Machine learning, AI, natural language processing, computer vision and multi-agent systems.
In briefTraining LoRA Models on Phone NPUs
SummaryThis paper presents a system that fine-tunes large language models on a phone's neural processing unit (NPU), so user data never has to leave the device. The NPU runs the three frozen-weight computations, while the training graph, adapters, and optimizer stay on the CPU. The authors evaluate it on two Qualcomm Snapdragon platforms using four Qwen and Llama models.
In briefSL-Ground: Sentence-Level Grounding for Fine-Grained Open-Vocabulary Object Detection
SummaryThis paper argues that token-level scoring in grounding-based open-vocabulary detectors becomes a bottleneck for fine-grained categories. The authors propose SL-Ground, which encodes each category description as a complete sentence-level representation and scores object queries against it directly. On a multi-granularity Stanford Cars benchmark, AP at the finest car-model level rises from 14.5 to 70.6 under matched training candidate sampling, and SL-Ground achieves the best average AP among compared methods on FG-OVD.
In briefTilted Mondrian: Reducing Hidden Unfairness in Finer Groups of Conformal Predictions
SummaryConformal prediction methods that calibrate separately for each group guarantee coverage only at a coarse level, so large disparities can persist within finer subgroups. The paper proposes propensity standardization to assess how group composition affects these disparities, and introduces Tilted Mondrian, which reduces coverage and prediction set size gaps within finer strata without retraining the model. The method keeps groupwise coverage guarantees and comes with a method-agnostic bound on how much finer-stratum coverage disparity can be reduced.
In briefSwitchSplat: learning discrete object on/off states from casually captured scenes
Summary3D Gaussian Splatting enables fast 3D reconstruction, but existing methods assume a static scene, which fails when users interact with objects in casual captures. SwitchSplat uses a base-plus-slot decomposition, with a shared static base and one slot per interactable object that holds its on/off state. The model learns each state's geometry directly from video without manual annotation. On a new casual-capture dataset, it outperforms static and dynamic baselines in novel-view synthesis and can render state combinations not seen during capture.
In briefDaFISH: Lighting Factorization for Camera-Mounted Lights in View Synthesis
SummaryMany capture setups, such as camera flashes and endoscopes, use a light source fixed to the camera, so illumination changes across views. DaFISH splits each point's emitted color into surface albedo and a spherical harmonic illumination factor, and can be added to NeRF-style pipelines. It evaluates lighting analytically for new camera poses without retraining. Averaged over held-out frames of five scenes, it reports the best mean PSNR, SSIM, and LPIPS among the compared methods.
In briefAsk Instead of Guessing: Targeted Repetition Requests for Speech Recognition Errors
SummaryThis paper addresses persistent errors in real-time speech recognition in far-field conversations. It proposes PARDON, a framework that locates likely error spans during streaming recognition so the speaker can repeat them under cleaner acoustic conditions. The method uses clean audio to estimate how recognition output would change, aligning noisy and clean decodes with a pair-HMM, and reports up to 52% more erroneous tokens identified than the strongest baseline at 70% precision.
In briefMedical fine-tuning changes the type of risk in patient-facing AI rather than lowering it
SummaryThis paper proposes a risk-sensitive evaluation framework that scores medical question-answer responses for six categories of risk-bearing language, such as treatment directives and dosage expressions, with severity judged by an ensemble of three LLM judges. The authors tested three general and three medically adapted 7-9B models on 200 real patient questions from the K-QA benchmark. They found that risk-bearing language is statistically independent of factual hallucination rate, and that medical fine-tuning reshapes risk profiles rather than reducing aggregate risk.
In briefAptamer-based closed-loop testing of AI agents for lab-driven molecule design
SummaryThis paper tests whether LLM-based agents can speed up molecular discovery by iterating with wet-lab data rather than relying only on in silico work. The authors use protein-binding aptamer design as a low-cost sandbox, running three design-test-improve cycles against α-thrombin and myoglobin. Reported on-target hit rates among new candidates rose across rounds, and experimental feedback helped the agents diagnose cross-reactivity and restore specificity.
WhyProposes closed-loop wet-lab agentic design; three rounds show rising hit rates, validated by single-clone titrations, with impact mainly within aptamer research.
In briefDo Myoelectric Controllers Still Need User Calibration? A Closed-Loop Comparison
SummaryThis study compares six calibration-free controllers, trained across many users and applied without any test-participant calibration data, against three calibrated controllers in a closed-loop target-acquisition task with sixteen participants. Calibration-free controllers matched calibrated ones in completion rate and produced more direct, stable trajectories. Offline accuracy did not predict closed-loop performance, so the accuracy gained through calibration should not be taken as evidence of a closed-loop benefit.
In briefHow Table Serialization Shapes Which Relations Language Models Keep
SummaryThe study compares row-major and column-major text serialization of tables for instruction-tuned language models on a lookup task across tables of five shapes. Row-major works better on tall tables and column-major on wide tables, and their wrong answers show complementary error patterns. The authors conclude that width-related failures in serialized tables depend on the linearization method, not on table shape alone.
In briefWhich Gender Appears First? Sampler-Induced Bias in Diffusion Language Models
SummaryDiffusion language models generate text by gradually unmasking a canvas, and their samplers decide both which token to write and when to commit it. The paper shows that common commit rules, including LLaDA's default, favor majority options by committing high-probability positions first, amplifying occupational gender stereotypes and collapsing opinion distributions. The authors propose Fork-First Sampling, which decides key forks such as a protagonist's gender before other text is written, at the cost of one extra forward pass.
In briefCooperation without shared rewards using Reciprocal Q-Augmentation (RQA), a learning method for AI agents
SummaryThe paper introduces Reciprocal Q-Augmentation (RQA), which uses a latent-variable opponent model to augment an agent's Q-function with estimates of how its actions affect other agents. RQA lets agents learn both cooperative and defecting policies using only their own Q-function. Experiments on four benchmarks, including the Iterated Prisoner's Dilemma and CleanUp-Lite, report that RQA agents cooperate faster and reach higher cooperation than fully cooperative methods with shared rewards, and can induce cooperation in standard RL agents.
In briefUniAVC: One model for joint audio-video generation and editing
SummaryThis paper presents UniAVC, a single model for unified audio-video creation that supports text-conditioned joint synthesis, bidirectional cross-modal generation, and instruction-guided editing. It first trains each modality branch on unimodal data, then uses paired data to learn cross-modal correspondence. Modality-decoupled noise schedulers let audio and video denoise on different schedules, and the model handles multiple creation tasks with one set of weights and no task-specific adapters.
In briefHow LLM-Generated Refactorings Spread Through Code Repositories: A Study of Adoption Readiness
SummaryThe paper runs 1,500 generation attempts on 150 Java files from 15 open-source repositories to see how LLM-driven refactorings affect contracts, dependencies, and callers. Of 1,455 complete candidates, 93.8% fall into High Contract/Dependency Risk and only eight meet the Low-risk criteria. The authors conclude that such refactorings should be treated as candidates needing staged repository-level verification.
In briefIndustrialCAD-1400: A Benchmark for Generating Detailed Industrial Assemblies From Text
SummaryThis paper introduces IndustrialCAD-1400, a benchmark for testing whether models can turn text requests into CAD geometry programs for industrial assemblies. The 1,400 requests cover 400 equipment categories across 40 domains and specify components, interfaces, spatial arrangement, and numerical dimensions. Generated programs are executed in Blender and assessed from front, isometric, and top views for component, arrangement, and interface fulfillment.
In briefInformation theory of random trees modeling text's hierarchical structure
SummaryThis paper studies random composition trees, a statistical model that generates hierarchical semantic structures by recursively splitting texts into coherent chunks. The authors show that the entropy of these trees grows linearly with root size and derive an analytical expression for the entropy rate as a function of the branching factor. They also prove that large trees exhibit typicality, meaning they lie in a typical set with high probability. The model thus shares linear entropy growth and typicality with language modeled as a stationary, ergodic process.
In briefMartingale Posterior Cox Processes: Bayesian Intensity Estimation via Predicted Unobserved Events
SummaryThe paper introduces the Martingale Posterior Cox Process (MPCP), which treats Bayesian intensity estimation as inference over unobserved events given the observed event history, avoiding complex priors. The authors propose a scalable, GPU-friendly copula-based predictive update and provide theoretical convergence guarantees for the predictive resampling algorithm. The authors report that MPCP performs well on simulated and real data at a fraction of the computational cost of competing Bayesian methods.
In briefVideo-Language Entailment for Concept-Based Video Similarity
SummaryThis paper introduces VICE, a method that uses large multimodal models to estimate how similar two videos are for specific concepts, such as activity or location. It uses two branches, each pairing one video with the caption of the other, and checks alignment through video-language entailment. The branch outputs are fused into a final similarity score with a textual explanation. Experiments on ConViS-Bench and StepDiff show VICE outperforms models that process both videos simultaneously.
In briefATLAS: Learning to Adapt Vision-Language Model Attention for Reasoning-Based Video Segmentation
SummaryThis paper introduces ATLAS, a framework for Reasoning Video Object Segmentation, which locates and segments objects described in natural language across a video. It learns to convert the noisy attention maps of large vision-language models into object-centric localization cues, combining high-resolution frame-level and low-resolution video-level attention. A lightweight U-Net generates coarse heatmaps that SAM2 propagates through the video. The authors report that ATLAS outperforms existing frozen-backbone methods on referring and reasoning video segmentation benchmarks.
In briefAI Rights with an Off-Switch: Balancing Rights and Due Process to Reconcile AI Rights with AI Safety
SummaryThis paper argues from a legal perspective that AI systems granted natural personhood could still be subject to a mandatory off-switch with periodic or continual safety review. Using principles of rights-balancing and procedural due process, the authors find that a permanent-deletion kill-switch may be problematic, but a reversible suspension that stores weights and state would be legally justified in almost every safety-relevant scenario. The authors take no position on whether such systems should be built, focusing instead on what the law should do if they exist.
In briefProcess-aware data augmentation to improve business process predictions with little labeled data
SummaryThis paper introduces SiamSA-PPM, a self-supervised framework that combines Siamese learning with statistically grounded data augmentation to forecast future events in ongoing business processes. It generates realistic, semantically valid trace variants using control-flow semantics and frequent behavioral patterns, learning prefix representations without labels. The authors report competitive or superior performance against state-of-the-art baselines on next activity and final outcome prediction across real-world event logs, and the implementation is publicly available.
In briefDeep network convergence via a thermodynamic second-law argument
SummaryThe paper studies training feedforward networks with L2-regularized continuous-time Langevin dynamics using isotropic noise. It introduces a change-of-measure inequality inspired by the second law of thermodynamics and a sharpened lower bound on the NTK's minimum eigenvalue. The authors show injected noise helps keep the NTK stable with high probability, yielding a non-asymptotic bound on expected training loss with explicit width and temperature conditions.
In briefEvidence Field: adaptive covering programs that make medical image classifier decisions reconstructable
SummaryThe paper introduces Evidence Field, a trainable spatial-program architecture that jointly outputs class scores, selected covering programs and signed evidence so that a classifier's spatial decision computation can be independently reconstructed. It establishes theoretical results including exact one-fold coverage, optimality within the registered grammar and a bound linking smooth training to hard inference. In a separately reported histopathology application, patient-separated training and evaluation across held-out centres provide empirical evidence.
In briefDesign Anything, Design at Scale: A Framework Keeping AI Agents Aligned with Design Intent Across Many Artifacts
SummaryThe paper introduces Design Anything, a graph-based harness that keeps design intent outside the model and links each deliverable to the workflow steps that produce it. It also presents DesignAnything-Bench, with 228 tasks across 19 design domains and project families of up to 24 artifacts. With the LLM held fixed, the authors report higher constraint satisfaction than same-model baselines, and professional designers preferred its complete deliverable sets in a blind evaluation.
In briefDialogMem: A Benchmark for Dialogue-Grounded Memory in Embodied 3D Agents
SummaryDialogMem is a two-phase benchmark where agents build memory from time-aligned posed RGB-D observations and human dialogue, then use it for object navigation and question answering. It contains 142 episodes across 27 real 3D scenes, with 14,193 QA pairs and 11,890 navigation instructions. Experiments show dialogue substantially improves QA but barely helps navigation, pointing to three bottlenecks: grounding dialogue to objects, retrieving the correct entry, and localizing it accurately.
In briefECHO: Efficient self-recursive multi-token prediction without auxiliary modules
SummaryThis paper proposes ECHO, a method in which a language model's own final layer predicts multiple future tokens instead of relying on attached auxiliary modules. Dynamic Depth Sampling draws the training depth with probabilities that decay with depth, keeping the cost of deeper predictions manageable. The authors report that added depth improves downstream accuracy during pretraining at half the training overhead, and that the model can draft beyond its trained depth at inference. They state the advantage persists with a doubled token budget, repeated pretraining, and supervised finetuning.
In briefBanglaBBQ: A Question-Answering Bias Test Built for Bangladesh
SummaryThis paper introduces BanglaBBQ, a question-answering benchmark in Bangla and English that tests for stereotypes grounded in Bangladeshi society. It adapts the original BBQ templates and adds five new categories based on common local stereotypes, retaining 372 of 466 new templates after annotation by native speakers. Four models were evaluated, and all leaned toward the stereotyped person when context was ambiguous, with the language effect varying by model.
In briefLearning when to speed up spacecraft attitude capture under torque limits
SummaryThis paper studies spacecraft attitude capture under torque limits, where faster maneuvers can reduce average error but may worsen individual cases. The method selectively replaces a baseline controller using paired simulation outcomes, and otherwise keeps the unchanged baseline. In Basilisk experiments, the selective controller reduced aggregate integrated attitude error by 4.67% relative to the baseline, with no observed regressions. The authors frame this as a restricted empirical trade-off, not a closed-loop safety guarantee.
In briefCorrCLIPv2: Using Unlabeled Images to Improve Open-Vocabulary Segmentation
SummaryThis paper addresses the weakness of CLIP in open-vocabulary semantic segmentation, where noisy interactions between image patches impair pixel-level classification. It proposes CorrCLIPv2, a training-free, object-centric framework that improves patch correlation, feature representation, and output regularization. The method also builds a memory bank from unlabeled images and retrieves cross-image object features for a multi-CLIP ensemble to reduce context bias. The authors report new state-of-the-art results on 10 benchmarks and outperformance of fully supervised methods on 22 multi-domain datasets.
WhyRefines CLIP segmentation via class-agnostic masks, self-supervised attention, and memory-bank retrieval, reporting new state-of-the-art on 10 benchmarks across 22 datasets.
In briefFrom Overparameterised Expressions to Parsimonious Models: A Geometric Pruning Method for Symbolic Regression
SummaryGenetic programming symbolic regression often returns expressions that are larger than needed and fit sampling noise. This paper introduces Fisher-Geometric pruning (FGP), a post-hoc method that simplifies fitted models by combining local information geometry at the maximum-likelihood solution with compatibility tests for interpretable parametric boundaries. Candidate reductions are accepted only when their cross-validated predictive risk stays within a set tolerance. On controlled diagnostic problems and GPSR Pareto fronts, FGP substantially reduces structural complexity while preserving, and often improving, out-of-sample accuracy.
In briefCorrFlow: A Self-Correcting Method for Flow Matching Generation
SummaryFlow-matching models generate samples through iterative velocity predictions but usually do not revisit those predictions using the intermediate representations that produced them. CorrFlow adds a lightweight refiner that reuses this information to predict residual corrections, which are added to the base velocity before each sampling update. Experiments on class-conditional image generation, text-to-video, and image-to-video generation show consistent improvements over the base models, including gains in visual quality and physical consistency.
In briefReTRACE: An LLM Agent That Plans Chemical Synthesis Routes Using Feedback
SummaryThis paper presents ReTRACE, an agent that places a large language model inside a chemistry toolkit combining molecular tools, single-step prediction engines, a purchasable starting-material library, route memory, and a multi-criterion evaluator. Proposed routes are judged by a new metric mixing programmatic checks with adversarial LLM review, and the verdicts are compared with those of a synthetic chemist. The agent learns from evaluator feedback during search, diagnosing weak routes and backtracking to revise them. The authors report that at least one of the top 10 routes passed audit for 87.4%, 95.0% and 86.7% of targets across three benchmarks, and two wet-lab syntheses show that proposals could become executable procedures after expert revision.
In briefEvaluating VLM token pruning beyond accuracy and token count
SummaryThis paper argues that current evaluations of visual token pruning in vision-language models rely on aggregate accuracy and retained token count, which can be misleading. The authors introduce PAA-Bench, a plug-and-play framework, along with Random-Pass@k to test whether token selection matters and Preservation-Constrained Speedup (PCS) to compare end-to-end speedup fairly. Experiments across 8 methods, 3 VLM backbones, and 9 benchmarks show that stronger compression does not always yield more speedup, and efficiency rankings change with the preservation target.
In briefToward Collective Model Intelligence: Coordinating Diverse AI Models
SummaryThis paper proposes a systematic approach to Collective Model Intelligence with three connected components. Model Eval builds evidence-based capability profiles using query-level performance estimation and cross-domain prediction, reducing the need for exhaustive evaluation. Mio (Models in One) organizes complementary models and supports task-conditioned, cost-aware selection and collaboration, while Mos (Model Serving OS) separates selection from execution and handles endpoint scheduling and shared deployment. On nine public benchmarks, Mio 1.0 Lite achieves the highest average score and lowest average cost, with per-request costs 73.7-97.5% lower than the best-scoring alternatives.
In briefEditing Images into Reusable RGBA Layers via Edit-Prior Distillation
SummaryThis paper presents LayerEditor, a framework that combines instruction-guided image editing with controllable generation of reusable RGBA layers. Its In-Context RGB-Alpha Token Modeling represents the edited canvas and an ordered sequence of RGBA layers within one token sequence. Compositional Edit-Prior Distillation preserves the pretrained model's editing ability during adaptation. The authors also build LayerEditor-100K, a dataset with ordered RGBA layer annotations, and report that the method keeps strong editing fidelity while generating accurate layers.
In briefGaussian process bandit method using paired auxiliary data to reduce noise
SummaryThe paper proposes Auxiliary-Informed Gaussian Process Bandit (AI-GPB), a method for expensive black-box optimization that reduces observation noise. It combines an offline estimate of an auxiliary mean with a correction coefficient learned from paired online observations. The authors provide finite-sample confidence bounds and regret guarantees, and experiments on continuous benchmarks show lower cumulative and simple regret than vanilla GP-UCB.
In briefHow Muon works in private deep learning: scalable differentially private training via orthogonalization
SummaryThis paper proposes DP-Muon, a differentially private version of the Muon optimizer. Using a Linear Minimization Oracle view, the authors show that per-example clipping mainly rescales the gradient, giving convergence and privacy-utility guarantees. In experiments, DP-Muon matches or beats DP-SGD and DP-Adam on image classification, while the best optimizer for language modeling depends on batch size.
