In briefAutoAssumption: a method that searches over assumptions to find machine learning systems algorithms
SummaryThe paper formalizes 89 algorithms from KV-cache eviction, weight pruning and speculative decoding as proof graphs with explicit assumptions, checked in Lean. Its harness, AutoAssumption, relaxes one assumption of a published method, re-derives and re-proves it, and generates code. In KV-cache eviction and weight pruning, it improved on every published algorithm it ran and produced state-of-the-art algorithms.
In briefSPIRE: Robot control that learns continuously by inferring situations and reusing motor skills without task labels
SummaryThe paper introduces SPIRE, a recurrent spiking system that infers latent contexts from action-outcome evidence and learns context-specific motor actors online. Mature actors become protected sources that combine with a target-context residual to drive motor populations. In coordinate, three-joint and hidden-force control tasks, SPIRE maintained control, inferred situations autonomously, retained prior behavior, and accelerated learning of an unseen source composition. The authors report it achieved 2.4 times the mean prior-learning advantage of the strongest eligible task-free alternative.
In briefSophon: A Benchmark Diagnosing Spatial Reasoning Failures in Vision-Language Models
SummaryThis paper introduces Sophon, an open-source diagnostic built from procedurally generated colored point-cloud images across three coordinate systems and four question types. Open-weight vision-language models score 47-52%, while frontier models reach 77-83% and humans 84-94%. Probing and fine-tuning experiments attribute the gap to a routing bottleneck in mid-stack layers rather than a missing capability, and targeted LoRA fine-tuning raises 7B-class models to frontier-level accuracy.
In briefLearning fast action checks for frozen robot vision-language-action models
SummaryThe paper presents HDH, a lightweight framework that uses experience from a frozen vision-language-action (VLA) model's own deployments to improve how its actions are verified. It first deploys the policy to collect autonomous rollouts, then trains a compact verifier that gives each action chunk a predicted task-success score. The frozen VLA generates candidate actions, and the verifier executes the highest-scoring one. HDH reports an 82.3% average success rate across four LIBERO-10 tasks and a 22.6% average success improvement over the baseline on three real-world manipulation tasks.
Wen-Han Hsieh, Ken Jiankun Zheng, Chia-Chien Chen et al.
In briefDo people trust AI assistants even when they are consistently wrong? A study of reliance in multi-turn conversations
SummaryThis paper examines how much people rely on advice from large language models across repeated conversations, comparing assistants that are always accurate, always inaccurate, or accurate half the time. Across four experiments, participants learned to trust accurate models more only when they received external feedback on accuracy. Without such feedback, a consistently inaccurate model still persuaded participants 48% of the time.
Cameron R. Jones, Julia Wong, Aaron S. Cohen et al.
In briefNeural Compression for Scientific Particle Data: How Learning Might Help
SummaryAs particle-based simulations grow larger, efficient compression of scientific data becomes critical. The paper argues that scientific particle data differs from computer vision point clouds in fractal dimension, since vision point clouds are locally planar while particles are not, which limits the octree structure behind most point cloud compression methods. Evaluations across cosmology, molecular dynamics, and plasma physics show that neural codecs achieve only marginal bitrate reductions over the traditional G-PCC baseline. The paper outlines where the structure of scientific data limits current neural approaches and where learning may help in the future.
In briefHow AI Language Models Could Make Teaching More Uniform
SummaryThis paper examines whether large language models produce less varied educational content, testing math item writing and essay feedback. It compares direct prompting, two diversity-prompting methods, and post-training against human-written reference content. Directly prompted outputs were markedly less diverse than human content, 78% higher on lexical similarity and 51% higher on semantic similarity, and covered fewer instructional forms. The authors find that diversity methods mostly change surface features rather than broadening instructional approaches, and models amplify the most common routines of human teaching.
In briefDischarge summary generation is more than summarization: a reality check on evaluation
SummaryDischarge summaries include both a review of the hospital stay and decisions made at discharge, the latter of which may lack support in the prior clinical notes. The authors introduce FairSumEval, which checks fact by fact whether reference and generated summaries are backed by the model's evidence, and computes two agreement metrics, Soft IoU-S and Soft IoU-NS. On 200 MIMIC-III admissions, overlap is much higher for evidence-supported content than for content without identified support, and conventional metrics do not capture both dimensions consistently.
Radhika Dua, Frederick Wolff, Kyunghyun Cho et al.
In briefFrom Classification to Generation: Vision-Language Models for Roman Coin Description
SummaryThe paper recasts the cataloguing of Roman Imperial coins as an image-to-text generation task rather than a classification task. The authors built about 30,000 image-text pairs from auction records and trained separate low-rank adapters on the open-weight Qwen3-VL-8B-Instruct model for four tasks: transcribing obverse and reverse legends and describing obverse and reverse imagery. The adapted models outperformed both the base model and GPT-4.1 on all four tasks, cutting legend transcription normalized edit distance by over 40% relative to GPT-4.1. Gains were smaller but still present on legends not seen in training.
In briefWhat Tokenizer Comparisons Really Measure: Granularity, Root Assignment, and Word-Level Readout in Amharic and Hausa
SummaryThis paper examines whether gains from morphology-aware tokenizers over byte-pair encoding come from their linguistic units. The authors use a lossless root-and-pattern tokenizer and controls that change one factor at a time, separating granularity, root assignment, and word-level readout in Amharic and Hausa. Real root assignment improves Amharic named-entity recognition over scrambled roots by 4.9 and 4.7 F1, while readout choice shifts F1 by up to 13.4 points in small models and can reverse comparisons. The authors recommend decomposing tokenizer effects and reporting the readout before attributing gains to linguistic structure.
In briefBirdDreams: A Driving World Model Generating Perspective and Bird's-Eye-View Video Together
SummaryBirdDreams is an autoregressive driving world model that jointly generates perspective video and a schematic bird's-eye-view video, making scene layout part of the predicted future. The BEV video shows roads and vehicles as simple shapes at a fixed metric scale, while perspective video keeps details such as traffic signals and object appearance. Ego control is achieved by translating and rotating the predicted BEV along an input trajectory. On PhysicalAI-AV and nuScenes, it follows input ego trajectories more faithfully than prior models, and human evaluation rates its surrounding-vehicle behavior as more realistic than a perspective-only baseline.
In briefTighter last-iterate convergence limits for stochastic gradient descent (SGD), with and without replacement sampling
SummaryThis paper studies last-iterate convergence bounds for stochastic gradient descent (SGD) on convex, β-smooth loss functions. In the interpolation regime, it improves the bound from O((log T)^3/T) to O((log T)^3/(T log log T)). The improvement comes from applying a horizon-dependent multiscale estimate of gradients when comparing trajectories. The bound holds both for sampling with replacement and for sampling without replacement within a single pass.
In briefGenPFN: A Training-Free Tabular Foundation Model Method for Expanding Small Engineering Datasets
SummaryThis paper introduces GenPFN, a training-free method that generates candidate engineering designs using a fixed pretrained TabPFN classifier guided by labeled desired and undesired designs. Generated candidates are evaluated against the problem's constraints and objectives, and the desired ones are added to the data for regression training. Across fourteen problems, GenPFN achieves higher class satisfaction and lower realism distance than several compared generators. Augmentation reduces estimated AutoGluon validation mean squared error by 62.1% relative to context-only training in observed paired fits.
In briefDistributionally robust BEST-RQ for learning speech representations
SummaryBEST-RQ is a self-supervised method that learns speech representations from large amounts of unlabeled audio using random projection and quantization. This paper proposes a distributionally robust extension that treats each codebook as a group and trains with group distributionally robust optimization, reweighting groups to emphasize poorly performing ones. A Kullback-Leibler divergence regularizer limits excessive distribution deviation. Evaluated on Multilingual LibriSpeech, the approach outperforms the conventional uniform average across multiple codebooks.
Muhammad Khan, Xiaodong Cui, Brian Kingsbury et al.
In briefWhen Proof-Step Verification Is Worth Its Cost: Operating Conditions for Certify-or-Decline Checking in Open-Weight Math Reasoning
SummaryThe paper evaluates a pipeline that rewrites a solver's answer into typed proof steps, then uses LLM judges to certify or decline them, testing open-weight models on GPQA, AIME-2025, and a MATH-500 subset. Certification raised answer-key-match precision by 31 points on a strong GPQA setup and 46 points on AIME, but answered only 35% and 21% of attempts. On near-ceiling MATH-500, gains were at most 5 points with a coverage loss of 13 to 27 points.
Abhinav Gupta, Ben Slivinski, Yashica Patodia et al.
In briefRobust state estimation from pretrained perception models that stays reliable under uncertainty
SummaryThe paper addresses uncertainty in state estimation by pretrained perception models such as Vision-Language Models, which often reduce predictions to a single best guess. It proposes a Minimal Belief Set that keeps the smallest group of likely states whose cumulative probability meets a set threshold, built from the model's own output probabilities. An algorithm called PRaVaSS constructs this set with correctness guarantees and polynomial complexity in the unconstrained case. Experiments with a VLM-based estimator show improved downstream decision-making in a task-planning pipeline.
In briefRE-PRESENT: Combining evidence from re-presented questions to make LLM ensembles more efficient
SummaryThis paper studies how to get reliable answers from an ensemble of different large language models. It finds that a simple single-round aggregation rule performs competitively with learned routing methods, but single-round responses offer limited power to identify reliable predictions. The authors introduce RE-PRESENT, a training-free framework that selectively re-presents disputed questions and aggregates candidate-level evidence across alternative views. On seven multiple-choice benchmarks and two model ensembles, it improves over strong single-round aggregation and offers a better accuracy-cost trade-off than self-consistency and multi-model debate.
In briefMach2: A Science Discovery Platform That Enforces Deterministic Execution Rules
SummaryThis paper presents Mach2, a platform that uses deterministic, lint-like checks to enforce operating rules on language-model agents' tool requests, source changes, acceptance tests, and artifact contracts. Domain requirements are encoded as executable predicates, and mandatory failures block the checked operation from advancing. The platform also combines test-driven development, AST analysis, and history-guided evolutionary search, and generates task-specific agent packages that run on independently chosen endpoints. The paper proposes measurable criteria for rule compliance, engineering validity, and cross-endpoint task continuity.
In briefSlipStream: A Prefetcher That Makes High-Bandwidth Flash Practical for GPU-Based LLM Inference
SummaryThe paper addresses the memory demands of large language model inference by considering High-Bandwidth Flash (HBF), which replaces HBM's DRAM with NAND flash for up to 16x more capacity but with roughly 20x higher read latency. It proposes SlipStream, a software-hardware co-designed prefetcher that treats prefetching as a demand-paced producer-consumer pipeline using page-level tracking and a staging warp. Across seven dense models spanning 16 to 800 GB, SlipStream with a 40 MB buffer reaches 92.4% of ideal HBF performance, compared with 43.1% for H3 even with a 320 MB buffer. It also supports Mixture-of-Experts models, reaching 91.2% of ideal performance.
Anish Saxena, Hritvik Taneja, Christos Kozyrakis et al.
In briefAdapting a Frozen LLM into a Word-Level Full-Duplex Spoken Dialogue System with Little Data
SummaryThis paper shows that a frozen general-purpose multimodal LLM can be adapted into a word-level full-duplex spoken dialogue system by training only lightweight adapters and a few new token embeddings (97M parameters). The system works on an 80 ms frame clock, emitting one token per frame, with pauses, yields and onsets encoded as tokens kept valid by a finite-state machine. A four-stage curriculum performs the conversion in about 30 GPU-hours, and the talker component is trained independently on the thinker's discrete tokens. The authors report that the system is competitive with open-source full-duplex systems trained on large-scale speech data on Full-Duplex-Bench and InteractionBench.
Tairan Chen, Richard Yucheng He, Yihang Liu et al.
In briefLeaderboard Hacking: Preference-Based Model Evaluations Can Be Manipulated
SummaryThis paper shows that arena-style language model leaderboards, which rank models through pairwise votes aggregated with the Bradley-Terry model, are vulnerable to manipulation. It examines three strategies: strategic voting, strategic prompting, and strategic nomination of irrelevant models. The authors demonstrate these effects on public leaderboard data and report that each attack can raise a model by up to 15-20 places, and up to 45 places combined.
Eve Fleisig, Joachim Baumann, Dylan Hadfield-Menell et al.
In briefHow Humans and AI Use Timing to Coordinate Socially in a Cooperative Card Game
SummaryThis study uses the cooperative card game The Mind, which forbids explicit communication, to test whether players can coordinate using only the passage of time. It evaluates 150 humans and two kinds of AI models: computer-use agents and text-prompted LLMs. Human behavior is best captured by a Bayesian strategy that infers a partner's belief from their timing, and frontier models show a similar pattern. However, none of the tested models adapted to their partners over the course of the game as efficiently as humans did.
In briefSynthCorp: A Platform for Building Benchmarks for Data Agents
SummaryThis paper presents SynthCorp, a platform that simulates how business events create and change corporate data over time to build customizable data estates and benchmarks. Each data estate includes relational databases, a document store, and a data catalog, along with evaluation tasks and executable gold programs. The authors used it to generate initial benchmarks with seven data estates and 256 tasks, and evaluated three models across four data-context settings.
In briefRECALL: Real-time recalibration of multi-camera rigs using learned factory-anchored corrections
SummaryRECALL is a feed-forward method that recalibrates a movable multi-camera rig in a single pass, using the factory calibration as an anchor and a brief observation of a known object. It represents each camera's calibration as a shared rig rotation plus a small per-camera drift, refined by a learned recurrent operator with differentiable reprojection and cross-camera attention. On a synthetic 10-camera benchmark it is more accurate than per-camera and joint bundle adjustment and runs in about 0.12 ms per rig.
Mario Alberto Tapia de la Cruz, Charalambos Poullis
In briefGraph Label Alignment: A Diagnostic Atlas for Graph Classification
SummaryThis paper proposes graph label alignment, a lightweight diagnostic that measures how accessible labels are without training a graph encoder, using two probes: structural label alignment (SLA), based on permutation-invariant topological and spectral summaries, and feature label alignment (FLA), based on pooled node features with edges removed. Across real graph classification benchmarks, the SLA–FLA plane strongly organizes test accuracy, and low-alignment datasets are where standard pooled pipelines struggle. The authors also introduce controlled interaction-dominated stress tests to probe beyond marginal accessibility.
Neelam Akula, Varun Shiralkar, Murat Kantarcioglu et al.
In briefA multilingual benchmark for testing how well vision-language models understand documents
SummaryThe paper introduces MDU, a benchmark of 23,146 examples covering 13 document understanding tasks in five languages: English, French, German, Italian, and Spanish. It evaluates 45 open-weight vision-language models (1.7B to 235B parameters) and one commercial system under the same protocol. The best average score is 69.0, and only eight systems reach 65, with the top two results coming from 27B models. The authors report that the gap between the strongest and weakest languages does not shrink with scale, and that per-task abilities are largely uncorrelated.
Matteo Gabburo, Daniela Baiamonte, Stefano Simonazzi et al.
In briefFrom Nodes to Pixels: Two-View Graph Imaging for Graph Learning
SummaryThis paper introduces G2Image, a framework that converts each graph into two complementary image-like views. TopoGrid encodes an intrinsic bifiltration as a compact multipersistence surface grid for a topological view, while GraphGrid aggregates edge densities between node groups for a structural view. Both views are permutation-invariant and come with resolution-controlled stability guarantees, and they are aligned via a supervised cross-view contrastive objective. The authors report strong average performance on graph classification and molecular prediction benchmarks.
Md Joshem Uddin, Soham Changani, Sai Karthik Navuluru et al.
In briefMatching Distributions Despite Unknown Measurement Noise Using a Small Audit Sample
SummaryMachine learning systems often work with noisy measurements, which can make standard distribution comparison tools like maximum mean discrepancy (MMD) biased. This paper proposes MEC-MMD, an audit-based estimator that corrects kernel evaluations using a small subset of ground-truth measurements and does not require a noise model. The authors show it is exactly unbiased for the oracle MMD under simple random auditing and is consistent for the population MMD at the standard rate when the audit size scales with the dataset. Experiments on parameter estimation and unsupervised tasks report that MEC-MMD outperforms MMD and other alternatives.
In briefUniPipe: A unified system for serving and fine-tuning neural network surrogates under strict deadlines
SummaryThis paper presents UniPipe, a system that runs inference serving and fine-tuning of a deep neural network surrogate in a single context on one edge GPU. It targets scientific workflows, such as Ptychographic imaging reconstruction, that must process high-rate sensor data within hard per-unit deadlines. UniPipe combines training and inference samples in one forward pass and uses a polynomial-time scheduler to choose batch composition. On a real Ptychographic workflow, it reduces reconstruction MSE by up to 30% and inference MSE by up to 50% versus a pretrained baseline, and cuts deadline-miss rate by 13%-22% versus a multi-context baseline.
Jaiaid Mobin, Avinash Maurya, M. Mustafa Rafique et al.
In briefWhen semantic priors should guide partially identified causal discovery
SummaryThis paper studies when and how semantic information from large language models should be used in observational causal discovery. It separates three decisions: where to allocate semantic queries, which proposed orientations to admit, and which causal graphs to retain. Experiments on benchmark networks show semantic orientations outperform coin-flip baselines, but targeting unstable edges shows no clear advantage over matched-random allocation. The tested admission certificate could not support nontrivial admission with the available calibration tasks.
Muhammad Ahmed Mohsin, Saad Bin Ather, Valentin Six et al.
In briefSyntheIce: A Synthetic Benchmark for Reconstructing Ice Layers in Radargrams
SummaryThis paper addresses the problem of extracting englacial layers, which record ice history, from radio-echo-sounding radargrams, where automatic detectors often produce fragmented horizons. The authors propose a slope-aware greedy algorithm that links partial fragments into full-width layers and assigns each reconstruction a content-based hash ID. They release SyntheIce, a synthetic dataset of 600 radargrams with exact ground-truth annotations, and extend the XCheck consistency framework to four consecutive tiles. Evaluation against expert-labeled real radargrams reports high precision, recall, F1-score, and mean IoU.
Muhammad Behroze Hassan, Bayu Adhi Tama, Sanjay Purushotham et al.
In briefFixing one unfairness can worsen another: spillover effects of fairness interventions across sensitive attributes
SummaryThis paper studies how correcting a classification or regression model for unfairness with respect to one sensitive attribute affects fairness with respect to other attributes, including ones the practitioner cannot observe. The authors use Wasserstein barycentric disparities to build a distributional framework that tracks both marginal and joint fairness measures. In a jointly Gaussian model they derive explicit disparity trajectories and sharp partial-identification bounds for unavailable attributes. Experiments show that interventions often increase unfairness with respect to the hidden attribute, and on ACSIncome, correcting for race reduces sex disparities while subsequently correcting for age increases them.
Fernandes Machado Agathe, Jamotton Charlotte, Charpentier Arthur
In briefRepresenting Relationships Beyond Foreign-Key Links in Relational Deep Learning
SummaryThis paper proposes a conceptual framework for building relationship-aware representations from relational databases, distinguishing schema-defined relationship semantics, association entities and relationship-level attributes, and contextual information about a connection. It introduces a Relationship Representation Construction (RRC) protocol that maps schema elements and records to node representations, relationship descriptors, and contextual representations. The representations can be incorporated into existing relational graph learning architectures through relationship-conditioned message passing. The paper also discusses the framework's foundations, illustrates it on a multi-table database, and outlines research questions for evaluating its benefits.
In briefCode as Explanations: Verifiable Code for Explaining LLM Behavior
SummaryThis paper proposes using executable programs to explain how large language models behave across a whole dataset of inputs, so the explanation can be checked automatically by running it. The authors define a Consistency metric, the fraction of inputs where the program reproduces the model's outputs, and use it to select candidate programs. Tested on three open-source code-capable models and eight tasks, code explanations reached high consistency in several settings and were 6 to 100 times shorter than chain-of-thought, but consistency was much lower on tasks with more complex language.
In briefTesting “What-If” Users Before Launch: Persona Simulation for Real Decisions
SummaryThis paper presents a framework and infrastructure that turns real industry questions into controlled experiments with simulated users, formally defining target cohorts, stimuli, and decision metrics. It supports large-scale concurrent runs across survey, web, chatbot, and OS app environments, with auditable logging and statistical reporting. Across twelve industry partner cases in five domains, the authors show that persona attributes materially shape simulation outcomes, and that the method yields directional evidence for design decisions.
In briefEvaluating and learning how to combine cell signals into reusable table embeddings
SummaryThe paper introduces a benchmark that measures table-level embedding quality through consistency under partial views, discriminability, robustness to perturbations, and efficiency. Among twelve methods, lightweight schema- and text-based fingerprints performed strongest, while tabular foundation models lagged under simple pooling. The authors propose SCDM, a self-supervised loss that learns only the compositional rule over a frozen ConTextTab encoder, raising correlation with partial-view overlap and improving retrieval and clustering.
Ali Younes, Saeed Ghoorchian, Maximilian Schambach et al.
In briefRisk-aware joint inference and resource control for AI devices in 6G networks
SummaryThis paper proposes RACE, a method for edge inference on Physical-AI devices in 6G networks. It sets adaptive signal-to-interference-plus-noise ratio (SINR) targets based on predicted action-failure probability and task criticality, combined with deadline-derived SINR requirements. Simulations report an 11% improvement in the average accuracy-latency-energy objective over fixed high-SINR control, with near-zero constraint outage.
In briefWhat Entropy-Based View Selection Misses: Auditing Test-Time Adaptation for Vision-Language Models
SummaryThis paper examines test-time adaptation for vision-language models such as CLIP, which create multiple augmented views of each test image and select which views to trust. The authors find that view selection accounts for nearly all accuracy gains, while entropy-based selection ranks correct views no better than chance when predictions are wrong. They propose ranking views by the fraction of the original image each crop retains, which improves accuracy on 14 of 15 benchmark datasets over entropy.
In briefTrait Hitchhiking: How RL Post-Training Can Shift Unrewarded Traits in Language Models
SummaryThis paper studies how traits evolve during reinforcement learning post-training of language models, using a formalism that recovers the Price equation from evolutionary biology. It models "trait hitchhiking," in which non-targeted traits are amplified or suppressed as training redistributes probability across completions. The authors evaluate this model on language models trained with GRPO and find that the Price estimates closely track sampled trait dynamics. The paper suggests this mechanism can silently amplify undesirable traits during RL post-training.
In briefTogether Compiler: A Continual Compiler for AI Inference Optimization
SummaryThis paper presents Together Compiler (TKCC), a continual agent loop that optimizes AI inference as workloads and models change. It introduces a Tiered Persistent Engine that separates engine launch overhead from forward pass computation, along with a cost model that selects parallelism configurations to seed optimization. The authors report higher throughput than hand-written Megakernels and leading open-source engines across several frontier models and hardware platforms.
In briefFast slice-wise deep learning for correcting motion in spinal cord fMRI
SummaryThe paper proposes a lightweight convolutional neural network that estimates the in-plane shift of each axial slice to align it with a reference, correcting respiratory and bulk motion in spinal cord functional MRI. Two models are trained without motion labels. On 3T and 7T datasets, the method outperforms the Spinal Cord Toolbox (SCT) on all reported metrics and runs in about 10 seconds per scan, roughly 30 times faster than SCT.
Thomas Dagonneau, Sandrine B\'edard, Nontharat Tucksinapinunchai et al.
In briefMITRA-Swarm: A Multi-Agent Method Adding Explicit Linguistic Structure to Foundation Models for Classical Sanskrit Translation
SummaryThe paper addresses three difficulties in translating classical Sanskrit: Sandhi sound fusion, dense compounding, and zero-anaphora forced by poetic meter. It proposes MITRA-Swarm, a multi-agent pipeline with a Sandhi-splitting agent, parallel generalist and specialist translators, and a discourse-aware ellipsis restorer. On a 1,000-verse stratified benchmark, the full system scores 85.90 MQM against 85.27 for a monolithic baseline, with reported statistical significance and ablations showing the gains are small.
Kayshav Bhardwaj, Pragun Seth, Kush Bhardwaj et al.
In briefWhere Should the Code Go? Locating Foundation-Model Errors in Bitemporal Inventory Simulation
SummaryThe paper studies how foundation models should handle inventory simulations where events and their reports carry separate timestamps, and reports can be revised later. Using an exactly solvable lost-sales inventory environment, the authors rescored 36,672 requests across twelve closed and open-weight model configurations to compare remedies such as canonical input, exact code before or inside the model call, and more reasoning. They find that weak models fail in separable ways, strong models are near ceiling on raw reports, and the tool-calling interface can confound reasoning evaluation for some models.
In briefOutcome or learning? How students use AI code assistants to write unit tests
SummaryThis paper reports a user study on how students use AI code assistants when writing unit tests with unrestricted access. Most students used a one-shot strategy, asking the AI to generate a complete unit test directly from the initial requirements. The authors argue this focuses students on completing the task rather than learning to write effective tests, and that students rarely critically evaluate or refine the output. The paper discusses how training, accountability, and learning task design could support more learning-centric, responsible AI use.
Daniel Hinbest, Jeremy S. Bradbury, Michael A. Miljanovic
In briefAbstention Is Not Threshold Compliance: Testing Whether Language Models Follow Stated Scoring Rules
SummaryThis paper examines whether open language models follow a scoring rule stated in the prompt, under which they should answer only when their probability of being correct exceeds a threshold t. It tests 12 checkpoints from four fully open 3-8B recipes, plus exploratory models up to 32B. Adding an abstention-format line raised abstention sharply, but changing the threshold had little effect within that format. The authors conclude that a high abstention rate does not show threshold compliance, and that evaluations should report abstention rates, threshold responsiveness, and which items are answered.
In briefClosing the Loop: Co-Evolving EM for Generating Regular Series from Irregular Time Series
SummaryThis paper addresses generating regular time series from irregular ones, arguing that the common approach of imputing missing values once and then freezing the completions is structurally flawed. The authors propose a co-evolving Monte Carlo EM framework that alternates between sampling missing values from the current generator's posterior and retraining the generator on those completions. They also introduce Posterior Sampling for Lifted Representations (PSLR) to handle conditioning between the lifted diffusion prior and time-series observations. The paper reports state-of-the-art results across ten datasets and four corruption regimes.
WhyThe authors replace one-shot imputation with a co-evolving Monte Carlo EM loop and PSLR, reporting gains over the strongest baselines across ten datasets and four corruption regimes with ablations.
In briefThompson sampling with diffusion transformers pre-fitted to priors
SummaryThe paper introduces prior-fitted diffusion transformers, diffusion models pre-trained with the prior-fitted network paradigm that perform Thompson sampling in-context at inference time. The authors develop practical non-Gaussian priors for general black-box functions, including priors over unimodal functions. Across global optimization benchmarks, models using these non-Gaussian priors outperform traditional Gaussian process-based pipelines.
In briefCleaning an Agent's History Can Hurt When the World Still Holds the Error
SummaryThis paper studies tool-using language-model agents whose histories of mistakes are compacted, repaired, or handed off while their effects remain in the world. Experiments on AgentCE-Bench and AppWorld freeze an agent mid-task, edit either the history or the world state, and let a receiver model continue. The results show that the same correct content can lower performance when the world has not been updated to match. The authors conclude that a history edit's effect depends on the world it describes.
In briefChoosing several candidates in random order when value predictions are available
SummaryThis paper studies the multiple-choice secretary problem, where candidates arrive in random order and at most k must be selected to maximize total value. Predictions of each candidate's value are available before arrivals begin. By retuning the threshold of an existing switching algorithm and analyzing it more sharply, the authors remove a logarithmic factor from the prior competitive loss bound and give a competitive ratio lower bound. With exact predictions, the algorithm matches the offline optimum on every arrival order.
In briefTesting how well LLMs can draft NIH data management plans
SummaryThis paper evaluates how well Large Language Models (LLMs) can draft data management plans (DMPs) that comply with the 2023 NIH guidelines. The authors compared outputs from Llama and GPT against NIH-provided DMP samples, and domain experts assessed the LLM-generated plans. The study concludes that LLMs can speed up initial drafts but still need expert review for clarity and completeness, and that open-source models may be less ready off-the-shelf than commercial ones for funder-compliant drafting.
In briefEvaluating AI Musical Creativity Using Value, Novelty, and Surprise
SummaryThis paper proposes a quantitative framework for evaluating the creativity of AI-generated music. Drawing on Margaret Boden's theory, it adapts the dimensions of value, novelty, and surprise to the musical domain. Rather than relying on a single measure, the authors introduce a set of complementary metrics that capture different aspects of musical creativity. The framework aims to give researchers a methodology for empirical evaluation and to support future generative models that use these metrics as learning objectives.
Hindy Rossignol, Giorgio Franceschelli, Mirco Musolesi
In briefMedical Image Segmentation: Comparing Adapted Foundation Models with Task-Specific Training
SummaryThis paper compares DINOv2, a foundation model pretrained on large natural-image collections, with U-Nets and ResU-Nets trained from scratch for medical image segmentation. Experiments cover colonoscopy, breast ultrasound and dermoscopy, with testing on both in-distribution and external datasets. The adapted foundation model generalised better on every external dataset, while task-specific models stayed competitive in-distribution when data were abundant. The authors also find that the required adaptation capacity depends on the imaging domain, and that a widely used polyp test set duplicates images from another benchmark.
In briefToward Foundation Detection Models via Example-Based Grounding in EM-DETR
SummaryThis paper presents an approach to building a multi-domain foundation model for lesion detection by combining medical image datasets annotated for both anatomical structures and abnormalities. A student-teacher training strategy is used to merge datasets with heterogeneous label spaces while mitigating catastrophic forgetting. The authors adapt their earlier example-based EM-DETR framework to learn jointly from multiple domain-specific datasets. The model reports image-level AUC gains of 1 point over prior state of the art on four chest X-ray findings and leading mAP50 on chest X-rays and mammography.
Sheethal Bhat, Bogdan Georgescu, Awais Mansoor et al.
In briefTorchCurves: A PyTorch library for learnable parametric curves
SummaryThis paper presents TorchCurves, a PyTorch library that treats vector-valued parametric curves as learnable building blocks. It maps scalar inputs to points on curves expressed in B-spline or Legendre polynomial bases, with gradients flowing through both inputs and coefficients. The coefficients can be optimized together with a model or generated by another network. Timing experiments compare its evaluation speed with NumPy and SciPy and its Kolmogorov-Arnold network training speed with EfficientKAN across several problem sizes.
In briefAutoScaffold: AI-Guided Training Hints for Improving Language Model Reinforcement Learning
SummaryThis paper introduces AutoScaffold, a framework in which a frozen teacher model periodically reviews a student model's training experience, diagnoses bottlenecks, and designs two kinds of training-time scaffolds: strategies shared across task classes and hints for individual hard tasks. The scaffolds are removed at evaluation. The authors test the framework on ALFWorld, search-based QA, mathematical reasoning, and GPU kernel generation, reporting that students trained this way outperform corresponding RL baselines when evaluated without scaffolds.
WhyThe authors report a frozen teacher auto-designing training-time scaffolds, with gains across four domains evaluated without scaffolds.
In briefContrastive Discovery: AI agents comparing rival explanations to solve science questions
SummaryThis paper addresses a limitation of LLM agents in data-driven scientific discovery, which test each hypothesis only against its negation. It introduces Contrastive Discovery, which replaces a single hypothesis with a set of mutually exclusive and collectively exhaustive competing explanations, and resolves a question when evidence concentrates belief on one of them. Across six domains, the approach resolves more questions to a specific alternative from the same evidence, and under the same experiment budget it resolves more questions than pointwise surprisal.
In briefPyrite: Benchmark and AI Agents for Ranking Drug-Protein Binding Poses
SummaryThe paper introduces Pyrite, a benchmark of 1,500 pose pairs across diverse protein-ligand complexes for testing how well candidate binding poses can be ranked. Expert chemists using 3D visualization and physics-based scores performed near chance. The authors develop two LLM-based agents, PyriteReasoner and PyriteFeaturizer, which reach up to 67% and 74% accuracy respectively, exceeding observed expert performance. LLM-guided feature discovery generally outperforms direct reasoning, and GRPO training helps open-weight models match or surpass closed-source frontier models.
Megha Srivastava, Michael H. Lim, Allison M. Keys et al.
In briefFreeDeshadow: Removing User-Chosen Object Shadows With Visual or Text Prompts
SummaryThe paper introduces an object-aware controllable shadow removal task, where users specify a target object with a simple prompt and only its cast shadow is removed. It proposes FreeDeshadow, a unified framework with an object-aware generator and a shadow-aware local refiner, supporting both visual and textual prompts. The authors also build PCSRD+, a dataset of 154,571 quadruples at 1024×1024 resolution covering six prompt types. Experiments reportedly show better shadow removal and controllability than existing methods.
In briefDoes the Real Wiring Win? Testing Connectome Structure–Function Claims Against Unmeasured Physiology
SummaryThe paper examines whether connectome-constrained models really show that measured wiring outperforms shuffled or rewired wiring, given that model physiology is often tuned on the real wiring. The authors compare three estimands and test them with synthetic ground truth, placebo nulls and three Drosophila circuit models. Some claims hold under every protocol, such as segregated dopamine improving learning, while a mushroom body advantage for real wiring reverses once physiology is constrained by independent data.
WhyThe authors propose a three-estimand protocol showing connectome structure-function conclusions can reverse with tuning, validated on synthetic, placebo and Drosophila models with confidence intervals
In briefContinual-ARC: Benchmark for Measuring Lifelong Skill Acquisition
SummaryThis paper introduces Continual-ARC, a lifelong learning benchmark built from three ARC-AGI editions whose tasks require no domain knowledge. Tasks recur after dormant gaps with fresh instances, and learners are scored by the share of their feedback budget they consume, allowing memory-based and weight-based systems to be compared on one scale. Experiments find that harness persistence generally does not improve memorization, and no harness-based method gains on unseen tasks, while weight-based learning shows substantial gains.
In briefBiological abilities that appear unexpectedly in a model of molecular interactions
SummaryThe paper presents Bonbon, a foundation model trained on about 3.2 trillion tokens of paired protein-ligand sequences. Without task-specific supervision or structural coordinates, it shows zero-shot emergent abilities at functional, residue, and bond levels, which generalize to out-of-distribution complexes. The authors report sharp performance jumps at the largest of six model scales, even as pretraining loss decreases smoothly. In prospective wet-lab tests across nine disease targets, de novo compounds reached a 61.5% dose-response hit rate.
WhyThe authors extend emergent capabilities to a molecular interaction foundation model and report sharp scaling jumps, plus wet-lab validation across nine targets.
In briefSpatial patterns of transport mode choice in Niterói: a synthetic-population simulation study
SummaryThe paper presents a computational framework that uses aggregated, static urban data to build a detailed, dynamic model of urban mobility. Applied to about 475,000 synthetic residents of Niterói, Brazil, the simulated neighborhood-level transit patterns correlate strongly with an empirical activity proxy. Walking is more common in compact central areas, car use rises toward the periphery, and transit use stays relatively high along bus corridors. Reducing walking times to and from bus stops by 30% raises the modeled transit-to-car ratio by 11.3%.
Sandro Martinelli Reia, Taylor Anderson, Dieter Pfoser
In briefConnectome-constrained fly vision model: robustness, efficiency, and linking adversarial and perceptual illusions
SummaryThis paper benchmarks FlyNet, a task-optimized continuous-time recurrent network constrained by the fly visual system's synaptic connectome, against CNNs, RAFT, and recurrent networks lacking connectome constraints on a collision-detection task. FlyNet is adversarially vulnerable but, relative to its compute, more robust to common video corruptions than RAFT, using up to five orders of magnitude fewer FLOPs. The authors show that the same low-level motion statistic underlies both adversarial texture failures and the reverse-phi illusion.
Alexandre Garcia-Duran, Alejandro Rodriguez-Garcia, Manuel Molano-Maz\'on et al.
In briefRS-JEPA: A Compact Self-Supervised Model for Matching Remote Sensing Images Across Different Sensors
SummaryThis paper presents RS-JEPA, a compact self-supervised framework for retrieving scenes across different remote sensing sensors, such as SAR and optical imagery, without manual labels. It uses a Vision Mamba encoder shared by both sensors and introduces scan-path views, which treat two reading orders of the same image as two views to predict hidden patches. Experiments on SAR-optical and PAN-MS benchmarks show the highest F1@5 among twelve re-implemented methods, with a retrieval network of 3.88M parameters.
Tanishq Rachamalla, Srishti Kaushik, Muhammad Attique Khan et al.
In briefComparing three ways to test AI analytics agents for failures and security risks
SummaryThis paper compares three evaluation frameworks for LLM-based analytics agents used in enterprises: trace-based LLM judging, text-only LLM judging with red-teaming, and deterministic heuristics. The study runs in three phases, measuring evaluation stability, detecting known risks on 72 ground-truth test cases, and discovering unknown risks. The authors conclude that the information available to the evaluation harness shapes evaluation quality, that functional and security evaluation are complementary, and they recommend combining trace-based functional assessment with adversarial red-teaming.
In briefReTrace: Turning Tool-Using AI Agents into Planner-Free Workflows
SummaryThis paper presents ReTrace, a method that converts a natural-language Standard Operating Procedure and a few logged agent runs into a reusable workflow in which code, not a large language model, chooses each next step. It extracts the condition behind each step from the agent's reasoning, merges these conditions into a compact rule document, and generates an executable workflow from it. The authors report that across several benchmarks the workflows achieve completion or success rates close to the agent's while lowering cost and running time.
Kaiyuan Hou, Yang Ping Kuo, Indranil Bhattacharya et al.
In briefTimeSeek: A Multi-Task, Multi-Domain Pretrained Model for General Time Series Analysis
SummaryTimeSeek is a general-purpose model for time series analysis that adapts to patterns from multiple domains. It uses a Multi-Domain Feature Extractor to capture temporal dependencies and channel correlations specific to each domain. Through multi-task collaborative training, it unifies forecasting, anomaly detection, and classification in a single pretraining stage. The authors report state-of-the-art results on these tasks in few-shot and zero-shot settings, with favorable inference efficiency.
In briefRe-Search: A two-player questioning game that dissects how LLMs generate research ideas
SummaryThe paper adapts the Twenty Questions game into Re-Search, a benchmark where a model must rediscover a published paper's core idea from its research problem by asking yes-or-no questions. Tested on 200 ICML 2026 papers across six recent LLMs, most progress occurs within the first 30 of 100 questions, after which most models plateau. The authors find the bottleneck is generation rather than judgment, and that models fail by either over-exploiting one direction or over-exploring.
In briefMechanistic Interpretability Needs a Logic of Inference: A paper proposing five principles and a recursive framework for drawing mechanistic explanations from neural network observations
SummaryThis paper examines how mechanistic interpretability (MI) researchers can move from observing internal neural network components to forming mechanistic explanations. The authors compare efforts in neuroscience and MI and identify recurring inferential bottlenecks when inferring mechanisms from complex neural systems. They propose five principles organized into a recursive logic for more rigorous mechanistic inference. The framework aims to turn local findings into cumulative knowledge and generate testable hypotheses beyond the models, but the paper concludes that evidence beyond MI is needed to establish what model mechanisms reveal about the world.
In briefPterodactyl: A Low-Cost System for Automated Translation Quality Evaluation
SummaryThis paper presents Pterodactyl, the system submitted by 1020 Group to the WMT26 shared task on automated quality evaluation. It targets subtasks 2 and 3, namely segment-level quality score prediction and detecting error-free segments. The system uses an LLM-as-a-judge approach, with prompt engineering and model selection guided by ROC analysis, and is designed as a highly customizable, ultra-low-cost binary classifier.
In briefFERN: Table Retrieval for Text-to-SQL Across Benchmarks Using Frozen Retrievers and Learned Ranking
SummaryThis paper presents FERN, a method for retrieving the tables needed by text-to-SQL systems from large pooled catalogs. It fuses three frozen retrievers into 100 candidates, describes each with up to 120 features, and reorders them with LightGBM rankers trained on BIRD, without generative LLMs. The authors compare its table recall against the published results of RASL across Spider, BIRD, Fiben, and BEAVER, reporting that FERN matches RASL on Spider at five tables and reaches 98.0 recall at fifteen tables on Spider's candidate pool.
In briefBoviShift: A Six-Year Multi-Attribute Audit of Cattle Vision Across Animals and Time
SummaryThe paper introduces BoviShift, a dataset rebuilt from six consecutive annual cohorts of a cattle marketplace from 2021 to 2026, containing 2,657 cattle, 10,610 still images, and 2,566 walk-around videos. It compares image-random and cattle-disjoint data splits and tests how a ResNet-50 model transfers to future cohorts. The evaluated attributes are weight, age, height, breed, and coat color.
In briefTanCRAFT: Predictive Feedback to Prevent Errors in Assembly Tasks
SummaryTanCRAFT is a four-stage real-time system that helps people avoid mistakes during physical assembly. It tracks the scene, reasons about the partial assembly, predicts where a manipulated piece will move, and intervenes when that placement conflicts with the target. The system was tested on a tangram assembly task through offline benchmarks and a within-subject user study. Predictive feedback significantly reduced errors compared with unguided and reactive sessions, and it produced the lowest mean perceived workload and shortest mean completion time.
Sassan Mokhtar, Lars Doorenbos, Marius Bock et al.
In briefC²-Routing: Complementary Cropping and Compression for Efficient Vision-Language Inference
SummaryThis paper addresses the heavy computational cost of visual question answering in vision-language models by proposing C²-Routing, which adaptively chooses between image cropping and whole-image compression based on the input image and question. A safety gate uses predicted joint correctness under original and reduced inputs to fall back when needed and protect accuracy. Across three VLMs and 18 datasets, on top of a 1,024 visual-token cap, it cuts end-to-end FLOPs by 36.53% and visual tokens by 42.36% on average, with only a 0.77% mean relative task-score drop.
In briefBayesian inference of epidemic threshold for a spatial SIS infection model
SummaryThis paper studies Bayesian inference of whether an infection persists, using early observations from a heterogeneous SIS reaction-diffusion model. The authors show that early prevalence depends on the unknown parameters only through the difference between transmission and recovery rates and the infected diffusivity, which also determines the principal eigenvalue governing persistence. The paper establishes well-posedness of the model and the Bayesian problem, and proves O(ε) Hellinger convergence of nonlinear posteriors to low-prevalence limits. Numerical examples illustrate threshold inference.
In briefA Theory of Policy Optimization in Post-Training: Coverage Collapse and Reference Resets
SummaryThis paper theoretically analyzes policy optimization in language model post-training, defining "coverage-preserving" algorithms whose sample and gradient costs depend on how readily the base model generates optimal responses. It shows that REINFORCE, the basis of PPO and GRPO, can drive the probability of optimal responses near zero even as reward rises, a failure called coverage collapse. The authors propose resetting the KL reference to the latest policy within an Approximate Policy Mirror Descent framework, and use it to prove global convergence for several widely used post-training algorithms.
WhyThe authors introduce coverage collapse and reference resets, giving first global convergence guarantees for softmax linear policies under APMD.
In briefA theory of how much cross-validation estimates vary
SummaryThis paper studies how much the estimate from cross-validation, a method that repeatedly splits data into training and validation sets to measure predictive performance, fluctuates. It proves an exact finite-sample identity for the estimate's variance, showing it depends on the weights that random splits give each observation. From this, K-fold cross-validation is first-order variance optimal, leave-one-out has the smallest variance among K-fold schemes under reasonable conditions, and leave-pair-out is optimal for AUC. The estimate is also shown to be asymptotically normal.
WhyThe authors derive an exact finite-sample variance identity for cross-validation, prove K-fold and leave-one-out variance optimality, and report asymptotic normality covering MSE and AUC.
In briefAccurate Prediction of Energetic Material Properties Using a Tabular Foundation Model
SummaryThis paper applies TabPFN, a tabular foundation model, to predict impact sensitivity, decomposition temperature, and crystal density of energetic materials. Without hyperparameter tuning, TabPFN outperforms the best published models and established methods such as tuned XGBoost, random forests, and Gaussian processes. The model also provides built-in uncertainty quantification because it predicts the full target distribution, and Shapley values are used to interpret feature importance.
Vagn Krister Nes Glomsrud, Mathias Bynke, Lars Henry Berge Olsen et al.
In briefWhen does text encoding matter for language modeling?
SummaryThis paper examines how text encoding choices affect the performance of byte-level language models, comparing UTF-8 with other Unicode-based, language-specific, and custom encodings. The authors propose Aether, a hybrid scheme that uses script-specific encodings for target languages while falling back to UTF-8 for full Unicode coverage, and use it to model four languages with diverse writing systems. They find that languages with information-dense characters, such as Chinese and Korean, benefit significantly from compressive encodings, while Georgian and Tamil are largely insensitive to encoding choice. The authors conclude that text encoding should be a language-specific design decision rather than a universal default.
In briefA lightweight differentiable method for real-time skin rendering and digitization
SummaryThis paper presents a differentiable Screen-Space Subsurface Scattering framework that directly estimates scattering parameters for real-time rendering pipelines. It jointly optimizes spatially varying base textures and scattering distances, estimating them across facial regions to produce a blended profile map. The method uses a low-cost single-light, single-view capture setup and is validated on a 45-person real dataset spanning the Monk Skin Tone Scale, plus synthetic, multi-light, and tilted-view experiments.
Mario Marquez-Balduque, Carlos Rodriguez-Pardo, Jorge Lopez-Moreno
In briefUGG: a unified generative model that jointly handles video and geometry for 4D scene understanding
SummaryUGG is a unified framework for 4D scene understanding that jointly models RGB video together with geometric representations such as depth and camera pose. By assigning different noise levels across modalities and time, it treats visual generation and geometric prediction as conditional completion within one generative model. The authors report that UGG outperforms non-unified baselines on camera-controlled video generation, depth estimation, and camera pose prediction, including sparse observation settings.
In briefMetric-aligned policy optimization for explainable deepfake detection
SummaryThis paper addresses deepfake detection that must explain the visual evidence behind each decision, not just classify real versus fake. It proposes a two-stage predict-then-explain pipeline, submitted to the ACM Multimedia 2026 Explainable Deepfake Detection Challenge and trained and evaluated on the XPlainVerse benchmark. A vision-foundation-model ensemble classifies each image, and Qwen3-VL-8B generates an explanation conditioned on that decision. Metric-aligned Group Relative Policy Optimization improves the explainer, raising complex BERTScore-F1 from 0.686 to 0.712, and the submission ranked first on that metric and seventh overall.
In briefOne Policy for 100 Household Tasks: Our 2026 BEHAVIOR Challenge Solution
SummaryThis paper describes a solution to the 2026 BEHAVIOR Challenge, which includes 100 long-horizon household tasks in the OmniGibson simulator. The team placed 9th using a single vision-language-action model with one checkpoint for all 100 tasks, which scored 0.392 overall. On 50 tasks shared with 2025, its Q rose from 0.260 to 0.371. For 14 tasks, control was routed to a GPT-6 Astra coding agent, which achieved a mean Q of 0.639 compared with 0.108 for the learned policy.
In briefTwo Fields, One Scheduling Problem: Comparing Branch-and-Bound and A* for Resource-Constrained Project Scheduling
SummaryThis paper compares Branch-and-Bound (B&B), a method from Operations Research, with A*, a method from Artificial Intelligence, as exact search approaches. It uses the Resource-Constrained Project Scheduling Problem as a test case and runs experiments on 480 benchmark instances. Graph-based A* substantially outperforms the B&B variants studied, proving optimality on all 480 instances with additional pruning rules, compared with 445 for best-first B&B under the same limits.
In briefMeasuring how stable time-series foundation models are when sensor sampling frequency changes
SummaryThis paper stress-tests four time-series foundation models (MOMENT, Chronos-2, Moirai-MoE, and TimesFM) to measure how stable they remain when sampling frequency changes. The authors use three new metrics, FSI, FDI, and CDI, to assess embedding shift, forecasting degradation, and downstream performance degradation. They argue that this invariance depends on the margin between a signal's native sampling frequency and the bandwidth its task requires, and that much of the performance loss reflects information loss.
Sourish Gunesh Dhekane, Abinaya Kumar, Torsten Woertwein et al.
In briefDnNP: Denoising Input Uncertainty in Neural Processes
SummaryNeural Processes are flexible meta-learning models that quantify uncertainty, but the standard framework assumes noise-free inputs. This paper introduces Denoising Neural Processes (DnNPs), which explicitly model input uncertainty and are trained via variational inference to denoise corrupted inputs while performing meta-learning. On synthetic data and PM2.5 spatial interpolation from real air quality sensors, DnNPs improve prediction and uncertainty estimates over standard NPs, reducing RMSE by up to 30% and NLL by up to 73%.
In briefEarlybird: A 29M-Parameter Model That Predicts Speech Words Before They Are Finished
SummaryThis paper presents Earlybird, a 29-million-parameter streaming speech recognizer that fits in 16.8 MB at 4-bit precision. It is trained on audio cut off at random points and predicts each word due by then, plus the next two words. On LibriSpeech test-clean, the median word is predicted 90 ms before the speaker finishes it, while released streaming models output it 90 to 400 ms after it ends.
In briefOnlyAct: Extracting action representations from pretrained models for world simulators
SummaryVideo world models need action labels, which are scarce and hard to scale. This paper starts from the pretrained foundation model V-JEPA 2, whose frame-pair embeddings already contain much action information, instead of learning actions from scratch. A label-free contrastive adaptation method pulls together frame pairs with the same motion in different scenes and pushes apart pairs with different motion in the same scene, extracting action while discarding the scene. Evaluated on robot manipulation, camera navigation, game, and driving benchmarks, the representations are more expressive than annotation-free latent-action methods while matching their scene invariance.
In briefTesting whether AI-generated world models are structurally faithful for neuro-symbolic reasoning
SummaryThe paper tests whether five large language models can turn natural-language scenarios into formal world models, represented as partially observable stochastic games. It uses a benchmark of 44 game, situational, and cultural scenarios and evaluates outputs for mechanical validity, structural agreement with references, and semantic quality. Most generated models run correctly, but few exactly match the reference structure, so executability alone does not ensure faithful models.
Souhaila Djaffal, Amna Jamal, Gagana Madhugiri Devaraja et al.
In briefPrecision on Demand: Raising Weight Bits Only at Uncertain Decoding Steps in a Small Local Model
SummaryThe paper studies a decoding scheme for small local language models that stores 8-bit weights but normally decodes with only the top 4 bits. A step is redone at 8 bits only when the low-precision output distribution is uncertain, judged by entropy or a small top-two margin. Experiments use Qwen3-0.6B on GSM8K, with cost measured in weight bytes read per token. The authors report replayed and live results and note that the live data do not clearly separate the gate from either all-high or all-low precision.
Mohammed Nayeem Khan, Logesh L, Aliyas Ahmad Misri
In briefPost-hoc disentangling of audio foundation models for music similarity
SummaryThis paper studies post-hoc disentanglement of pretrained audio foundation models such as CLAP, MuQ, and MERT for music similarity. It adapts Conditional Similarity Networks to gate frozen embeddings into attribute-specific subspaces. Across two datasets, minimal trainable adaptation improves attribute-specific disentanglement and retrieval performance, while pooling, masking, and depth choices add limited gains.
Lucas de Souza Brand\~ao, Gabriel da Costa Merlin, Diego Furtado Silva
In briefBits or permutations? Budgeting weight precision versus option-order averaging in small local models
SummaryThis paper asks whether reducing the weight precision of small local language models hurts their sensitivity to answer-option order, and which is the cheaper way to spend a fixed inference budget. The authors test Qwen3-0.6B and 1.7B with simulated round-to-nearest quantization at 8, 6, 4 and 3 bits on ARC, MMLU and HellaSwag. They find 6-bit weights match fp16 accuracy, and in cost-matched comparisons extra option orders beat extra bits only when weights have 6 or more bits.
Mohammed Nayeem Khan, Logesh L, Aliyas Ahmad Misri
In briefCausal Concept Explanations for Deep Neural Models: A Training Method and Measure for Causal Concept Effects
SummaryThis paper argues that concept-based explanations of neural networks are often only correlational, so they can look correct while not being necessary or sufficient for a model's output. The authors propose CCW-Net, which uses interventional mediation analysis as a training objective to make concept representations necessary, sufficient, and independent. They also use mediation analysis as an evaluation tool to quantify each concept's causal contribution to a model's output. The method is evaluated on aircraft control, driving, and image classification.
In briefEquations Found for South Asian Monsoon Rainfall Thermodynamics
SummaryThis paper uses symbolic distillation, combining neural networks and symbolic regression, to derive interpretable precipitation equations from 21 years of June-August reanalysis data. The best equation explains 1.5 times as much precipitation variance as a widely used buoyancy-based diagnostic. It recovers about 85% of neural network skill while using roughly a thousand times fewer parameters.
Savannah L. Ferretti, Jane W. Baldwin, Michael S. Pritchard et al.
In briefREAL Loss: Training Speculative Decoding for Real Acceptance Length
SummaryThe paper addresses speculative decoding, where a draft model proposes tokens that a target LLM verifies, with speedup governed by expected acceptance length (EAL). It argues that common training objectives and recent block-level surrogates do not match EAL exactly, creating a covariance gap that can reverse gradient direction. The authors derive REAL, a loss using path weights for unbiased estimation and Rao-Blackwellization for variance reduction. In controlled comparisons it outperforms several baselines across tasks and seeds, without changing the drafter architecture or inference procedure.
In briefKernelMosaic: An Open Dataset and Agent Framework for Generating CUDA Kernels
SummaryThe paper presents KernelMosaic, an open dataset that pairs diverse CUDA tasks with verified solutions and iterative optimization trajectories, along with an agentic framework that guides a strong LLM through implementation, correctness repair, and performance optimization. Supervised fine-tuning of a small open-source model on the dataset raised correctness on KernelBench L1, L2, and L3 from 4%, 0%, and 0% to 88%, 82%, and 16%. The authors also report that training on curated trajectories outperforms training on direct responses.
In briefA margin-based analysis of LoRA: resistance to forgetting and adapter merging (MaLoRA)
SummaryThis paper explains, through a margin-based theoretical framework, why LoRA fine-tuning resists catastrophic forgetting and why independently trained adapters can be merged. Analyzing multiclass linear classification with ℓ2 regularization under a near-orthogonal task setting, the authors show that in an intermediate regularization regime the optimal adapter aligns with the max-margin solution on fine-tuning data. They propose MaLoRA, a training-free merging rule, and report experiments on vision and language classification tasks where it matches or outperforms several adapter-merging baselines.
Ziqing Xu, Hancheng Min, Lachlan Ewen MacDonald et al.
In briefBlurring Block Boundaries: Overlapping Block Denoising for Faster Diffusion Language Models
SummaryThis paper proposes B3 (Blurring the Block Boundaries), a method that speeds up inference in block-diffusion diffusion language models. It refines two neighboring blocks in parallel within a single forward pass and decides on the fly when the next block can start. The method requires no training or hyperparameters and can be added to any self-editing block diffusion model. Across 10 datasets, the authors report generation quality on par with the base model and a 11.5% to 14.5% speedup.
In briefAn Energy-Based View of Machine Learning: A Unifying Tutorial
SummaryThis tutorial introduces energy-based models (EBMs), which define probability distributions through a scalar energy function. It treats EBMs as a unifying lens, arguing that any EBM is specified by its energy, inference procedure, and learning method. The paper shows that many methods, including discriminative and generative models, representation learning, reinforcement learning, and planning, can be viewed as EBMs that differ in these choices. It also proposes that replacing one component can suggest new algorithms, with energy composition offering a way to combine learned models at inference time.
In briefMapping How Large Language Models Learn by Internal Thinking
SummaryThis paper proposes a Cognitive LbT Framework that maps four mechanisms of human-like learning by thinking (Reflective Reasoning, Synthetic Memory Formation, Internal Simulation, and Conceptual Self-Organization) onto computational processes in large language models. It analyzes how methods such as chain-of-thought prompting, self-consistency, and Self-Instruct approximate these mechanisms. The paper also examines their characteristic failures, such as hallucination diffusion and model collapse, and the structural asymmetries that cause them.
