In briefFrom Knowledge Access to Source Learning: Building Skill With a Specific Source
SummaryThis paper studies how LLM agents can build reusable understanding of a persistent authoritative source over repeated use, rather than just re-accessing it. The authors propose SourceLearn, which combines self-directed and task-guided learning to construct and refine a persistent source model. Across five benchmarks and three LLM backends, it achieves the best results in 13 of 15 settings.
In briefOmni-Embed-Mini: a compact model that embeds text, audio, images, and video together without losing text retrieval quality
SummaryThe paper introduces Omni-Embed-Mini, a 0.9B-parameter model that maps text, speech, audio, images, video, and visually-rich documents into one shared cosine embedding space without changing any text-side weights. Each media sample is paired with a dense caption, and the teacher target is the frozen backbone's embedding of that caption, trained with a Matryoshka SigLIP contrastive loss and online hard-negative mining. The authors report unchanged text retrieval (49.57 nDCG@10 on MTEB-v2 BEIR-8) and a model that is roughly 2.7x to 9.5x smaller than the open omni embedders they compare against. A 2.3B variant is also described as competitive with a closed commercial embedder.<br>
Mohammed Irfan Kurpath, Jaseel Muhammad Kaithakkodan, Sahal Shaji Mullappilly et al.
In briefFaynt: Scaling and Optimizing AI Policies for Competitive Super Smash Bros. Melee
SummaryThe paper introduces Faynt, a pair of Transformer-based game-playing models with 10 million and 75 million parameters that control all 26 characters in Super Smash Bros. Melee with one checkpoint. After reinforcement learning, the 10M model won 240 of 244 same-character games against fourteen rival models. The authors study architecture, optimization, scaling, and pretraining on about 840,000 human replays, using curricula, distillation, and Fox mirror-match reinforcement learning. They also release the weights, benchmarks, and a tournament platform.
In briefKeyword tests can overrate small models' tool use; a cheap diagnostic ladder
SummaryThe paper argues that keyword-matching benchmarks can credit small language models with tool use they do not actually perform. Two architecturally matched Spanish security models score almost identically on lenient metrics, but only one passes strict verbatim-reproduction checks that expose the other's failure. The authors localize the failure to a missing prior on the tool-call token, repair it with a targeted fine-tuning recipe, and recommend such cheap strict diagnostics before making tool-use claims.
In briefFinetuning with Sampling: Making supervised training learn more effectively
SummaryThis paper proposes a Markov chain Monte Carlo (MCMC) sampling method that gradually shifts off-policy expert traces toward the model's own distribution before supervised finetuning (SFT). Across scientific skill acquisition, mathematical reasoning, and open-ended expertise tasks, the approach lets SFT rival leading posttraining methods, often generalizing better and forgetting less than strong on-policy baselines. The authors also report that the resulting models can learn beyond sharpening the base model's distribution.
In briefMIRTO: a registration-gated, multiverse-tested protocol for evaluating unsupervised brain MRI anomaly segmentation
SummaryThe paper presents MIRTO, an evaluation protocol for unsupervised anomaly detection in brain MRI that makes often-unreported choices, such as registration, thresholding and false-positive budget, explicit and measures their effect. Applied to four methods on 312 BraTS 2020 subjects, it shows that an axis-order mismatch substantially lowers a diffusion model's voxel AUROC and that a Dice advantage can vanish under equal realised false-positive burden due to threshold transfer. Because the same cohort was used to develop the protocol, all inference is exploratory.
In briefLinear programming and fast exact algorithms for robust decision-making under uncertain Markov models
SummaryThis paper studies linear programming formulations and strongly polynomial algorithms for robust Markov decision processes whose rewards and transitions have rational polyhedral uncertainty. It builds a single linear program, encoding a finite number of robust policy-iteration steps, whose optimal solutions recover the robust optimal value and optimal stationary randomized policies. The authors also analyze the complexity of robust policy iteration and derive new strongly polynomial bounds for several uncertainty sets.
WhyEncodes finitely many robust policy-iteration steps into one LP and proves strongly polynomial bounds for several uncertainty sets, a substantial theoretical advance.
In briefSample complexity limits for sampling categorical Markov random fields with discrete diffusion
SummaryThe paper studies sampling from high-dimensional categorical distributions with local dependence, modeled as low-order Markov random fields, using discrete diffusion with uniform noising. It introduces a pinning decomposition of the discrete score and a weight-sharing neural score learner combined with τ-leaping, yielding end-to-end sample complexity bounds that depend explicitly on vocabulary size, interaction order, and sample size. Experiments on Potts, Ising, and tree-structured models show that weight-sharing score networks outperform fully connected ones for sampling long sequences.
In briefLocal Support Learning: a method to reduce catastrophic forgetting in large models
SummaryThis paper studies catastrophic forgetting in large pre-trained models by treating it as a geometric problem in each weight matrix's input space. It proposes Local Support Learning (LSL), which pairs a standard weight adapter with a Gaussian Mixture Model gate that enables the adapter only on inputs from the current training distribution. The authors report that this post-training approach preserves pretrained and finetuned capabilities across multiple phases in LLMs of up to 7 billion parameters, with efficient memory and compute.
In briefUsing domain experts in multimodal mixture-of-experts models to adapt faster
SummaryThe paper studies whether sparse routing in multimodal Mixture-of-Experts models leads experts to specialize semantically. It introduces ExpertLens, a data-free method that finds domain-specialized experts by decoding router weights into vocabulary tokens. Fine-tuning only those experts, which is 21.7% to 47.0% of parameters, matches or beats full fine-tuning on math, medical, and remote sensing tasks, with a 4.0x average training speedup and better results than LoRA.
In briefArgo-Bench: Testing AI Data Agents on Realistic Enterprise Workflows
SummaryThis paper introduces Argo-Bench, a framework of 210 data science and analytics tasks built on a simulated food delivery business in New York City with 81 million orders in 2024. The simulated world is exported as a 235-table, 7.5 billion-row enterprise data warehouse, and agents must reconstruct hidden facts before taking actions like banning fraudulent accounts, graded by their effects in the simulator. The strongest of 14 models scores 95 or higher on only 34.8% of tasks, averaging 59.5 points.
WhyBuilds an enterprise-scale agent benchmark with action-based grading across 235 tables, validated on 14 frontier and open models.
Gabriel Tomitsuka, Arman Raayatsanati, Emma Xing et al.
In briefPower-optimal decentralized coordination of magnetically actuated spacecraft swarms
SummaryThis paper proposes a decentralized, power-optimal coordination framework for spacecraft swarms that use time-varying magnetorquer actuation instead of propellant. The framework jointly designs the interaction graph, frequency grouping, and controller gains while preserving angular momentum. In a closed-loop simulation of a thousand spacecraft, the relative position, attitude, and reaction-wheel momentum errors converge to the desired states.
In briefWhere-OPD: Training multimodal AI on synthetic scenes by guiding it to find relevant image regions
SummaryThis paper introduces Where-OPD, a method that improves multimodal large language models by giving a teacher version textual guidance about which image regions are relevant to a question. The student model learns to reproduce this behavior using only the image and question. Training uses procedurally generated synthetic scenes with automatically known object identities and coordinates, so no human annotation is needed. The method improves counting and document and chart understanding, and gains of 3.23 points on average transfer to real-world perception benchmarks.
Sophia Sirko-Galouchenko, Monika Wysoczanska, Andrei Bursuc et al.
In briefComparing Explanation Methods for Auditing DeBERTa-v3 in Medical Abstract Classification
SummaryThis paper presents a framework for comparing explainability methods when auditing DeBERTa-v3 on zero-shot classification of medical abstracts. It compares five attribution methods, including SHAP, LIME, occlusion, Input x Gradient and Attention x Gradient, using Jaccard agreement. The authors find accuracy is high for well-defined clinical categories but drops under semantic ambiguity, and explanation stability tracks predictive certainty.
Javier Diaz Esteban-Herreros, David Mu\~noz-Valero, Raquel Mart\'inez-Espa\~na et al.
In briefCombining brain MRI surface and volume features with self-supervised learning for gene discovery
SummaryThis paper introduces MEVA (Mesh-Enhanced Volumetric Autoencoder), a self-supervised framework that encodes voxel intensity together with cortical mesh geometry, including curvature and cortical thickness, into one shared set of imaging features. In the UK Biobank, these features yield more genome-wide significant loci in GWAS than features learned from volumes alone or meshes alone. The authors suggest adding cortical surface geometry captures additional heritable brain variation.
In briefAre Circuit Evaluations Measuring the Right Thing? Objective-Level Recovery Gaps in Interpretability
SummaryThis paper examines whether intervention-based faithfulness metrics reliably identify the circuits that reproduce a model's behavior. The authors show that these metrics can distort the computational context by altering signals outside the circuit, sometimes preferring a less faithful, equally sized circuit. Testing on pretrained language models with known mechanisms, they find such misrankings occur in outputs of automated discovery methods, in 13.2% to 24.4% of candidate pairs. Restoring selected intact-model signals reverses 96 of 100 persistent misrankings, and the authors call for rethinking both circuit discovery and its evaluation.
WhyShows intervention-based faithfulness misranks circuits due to context distortion, validated via controlled edits, multiple discovery methods, and reversal experiments.
In briefA scalable, transferable meta-network for data selection, and why it needs a different loss
SummaryThis paper addresses how to choose training data for large language models using meta-learning. The authors propose TESS, a framework that replaces per-sample weights with a selection network and uses a Pointwise Value Matching objective to avoid unstable optimization. Experiments on LLM safety and targeted instruction tuning show the method transfers across dataset scales and model sizes.
In briefGeoLatent: Geometry-Guided Latent Structuring with Routed Optimization for 3D Reasoning
SummaryThis paper addresses the difficulty of 3D spatial reasoning from 2D images. It proposes GeoLatent, which uses Common–Residual Geometry Alignment to separate shared and residual geometry and routed optimization to train geometry and language jointly. The authors report 73.0% on SPAR-Bench and 72.1% on SPBench, outperforming previously reported methods on both.
In briefHumanoidToolBench: Testing How Humanoid Robots Choose and Use Tools
SummaryThis paper introduces HumanoidToolBench, an 18-task benchmark covering three scenarios, three execution levels, and two tool-set modes, along with ToolBook, a dataset of 3.1k demonstrations collected in simulation and on a real Unitree G1 robot. The authors evaluated seven policies in simulation and three on the real robot, finding substantial gaps between choosing a suitable tool and completing the task. Focused tests on GR00T N1.7 showed reduced selection accuracy on unseen tools and continued task execution under unrelated instructions.
In briefKolmogorov-Arnold Networks for Solving Free-Boundary PDEs
SummaryThis paper proposes a physics-informed approach using Kolmogorov-Arnold networks (KANs) to solve free-boundary problems, incorporating obstacle constraints, complementarity conditions, and boundary conditions through residual-based losses. It tests three cases: a linear elliptic obstacle problem, a nonlinear p-Laplacian obstacle problem, and a time-dependent one-phase Stefan problem. Compared with PINN and residual-network baselines, the KAN solver achieves low relative L2 and L-infinity errors while accurately resolving contact regions and moving interfaces.
In briefWhy Gradient-Flow and Forward-Diffusion Samplers Struggle with Multimodal Distributions
SummaryThis paper examines sampling methods built on Wasserstein gradient flows and forward-only diffusion processes, which often come with guarantees of fast convergence. The authors show these methods share density evolution with overdamped diffusion, so they inherit metastability and slow mixing. Using spectral and mean first-passage time analysis, they find that well-separated modes can cause exponentially long mixing times. For the log-linear annealing schedule studied, adding intermediate distributions does not remove this exponential scaling.
Daniel McBride, Pratik Khandagale, Cristina Garcia-Cardona et al.
In briefLLM-as-Jev: Off-the-Shelf LLMs Already Work as Jev-Style Decision Models, and When to Fine-Tune
SummaryThis paper studies whether general-purpose large language models can already act as Jev-style decision models, which output probability distributions over predefined options instead of free-form text. The authors propose LLM-as-Jev, a framework offering a training-free inference method and a fine-tuning objective with KL-divergence anchors. Experiments on Qwen3.5-4B and Qwen3-0.6B show that frozen off-the-shelf models match or exceed existing training-free approaches and handle multimodal decisions, while fine-tuning mainly helps smaller models and complex routing.
In briefPyPottery: An AI tool suite that speeds up pottery drawing processing and publication
SummaryThis paper presents PyPottery, an open-source suite of four AI-assisted modules that semi-automate the documentation pipeline for ceramics, covering image extraction, handwriting recognition, inking, vectorization, and layout generation. It was evaluated on 50 hand-drawn sheets containing 240 pottery drawings from the Terramara di Montale site in Italy. Usability study participants reported a median perceived speedup of 40 times over traditional workflows, with a range of 17.5 to 120 times.
In briefCausal Memory Policy: Making Memory Utility Measurable by Intervening on Retrieval
SummaryThe paper notes that memory-augmented language models cannot estimate the utility of memories that are never retrieved. It proposes Causal Memory Policy (CMP), which reserves a fixed number of context slots for memories sampled with known propensities and estimates utility via self-normalized inverse propensity weighting. The authors report that identification fails for 54% of required memories on LongMemEval and 67% on LoCoMo, and that CMP raises the AUC for separating required from non-required memories from 0.54 to 0.66.
WhyProposes intervening on retrieval to restore identifiability of memory utility, with proofs and experiments on two benchmarks plus a deployed system.
In briefAI emulator of stratospheric sudden warmings with interpretable latent clusters
SummaryThe paper builds a probabilistic deep learning emulator for the stochastic Holton–Mass model, which has strong and weak polar vortex regimes with rare transitions. The emulator reproduces short-term dynamics, steady-state distributions, persistence, transition rates, committor function, and expected lead time of the physical model. Principal Component Analysis of its 32-dimensional latent space reveals four physically interpretable clusters found without supervision.
C. Daniel Boscu, Daniel Hernandez, Fabio Alvarez Ventura et al.
In briefRebalancing LoRA Adapter Gains After Training Without Losing Learned Directions
SummaryThe paper identifies adaptation imbalance in LoRA fine-tuning, where a few singular directions dominate the learned update and make performance sensitive to how gains are allocated. It proposes LoRA-Norm, a post-training method that keeps the learned directions while rebalancing their gains through spectral rebalancing and nuclear-norm restoration. The method needs no calibration data or extra training and adds no inference overhead. Across two backbones and three tasks, it improves average specialization and capability retention over the evaluated post-hoc spectral pruning and gradient-guided editing baselines.
In briefCross-model span-level hallucination detection using hidden state probing
SummaryThe paper introduces a method that inspects an LLM's internal hidden-layer activations to locate the exact onset and continuation tokens of hallucinated spans, rather than classifying each token as a simple binary label. Experiments report substantial gains in Precision-Recall AUC over random baselines despite extreme class imbalance. The authors also propose a cross-model setup in which one model probes another model's internal representations, finding that an external observer, including a smaller model, can match or exceed the generator's self-detection of hallucination onsets.
In briefBasic time-series patterns that foundation models still miss
SummaryThe paper introduces SimpleTimeBench, a diagnostic suite of univariate and multivariate tests covering monotonic trends, periodic signals, and leading-indicator covariates. Multivariate time series foundation models such as Chronos-2, Moirai, and Toto often give suboptimal zero-shot forecasts on these simple cases. Fine-tuning Chronos-2 improved some tasks but degraded performance on other fundamental patterns. The same failures appear in real-world sensor forecasting, where models underuse available leading indicators.
Nafiseh Ghoroghchian, Haipeng Zhang, Shuyi Han et al.
In briefLearning from Failure: Using Unreliable Predictions to Remove Adverse Weather from Real Images
SummaryThis paper addresses restoring images degraded by rain, haze, snow and similar weather, aiming to make outdoor vision systems more robust. It proposes a student-teacher semi-supervised framework that stores failed teacher predictions as negative samples and high-quality predictions as positive samples for contrastive learning. It also adds a phase spectrum-based semantic constraint and an adaptive phase consistency loss. The authors report better restoration quality and perceptual fidelity than existing state-of-the-art methods on real-world benchmarks.
In briefHydroJEV: A training-free, one-second screen that sorts water network alarms into cyberattack, fault, or other causes.
SummaryThe paper tests Jev, a training-free model that returns class probabilities in about one second, as a first-stage triage tool for SCADA alarms in water distribution networks. On a four-class cause-attribution benchmark built on the C-Town network, Jev matched a hand-written rule tree and outperformed a supervised classifier on event subtypes absent from its labels. Jev was 20 to 40 times faster than frontier LLMs. The authors report that passing only benign verdicts confirmed by the rule tree to an LLM reviewer spared that reviewer 35 to 38 percent of windows without loss of macro-F1.
In briefForm and Void: AI agent that stages positive-negative space image composition
SummaryThis paper introduces the Form and Void Agent (FaV-A), a multimodal AI agent that generates images with positive and negative space in stages. It first creates a base object, then analyzes its shape and spatial structure to find candidate negative-space meanings, and finally writes instructions for the final image. The authors report that experiments and ablation analyses show it outperforms direct zero-shot multimodal LLM baselines in visual coherence and semantic alignment.
In briefDiDE: Training-free 3D stylization that controls color and texture separately
SummaryThis paper introduces DiDE, a training-free framework for 3D stylization that transfers color and texture from reference images independently, rather than jointly as existing methods do. It exploits the observation that the latent space of image-to-3D models has spare channels beyond those used for texture, allowing color to be encoded in a free subspace. On the authors' newly collected Disen3D-Bench, DiDE reportedly outperforms 2D and 3D baselines in color fidelity, texture transfer, and content preservation.
In briefHakemBench: A Turkish Benchmark for Structured Decisions
SummaryHakemBench is a Turkish benchmark in which models read a text, a question and fixed options, then assign a probability to each option. Version 1.0 is released under CC BY 4.0 with 2,346 items and 4,275 questions across seven tracks. The harness scores decision quality, calibration and selective automation, and reports bootstrap intervals. Most gold labels come from AI model votes rather than human verification, and some results are flagged as not blind.
In briefDistributionally Robust Schrödinger Bridge: A Controller for Shifting Initial Distributions
SummaryThis paper introduces the Distributionally Robust Schrödinger Bridge (DRSB), which learns one controller that stays robust when the initial distribution varies within an ambiguity set. Its objective combines control energy with a KL penalty between the terminal and target distributions, optimized by an alternating algorithm with Wasserstein and Sinkhorn variants. Experiments on 2D transport and image-to-image translation show better robustness to input perturbations, at some cost to nominal performance.
Jinhwan Sul, Panagiotis Theodoropoulos, Vincent Pacelli et al.
In briefTesting whether stack-based language models mirror typological word-order patterns in artificial languages
SummaryThis paper studies whether the word orders common across natural languages may stem from language models' learning biases. The authors test stack-based language models on cross-serial dependencies, a structure marking the upper limit of attested syntactic complexity, across diverse artificial languages. They find such models struggle with these constructions. However, models with limited working memory generalize better, suggesting a possible inductive bias behind some typological commonalities.[](https://arxiv.org)
In briefCARM: A Masking Method That Stops Probability Changes from Canceling Out in LLM Reinforcement Learning
SummaryThis paper addresses off-policy mismatch in reinforcement learning for large language models, which arises from policy updates and differences between rollout and training engines. It proposes Cancellation-Aware Response Masking (CARM), which takes the absolute value of each token's log-probability ratio before averaging so that opposing changes do not cancel. The authors prove a joint bound for accepted responses. Experiments on math reasoning and code generation report gains over geometric-mean masking.
In briefMimir: Physics-grounded AI agent for long-term irrigation control
SummaryThis paper presents Mimir, an LLM-based agent for controlling irrigation over long growing seasons while respecting physical rules. A deterministic simulator checks and revises the model's proposed irrigation actions before execution, and recurring failures are stored as persistent guiding principles, while the physical model stays fixed. In a retrospective evaluation across multiple sites, crops, and years, Mimir reportedly achieved the lowest aggregate control cost among the references and used about 51% less irrigation than historical schedule replay.
In briefGlobal Coherence: When each AI agent is right but the team is wrong, and a framework to fix shared state
SummaryThe paper argues that multiple AI agents can each make locally valid decisions yet jointly produce an invalid result, a failure of shared state rather than of intelligence. It proves an impossibility theorem showing when a policy can guarantee a valid action, and proposes a runtime framework in which a harness, not the models, owns shared state and governs commits. Experiments across several benchmarks report that missing global state cannot be compensated by local intelligence..
In briefRelative Transitions Instead of Absolute Destinations: A Framework for Generating Human Trajectories Without Target-City Trajectory Data
SummaryThe paper addresses generating human mobility trajectories for a target city where no trajectory data exists, using only the city's POI coordinates and categories. The proposed Nomad framework learns a transition prior over semantic displacement, geographic displacement, and elapsed time from source-city trajectories, then grounds sampled transitions onto the target POI map via a behavior graph and an exploration-return walk. Across ten cities and 14 transfers, Nomad outperforms adaptation baselines, reducing average error relative to the best baseline per metric by about 15% in distributional fidelity and about 3% in downstream utility.
In briefUsing diffusion-generated synthetic sensor data to pretrain models for eating and drinking recognition
SummaryThis paper improves wearable activity recognition of eating and drinking by pretraining a CABiGRU model on synthetic smartwatch sensor windows generated by a diffusion model. The model is then fine-tuned on real data to address underfitting in minority classes. On the DEO dataset, the approach reaches 90.6% balanced accuracy, outperforming a strong supervised baseline.
E. Riveros (1), D. Vega-Oliveros (2), A. Soriano-Vargas (3) et al.
In briefSPHERE: Adaptive VR indoor scene generation using LLM-enhanced spatial preference learning and human-in-the-loop RL
SummaryThe paper presents SPHERE, a VR framework for generating 3D indoor scenes that retains user spatial preferences across sessions. It extracts preferences from speech and controller edits and turns them into hierarchical spatial constraints, while a human-in-the-loop reinforcement learning step updates retrieval based on the user's final edited scenes. A mixed-design user study with 42 participants and an offline ablation report fewer corrective edits and lower physical demand.
In briefOld Ideas, Novel Problems: Instability of LLM-Based Novelty Evaluation
SummaryThis paper studies how well large language models judge the novelty of AI-generated research ideas. The authors built an evaluation set by mining OpenReview for passages where reviewers unanimously affirmed or disputed a paper's originality, and paired these with ideas from a vanilla LLM generator. Across six judges, small changes in prompt design substantially changed verdicts, sometimes dropping accuracy below chance, while retrieval and larger reasoning budgets helped little. Two purpose-built novelty evaluators were outperformed by the cheapest prompted baseline.
In briefAdapting 2D Vision-Language Models into 3D Programmers for Task-Specific Tasks
SummaryThis paper presents 3D-Prog, a framework that lets existing 2D vision-language models handle 3D tasks without retraining. It introduces Canonical Coordinate Framing (CCF), which maps inputs and outputs to a shared Euclidean coordinate system to address axis ambiguity and inconsistent scale, and Task-Adaptive Feedback (TAF), which closes the reasoning loop with dynamic feedback for open-vocabulary tasks. The authors report consistent, interpretable results across object-level and scene-level 3D understanding, manipulation, and generation.
Arman Raayatsanati, Sombit Dey, Anna-Maria Halacheva et al.
In briefControllable multi-label video safety detection using adaptive Tversky policy optimization
SummaryThis paper addresses automated detection of harmful content in social media videos, noting that existing systems often treat the task as binary classification and use fixed training objectives that cannot adjust the precision-recall balance. The authors propose ATPO, a reinforcement learning framework, with an Adaptive Tversky Reward that dynamically adjusts false-positive and false-negative penalties during training. Experiments on SafeWatch-Bench and XD-Violence report that ATPO substantially improves multi-label performance, raising the Jaccard Index on SafeWatch-Bench-Real from 40.66 to 75.44.
In briefStudy on language drift during RLVR post-training of LLMs
SummaryThis paper examines why large language models develop unusual, non-standard language in their chains of thought during reinforcement learning with verifiable reward (RLVR) post-training. The authors prove theoretically that RLVR optimization pressure permits unbounded language drift, while supervised fine-tuning does not. Experiments show drift arises mainly on novel reasoning tasks where the target behavior cannot be drawn from the base model. They also prove that constraining drift is impossible without constraining expected reward.
In briefFrom Atoms to Processes: How AI and Machine Learning Are Used in Chemical Engineering
SummaryThis perspective reviews recent AI and machine learning methods and representative applications across chemical engineering, including atomic-scale simulation, materials and catalyst discovery, and industrial operations. It highlights a shift from purely black-box models toward hybrid, physics-informed frameworks that respect conservation laws and thermodynamic consistency. The authors conclude that AI and ML amplify, rather than replace, the core principles of chemical engineering.
Michael Baldea, Linda J. Broadbelt, Marianthi G. Ierapetritou et al.
In briefBranchIP: Learning adaptive computation to speed up interatomic potentials
SummaryThis paper presents BranchIP, a single-model framework that learns adaptive tensor product computation for equivariant machine learning interatomic potentials, trained with a novel distillation loss. On a heterogeneous catalysis system and a proton-conducting solid acid electrolyte, it reports speedups of up to 2.4× and memory reductions of up to 2.6× across model sizes while maintaining physical fidelity. The learned adaptive computation also indicates which interactions need deeper computation, offering some interpretability.
In briefSPIRIT: Physics-based 3D reconstruction for sparse radar imaging
SummarySPIRIT is a physics-informed inverse rendering framework for sparse-aperture synthetic aperture radar (SAR). Instead of enhancing formed images, it reconstructs an explicit three-dimensional surface by modeling visibility, phase coherence, and wave propagation. Experiments on synthetic and real-world data report accurate geometry and signal recovery with over 90% less data collection, outperforming baselines.
In briefBellman Meets Lyapunov: Unsupervised Reinforcement Learning Through Mastering Chaos
SummaryThis paper proposes Forward CIP (F-CIP), an unsupervised reinforcement learning objective that generates intrinsic rewards from agent-environment interaction without requiring humans to design task-specific reward signals. F-CIP is defined solely by the system's dynamics, so it needs no selection of information variables. Agents trained with it discover basic behaviors such as balancing and maintaining controllability, and combined with a simple forward-velocity reward they learn coordinated gaits like hopping and running.
Tristan Shah, Wooyoung Chung, Volodomyr Makarenko et al.
In briefTesting whether AI image watermarks survive latent frequency masking attacks
SummaryThis paper introduces Latent Frequency Masking, an attack that erases watermark evidence by replacing selected Fourier coefficients in a watermarked image's latent representation. The authors evaluate it against six diffusion watermarking methods on images generated from DiffusionDB and MS-COCO prompts. The attack removes or substantially weakens several watermarks while preserving perceptual quality, and the authors argue such attacks should be included in robustness evaluations of generative image watermarks.
Kirill Aistov, Khaled Abud, Irina Serzhenko et al.
In briefCounting Moves, Weighing Voices: Bayesian Dialectical Argumentation for Calibrated Multi-LLM Councils
SummaryThis paper proposes Bayesian Dialectical Argumentation (BDA), a method for aggregating answers and confidence from a council of several large language models. It treats the typed moves agents make, such as proposing, challenging, or conceding answers, as observations from an annotator model with per-agent reliabilities, and infers each agent's reliability from the deliberation trace. BDA yields calibrated posterior probabilities without additional LLM calls and remains more robust when some agents are persistently adversarial.
In briefMem++: A non-destructive memory system that keeps all document versions for organizational AI agents
SummaryThis paper introduces Mem++, a memory framework for LLM agents handling long-term organizational work, where decisions change across documents over time. Mem++ stores every document whole with its date and author, avoids generative models at write time, and at read time retrieves only documents dated up to the question's time, fusing lexical and semantic rankings. On the OrgMemBench benchmark it outperforms the strongest memory baseline by 8.0 to 13.1 points across two answering models. It also achieves the best average LLM-judge score on LoCoMo and ranks second on LongMemEval-S.
Ahmad Yehia, Aly O. Abdelkareem, Islam Ahmed et al.
In briefMingbird: A local-first agent harness helping small open models finish real tasks
SummaryThe paper argues that small open-weight models (2–9B) often fail real tasks under cloud-scale agent harnesses, and that much of this failure comes from the harness rather than the model. The authors present Mingbird, a local-first agent harness for Windows and Ollama, with ten mechanisms that compensate for small-model failure patterns. In a controlled comparison of 4 harnesses, 4 open models, and 18 real tasks, Mingbird scored 0.886 overall versus 0.631, 0.479, and 0.405 for the other harnesses. The authors note limits including a self-built benchmark, a single machine, and single-trial scoring.
In briefDomain-adapted weather recognition for street-view images, using a weather-conditioned adversarial method
SummaryThis paper addresses recognizing multiple weather types from street-view images when most training data come from non-street-view sources. The authors propose WA-ADDA, a method that conditions the domain discriminator on predicted weather to learn features that are both domain-invariant and weather-sensitive. They also build a multi-dataset benchmark with a standardized protocol using macro accuracy. Across ResNet-50, EfficientNet, VGG, and DenseNet backbones, the method consistently improves street-view performance.
In briefFixing Spatial Reasoning Errors in Video Models Without Retraining
SummaryThis paper argues that overall benchmark scores hide why vision-language models fail at spatial reasoning, and it identifies four recurring error sources by comparing predicted and ground-truth spatial context. The authors introduce CROSS, a training-free library of typed geometric operators and spatial skills. On ReVSI, the average score rises from 55.9% to 60.2%, and on DSI-Bench the SpatialClaw result improves from 62.8% to 66.3%.
Pengzhan Sun, Junbin Xiao, Ramanathan Rajaraman et al.
In briefGradient boosting versus GOES FDC for wildfire detection
SummaryThis study trains a CatBoost model on GOES ABI satellite imagery, using over 300,000 matching VIIRS fire detections as reference data. Tested on separate data from five regions, the model outperformed the operational GOES Fire Detection and Characterization (FDC) product in precision, recall, and F1 score, both inside and outside the training area. The model also detected fires at night, where GOES FDC recall was very low, around 0.03.
In briefContinual concept erasure in diffusion models by limiting interference between edits
SummaryThis paper studies removing unwanted concepts, such as copyrighted styles or private identities, from text-to-image diffusion models when erasure requests arrive over time. The authors observe that existing methods let residual perturbations accumulate across edits, which can degrade unrelated generations and collapse previously erased targets into noise. They propose CEASE, a training-free method that adds two subspace constraints to a closed-form solver, and a closed-form decomposition attributes the interference to two sources that these constraints suppress. Experiments on continual erasure of celebrities, artistic styles, and instances report that CEASE gives the most consistent balance between erasing targets and preserving other content.
In briefUniversal Byte-Level Encoding (UBE): Routing UTF-8 and UTF-16 to Narrow Cross-Script Token Budget Gaps
SummaryThe paper notes that UTF-8-based byte-level BPE tokenizers cost more for many scripts than for English, raising token counts and shrinking usable context. It proposes UBE, which keeps 1-2 byte UTF-8 characters on the UTF-8 path and routes 3-4 byte characters through UTF-16, lowering the encoding floor for high-premium scripts while preserving exact decoding. Reported experiments show UBE matches BBPE's language model quality and reduces token counts most for high-premium scripts.
In briefUniversal interpolation with deep residual self-attention networks
SummaryThis paper studies whether deep self-attention models can interpolate between arbitrary collections of token sequences using depth rather than width. The authors show that two frozen single-head attention blocks with Gaussian-initialized projection matrices suffice, with only the order, signs, and durations of their application depending on the task. The result holds for both continuous and finite depth. The paper also examines the restrictions imposed by causal masking and gives corresponding universal interpolation guarantees.
WhyProves universal interpolation using only two frozen single-head attention blocks in deep residual softmax attention, a strong theoretical result with mainly theoretical impact.
In briefCurvature of Regret in Contextual Linear Optimization: A Theory for Decision-Focused Learning
SummaryThis paper studies decision-focused learning for linear optimization, where small cost errors can leave the optimal decision unchanged or shift it to a different vertex, creating discontinuities. The authors show that after averaging over the data distribution, this behavior becomes locally quadratic, and they derive the curvature in closed form as a matrix-valued measure on the walls of the normal fan. They propose a tractable approximation requiring one projection onto the feasible set and prove it weakly converges to the true curvature. Experiments on battery arbitrage report a 30.8% regret improvement over uniform allocation.
WhyShows regret becomes locally quadratic after distributional averaging with a closed-form curvature measure; novel theory, supported by convergence and battery-arbitrage experiments, useful for a subfi
Konstantinos Ziliaskopoulos, Alexander Vinel, Alice E. Smith
In briefJoint Simulation-Experiment Training for Balanced Physical System Prediction
SummaryThis paper studies how to learn spatiotemporal physical systems using both simulation and experimental data. The authors compare simulation-only, experiment-only, sequential Sim→Exp fine-tuning, and joint training on four fluid systems from RealPDEBench. Joint training gives the most balanced performance across a broad range of evaluation weightings and retains simulation-domain accuracy much better than Sim→Exp.
Mahindra Rautela, Alexander Scheinker, Ayan Biswas et al.
In briefToken-level reinforcement learning for video generation (TVRL)
SummaryThis paper proposes TVRL, a reinforcement learning framework for video generation. Standard methods assign one reward to a whole video, which cannot pinpoint flawed regions. TVRL uses a frozen vision-language model to derive credit for each video token, so optimization targets the tokens that need correction. On VBench-2.0, TVRL reaches an Overall score of 57.69, 3.60 points above the base model, and it also improves matched GRPO baselines across several samplers and reward models.
In briefRASteer: Keeping Other Concepts Intact While Erasing Targets in Diffusion Models
SummaryThis paper proposes RASteer, a training-free method for removing a target concept, such as a copyrighted style, a character, or unsafe content, from a pretrained text-to-image diffusion model. It builds a subspace from concepts to preserve and removes the overlapping components from the erasure direction, then uses an overlap-adaptive calibration to balance erasure strength against preservation. Experiments on unsafe-content, instance, and artistic-style erasure show it matches or outperforms the evaluated activation steering and weight editing baselines.}
In briefFastCI: Making GPU-Heavy CI Faster for LLM Training Frameworks
SummaryThis paper presents FastCI, a framework that makes continuous integration more efficient for LLM training frameworks, whose GPU-intensive tests often slow development. It uses runtime evidence to select affected tests, prune redundant ones, prioritize high-risk tests, and optimize test workloads. On the CI workload of an LLM training framework, FastCI reduced CI latency by 77.5% and GPU usage by 63.9%, while improving modified-code coverage retention by 3.2%. It has been integrated into ByteDance's CI pipelines.
In briefAuditing Bias in Open-Source LLMs for ER Triage
SummaryThis paper compares ten open-source large language models on predicting pediatric Emergency Severity Index (ESI) levels. The researchers built paired clinical cases that change only one demographic, socioeconomic, or other variable while keeping the clinical presentation the same. They found that sensitivity to these changes varied widely and did not consistently decrease with larger model size or medical-domain pretraining. The fine-tuned Qwen2.5-7B showed the lowest sensitivity.
Manar Aljohani, Brandon Ho, Kenneth McKinley et al.
In briefSIEVE: A Method for Selectively Unlearning Personal Information in Vision-Language Models
SummaryVision-language models can link faces to biographical details, so there is a need to remove personally identifiable information while keeping permitted knowledge about the same person. The paper proposes SIEVE, which suppresses attention values toward zero for examples to forget, while matching retained examples to a frozen reference model, combined with sequence-level forget and retain supervision. Experiments report state-of-the-art unlearning across multiple model and modality settings with competitive retained utility. Ablations indicate that value suppression and negative cross-entropy provide complementary forgetting signals, and reference-based matching reduces utility loss.
Si Qi Goh, Cap Dang Xuan Kiet, Tat-Jen Cham et al.
In briefTraining-free diffusion motion planning using analytical local scores
SummaryThis paper presents a diffusion-based motion planner that requires no training, replacing learned global trajectory scores with analytical local scores based on obstacles, smoothness, velocity, and inter-agent feasibility. The authors observe that a trajectory's score can be reconstructed from local interactions between neighboring waypoints and nearby constraints, yielding a decomposed denoising procedure. Experiments in complex environments and large multi-agent tasks report smooth, feasible trajectories, such as paths for over 300 agents with more than 100 obstacles in under 6 seconds on a GPU.
Michael Y. Fatemi, Jinhao Liang, Ferdinando Fioretto
In briefEndoLive: Real-time style transfer converting cadaver endoscopic skull base surgery video to live-patient appearance
SummaryThis paper presents EndoLive, a framework that performs real-time style transfer between cadaveric and living human endoscopic surgical video. It combines the ConStructS GAN for realistic surgical style transfer with the HyPER-GAN model for fast complex translation, trained on unpaired cadaveric and live endoscope images. Experiments report translation speeds well above the minimum required for real time while preserving the semantic consistency of critical anatomical structures.
In briefDo Your Own Research: Learning to Forecast by Learning to Search
SummaryThis paper introduces an agentic forecasting environment, dataset, and harness built from over 2,100 resolved Polymarket questions. During rollouts, the agent gathers its own evidence through web search, page reading, and financial time series, with leak filtering restricting information to before each question's cutoff. Training Qwen3.5-35B-A3B with GRPO under a Brier-score reward improves calibration by 30–40% and reduces search attempts per rollout from 3.8 to 2.25. The trained model outperforms four frontier models, including Claude Opus 4.5, in the same harness at about 5% of the inference cost.
In briefPrecise non-asymptotic analysis of the penalized challenger in β-EB-TCI for Bernoulli bandit identification
SummaryThis paper analyzes β-EB-TCI, a top-two algorithm for fixed-confidence best-arm identification in Bernoulli bandits. It proves that once the empirical leader is the true best arm and its sampling fraction stays near β, the stopping time matches T*β(μ)log(1/δ) up to lower-order terms. The results give a non-asymptotic high-probability bound and sharp expected sample complexity under a unique best arm.
In briefLatent JEPA: Predicting Abstract Futures to Guide Chemical Reasoning in Latent Space
SummaryThis paper introduces Latent JEPA, a framework that combines autoregressive learning with joint-embedding prediction of one or more future views. It trains continuous latent thoughts to anticipate informative aspects of chemical solutions without verbalizing every intermediate step. On ChemCoTBench, the authors report gains in molecular optimization and several editing and reaction metrics. Representation analyses indicate the latent thoughts become more informative about molecular outcomes and better aligned with chemical structure.
In briefAnti-Persona: Blocking Unauthorized Identity Linking and Recognition in Personalized Vision-Language Models
SummaryThe paper notes that personalized vision-language models can link a person's identity from a few reference images and then recognize that person in new images, creating a privacy risk. The authors propose Anti-Persona, which adds subtle image perturbations that disrupt identity prototype alignment in the vision-encoder space. Experiments on two representative personalized models report protection rates of up to 95.0% while preserving visual fidelity.
In briefLatent-Foresight: End-to-end learning of predictable representations for latent world models
SummaryThis paper proposes Latent-Foresight, a framework that jointly trains a latent tokenizer and a flow-based generative dynamics model so the representation is shaped for temporal predictability. The authors report that it learns more temporally coherent latent representations and outperforms two-stage baselines across several future scene understanding tasks and prediction horizons. The approach removes separate training stages, including during high-resolution adaptation.
In briefWhen can an LLM calibrate simulators? Physics-labeled prompts for budget-limited calibration
SummaryThis study tests whether large language models, given physics-based descriptions of simulation errors, can propose parameter updates more efficiently than standard derivative-free optimizers. It evaluates the approach on a cylinder-wake large-eddy simulation and a damped-oscillator inverse problem. Physics-informed prompts yield a large early advantage that fades by fifteen evaluations, and the authors propose a hybrid method that uses the LLM as a warm start before a conventional optimizer, reaching target accuracy in 9 simulator runs.
In briefFewer Tokens, Better Action: GPT-6 Astra Robot Agents Gain 14% Success With 65% Fewer Tokens
SummaryThis paper presents PyRUA-Lean, a framework where a vision-language model agent composes robot primitives, including vision-language-action policies, into Python cells that check and retry locally and return only needed images and state. Across 700 simulated tasks, it raises success from 63.1% to 71.7% over tool calling with the same planner, using 49% fewer LLM calls and 65% fewer input tokens on jointly solved instances. The authors also train an open 4B model, PyRUA-Lean Coder v0.1-4B, which improves from 18.8% to 44.1% on unseen task instances.
In briefA Map of Rubrics for Judging LLM Clinical Reasoning: What Exists and What's Missing
SummaryThis structured narrative review surveys medical education instruments, clinical LLM benchmarks from 2023 onward, and general-domain long-form evaluation methods. It examines six dimensions of clinical reasoning, finding that no single instrument covers all of them and that faithfulness is the weakest. The authors propose combining existing tools with binary rubric items, separate completeness and correctness scores, and chance-corrected reliability reporting. The paper offers a design rationale rather than a validated instrument.
In briefGraph Representation Learning Using Discrete Morse and Cobordism Theory
SummaryThis paper brings tools from low-dimensional topology, namely discrete Morse theory and cobordism theory, into machine learning. It proposes the MG-Diff pipeline to improve graph diffusion models. The authors also provide theoretical guarantees showing the Morse-theoretic tools and their diffusion guidance are stable under small perturbations when a positive decision gap holds. Applications to spatio-temporal graph forecasting and graph regeneration are presented.
Jennifer Rozenblit, Chenguang Yang, Yuxin Liu et al.
In briefRethinking Knowledge Retrieval for Generation: A Survey of RAG Architectures and Applications
SummaryThis survey reviews Retrieval Augmented Generation (RAG), a method that combines large language models with external retrieval so that generated answers are grounded in verifiable sources. It describes the three core components of retrieval, generation, and augmentation, and surveys retrievers, fusion strategies, embedding optimizations, and reinforcement learning based retrieval policies. Recent work is grouped around efficiency, security, user-centric interactivity, and complex reasoning, alongside evaluation protocols, domain applications, and architectural variants. The paper concludes by identifying persistent challenges and future research directions.
Meghana Sunil, Shravya V, Shravan Venkatraman et al.
In briefLearning Distributions over Weight Updates for Test-Time Adaptation
SummaryThis paper studies how to generate LoRA parameter updates for a large language model using only the input query, and introduces distributional hypernetworks that output a distribution over possible updates rather than a single point estimate. The authors train them with a differentiable Monte Carlo loss and compare several distribution parametrizations. Results show the distribution's mean can outperform deterministic hypernetworks, sampling more weight updates improves performance, and generated updates can transfer across queries.
Azal Ahmad Khan, Keshav Ramji, Tahira Naseem et al.
In briefError-corrected sampling method for imperfect diffusion models
SummaryThis paper addresses the problem that pretrained diffusion models are often imperfect, which causes existing inference-time scaling methods to inherit their errors. It introduces the Energy-based Feynman-Kac Corrector (EBFKC), which corrects errors on the fly using a reference energy, combining Feynman-Kac dynamics with sequential Monte Carlo and variance-controlling guidance. Experiments on Gaussian mixture models, particle systems, and alanine peptides show that EBFKC more closely matches target distributions and free-energy profiles than standard baselines, which retain substantial sampling errors.
WhyProposes Feynman-Kac correction for imperfect energy-based diffusion models, outperforming standard baselines on Gaussian mixtures and alanine peptides.
In briefCLoSeR: Loop Closure for Drift-Free Long-Sequence 3D Reconstruction
SummaryFeedforward 3D reconstruction models suffer large tracking drift over long streaming sequences because errors accumulate. The method detects loop candidates via global descriptor retrieval and estimates relative poses between looped frames using loop-conditioned windows. All frame poses are optimized on the SE(3) manifold with sequential and loop closure constraints. Experiments on kilometer-scale sequences show reduced drift and consistent geometry, outperforming prior state-of-the-art methods.
In briefLAST: A looped audio transformer that reuses its blocks to refine classification at lower cost
SummaryThis paper proposes LAST (Looped Audio Spectrogram Transformer), which processes all audio tokens once, then reuses the same blocks to refine only the class token over fixed features. On AudioSet, ten-pass LAST reaches 0.345 mean average precision, 2.1% relative above a twelve-layer sequential transformer, with 49.4% fewer parameters, 42% fewer multiply-accumulate operations, and 9.8% higher measured throughput. The authors also report improved robustness to temporal masking and other auditory augmentations, and better generalization on music, environmental, and event sound classification.
Haider Al-Tahan, Sean O'Brien, Anastasia Razdaibiedina et al.
In briefAligning languages in decoder-only LLMs by matching MoE router outputs
SummaryThis paper addresses the difficulty of explicit cross-lingual contrastive learning in decoder-only LLMs, where varying multilingual tokenization limits alignment. Instead of an auxiliary loss on hidden states, it proposes using mixture-of-experts (MoE) router outputs as the alignment target, since they pool more reliably over tokens. Controlled continual pre-training on four open-source MoE models shows the routing loss also aligns hidden representations across languages and improves multilingual performance.
Lucas Bandarkar, Clark Peng, Ahmed Haj Ahmed et al.
In briefMoLE: Mixture of Latent Experts for Complementary Visual Reasoning
SummaryThis paper addresses latent visual reasoning in vision-language models, where multiple latent tokens often access the same visual evidence and yield redundant representations. The authors propose MoLE, which isolates latent visual experts so each extracts complementary visual information, with dedicated summary experts aggregating their outputs. Across five visual reasoning benchmarks, MoLE averages 78.6, outperforming data-matched supervised fine-tuning and the strongest latent baseline at the same latent budget.
In briefDecomVoxel: Splitting a scene into objects and background using 3D-aware generative guidance
SummaryDecomVoxel is a method for reconstructing objects and background separately from scenes, especially under heavy occlusion. It frames object completion as guided in-situ denoising optimization, combining 3D-native priors with neural scene reconstruction. The approach adds a reformulated epsilon-based distillation loss and adaptive spatial guidance with temporal annealing. The authors report better results than prior methods on Replica and ScanNet++ while preserving the original spatial layout.
In briefSame Reward, Different Skills: When Multimodal RL Learns to Look
SummaryThis paper examines how reinforcement learning with verifiable rewards (RLVR) affects vision-language models. The authors propose a design rule, visual resolvability, requiring that visual evidence be necessary for a correct answer while the task stays learnable. On counterfactual coordinate scenes, a 7B model raised its target-discovery accuracy from 0.425 to 0.875, while training on gray canvases yielded essentially none of this gain. The results indicate that what the reward requires determines what RL learns.
In briefMapLightning: Online Vectorized HD Map Construction with Compact 1D Map Tokens
SummaryThis paper proposes MapLightning, a method for building vectorized high-definition maps online for autonomous driving. Instead of dense bird's-eye-view grids, it uses a compact set of 1D learnable map tokens that are updated jointly with image features via self-attention. The authors report state-of-the-art accuracy and efficiency on nuScenes and Argoverse 2, with the lightweight variant outperforming MapTRv2 and running 1.73 times faster using 53% less memory.
In briefHigher-order positional encodings that add topological structure to graph learning models
SummaryThis paper proposes a representation learning approach that enriches graph representations with higher-order positional encodings, so standard graph models can use higher-order structure without changing their backbone. The authors give a theoretical account of the expressivity of these encodings and implement them with Hodge Laplacians derived from clique complexes. Experiments on ZINC and controlled synthetic benchmarks report improved predictive performance with Graph Transformers.
In briefAsynchronous LLM Post-Training: Group-Mass Capping and Convergence Analysis
SummaryThis paper studies how stale rollouts generated by earlier policies affect convergence in asynchronous reinforcement learning for large language model post-training. It derives a convergence bound for GRPO-style algorithms and proposes group mass capping GRPO (GMC-GRPO) to reduce bias from delayed data. Experiments on Qwen3 models and reasoning benchmarks report improved robustness to large rollout delays, with GMC-GRPO performing best among stable baselines.
In briefA foundational study systematically analyzing Transformers and RNNs for brain fiber tractography
SummaryThis paper systematically evaluates recurrent neural networks and Transformer models for iterative tractography, a method for mapping white-matter fiber pathways in diffusion MRI data. The authors introduce a generation-validation phase that allows supervision at the streamline level during training, and their models reach the highest reported performance on the ISMRM2015 tractography challenge dataset. Controlled experiments quantify the effects of missing bundles, noisy training streamlines, and invalid fibers, and the best models are also applied to in vivo data from the Tractoinferno database.
Emmanuelle Renauld, Philippe Poulin, Hugo Larochelle et al.
In briefSelection-Based Structured Reasoning: Making Multimodal Search Agents Faster
SummaryThis paper proposes Selection-based Structured Reasoning (SSR), which replaces free-form reasoning in multimodal agents with choosing among pre-specified reasoning candidates ranked by their likelihood in the current context. It was evaluated on seven multimodal search benchmarks using 2B and 4B models. According to the authors, SSR keeps task performance competitive with leading search agents of the same scale while cutting per-turn reasoning latency by over 90% and total per-question inference latency by 28-54%.
In briefUnsupervised domain adaptation with conditional flow matching to improve rain estimation from radiometer images
SummaryThe paper addresses the scarcity of labelled data and domain shifts in satellite radiometer imagery. It proposes an unsupervised method that uses parts of the deterministic ordinary differential equations in flow matching models, conditioned on different satellite instruments, to align domains. Experiments on the GPM-Core constellation show that this conditional domain adaptation improves rain precipitation estimation from radiometer images.
Victor Enescu, Assaad Zeghina, Matthieu Meignin et al.
In briefStochastic vs. nearest rounding in low-precision transformer inference
SummaryThe paper studies whether low-precision transformer inference should use stochastic rounding or round-to-nearest, varying only the rounding rule at individual operation sites while keeping the numerical format fixed. It extends the PRISM rounding library to arbitrary virtual precision and derives a forward-error bound for linear projections, along with a decomposition of cross-entropy loss change at the output softmax. On DistilGPT-2 at 6 significand bits, SR in the MLP yields 1.15x the full-precision perplexity versus 2.21x for RN, while the ordering reverses at the language-model head. A mixed configuration using SR in the MLP and RN in the head reaches 1.10x of full-precision perplexity, a 28% reduction over matched-bit RN.
Yohan Chatelain (1), Pablo de Oliveira Castro (2) ((1) Krembil Centre for Neuroinformatics, CAMH et al.
In briefTRACE: Agentic Heuristic Design for Real-World Resource Assignment
SummaryThis paper addresses dynamic resource assignment, which allocates task streams to processing nodes in computing infrastructure. It proposes TRACE, which pairs an evolutionary LLM-based heuristic design loop with agents that analyze system logs, form hypotheses, and write code to test them. The authors evaluate TRACE on a synthetic cloud benchmark and a 5G vRAN scenario, reporting that it outperforms state-of-the-art methods while adding under 2% overhead.
Jose A. Ayala-Romero, Andres Garcia-Saavedra, Xavier Costa-Perez
In briefGenerating B-roll shots grounded in a user's own video collection
SummaryThe paper introduces MemComposer, a three-stage system that creates multi-shot B-roll sequences from a user's video collection, guided by a natural-language directive and target duration. It builds an entity-centric visual memory offline, then plans and retrieves reference frames for each shot, and iteratively generates and critiques the sequence for consistency. In a user preference study, it outperformed an ungrounded text-to-video planner on prompt adherence and visual alignment.
Cusuh Ham, Fabian Caba Heilbron, Josef Sivic et al.
In briefFaster multi-agent coordination via one-step flow policies (OMAF)
SummaryThis paper proposes OMAF, an online multi-agent reinforcement learning framework that uses a one-step flow model to generate actions without iterative sampling. It employs a Transformer-based flow policy to capture complex coordination behaviors, optimized jointly with softmax Q-value estimation. Across 10 standard MPE and MAMuJoCo tasks, the authors report up to 3.4x higher returns and 10.5x better sample efficiency than baselines.
In briefEvenSplat: A Gaussian Splatting method that separates lighting effects from scene geometry under uneven exposure
SummaryThis paper presents EvenSplat, a framework for 3D scene reconstruction from multiple photos taken under changing or uneven lighting. It links an image-space lighting decomposition with a lighting field carried by the Gaussians, and uses a camera-response network and a local exposure-compensation module to handle remaining differences between training images. Experiments on multiple datasets and several kinds of uneven illumination report that EvenSplat generally outperforms state-of-the-art methods, especially under high-contrast lighting.
In briefWhere LLMs Struggle When Writing Visualization DSL Code
SummaryThis paper studies where and how large language models fail when they write charts in visualization domain-specific languages (DSLs). The authors evaluate 10 JSON-style visualization DSLs on 41 tasks using 3 LLMs, checking outputs with JSON and rendering tests plus qualitative coding of failures. They identify four recurring failure patterns, link each to specific DSL features, and discuss design considerations for future DSLs.
In briefFrom Pixels to Policy: A Multi-Agent System for Urban Intervention Design
SummaryThis paper introduces VIDA-Geo, a multi-agent system that searches for interventions to improve urban indicators in aerial and street-view images. It coordinates segmentation, diffusion-based inpainting, and indicator scoring models to produce edits that are perceptually realistic and aligned with real-world policies. The authors evaluate it on 8 indicators and report outperforming baselines in many cases. The system offers multiple candidate interventions for expert city planners to review.
In briefFrom Isolated Features to Orbits: Finding Music Concepts by Aligning Multiple SAEs
SummaryThis paper studies how to understand what a music foundation model has learned internally, shifting focus from isolated features to structured groups of features. The authors generate pitch-shifted input pairs and align their Sparse Autoencoder representations to find ordered orbits of pitch-related features. Experiments on two state-of-the-art music foundation models show that the method recovers structures corresponding to chords, keys, and melodic patterns, needing only a few anchor examples to interpret whole concept families.
