a kernel-based distance between probability distributions. It argues that MMD is both a statistically meaningful distance and a tractable objective for designing practical, efficient algorithms. The thesis summarizes recent theoretical and empirical advances in these minimum-MMD-based algorithms.
In briefWhat LOL Measures: Hidden Label Shifts in Low-Light Enhancement Benchmarks
SummaryThis paper audits the standard low-light image enhancement benchmarks LOL-v1 and LOLv2-Real using 18 public checkpoints from eight methods. It finds that test labels differ from training labels, with LOL-v1 test labels rendered differently and LOLv2-Real test labels darker than its training labels. The authors show that test-set checkpoint selection and released test-time settings affect reported scores and method rankings, and they provide reporting recommendations.
WhyThe authors audit 18 checkpoints of eight methods, report hidden label shifts in LOL benchmarks, and quantify their effect on rankings.
In briefPost-training LLMs with LLMON for safer instruction handling
SummaryThis paper proposes a post-training method using LLMON (LLM Object Notation), which separates executable instructions from passive data and requires identifier-based execution. The authors built a 4.6B-token structured training corpus and post-trained Granite-4.1 3B and 8B base models with supervised fine-tuning and reinforcement learning. They report that attack success rates on indirect prompt-injection benchmarks fell from 53-100% to 0.02-5.6% while general capabilities were preserved.
WhyThe authors report that LLMON-structured post-training cuts indirect prompt-injection success rates substantially on two Granite models, and introduce attention-based interpretability metrics.
In briefToo Hard, Too Easy, Just Right: Adaptive Challenge in VR Firefighter Training
SummaryThis paper presents ProACT, a performance-driven adaptive training system for virtual reality firefighter training. It adjusts scenario difficulty through closed-loop interaction between trainee performance and simulated fire dynamics. A global module updates scene-level fire spread in real time, while a local module periodically updates object-level flammability. Compared with adaptive and non-adaptive baselines, ProACT improves both objective training performance and subjective user experience.
In briefBeauty Is Human: Liking AI Art Leads People to Think a Human Made It
SummaryIn a public online AI Art Turing Test, 11,123 people judged whether 50 pictures, 26 of them AI-generated, were made by humans and picked their favourites. Holding respondent and picture fixed, a picture was 22.1 percentage points more likely to be judged human-made when it was the respondent's favourite, with a similar pattern for AI-generated and human-made favourites. A reanalysis of a published AI poetry experiment found the same pattern, suggesting that when provenance is hidden, liking and perceived human authorship go together.
In briefFrom Gettier Machine to Justified True Output: Rational Trust in Generative AI
SummaryThe paper asks when a user who takes a proposition from generative AI output actually knows it and when trusting it is rational. It argues that failures of AI-mediated inquiry match the problems that the justified-true-belief tradition and its post-Gettier anti-luck refinements were built to diagnose. It proposes a three-step pattern of use (acquisition, understanding, independent verification) and a stakes-indexed criterion called justified true output (JTO), with the everyday test: "Had this been false, would I have noticed?"
In briefWho Called? V33DA: A Verified Multimodal Benchmark for Identifying Zebra Finch Callers
SummaryThis paper introduces V33DA, a benchmark dataset for determining which zebra finch in a group produced a vocalization and where it was made. It contains 33,625 events that pair multichannel audio, three camera views, calibrated 3D bird poses, and radio telemetry for 10 birds, with caller identity verified by an on-body accelerometer that is withheld from the models. The authors test methods under three regimes and find that approaches relying on familiar vocal identity perform well on known birds but degrade on unseen ones, while candidate-aware spatial methods transfer better. The dataset, evaluation protocols, and code are released publicly.
In briefTGTV: An Interpretable Framework for Drone Fault Detection and Tracing Using Graph Learning and LLMs
SummaryThis paper presents TGTV, a framework for detecting and tracing faults in fixed-wing unmanned aerial vehicles. It combines spatiotemporal graph learning with a large language model, using temporal and graph networks plus a Transformer-VAE to flag anomalies and continuous-time Shapley attribution to identify critical sensors and intervals. On the public ALFA dataset with cross-flight holdout, it reports a window-level PR-AUC of 0.976 and a flight-level fault diagnosis F1 score of 0.944.
In briefSame Label, Different Interface: Robot Learning Action Labels Hide Key Differences
SummaryRobot-learning studies often compare position and velocity actions as if the label alone defined the experimental condition. This paper tests five joint-space interfaces on three contact-rich Factory tasks (PegInsert, GearMesh, NutThread) under one shared learner and training budget. It finds that omitted properties such as physical action scale, filtering, and feedback affect results, and that even a calibrated velocity scale leaves residual performance gaps.
In briefThe Prestige: A Benchmark for Visual Reasoning Using Magic Tricks
SummaryThis paper introduces The Prestige, a video understanding benchmark built from professional magic performances that use deliberate misdirection. It tests models on tracking events and explaining unexpected outcomes, with open-ended questions covering temporal and causal reasoning and robustness to adversarial questions. Evaluations of leading vision-language models show a substantial gap between even the strongest model and human performance.
In briefAn AI-driven digital twin framework that optimizes warehouse automation
SummaryThis paper proposes a framework in which a reinforcement learning agent, operating within a FlexSim-based digital twin, optimizes product placement in real time. The agent learns continuously from streaming operational data to reduce picking time and raise throughput. The simulated environment models a four-zone warehouse with forklift handling and uses the Asset Administration Shell framework to link the agent, physical asset, and digital twin. Reported results show a 29.66% drop in average picking time, a 35.42% throughput increase, and an 18.54% reduction in travel distance.
Md Mahinur Alam, Kanita Jerin Tanha, Md Tanjin Alam et al.
In briefTesting Whether LLMs Really Reason Causally: A Case Study on CLadder Benchmark Shortcuts
SummaryThis paper examines what can reliably be concluded when large language models perform well on test sets generated by the CLadder benchmark framework. The authors find that the generator embeds superficial, highly predictive patterns, and even simple random forest models perform well on its data. They also show that many commercial LLMs are brittle on these test sets, with performance barely changing when text critical for formal causal reasoning is removed. The authors argue that strong LLM results most likely reflect exploitation of these surface patterns.
In briefLanguage models' limit calculations and feasibility rejections diverge under mechanical tests
SummaryThe paper tests 12 language models on engineering problems such as beam loads and disk speeds set above or below known plastic collapse or burst limits. Across 4,860 calls and follow-up experiments on new designs, it finds that models can reject impossible designs when asked only about feasibility, yet may accept the same designs and report them as solved when asked for a full answer. Some models accepted beams exceeding collapse limits despite calculating those limits correctly in separate questions.
In briefNemotron 3.5 Super and Lightning: Open, efficient AI models for agent tasks
SummaryThe paper introduces Nemotron 3.5 Super, a 120B-parameter multimodal model with 12B active parameters, and Nemotron 3.5 Lightning, a 30B-parameter text model with 3B active parameters, both built for reasoning, long-context tasks, and agentic workloads. The models combine large-scale distillation, long-context training, and native speculative decoding, and Super adds Indic pretraining and software-engineering mid-training. After post-training with supervised fine-tuning and reinforcement learning, the post-trained Super model reaches 51.78% on SWE-bench Pro, and inference optimizations reach up to 4,856 output tokens/s/GPU on GB200.
NVIDIA: Aastha Jhunjhunwala, Aayushi Ashokkumar Srivastava, Abhibha Gupta et al.
In briefPredicting Which Pages AI Answer Engines Cite, with Open Models and a 4B Student
SummaryThis paper studies how to rank candidate web pages by how often answer engines will cite them in generated responses. The authors evaluate pretrained rerankers, a hosted decision model, eight open decision models, and distilled students on GEO-Bench queries across ten engines. A 4B student model fine-tuned from Qwen3-Reranker-4B performs comparably to the strongest scorers, and the model, benchmark, predictions, and code are released.
WhyThe authors report comparing rerankers and distilled 4B students against teacher decision models on GEO-Bench across ten answer engines.
In briefWhat Firms Can Do About Model Distillation: an economic analysis of distillation attacks and countermeasures.
SummaryThe paper examines how smaller firms can distill training data from conversations with powerful closed-weight AI models to build their own models at much lower cost. Using a two-player model of an incumbent and a distilling competitor, it studies effects on market competition, user surplus, and the incumbent's utility. It finds that penalty-based regulation can fully deter distillation when attacks are detectable, though it may harm competition and users, while self-distillation can mitigate undetectable attacks and help preserve competition.
In briefFrom Next-Day to Multi-Day Wildfire Spread: A New Multi-Continent Active-Fire Dataset for Sequence Forecasting
SummaryThis paper introduces AFSeq, a multimodal dataset of 1,194 fire events from 2023 to 2026 across eight countries in three regions, focusing on short, small fires that existing benchmarks underrepresent. It also presents UTAE-AR, a U-Net-based model with a Temporal Auto Encoder and recurrent modules that forecasts a whole sequence autoregressively by feeding its predicted fire masks back day by day. The authors report that UTAE-AR outperforms prior state-of-the-art methods on next-day accuracy across WSTS, WSTS+ and AFSeq while using about 21.5% of their parameters.
In briefFinding Interpretable Features in Text Using LLM Comparisons and Choice Models
SummaryThe paper proposes a method that has large language models compare pairs of items and explain in interpretable terms what distinguishes them, then uses a statistical choice model to combine these explanations into relevant features and per-item feature scores. On a synthetic hiring dataset of cover letters with known ground-truth features, the method significantly outperforms direct LLM feature extraction in recovering those features. On the ChangeMyView Reddit debate dataset, it discovers several features associated with persuasion while reaching predictive accuracy comparable to established baselines.
Baptiste Arnaudo, Amine Bennouna, S\'ebastien Martin
In briefTesting Monocular Depth Estimation for Lunar Surface Imagery
SummaryThis paper tests how well several state-of-the-art monocular depth estimation methods work on lunar terrain, which lies outside the terrestrial data they were trained on. The authors use the NASA Ames POLAR stereo dataset and Apollo 17 surface images, comparing results against ground-truth physical analogs and historical field data. They report that some of these techniques can generalize to produce high-fidelity relative 3D reconstructions of the unstructured lunar surface.
In briefNeural compression of long ADMM trajectories for solving multiparametric quadratic programs
SummaryThis paper presents TQPNet, a self-supervised neural framework that compresses long ADMM optimization trajectories into a compact neural warm start followed by a few refinement iterations. The network combines a ReLU multilayer perceptron with unrolled ADMM(ReLU) refinement layers and is trained offline using an optimality-condition residual loss, without solver-generated labels. Experiments on large-scale multiparametric quadratic programs and real-time model predictive control show high accuracy, strong generalization, and much faster inference than conventional optimization methods.
In briefPrompt Lottery: How Rewording a Judge Prompt Can Change LLM Evaluation Results
SummaryThis paper studies whether using a large language model as an evaluator gives consistent judgments when the judge prompt is reworded without changing its meaning. The authors name the effect "Prompt Lottery," showing that paraphrased prompts can change evaluation outcomes and, in some settings, reverse model rankings. A user study with 10 human annotators compares agreement under rephrasing, and the authors show that aggregating judgments across multiple prompt variants improves robustness.
Ilija Subasic, Brett Levert, Andrew Rabinovich et al.
In briefFrom Error Prediction to Causal Influence: Auditing SAE Features in Multiple-Choice QA
SummaryThis paper examines whether sparse autoencoder (SAE) features that predict model errors also causally influence answers in multiple-choice question answering. Using Gemma 3 1B and 4B with Gemma Scope SAEs across eight datasets, the authors filter out features tied to prompt and answer-order artifacts before testing causal interventions. SAE activations predict wrongness and uncertainty, and the filtered features shift answer margins more than random features, yet answer flips are rare and effects are inconsistent across meaning-preserving changes. The features therefore appear to reflect context-dependent answer competition rather than the mechanism behind incorrect answers.
Aria Nourbakhsh, Adelaide Danilov, Oleksandr Marchenko Breneur et al.
In briefVision Tomography: Portable 3D Face-Skull Reconstruction from Visual Data
SummaryThis paper introduces Vision Tomography (VT), a framework that reconstructs paired 3D face and skull geometry from visual inputs such as images or sketches. The authors built a dataset of over 200 cases with CT skulls, CT faces, and textured face scans, and proposed a 3D Bidirectional Face-Skull Morphable Model (BFSM) that infers shape between face and skull via a shared coefficient space. The framework is demonstrated in forensic and medical applications, including paired reconstruction and predicting surgical outcomes through joint face-skull editing.
WhyThe authors propose a bidirectional face-skull morphable model and a 200-plus sample dataset for paired 3D reconstruction from visual data, reporting experiments in forensic and medical applications.
In briefCoordinating EV Charging Across Stations with Multi-Agent Reinforcement Learning: QMIX vs. Independent Q-Learning
SummaryThe paper models the assignment of electric vehicles to several charging stations as a decentralized partially observable decision problem and compares QMIX, a centralized-training method, with an independent Q-learning baseline. Tested on a six-station urban network simulated in SUMO under four traffic flow levels, QMIX achieved higher mean reward and acceptance rate. An ablation study suggests the gain comes mainly from the training architecture rather than the cooperative reward terms or experience sharing.
In briefConformal prediction distributions for contextual order-fulfillment optimization
SummaryThis paper builds conformal predictive distributions for delivery time using Conformal Predictive Systems for regression and the cross Venn-Abers predictor for multiclass classification. These distributions feed a contextual optimization model for choosing fulfillment locations and carriers, solved with contextual sample-average approximation. On over five million proprietary order records, CVAP raised exact-day accuracy from 49.5% to 64.1%. In simulator experiments over six test dates, the methods reduced mean total cost by 7.1% and 6.7% and cut the late-delivery rate from 5.7% to below 4.0%.
In briefClassifying Urban Locations by Cyclist Facility Use and Travel Direction
SummaryThe study analyzes footage from motion-triggered cameras at 17 locations in Tempe, Arizona, in 2023 to see how cyclists use bike lanes and sidewalks. It finds that higher bike lane usage is associated with less wrong-way riding (Spearman rho = -0.87). Deployments are classified as bike-lane dominant, sidewalk dominant, or mixed.
In briefPromptSense: A Mental Health Chatbot That Adjusts Its Prompts Based on Detected Emotion
SummaryThis paper proposes a framework that analyzes sentiment and depression before model inference, applies a rule-based filter, and dynamically builds prompts from the results. It aims to make responses from mental health chatbots built on large language models more contextually relevant, empathetic, and risk-aware. The authors compared four model strategies and two prompt-engineered variants on qualitative measures such as empathy and safety, plus toxicity scores from the Perspective API. The prompt engineering approaches outperformed DSFT and RLAIF by 1.8% and 1.1%, respectively.
In briefCan Event Detectors Work Beyond Broadcast Cameras? A Tennis Study
SummaryThe paper releases a tennis benchmark of 345 frame-level annotations of ball bounces and racket hits across eight clips from broadcast, fixed elevated, phone and court-level viewpoints. It compares supervised spotters trained on broadcast footage, two zero-shot vision-language models, and RallyArb, a rule-based arbitration method that fits no parameters on event labels. Supervised spotters perform near-perfectly on broadcast video but poorly on phone and court-level footage.
In briefCARF: A Framework Using Expert-Corrected Reasoning to Screen Scientific Evidence
SummaryThis paper introduces CARF, a human-in-the-loop framework that adapts language models to screening studies for systematic reviews by learning from expert feedback on their generated reasoning. Systematic reviews require experts to manually screen thousands of studies against eligibility criteria. The authors evaluated CARF on reviews of verbal autopsy and cause-of-death methods used in global public health mortality surveillance. CARF showed competitive screening performance with high sensitivity and was more inclined to keep ambiguous, context-dependent studies for human review.
In briefEARS-Info: Information-Aware Event-Driven Reallocation for Multi-Robot Exploration
SummaryThis paper presents EARS-Info, an information-aware extension of Event-Driven Ant Reallocation Search for multi-robot exploration on grids derived from OpenStreetMap. Across 168,000 episodes on eight city maps, it achieved the highest mean operational score in every evaluation panel. With eight robots and 3% communication dropout, it improved the score over EARS-Safe by 5.43% and raised weighted discovery from 88.88% to 98.46%.
Kabeh Mohsenzadegan, Vahid Tavakkoli, Kyandoghere Kyamakya
In briefFaster deep Gaussian process surrogates using mini-batch latent MAP estimation
SummaryThis paper proposes the latent maximum-a-posteriori deep Gaussian process (LaMP-DGP), which uses gradient-based optimization to estimate latent representations and performs probabilistic prediction through linked-GP calculations. To scale to large datasets, the authors develop a mini-batch optimization algorithm and establish a stationarity bound for latent-state optimization. Experiments report predictive performance competitive with several existing inference methods, and substantially faster computation than the Vecchia MCMC approach.
In briefMeasuring Deep Learning Performance: How Results Vary Across Architectures and Tasks
SummaryThis study examines how deep learning model performance varies due to randomness, using 186 experiments with 100 random seeds each, totaling 18,600 runs, across image classification and time series forecasting. Performance distributions are often non-Gaussian, and time series models produce nearly three times more underperforming outliers than image classification models. The authors conclude that mean performance alone is insufficient for model selection, and that spread, symmetry, and tail risk should also be assessed.
In briefIntegral Probability Boundaries: Tighter Non-Linear Bounds Than Integral Probability Metrics
SummaryThis paper introduces Integral Probability Boundaries (IPBs), a non-linear generalization of Integral Probability Metrics (IPMs). IPMs reduce the gap between two distributions to a single worst-case difference in expectations, which can yield loose or vacuous bounds for rare events, safety failures, or under-represented groups. IPBs are designed to remain tight at these extremes. The authors derive their basic properties, present estimation algorithms, and demonstrate use in alignment, judge-free performance estimation, and fairness, including more precise fairness-gap estimates for rare groups in a generative modeling task than prior methods.
In briefLearning from relative action credit to escape local minima in robot reinforcement learning
SummaryThis paper addresses why on-policy methods like PPO often stall on long-horizon, contact-rich robot manipulation tasks. The authors argue that state-value critics judge progress only by episode outcomes, so they cannot distinguish better actions once the policy reaches a partial solution. They extend PPO with an action-conditioned value critic that compares the executed action with alternatives sampled at the same state, combined with Generalized Advantage Estimation and a bound on exploration noise. The extended method escapes the plateau and completes precise peg insertion within a training budget where PPO with identical exploration fails.
Giovanni Minelli, Giulio Turrisi, Lorenzo Amatucci et al.
In briefMetatron: A Promptable Whole-Body Execution Layer for Humanoid Robots
SummaryThis paper presents Metatron, a whole-body control layer for humanoid robots that pairs a generative base policy with a programmable skill layer. The base policy takes partial goals, such as timed body-part targets, velocity commands, and style cues, and produces diverse whole-body motions that match them. After training, the policy is frozen, and skills written as code issue prompts at each control step. On a physical Unitree G1, one frozen checkpoint supports walking, recovery, teleoperated object pickup, and sitting on furniture, with higher reaching success and push robustness than the tested reference-tracking pipelines.
In briefHow Users Explain Audience Change: Causal Attribution and Boundary Management on Xiaohongshu
SummaryThis paper examines Xiaohongshu discussions about a perceived audience influx linked to the World Cup, studying how users explained community change and how those explanations related to boundary-management strategies. The authors analyzed 262 comments from eight high-engagement threads. They found that algorithmic explanations were mostly tied to technical boundary management, while explanations centered on social actors more often aligned with social enforcement or normative socialization.
In briefTesting whether language models can handle fusion physics questions
SummaryThis paper evaluates language models on inertial confinement fusion using a 250-question benchmark covering qualitative reasoning, numerical calculation, symbolic derivation, and equation formulation. It compares general models with Prospero, a model specialized for ICF through domain pretraining and fine-tuning, and tests retrieval-assisted answering. Prospero improved its aggregate score by 10.7 percentage points over its base model, while frontier models scored highest but still made critical physics errors in numerical calculations.
P. B. Radha, D. D. Meyerhofer, M. Bhattarai et al.
In briefDetecting brain tumour progression from wearable accelerometer signatures
SummaryThis study examines patients with high-grade gliomas, using wrist-worn accelerometer recordings of walking converted into path segments. It applies rough path signatures and signature kernel maximum mean discrepancy to measure changes in walking behaviour. Normalising each patient against their own stable-period variation improved separation between stable and progressive disease. The authors present proof-of-concept evidence and call for larger longitudinal studies.
In briefBeyond Aggregate WER: Domain Adaptation and Learner Speech in Spanish Classroom ASR
SummaryThis paper evaluates automatic speech recognition on online L2 Spanish classes where English-speaking learners interact with instructors and tutors. It compares platform, open-weight, adapted, and commercial systems, and adapts Whisper and MOSS with LoRA using human-corrected classroom data. Domain match mattered more than data volume, and learner, English-only, and code-switched speech remained harder to recognize than instructor and Spanish-only speech.
Olga Kellert, Samhitha Harish, Johnatan E Bonilla et al.
In briefWhen Model-Derived Measurements Can Be Corrected or Compared
SummaryThe paper studies numbers computed from trained models, such as example difficulty or benchmark leads, which change with random seed and prompt. It shows that the textbook correlation correction can return values above one and proposes correcting only when both reliabilities pass a significance test, dividing by their upper confidence bounds. It also reports that prompt settings can reverse model rankings and compares models within each prompt setting.
In briefTransferable black-box attacks on MLLMs guided by semantic vulnerability (VISTA)
SummaryThis paper presents VISTA, a black-box framework that finds semantically vulnerable image regions from an multimodal LLM's responses and focuses adversarial perturbations there, without using gradients, logits, or parameters. It offers VISTA-Proxy, which probes an open-source model and transfers examples to commercial models without querying them, and VISTA-Victim, which queries the target model. The authors report attack success rates up to 97% on GPT-5, Gemini-3.5-Flash, and Claude Sonnet 4.6 while keeping PSNR above 30 dB. VISTA-Proxy reaches up to 93% success without querying the final victim.
In briefStatistical mechanics of scale-invariant learning: spectral equations of state and curvature transport
SummaryThis paper studies thermodynamic-like relations in discrete stochastic gradient learning dynamics that follow from scale symmetry rather than equilibrium assumptions. It explains an entropy-scaling relation through an exact rescaling of trajectories and derives a spectral equation of state whose virial coefficients depend on curvature-spectrum moments under a local harmonic approximation. The authors also present an exact curvature-transport identity for quantifying deviations from the ideal-gas description, tested on sphere models, teacher-student networks, and convolutional networks on CIFAR-10 and CIFAR-100.
In briefUsing large language models to accurately extract information from medical charts for stroke research
SummaryThis paper evaluates a prompting-based large language model framework for extracting eight clinical variables from 2,416 acute stroke intervention reports across three hospitals. Twenty-two open-source models were compared, and LLaMA3.3-70B with Chain-of-Thought prompting achieved the highest overall accuracy at 94.8%. The authors conclude that locally deployed LLMs can support scalable clinical data extraction and stroke registry curation, with expert oversight for ambiguous cases.
In briefMaking Smart Homes Suit Autistic Children: Family Work and Monitored Independence
SummaryThis qualitative study examines how general-purpose smart-home technologies affect autistic children and their families, drawing on five caregiver interviews and posts from Xiaohongshu and Weibo. Families adapt environmental outputs, prompts, and interaction channels to create support, while sensory intensity, communication rules, and opaque operating states can turn the same technologies into barriers. Remote monitoring enables conditional independence but creates tensions among safety, privacy, and control.
Tingting Xiang, Weihan Jiang, Hrishitva Patel et al.
In briefFrequency-adaptive Gaussian grids for compact image representation
SummaryThe paper introduces Gaussian-Grid, a hybrid neural image representation that combines structured feature grids with unstructured 2D Gaussians. It divides computation between the two based on local frequency content, adapting its architecture to the complexity of each image. On high-resolution natural images, it reports 1.24 to 2.14 times faster decoding and 23 to 176 times lower PCIe data traffic than JPEG plus ZIP archiving, while keeping memory use small and reconstruction fidelity high.
Anatolii Evdokimov, Stavros Diolatzis, Bingxuan Li et al.
In briefHealthcare Professionals' Views on Algorithmic Bias in Medical AI
SummaryThis paper examines algorithmic bias in machine learning and AI applications within healthcare information systems. Drawing on a survey and interviews with healthcare professionals in different job roles, it studies how bias, perceived fairness, and acceptance of ML outcomes are related. The authors find that opinions on bias causes, assessment criteria, fairness, and mitigation vary by role and algorithm characteristics. The paper concludes that further research is needed to ensure fairness and equality in healthcare.
In brief"It Would Make a Geordie Feel at Home": How Users Perceive and Adapt to Regional Accents in Voice AI
SummaryThis study examines Newcastle English (Geordie) speakers' reactions to a simulated voice assistant speaking either Geordie or Standard Southern British English, and their own speech when voice technology fails to recognise them. The regional voice was linked to familiarity, warmth and local belonging, but also to concerns about professionalism and authority, while the standard voice seemed more neutral and credible. Participants often shifted toward a mainstream accent, enunciated more clearly or spoke more slowly after recognition failures.
In briefAre seismic foundation models more than advanced feature extractors?
SummaryThe paper examines foundation models for seismic waveforms, training nine designs on one Mamba encoder-decoder with a fixed corpus of 3M windows while varying the self-supervised objective, quantization bottleneck, and mixer. Frozen evaluation on three downstream tasks shows that changing the objective redistributes skill rather than raising it, and no checkpoint is best everywhere. The authors argue that the learning targets are defined over single fixed-length windows and miss the event-level occurrence process, so these models behave as feature extractors rather than foundation models.
In briefNew multiple-descent effects caused by heterogeneous data
SummaryThis paper studies how mixing datasets with different covariance structures can produce multiple-descent behavior in linear random-feature models. The authors derive exact asymptotic test-error expressions for each data source in the high-dimensional regime, showing that heterogeneity in covariance scale and spectrum introduces group-dependent interpolation thresholds. These effects appear even with noiseless labels and linear activations. The theory is validated on synthetic Gaussian mixtures and on MNIST and CIFAR-10.
In briefKeeping Safety Intact During Sequential Fine-Tuning: Replay Analysis
SummaryThis paper studies how large language models lose safety alignment when fine-tuned in sequence and how to prevent it. It separates two strategies, supervised replay of refusal and compliance examples and logit distillation from the base model, and tests them on Mistral-7B and Llama-3-8B. Supervised replay cuts attack success most but causes heavy over-refusal, while distillation largely prevents safety erosion at a much lower cost. Geometric analysis finds that preserving the base refusal direction is not a reliable indicator of behavioral safety.
In briefQMUL_AK's Arabic Function-Calling System for AISA-ArabicFC 2026 Using ALLaM with Schema Constraints
SummaryThis paper describes the QMUL_AK team's submission to the AISA-ArabicFC 2026 shared task. The system fine-tunes the Arabic-English ALLaM-7B-Instruct model for function calling using low-rank adaptation (LoRA), combined with an Arabic schema-aware prompt, dialect and no-call resampling, a token-weighted training objective, and deterministic validation of generated calls. The authors report that function selection is nearly saturated, while argument extraction remains the main source of error.
In briefGeneralizing Pretrained All-Atom Representations to Unified Binder Design
SummaryThis paper introduces PairFuse, a parameter-efficient method that fuses pretrained all-atom representations into an independently pretrained geometric latent diffusion model for unified binder design. It uses lightweight layerwise adapters to enrich molecular block features and to incorporate pairwise structural information into the denoiser. The authors report that across antibody, peptide, and small molecule binder-design benchmarks, PairFuse outperforms the pretrained diffusion model and existing representation-fusion baselines.
In briefDeep learning for screening non-transvenous defibrillators to reduce inappropriate shocks
SummaryThis paper studies two non-transvenous implantable defibrillators, the Subcutaneous ICD and the Extravascular ICD, which avoid lead complications but can oversense heart signals and trigger inappropriate shocks. The authors present a 24-hour ambulatory screening framework and introduce AmpRatioNet, a 1D convolutional neural network that operates directly on ECG signals. Evaluated on 4,605 clinician-annotated 10-second segments from 89 patients, AmpRatioNet-SE outperforms the best existing approach on every metric.
Bartosz Morawski, Samuel J. Ward, Alex Touze et al.
In briefGAUSS: A Workload-Aware Simulator Predicting LLM Inference Latency and Capacity
SummaryThis paper presents GAUSS, a workload-aware analytical and Monte-Carlo simulator for LLM inference serving. It predicts the distributions of time-to-first-token and time-between-tokens, along with sustainable request rate, for a given workload, engine, and accelerator. GAUSS combines phase-specific latency models for prefill and decode with a Lindley-recurrence queueing model, using ensemble Monte-Carlo sampling to capture latency tails that trace-driven simulators under-sample. The authors show it supports operating-point selection, traffic segmentation, and shortest-remaining-processing-time scheduling analysis.
Srikant Bharadwaj, Ankur Mallick, Spyridon Mastorakis et al.
In briefPredicting sparse flow measurements with coarse-mesh graph neural networks
SummaryThis paper proposes reduced-order models that use graph neural networks on coarse meshes to predict sparse probe locations and integrated quantities, avoiding full flow-field reconstruction. The models are applied to parameterized 2D and 3D confined cylinder flows and an unevenly heated cavity, and they generalize to unseen conditions. Coarse-mesh models match or outperform finer-mesh surrogates while reducing training time and memory.
Brener Ramos, William R. Wolf, Tarc\'isio C. D. Oliveira et al.
In briefWho Else Is In? How AI Agents Might Start Secret Collusion Using Vague Messages
SummaryThis paper studies how multi-agent LLM systems could begin collusion, with agents using vague public messages to find hidden partners while appearing to do ordinary work. The authors run a mutual identification game in which two agents must recognize each other while solving a shared task, and some models succeed without explicit guidance, often evading monitors at low false positive rates. They report that as messages grow more indirect, monitor punishment can fall faster than partner coordination, creating a profitable middle ground.
WhyThe authors introduce a mutual identification game showing strategic vagueness can initiate collusion past monitors; the abstract reports successful handshakes and an asymmetric monitor-punishment ver
Pramod Kaushik, Sarath Sivaprasad, Laura Ferrarotti et al.
In briefA hierarchical diffusion method for camera-controlled long video generation.
SummaryThis paper addresses generating long videos from a single image under camera control. It proposes a lightweight framework combining autoregressive generation (WanAR) with hierarchical long-context denoising (FixAR), and reformulates video generation as a warping-and-inpainting problem to keep scene geometry consistent with input cameras. Experiments on RealEstate10k and DL3DV show the method outperforms prior approaches, producing results that are temporally stable and consistent with input camera trajectories.
In briefSeparability Is Not Alignability: Representations for Spectrum-to-Molecule Retrieval
SummaryThis paper studies identifying small molecules from tandem mass spectra, a task limited by scarce paired data. It examines aligning separately pretrained spectrum and molecule encoders, testing this approach on MassSpecGym. The authors find that the objectives training spectrum encoders never account for molecular structure, so the distance metric needed for alignment is not guaranteed, and a candidate-based objective rewards a spectrum-independent prior. They conclude that simulated spectra outperform frozen geometry as a remedy for data scarcity.
Karol Musielinski, Fanurs C.E. Teh, Takahiko Koyama et al.
In briefSolo movement style predicts group foraging, but whom to follow appears only in groups
SummaryThe study uses three public datasets of human collective foraging, with 621, 128 and 1,133 participants, to test whether solo behaviour predicts behaviour and performance in a group. Movement measured while alone predicted behaviour inside the group, and group performance was predicted from individually measured performance. Selective social learning, following those who succeeded or stopped, showed stable individual differences within groups but was not predicted from the same person's solo behaviour.
In briefSometimes less is more: how MLPs and Transformers learn noisy k-parity features differently
SummaryThis paper compares how MLPs and Transformers learn from synthetic datasets built with noisy k-parity features, which allow precise control over each feature's predictivity and availability. Transformers handle noisy features more robustly but struggle with complex, low-availability ones, while MLPs show the opposite pattern and keep improving with more data. Under a fixed training budget, adding more unique training examples can slow Transformer optimisation enough to lower final accuracy, so a smaller dataset may perform better.
In briefFORGE: Online reference generation and full-body control for agile humanoid robot movement
SummaryThis paper proposes a two-stage learning framework for agile, versatile humanoid locomotion. The first stage trains a command-conditioned whole-body reference generator using single rigid body dynamics and full humanoid kinematics. The second stage runs the frozen reduced-order policy online to produce references for training a whole-body tracking policy on the full-order model. Demonstrated on a Unitree G1 robot, the approach reaches longitudinal speeds from -2.0 to 2.8 m/s, lateral speeds up to ±1.0 m/s, and one-legged hopping.
In briefMerge and Branch: efficient tree-structured fine-tuning of LLMs for recommendation
SummaryThis paper studies LLM-based generative recommendation, where models read long interaction histories and generate short semantic identifiers. It proposes Merge and Branch (M&B), which compresses each historical item into a few learned tokens and uses this compressed history to predict multiple target identifiers arranged in a shallow, wide tree. The method introduces tree FlashAttention and a tree-structured key/value cache to share history computation. Experiments on public benchmarks report that M&B outperforms strong recommendation baselines while improving training and decoding efficiency.
In briefApplying PDE emulators to a multi-material, multi-physics simulation problem
SummaryThis paper tests several computer vision-based machine learning architectures as surrogate models for a computational physics task involving multiple materials. It finds significant implementation difficulties in adapting some pre-trained models, including transformer-based ones. The authors argue that multi-physics, multi-material datasets call for learning methods suited to sparse and imbalanced data, adapted spatial encodings and physical consistency metrics, and lower-cost ways to generate training data.
Diane Oyen, Morgan Jones, Adithan Kathirgamaraju et al.
In briefMeta-initialization to speed up deformable medical image registration with implicit neural representations
SummaryThis paper addresses the slow per-image-pair adaptation of implicit neural representation (INR) methods for deformable medical image registration. It examines whether initializations learned from previously solved registration tasks, via Reptile meta-initialization and STRAINER transfer of a coordinate network's shared component, can accelerate early adaptation, using KAN-IDIR as the base model on OASIS-1. Both schemes achieve higher Dice and lower HD95 at early iterations than random initialization, with Reptile numerically better under a small iteration budget though pretraining budgets differ. The gap between the schemes narrows as iterations increase.
In briefTime-series world model conditioned on self-supervised latent actions
SummaryThis paper proposes ACTS-WM, a framework for time-series analysis that infers the information behind changes between observations as latent actions, without requiring explicit action annotations. It combines a JEPA-based encoder with an action-conditioned world model that predicts target time-series representations. The method is evaluated on 22 datasets from diverse domains across four tasks, including univariate and multivariate forecasting and classification. The authors report that it improves downstream performance over pretrained JEPA-based encoders by learning complementary representations.
In briefStatistical tests for position bias in LLM-based recommenders
SummaryThis paper studies positional bias in LLM-based recommender systems, where the recommended item depends on the order in which candidates appear in the prompt. The authors propose two statistical detection methods and run experiments on a real product catalog with 40,000 instances. Both methods find significant positional bias in all evaluated state-of-the-art models, which the authors describe as structural and model-dependent. They recommend routine auditing and controls such as order randomization.
Carlos Pel\'aez-Gonz\'alez, Francisco Herrera, Andr\'es Herrera-Poyatos
In briefSAG-Sep: Sparse Augmented Graphs and Onion-Guided Search for Rounded Capacity Cut Separation
SummaryThis paper presents SAG-Sep, a neural separator for rounded capacity cuts (RCCs) in capacitated vehicle-routing relaxations. It encodes a fractional LP solution as a sparse augmented graph and scores customer membership in one encoder pass. A training-free Onion-Sweep decoder then turns these scores into nested RCC candidates. On the N=1000 NeuralSEP benchmark, SAG-Sep with Onion-Sweep reduces the average root gap by 17.7% over the strongest NeuralSEP baseline, and the single-pass separator is roughly 30 times faster per iteration.
In briefFederated learning for semantic communication over shared wireless channels
SummaryThis paper proposes a federated learning framework for semantic communication over a multiple access channel, where several distributed transmitters send learned semantic information to one receiver at the same time. Feature extraction, compression, wireless transmission, and semantic inference are optimized jointly. Evaluated on CIFAR-10 over fading channels, the approach keeps high inference accuracy and improves performance under severe data imbalance, while reducing training overhead.
