In briefHiPPT: Learning Reusable Wearable PPG Representations with Multi-Scope Text Alignment
SummaryThis paper introduces HiPPT, a framework for learning reusable representations from wrist-worn PPG signals that matches each supervision source to the waveform evidence scope it suits. It builds a PPG-Text corpus of over 4 million recordings from 202K users, aligning individual recordings with pulse morphology and recordings sharing metadata with context. The authors report the best primary linear-probing results across six downstream tasks, including blood pressure and heart-rate estimation, and zero-shot profiling of age group, sex, and BMI category.
In briefAlignPoison: Conditional targeted poisoning of generative recommenders using decoding-aligned prefixes
SummaryThis paper studies how vulnerable generative recommendation models based on autoregressive decoding are to data poisoning. Under a black-box setting where an attacker can inject limited malicious interactions and query the recommendation API, injecting fewer than 1% synthetic interactions significantly raises the chance that malicious targets are promoted under a trigger. The authors propose AlignPoison, a framework that finds context-aligned prefixes from API outputs, and report attack success rate gains of up to 16.3 percentage points over baselines on unimodal and multimodal benchmarks.
In briefControlled LLM workflow for explaining risk score gaps with evidence
SummaryThis paper proposes a controlled large language model workflow that explains differences in risk scores using similar historical transactions. The workflow combines reference retrieval and contrastive SHAP evidence with reference-based LLM reasoning and deterministic validation. Experiments on a test dataset show it yields more consistent, evidence-grounded, and traceable diagnoses than the same LLM without structured workflow guidance.
In briefOptimize Once, Execute Fast: Latency-Aware Multi-Agent Workflow Learning for Recurrent Queries
SummaryThis paper addresses the latency that builds up when LLM-based multi-agent workflows are reused for recurring queries, where most automated design methods only optimize downstream performance. It defines latency-aware workflow generation, which aims to minimize workflow makespan while retaining task performance. The authors propose LAWA, which reduces makespan through graph edits targeting critical-path bottlenecks and supports both new workflow generation and refinement. Results on seven benchmarks reportedly show significant makespan reduction with competitive task performance.
In briefAphaVoice: A Text-to-Speech Model for Simulating Aphasic Patient Speech
SummaryThis paper introduces AphaVoice, a text-to-speech framework designed to synthesize realistic speech from people with aphasia. Compared with state-of-the-art TTS systems, it better captures dysfluency and performs better on temporal metrics while keeping speaker identity. By shifting the latent space, the model can also generate multiple speakers while preserving phonetic realism. The authors position it as a scalable tool for medical training and clinical simulation related to aphasia.
Al Varrel Putra Kusuma, Fernando Bello, Joshua Brown
In briefHow Far Large Language Models' Internal Representations Agree, and Why
SummaryThis paper compares the internal representations of several large language models using two new methods: Neighborhood-Weighted RSA (NW-RSA), which weights agreement by proximity, and Counterfactual RSA (CF-RSA), which tests which linguistic changes affect alignment. The models strongly agree on coarse global geometry but diverge substantially in local neighborhoods. Changing logical scope reduces alignment far more than changes to syntax, event roles, or discourse structure, so overall similarity does not guarantee agreement on finer distinctions.
In briefEdgeLLM-Nano: Testing KV-cache compression methods on a 33.6 MB on-device language model
SummaryThis paper measures KV-cache compression for edge devices using a tiny 8.39M-parameter decoder-only language model as a testbed. It finds that a polar-guided key-cache quantizer causes far more damage than a round-to-nearest control, and that a consistency-anchored progressive pruner does no better than one-shot magnitude pruning. The authors propose Long-Context Decode Fidelity (LCDF), which separates compression methods that perplexity cannot, and report on Raspberry Pi 5 that all tested compression backends use more energy per token than uncompressed FP16 inference.
S M Mahbub Murshed, Hayden Bright, Rezaul Karim et al.
In briefRES-MR: A risk-aware reasoning framework for explainable, safer drug recommendations
SummaryThis paper proposes RES-MR, a framework using large language models to recommend medications in a more explainable and safe way. It works in two stages: first distilling clinical reasoning grounded in graph-retrieved patient knowledge, then optimizing policy with patient-specific risk boundaries and reward shaping. The authors report that RES-MR outperforms state-of-the-art baselines on benchmark datasets in accuracy, safety, and explainability.
In briefDiffusion-based multi-winner auctions for banner ads
SummaryThis paper proposes a banner ad auction framework that uses diffusion models so that multiple bidders can co-win one ad slot and influence the displayed image according to their bid strength. Bids and textual descriptions are combined into a score that guides sampling of candidate images, and the welfare-maximizing candidate is selected. Experiments evaluate allocation efficiency against single-winner auctions, monotonicity, and approximate incentive compatibility with Myerson- or Clarke-based payments, reporting improvements as Monte Carlo samples increase while preserving image quality.
Lilian Sun, Henry Huang, Fucheng Warren Zhu et al.
In briefCo-adaptive rule routing and constraining for structured knowledge retrieval
SummaryThis paper proposes CoRE, a method for retrieving task-relevant evidence from structured knowledge sources. It jointly learns which reasoning rule to assign to each query and how to retrieve evidence under that rule, without predefined query-rule pairs. Routing and retrieval are trained alternately using reinforcement learning and contrastive learning, each providing supervision to the other. Experiments on STaRK-QA datasets across several domains demonstrate its effectiveness.
In briefBias Testing of Chinese LLMs: How Task Design and Ironic Wording Affect Results
SummaryThis preregistered study audits several Chinese large language models on identifying toxic and discriminatory content across different tasks and writing styles. In an echoic-irony register (yinyang), the core panel's classification error rate rose from 10.9% to 23.5%, and irony-aware prompts reduced the intervention panel's error by 9.6 percentage points. The authors find that how a bias task is set up measurably changes results, and argue that safety evaluations should include ironic expression as a test condition.</summary>
In briefRe-running coding-agent tests: decisions unchanged, policy risk still open
SummaryThis paper reports two preregistered re-execution studies that check how executable tests grade LLM coding-agent patches, and both returned negative results. In SWE-bench Pro v1's offline option, none of 59 accepted Go patches with changed grading configurations were rejected, though a 3% exposed-patch risk remains unresolved. Re-executing 1,929 selected evaluations from the Darwin Gödel Machine archive preserved best-agent selection and system-variant ranking.
In briefLearning to Extrapolate in Depth with Looped Transformers
SummaryThis paper studies how far a looped transformer block trained at one depth can be extrapolated to greater depths, using all-pairs graph connectivity as the test problem. The authors derive a finite extrapolation window governed by a single effective parameter and bound the graph hardness solvable beyond training. Experiments validate the predicted windows and show how compute budget and initialization affect the optimal training depth.
In briefJevolution: An Inspectable Framework for Studying AI Agent Communication and Collective Dynamics
SummaryThe paper presents Jevolution, an instrumented spatial predator-prey simulation framework for experiments with remote decision-making agents, separating the evolving world, asynchronous scheduling, controller adapters, and experiment records. Synthetic fixtures validate the specified software behavior rather than model intelligence or ecological realism. A small communication case study across 6 habitat-seed pairs found a mean signals-on minus signals-off difference in normalized rabbit abundance of -0.0095, with a confidence interval spanning zero, so the direction of the signal effect is unresolved.
In briefKV4-RL: Faster RL training rollouts via INT4 KV cache quantization with selective INT8 protection
SummaryThis paper studies asynchronous rollout in reinforcement learning with verifiable rewards (RLVR) and proposes KV4-RL, which combines calibration-free INT4 KV cache quantization with INT8 protection of attention sinks and recent context. Across 400 iterations of Qwen3-8B-Base GRPO training, it reports a 1.80× cumulative rollout speedup over BF16 rollout. Its final six-benchmark mean accuracy is within 0.5 percentage points of BF16-rollout training for both Qwen3-4B-Base and Qwen3-8B-Base.
In briefFreeCHROMA: Generative Chroma Keying from Unaligned Background Backgrounds
SummaryFreeCHROMA adapts a pretrained video diffusion model to composite a foreground performance into one to three unaligned background references. It integrates and harmonizes the foreground and background in a single generation process, with optional text instructions for spatial and camera guidance. Training uses pseudo-paired data built from real monocular videos through foreground-background separation and augmentation. The paper reports consistent gains over recent baselines, especially under camera motion.
In briefSpecification-Based Testing of Large Language Models (LEST Framework)
SummaryThis paper introduces LEST, a framework that treats safety claims documented in model and system cards as specifications for black-box testing. It mutates benchmark prompts to search for counterexamples, manually validates them, and traces confirmed cases back to the claims they contradict. Evaluated on Claude Opus 4.8 and GPT-5.5, the study finds automated detection substantially overestimates failures, with only 6.1% of candidate transitions confirmed by humans.
Kausar Moshood, Jace Bolante, Roshan Aditya Murali et al.
In briefTesting whether LLMs know semantic relations via bidirectional gloss-pair prediction
SummaryThis paper proposes a task in which large language models predict the semantic relation between two sense glosses presented in both orders, checking whether predictions respect the symmetry or asymmetry of each relation. The authors built a dataset of 3,069 sense pairs verbalised in English, Greek, Romanian, Turkish, and Japanese for monolingual and cross-lingual evaluation. Across 17 LLMs in zero-shot and five-shot settings, performance was generally poor, with few models significantly beating a random baseline or a neural baseline.
In briefTETPO: Smoothly controlling gradient variance in off-policy LLM reinforcement learning
SummaryThe paper addresses heavy-tailed importance weights in asynchronous reinforcement-learning post-training of large language models, where stale rollouts can collapse the effective sample size. It proposes TETPO, which raises each weight to a data-adaptive exponent chosen per prompt group so that a target ESS fraction is met, tempering more strongly when weights are more concentrated. The authors report that TETPO outperforms strong on- and off-policy baselines on mathematical-reasoning benchmarks under high staleness.
Shayan Mohajer Hamidi, Kayhan Behdin, Wenhui Zhu et al.
In briefSceneForesee: Learning Scene Evolution for Spatiotemporal Reasoning
SummaryThis paper introduces Scene Evolution Understanding, a task where models predict how entities move and what events follow from observed video, with supporting visual evidence. The authors build the SceneForesee benchmark with 144.57K temporal pairs, 1.05M entity correspondences, and 272.49K events. They also propose EvoSee, a reasoning framework that outperforms Qwen3.5-9B by 10.45 points in Combined Point-F1 and 17.24 points in Event Global F1 on this benchmark.
In briefNeither True nor False: Can LLMs Handle Truth-Value Gaps in Plural Claims?
SummaryThis paper examines whether large language models can represent truth-value gaps in plural predication, such as claims about groups where only some members have the property. The authors introduce HomogeneityNLI, a diagnostic dataset of 1,000 minimal pairs, and evaluate seven open-weight models. They find a widespread Bivalence Bias, with models treating plural definites as fully true or false, and negated definites judged False in 84% of cases in the zero-shot setting.
In briefFrom Transfer Probes to Expert Grouping: LoRA Post-Training Evidence
SummaryThe paper asks whether transfer probes, which train on one data source and measure paired changes across all sources, can discover how business lines should share LoRA experts when no taxonomy is given. Using seven public sources with a label tree as a held-out reference, it compares a probe-derived grouping with a uniform default under fixed budgets. The probe-derived grouping did not reach the prespecified 2-point gain and was indistinguishable from the uniform default in a post hoc comparison.
In briefDesign choices for Transformer models using incomplete, multi-visit healthcare data
SummaryThis paper systematically studies design choices for multimodal Transformer models on healthcare data from three cancer datasets. It compares modality-level and feature-level tokenization and evaluates attention masking and imputation for handling missing information. The authors also examine how to encode time for discrete-time survival prediction. Results show multimodal models outperform unimodal ones, but missing data strongly affects predictions, making tokenization granularity and intra-modal missingness handling essential.
Rita Cordeiro Mendes, Maria Manuel Pereira, Carlos Santiago et al.
In briefA Simple Agentic RAG Approach Using Portable Skills for Customer Support Workflows
SummaryThis paper reports a deployment case study in which a portable skill, comprising a natural-language instruction, a retrieval API and format-checking scripts, and an LLM-generated navigation index, replaces a complex retrieval pipeline for customer support resolution workflows. The approach reached 99.2% procedural equivalence to human-approved ground truth and 83.3% evidence sufficiency. Each case took about 90 seconds, roughly ten times slower than the production assistant (about 8-9 seconds), so the authors claim no latency advantage. The claims are limited to settings with a well-curated corpus, latency-tolerant workflows, and human review of every case.
In briefScaling laws for mixing text, vision, and audio data when pretraining multimodal AI models
SummaryThis paper studies how much text, vision, and audio data to mix when pretraining multimodal large language models. The authors ran 268 experiments with Mixture-of-Experts models from 457M to 8.3B parameters, using up to 150 billion tokens. They find audio favors scaling model parameters over data, while text and vision need balanced scaling. Repeating vision or audio data beyond 2x gives negligible benefit, and vision strongly interferes with audio learning, whereas audio's interference with text and vision is negligible.
WhyDerives data-mixing scaling laws for text, vision, and audio, validated by 268 MoE experiments, offering actionable guidance for multimodal LLM pretraining.
Aditi Khandelwal, Ayush K. Tarun, Yixuan Xu et al.
In briefCoarse-graining recursive expert computation: evaluating replacements by what follows them
SummaryThe paper proposes a framework for judging a cheaper module that replaces a sequence of computation, based on the computation that follows it, using downstream task quality and total execution cost as joint criteria. Experiments on two frozen pretrained language models show that representation fidelity and task loss can favor different fitting objectives. Under distribution shift, the reported ranking between objectives reverses in all nine paired comparisons.
In briefPotentialEP: training CIFAR-scale generative models without backpropagation using Equilibrium Propagation
SummaryThe paper proposes generating images by energy relaxation instead of hundreds of sequential diffusion or flow-matching steps, a form suited to analog neuromorphic hardware. The authors prove that a predictive coding network with one added linear injection term relaxes to a state whose visible layer carries the needed gradient. Trained with Equilibrium Propagation, the model reaches an FID of 3.64 on unconditional CIFAR-10, compared with 3.54 for backpropagation.
WhyAdds a linear injection term so predictive coding energy relaxation yields gradients, and trains CIFAR-10 generative models via Equilibrium Propagation with competitive FID.
In briefMemory without curation: multi-stream latent memory for robot policies
SummaryThis paper proposes IMPRINT, a framework that adds multi-stream latent memory to vision-language-action models without explicitly deciding what the policy should remember. It compresses observations, actions, and proprioception into a latent memory updated incrementally during inference. Training uses two stages: first reconstructing past streams and estimating task progress, then end-to-end finetuning with the policy on action prediction. The authors report at least a 26% average success gain over the base VLA in simulation and real-world tasks, with the lowest inference latency among memory-augmented VLAs.
In briefSPROUT: Scaling Agentic RL for Long-Horizon Software Tasks
SummaryThis paper presents SPROUT, a reinforcement learning rollout framework for coding agents that resumes and branches from intermediate states on long-horizon software engineering tasks. It checkpoints both the filesystem and agent context after each assistant turn, and uses a verifier-guided backtracking policy to choose where to resume a failed attempt. The authors report up to 23.6% improvement over pass@k across various LLMs while using 32% to 52% of their tokens.
In briefVisualCoding: Learning Executable Skills for Reliable Visual Reasoning
SummaryThis paper proposes VisualCoding, a method in which visual reasoning agents learn reusable visual programs guided by execution feedback and directly run retained solvers to answer applicable inputs. The approach is evaluated on four models and four benchmarks covering 69 task families, with reported gains in mean accuracy over the strongest baseline. Ablations indicate that direct execution outperforms textual or callable reuse of the same programs.
WhyExecution-refined reusable visual solvers with direct answer execution outperform textual or callable reuse, validated across four models, four benchmarks, 69 task families.
In briefSAT-RL: Training language models to reason logically with symbolic solvers
SummaryThis paper proposes SAT-RL, a reinforcement learning framework that trains large language models to call symbolic solvers such as Z3 over multiple turns. It combines synthetic data with verifiable solutions, supervised warm-up on solver-augmented traces, and reinforcement learning with tool-aware rewards. The authors report that SAT-RL substantially improves tool-based symbolic reasoning on both in-domain and out-of-domain logic tasks.
In briefSparse-TD-JEPA: Sparse latent task representations for better zero-shot reinforcement learning
SummaryThis paper addresses zero-shot reinforcement learning methods that pre-train a task encoder on reward-free data and then solve linear regression at test time. It finds that dense task representations spread the inferred reward across states, and adds an l1 sparsity penalty during training to concentrate reward-relevant structure. Evaluated on 65 tasks across DMC and OGBench, Sparse-TD-JEPA outperforms TD-JEPA in all 13 domain-level comparisons, raising average DMC return from 661 ± 6 to 714 ± 3 and average OGBench success rate from 38.0 ± 0.8 to 48.4 ± 2.3.
In briefClinical value claims for multimodal AI outpace the supporting evidence
SummaryThe authors audited 177 studies comparing multimodal and unimodal prediction models, drawn from a systematic review and a probability sample of recent PubMed records. Studies that fitted a separate model for every data source reported smaller gains in AUROC. The share of studies performing complete benchmarking fell from 69.1% in review studies to an estimated 46.0% in recent ones. None evaluated a model prospectively, although most claimed potential clinical usefulness or improved care.
In briefWhy large recommender models diverge during training, and how to stabilize them
SummaryThis paper studies training instability in large-scale short-form video recommender systems that handle billions of daily interactions. It analyzes a gradient-to-weight-to-loss explosion cascade (G-W-L) and proposes a framework with interventions targeting each stage. The authors report empirically validating these strategies on a major industrial short-form video recommendation system.
In briefSciCurateBench: Verbal-action agents that curate experimental data from scientific papers
SummaryThis paper introduces SciCurateBench, a benchmark of 200 expert-annotated life-science papers that requires verifying conjunctive experimental criteria from raw literature. The authors report that plain LLMs and common agent approaches struggle with this task. They propose LangPoli, in which a policy agent writes natural-language action guidelines that actor agents turn into executable tools, reaching 83.16% F1 on paper classification versus 58.22% for its base LLM.
In briefMeasuring the energy cost of open-weight LLM inference across tasks, reasoning modes, and hardware
SummaryThis paper directly measures the trade-off between benchmark score and energy consumption for 86 open-weight LLM checkpoints, tested on seven benchmarks and six hardware platforms from a consumer GPU to an eight-GPU server. It finds large energy-efficiency gaps even among models of similar capability, with most configurations off the score-energy Pareto frontier. The authors argue that benchmark scores should be reported together with measured energy use under a specified serving regime and output budget.
In briefDecomposing Neural Learners: A paper on separating the predictable part of a model from its reliability
SummaryThis paper proposes a probabilistic decomposition of neural learners to account for uneven predictability across the feature space. It combines the EM algorithm, stochastic gradient descent, and maximum likelihood estimation to jointly learn a predictable component and an input-dependent reliability score. The framework is extended to include localized feature transformations and is tested on synthetic and real-world datasets.
In briefA control-theory overview of deep learning models and training
SummaryThis paper surveys recent deep learning developments from a control theory viewpoint, covering feedforward and residual networks, transformers, large language models, and selective state space models. It explains that residual networks can be read as discretizations of continuous-time control systems, enabling control-theoretic analysis. Two results are highlighted: residual networks' approximation of continuous functions as a nonstandard controllability problem, and transformers as discretized consensus algorithms. The paper also relates backpropagation to optimal control and lists open research problems.
\'Alvaro Rodr\'iguez Abella, Jo\~ao Pedro Silvestre, Bahman Gharesifard et al.
In briefANASSA: From High-Level Framework to Low-Level Architecture for Spatial Agentic AI
SummaryThis paper refines the earlier ANASSA framework, which governs AI agents working with geospatial data, into a detailed low-level architecture. Using Design Science Research, it defines 47 logical roles, versioned task graphs, typed contracts, spatial validation, governance controls, failure recovery paths, and provenance records. The design is demonstrated through an end-to-end execution trace. It is analytically evaluated for requirements traceability, internal consistency, responsibility preservation, and technology independence.
In briefACED-Bench: Testing How LLM Agents Gather and Use Causal Evidence
SummaryThis paper introduces ACED-Bench, a benchmark for checking whether LLM agents can actively investigate a system rather than just answer a prompt. LLMs write the readable scenarios, while hidden probabilistic graphical models define the causal graph and the correct answers, enabling scalable automatic scoring. The main leaderboard, ACED-CORE, contains 180 tasks, and seven contemporary and two earlier models were evaluated. Results show agents often test the wrong intervention, lose relevant evidence, or keep going after enough evidence has appeared.
Yiwei Chen, Amila Weerasinghe, Siddharth Biswal et al.
In briefDPAL: A Test-Time Adaptation Loss for Financial Price Forecasting
SummaryThis paper proposes DPAL, a loss function for test-time adaptation that lets financial time series forecasters update after deployment. Besides fitting returns at each time step, it adds terms for return direction and for the price paths rebuilt from return sequences. DPAL is applied within existing TSF-TTA frameworks without changing the source forecaster or adapter. Across diverse financial datasets, it reduces reconstructed-price mean squared error in most evaluated settings.
In briefDexFLEX: a contact-aware controller that turns rough fingertip commands into stable grasping motions for dexterous robot hands
SummaryDexFLEX is a contact-aware foundation controller for dexterous robot hands. It treats upstream fingertip-motion drafts as intent, proposes feasible short motion chunks from tactile and proprioceptive state, and selects the option that follows the command while keeping future contact stable. The paper reports that real-world teleoperation success rose from 29.2% to 78.3% without retraining.
In briefMore memory can fail: hidden feasibility holes in dynamic tensor rematerialization
SummaryThis paper studies Dynamic Tensor Rematerialization (DTR), which trains neural networks under a memory budget by evicting and recomputing tensors. The authors replay public DTR training traces in the simrd simulator and find budgets where training runs out of memory even though both smaller and larger budgets complete, which they call feasibility holes. All results come from simulation, and GPU behaviour is untested.
In briefHistory, Adaptation and Hidden Gain: Testing Learned Predictive Control in a Simulated Neuron Population
SummaryThe paper studies learned one-step predictive control of 64 simulated leaky integrate-and-fire neurons whose light sensitivity is scaled by a hidden factor. Two prospective tests on plants simulated after the analysis was frozen show that removing the 200 ms input-output history degraded a neural predictor at low and moderate gain. Re-estimating one multiplicative input gain online recovered 87% of the benefit of full adaptive ARX, and supplying this estimate to the network reduced high-gain tracking error. Results are limited to this synthetic plant, and the effects shrank markedly under high process noise.
In briefConformalGuard: A Method for Controlling False Positives in Language Model Classifier Action Gating
SummaryThe paper presents ConformalGuard, a model-agnostic calibration layer that wraps a fixed scoring function to give finite-sample control over how often legitimate inputs trigger an intervention. The authors compare temperature scaling, isotonic regression, and conformal calibration on public email corpora using fine-tuned BERT, Claude Sonnet 4, and GPT-4o. For Claude, the false-positive rate drops from 0.040 to 0.008, while recall falls from 0.938 to 0.875.
Bishnu Bista, Dhanraj Haibatrao Jadhav, Vibha Choudhary et al.
In briefA configurable ETC2 texture encoder that learns how to spend its search effort
SummaryETC2 fixes the texture format and bit rate, so an encoder can only trade search time against distortion. The authors propose a configurable CPU encoder where rules or small neural networks decide which conventional ETC2 searches to run. On 64 color textures, the fastest configurations matching Etc2Comp's PSNR run 30.7 to 40.9 times faster. Outputs remain standard ETC2 and decode in hardware.
In briefSample Complexity of Multicalibration Using Only an ERM Oracle
SummaryThis paper studies how many samples are needed to learn a multicalibrated predictor when the learner can access the function class only through an empirical risk minimization (ERM) oracle. The authors give an algorithm using about d/ε³ samples (up to log factors) and polynomially many ERM calls, where d is the VC dimension, and prove matching lower bounds. They also show that estimating ordinary multicalibration error is computationally harder than learning a calibrated predictor for some classes.
WhyGives near-optimal sample complexity bounds for multicalibration under an ERM oracle, with matching lower bounds and a clear computational separation.
In briefWhen user pleasing undermines care: AI sycophancy in caregiver-mediated healthcare
SummaryThis paper studies how AI advisors behave when a caregiver, such as a parent or pet owner, makes decisions for someone else, such as a child, animal, or older adult. The authors built a benchmark across veterinary, pediatric, and older-adult care and tested whether models keep appropriate care when users express clinically inferior preferences. They report that many frontier and open-weight models shift toward inappropriate care in these cases, and that these failures build up over multiple turns. The authors conclude that current models respond strongly to expressed preferences but are less sensitive to whose preferences they are and who bears the consequences.
In briefActionable Meta-Requirements for AI Robustness Auditing
SummaryThis paper addresses how to audit whether AI-enabled systems remain dependable when operating conditions change, such as input perturbations, distribution shifts, or external dependency changes. The authors derived 43 meta-requirements from a systematic review of 74 papers and validated them with 128 practitioners, producing a final set of 44 across 16 categories. The meta-requirements are intended to help organizations implement, assess, and refine AI robustness auditing frameworks.
Joshua Owotogbe, Roya Nasiri, Indika Kumara et al.
In briefClass-conditional conformal prediction for interpretable time-series logic formulas with many classes
SummaryThis paper studies Signal Temporal Logic (STL) inference, which produces interpretable formulas for classifying time series into many classes. The authors introduce two metrics, positive satisfaction rate and codeword satisfaction rate, and use a two-stage procedure that learns discriminative formulas and then refines predicate thresholds. The decoding loss is calibrated separately for each class to give finite-sample class-conditional coverage under intra-class exchangeability. Experiments on naval, synthetic, and robot manipulation trajectories show that classification accuracy and formula satisfaction can diverge.
Yixuan Wang, Danyang Li, Matthew Cleaveland et al.
In briefLearning Transferable World-Action Models from Task-Paired Human Videos (WATCH)
SummaryThe paper proposes WATCH, a cross-video pre-training method that learns from task-paired human demonstrations recorded across diverse environments. Given one demonstration's video and actions, the model jointly predicts another demonstration's actions and future frames, encouraging transferable task interactions. The authors report that this can be learned from human-human pairs alone, without matched human-robot data. After robot post-training, they report MimicDroid in-context success rising from 40.9% to 68.3% with Cosmos Policy and from 59.8% to 77.7% with Cosmos3.
In briefJoint Multimodal Generation via Continuous Bitstream Diffusion
SummaryThe paper adapts continuous bitstream diffusion (CoBit), originally built for language modeling, to joint and conditional generation across image-text, speech-text, and protein sequence-structure data, using one shared training and sampling protocol. For speech-text, generated text versus transcripts of generated speech reaches a 3.5% word error rate. For proteins of 100 to 500 residues, refolded CoBit sequences match the generated backbones with a mean self-consistency TM-score of 0.911.
Georgios Batzolis, Georgios Kontos, Bruno Trentini et al.
In briefEvaluating LLMs as Workflow Executors for Customer Support Tasks
SummaryThis paper studies how large language models perform narrow, tool-using customer-support workflows. The authors build a benchmark by mining recurring workflows from real transcripts, turning them into skill descriptions, and simulating environments to generate multi-turn tool-using dialogues. They evaluate 17 models zero-shot using five LLM judges and a rule-based evaluator. GPT-5.4 ranks highest, and QWEN3.5-9B is the strongest open-weight model under 20B parameters, while judge agreement (average Kappa 0.524) is highlighted as an important factor in reliable evaluation.
Md Tahmid Rahman Laskar, Gundeep Singh, Xue-Yong Fu et al.
In briefDepth transfer of neural optimizers via L-smoothness
SummaryThis paper studies how hyperparameters transfer across network depth by analyzing how the Lipschitz and L-smoothness constants of the loss scale with depth K. It shows that scaling residual branches as Θ(1/K) at fixed width makes both constants depth-independent, for general residual networks and several optimizers including AdamW, Muon, and MOGA. Learning-rate sweeps on GPT-2 (depths 2–128) and CIFAR-100 ResNets (8–32 blocks) indicate that Muon and MOGA transfer a common optimal learning rate across a 64-fold depth increase.
In briefContinuous-time cloth dynamics modeling with physics-informed neural ODEs
SummaryThis paper presents ContinuousCloth, a continuous-time neural solver for deformable cloth that embeds a spring-mass ODE in a neural trajectory model. It can query displacement, velocity, and acceleration at arbitrary timestamps via automatic differentiation. In physics-only tests it matches high-precision solvers' trajectory accuracy while running up to 17.4 times faster. When combined with sparse, noisy observations, it improves interpolation and parameter estimation, reducing errors relative to baselines in simulation and real-world cloth sequences.
In briefSTREGO: A Scalable Trust-Region Method for Noisy Bayesian Optimization
SummaryThis paper extends TREGO, a trust-region Bayesian optimization method for expensive black-box problems, so that it can handle noisy function evaluations, and provides a convergence analysis for sufficiently accurate noisy estimates. It then introduces STREGO, a scalable implementation that selects candidate batches globally using a bi-objective problem and performs local refinement with a second Gaussian process. The authors compare STREGO with TREGO, TURBO, BAXUS and NSMA-BO on problems from dimension 4 to 124, reporting the best average results on all synthetic settings tested and the best or tied-best solutions on four of six real-world problems.
Youssef Diouane, Issmail El Hallaoui, Meisseme Kadri et al.
In briefWhich Metadata Affects LLM Tool Selection in MCP? A Controlled Study
SummaryThis paper studies how large language models choose tools in the Model Context Protocol (MCP) based on tool metadata rather than implementation. The authors analyze metadata for 367,174 MCP tools and find that functionally similar tools often have different parameter schemas. Using controlled experiments across five LLMs and five tasks, they find that parameter schema changes have the largest effect on tool selection, descriptions have a moderate effect, and names have weaker effects.
In briefDecentralized bandit learning of equilibrium strategies in repeated games
SummaryThis paper studies how agents in discounted, infinite-horizon general-sum Markov games with continuous actions can learn without coordination. Each agent sees only the public state and its own realized rewards, and updates its action using simultaneous perturbation stochastic approximation. Under strong monotonicity and small step sizes, the authors prove almost-sure convergence to the unique stationary Markov perfect equilibrium at an O(n^{-1/2}) rate, and extend the approach to episodic finite-horizon settings.
In briefTraceSampler: Making Accelerated Generative Sampling Errors Predictable and Correctable
SummaryTraceSampler is a post-training method that corrects endpoint errors in low-step sampling of generative models without changing the pretrained model. It treats the error as a sample-specific prediction problem, using information from the sampling trace and a two-stage learned residual correction. Evaluated on CIFAR-10, ImageNet-256, and SD3.5 Medium, it reduces trajectory error at 10 steps and approaches the quality of 120-step Heun references with little extra inference cost.
George Matthew (1), Guoqiang Zhang (1), Kenta Niwa (2) ((1) University of Exeter et al.
In briefReal-time traffic density estimation using combined vision detection and forecasting
SummaryThe paper presents a two-stage pipeline that detects and tracks vehicles with a road-mask spatial filter, then forecasts five future steps from ten-frame windows using LSTM, GRU, or TCN models. Ten detector variants were benchmarked on 5,716 annotated frames, showing a speed-accuracy trade-off between models. On 29,243 held-out windows, GRU, LSTM, and TCN reached aggregate R-squared values of 0.858, 0.856, and 0.840, though these measure detector-derived telemetry rather than absolute traffic density accuracy.
In briefLearning to track positions on curved spaces with recurrent networks
SummaryThis paper studies position tracking on general manifolds, formalized as a velocity-driven task. The method predicts trajectories with GRU networks in a Euclidean representation, maps them onto the manifold via an immersion function, and trains with custom loss functions. Experiments on analytic and numerical manifolds and a LiDAR-reconstructed surface show that recurrent models outperform Euclidean baselines, and some custom losses improve long-horizon accuracy.
Sarah Harkins Dayton, Sathya Rengaswami, Vasileios Maroulas et al.
In briefEV-JEPA: A Video Foundation Model That Runs Flexibly on Any Device and Splits Across Devices
SummaryThis paper introduces Elastic V-JEPA (EV-JEPA), a single model trained once that can be deployed across devices from smart glasses to servers. Its width can be adjusted, and it can be split across devices during inference. A training method called Transition-Aware Triplet Training lets the same weights handle both single-device and cross-device execution. The authors report accuracy within 1 point of independently trained baselines.
In briefMvFFN: A Feed-Forward Network That Rebuilds Multi-Room Floor Plans from Unposed Panoramas
SummaryThe paper addresses reconstructing camera poses and multi-room floor-plan layouts from sparse, unposed, wide-baseline RGB panoramas. It introduces Multi-view Ordered Wall Instance Segmentation (MOIS), in which each panorama column predicts only its observed wall and the next two walls clockwise, with the global structure obtained by deterministic aggregation. Building on this, MvFFN jointly predicts coarse poses, dense global geometry, and per-wall semantics, and a Layout Chaining post-processing step assembles them into floor plans. The authors also release the ZInD-CrossView dataset.
In briefASSET: Learning to Cut Speech into Fewer Tokens with Reinforcement Learning
SummaryThis paper introduces ASSET, a reinforcement learning framework that adaptively compresses the outputs of speech encoders for speech large language models. It splits encoder outputs into variable-length segments and pools each into one continuous audio token, training the segmenter with GRPO using task performance feedback and a token-rate penalty. In speech recognition experiments, ASSET matches a transcript-assisted character-boundary oracle on English LibriSpeech with about 25% fewer audio tokens, and outperforms a phone-boundary oracle at a similar token rate on Mandarin AISHELL.
In briefConicFlow: a method for constrained, inference-time optimal control of flow matching models
SummaryThis paper proposes ConicFlow, which treats constrained sampling from a flow matching model at inference time as an optimal control problem. It balances staying close to the pretrained trajectory and maximizing reward while enforcing hard terminal constraints, using receding-horizon control and a primal-dual algorithm. Experiments in low- and high-dimensional settings with a recent large-scale model show high reward, strong constraint satisfaction, and less deviation from the pretrained flow than recent baselines.
In briefLearning to See through Action: Predictive Curiosity for Robots
SummaryThis paper introduces DySCOVER, a self-supervised framework in which robots actively gather experience through curiosity-driven interaction to learn vision. It combines a visual autoencoder with a recurrent world model, and the policy is trained only on intrinsic rewards from reconstruction error and prediction error. Experiments with robotic arms report that adding predictive curiosity improves the vision model over visual curiosity alone, across varied conditions.
Stelios Zarifis, Panagiotis Filntisis, George Retsinas et al.
In briefRUBATO: Letting Speech Models Think at Hardware Speed
SummaryThis paper addresses internal thinking streams in speech full-duplex models, which currently run at a fixed decoding rate. The authors propose RUBATO, a multistream positional encoding that gives each stream an independent clock so it can decode at its own hardware-adaptive speed. Training on only two extreme decoding speeds lets the model zero-shot adapt to intermediate rates. Across five spoken reasoning tasks, RUBATO maintains accuracy across rates and during user interruptions, where the baseline fails.
In briefDiagnosing Where AI Vulnerability Detectors Fail Using Calibrated Embedding Distances
SummaryThis paper examines why learning-based vulnerability detectors perform poorly under distribution shifts despite high benchmark scores. Using CodeBERT and Qwen2.5-Coder on paired vulnerable and patched code, the authors show that whitened intermediate-layer embeddings with nearest-neighbour scoring separate unfamiliar samples (mean OOD AUROC 0.739). Distance scores track accuracy drops by shift type, with cross-project shifts degrading most, and false positives on patched code rise with proximity to vulnerable training neighbourhoods.
Konstantin Ievtushenko, Ralf L\"ammel, Jan J\"urjens
In briefControllable Diverse Sampling via Pullback Directions
SummaryConditional diffusion models generate high-quality images that match a prompt, but different random seeds often yield visually similar results, limiting creative exploration. The authors propose a training-free method that uses the local pullback geometry of the mapping from conditioning embeddings to image space to find directions of maximal output sensitivity. Steering sampling trajectories along these directions encourages distinct generations while staying consistent with the conditioning signal. Experiments on MS-COCO and BrushNet show the method improves diversity in prompt-conditioned generation and inpainting.
Sylvain Dehayem Kenfouo, Aaron Zweig, Mingxuan Zhang et al.
In briefGeoFix: A Framework That Repairs Geometry for Consistent 4D Video Generation
SummaryVideo diffusion models are often used as priors for 4D generation, but their outputs frequently lack geometric consistency. GeoFix is a unified framework that uses a single geometric reward to improve both geometric consistency and visual quality, applicable during training and inference. The reward measures cross-view depth consistency in static scene regions and is calibrated against the camera field-of-view overlap trend.
In briefLayer-wise Input Recoverability in Deep Neural Networks
SummaryThis paper studies how much of a convolutional layer's input can be recovered from its output. The authors evaluate recoverability on sampled input patches using a reconstruction probe, measured as R² across nine ImageNet CNNs. Recoverability generally decreases with depth and is more consistently predicted by the layer's compression. Single-property random controls show that each pretrained property shifts recoverability in a consistent direction across models.
Maxwell Spiridonov, Alireza Rahmansetayesh, Pierre Baldi
In briefTGSR: Text-Grounded Diffusion for Faithful Image Super-Resolution
SummaryDiffusion models improve image super-resolution but often hallucinate non-existent or distorted text, altering scene meaning. This paper proposes TGSR, a test-time framework that progressively detects and suppresses likely hallucinated text regions during generation while fully super-resolving other content. Dedicated detectors were trained on a large dataset of diffusion-based SR outputs. TGSR improves text fidelity across multiple backbones, raising recognition precision by up to 17.68 points without degrading perceptual quality.
In briefThe RNN Homunculus: How Gated Networks Learn to Control Long-Term Memory
SummaryThis paper studies how gated recurrent networks such as GRUs and LSTMs learn the gate values that control long-term memory. Through controlled sequence classification tasks, the authors show that gated models suffer from vanishing gradients when every training sample needs long-range memory. Adding short-dependency samples helps the models learn solutions that handle long dependencies and generalize across class labels, matching or exceeding modern architectures such as Mamba and xLSTM.
Zeyu Fan, Alexandru-Dragos Manolache, Jan van Gemert
In briefEcho: Error-compensated 4-bit communication for faster tensor-parallel LLM training and inference
SummaryTensor parallelism requires an all-reduce after each layer, and this communication cost grows as compute outpaces interconnect bandwidth. Echo sends data in 4 bits, sums in high precision, and feeds quantization errors back into later calls so they cancel rather than accumulate. The authors prove the accumulated error stays bounded regardless of depth. On A100 GPUs it is 2.09x faster than 16-bit NCCL all-reduce, with accuracy close to 16-bit inference and training.
WhyApplies error feedback to 4-bit tensor-parallel all-reduce with a proven depth-independent error bound, validated by A100 speed and accuracy experiments on LLM training and inference.
In briefThinking leakage: why RL gains in no-think mode may come from reactivated reasoning
SummaryThis paper studies why hybrid reasoning models such as Qwen3 improve when reinforcement learning is applied in NO-THINK mode. The authors argue that much of the gain comes from thinking leakage, where reasoning abilities from the original THINK mode are partly reactivated rather than newly learned. They support this with analyses of weights, activations, and attention heads.
In briefAdaptive client-side synchronization for edge federated learning
SummaryThis paper addresses synchronization bottlenecks in edge federated learning, where fixed schedules are common. It proposes a reinforcement learning framework in which each client decides during training whether to upload, download, synchronize bidirectionally, or defer communication, using a lightweight scheduler beside its local model. Evaluated on spatiotemporal traffic forecasting and image classification, the adaptive approach improves the accuracy-cost trade-off, and the authors also examine privacy-constrained settings.
Ebasa Temesgen, Ammar Ahmed, Nathan Lechamo et al.
In briefVisual reasoning with latent evidence, without pixel coordinates
SummaryThis paper examines the "thinking with images" approach in multimodal large language models, which relies on explicit coordinate-based grounding, and identifies a semantic gap and a resolution gap as its limitations. The authors propose TiCO, which frames visual reasoning as semantic evidence discovery in latent space and injects evidence as token-level regions without pixel coordinates. The method uses a probe token and importance sampling over spatially connected components. The authors report that TiCO outperforms coordinate-based methods on several high-resolution and general visual reasoning benchmarks with better inference efficiency.
In briefBigCell: Generating Gigapixel Whole-Slide Pathology Images
SummaryWhole-slide images are gigapixel digital scans of tissue used for cancer diagnosis, while existing generators only produce small 1024×1024-pixel tiles. BigCell trains a flow-matching transformer on over 120,000 public slides and uses a sliding-window pipeline to generate coherent images up to 32,768×32,768 pixels. The authors also report an adaptive scheduling strategy that speeds generation by over 3 times and that adding synthetic samples improves slide-level few-shot classification by up to 18%.
WhyFirst gigapixel whole-slide histopathology generation, validated on over 120,000 WSIs, with synthetic data improving few-shot classification by up to 18%.
Srikar Yellapragada, Alexandros Graikos, Zilinghan Li et al.
In briefVariance Fingerprints: Identifying Individual Content Factors by Designing Augmentations
SummaryThe paper proposes a self-supervised learning method in which content factors fluctuate across augmented views while their per-instance mean is preserved, with augmentation strengths recorded for each instance. Under suitable decoder and support assumptions, the authors prove that differences in how factors respond to these strengths identify individual factors up to permutation, scaling, and translation. Synthetic and controlled dSprites experiments show near-perfect factor recovery and improved coordinate alignment.
In briefDo Neural Networks Need Dendritic Nanotubes? Why Importance Estimates Shrink in Continual Learning
SummaryInspired by dendritic nanotubes, the paper lets temporary lateral links between hidden units diffuse importance tags and explains the statistics behind this effect. Random links improved final accuracy on 20-task Permuted MNIST, which the authors show is equivalent to linear shrinkage of importance toward the layer mean. The study compares EWC, SI, and MAS and examines how the number of importance samples affects each method's gains.
In briefImage diffusion models for generating 3D human motion
SummaryThis paper tests whether image diffusion architectures can generate 3D human motion. Root-centered joint trajectories are arranged into topology-informed joint-time images, and global root dynamics are modeled with a separate latent stream. Dataset-specific models reach state-of-the-art results on HumanML3D and KIT-ML, and a single shared model trained across both datasets, which keep their own skeleton definitions, improves performance on both.
In briefSliderEdit++: Improved Slider Trajectories for Instruction-Based Image Editing
SummaryThis paper addresses slider-based control in instruction-based image editing, where existing methods can show abrupt or non-monotonic changes and interference between sliders. The authors introduce SlideBench, a benchmark with metrics for monotonic progression, perceptual continuity, and inter-slider disentanglement. They propose SliderEdit++, which regularizes intermediate editing trajectories and adds a disentanglement objective, achieving state-of-the-art results on SlideBench and running about 3x to 20x faster than existing slider-based approaches.
In briefAdaptive-Depth Is a Finishing Stage: Train at Fixed Depth First, Vary Depth Only Briefly at the End
SummaryThe paper studies looped transformers, which reuse the same layers so inference compute can be adjusted. The authors propose fixed-first training, which trains at a fixed depth and varies depth only during a short final phase. At 200M parameters, training at depth 2 for the first 67% of the run cut modeled training compute by 45% while keeping mean NLL across inference depths within 0.5% of full-range training. At 1B parameters, compute fell by 31% to 37% and wall-clock time by 32% to 39%, with final NLL within 0.8%.
In briefAkule: Fast, scalable planning for many robots using sparse interaction diffusion
SummaryThis paper introduces Akule, a framework for multi-robot motion planning that builds on a frozen single-robot diffusion prior. It adds learned modules that correct pairwise robot interactions, select which interactions to evaluate, and combine their effects into coordinated trajectories. The joint trajectories are then validated and repaired by a downstream search-based planner. The authors report competitive success, planning speedups of up to 25 times, and planning with up to 256 robots.
Manav Vora, Gokul Puthumanaillam, Aditya Penumarti et al.
In briefRecovering Non-Gaussian Structural Models Using Conditional Risk and Recursion Direction
SummaryThe paper extends variance-based methods for recovering linear structural equation models to non-Gaussian noise by replacing conditional variance with conditional risk, defined via the negative log-likelihood of the noise distribution. Least-squares methods become a special case. Matched-loss procedures recover the structure more accurately than least-squares and LiNGAM methods. Under non-Gaussian noise, the top-down method is correct for any noise density with finite differential entropy, while the bottom-up method needs a regularity condition and can fail without it. Simulations confirm these results.
In briefAgent-driven data engineering for LLMs: system design and a controlled study
SummaryThe paper proposes Agentic Data Engineering (ADE), a framework where agents use cumulative experimental feedback to revise strategies for data selection, reward design, and curriculum learning for LLMs. Under fixed training and evaluation protocols, selected strategies outperform generic baselines across math and code tasks, with an average gain of 5.85% across four held-out benchmarks. The authors also find that in-loop score gains do not always translate into better generalization on held-out data.
In briefLatent world action model for GUI agents: predicting action-caused changes in latent space instead of reconstructing full screens
SummaryThe paper proposes a latent world model for GUI agents that predicts action-induced future information in latent space rather than reconstructing full screens in pixel space. It uses a large multimodal model with two heads, one for action selection and one for transition prediction, and adds an action-aware contrastive objective to avoid predictions that simply copy the current state. The authors report relative success-rate gains of up to 31.9% offline and 40.6% online over pixel-space world models, with inference latency reduced by up to 98.9%.
WhyShifts GUI world modeling from pixel reconstruction to latent action-induced prediction with a contrastive objective against identity shortcuts; extensive offline and online experiments show clear gai
In briefCAPTOR: A Vision-Language Model for Oncologic Chest-Abdomen-Pelvis CT
SummaryThis paper presents CAPTOR, a contrastive vision-language foundation model for interpreting oncologic chest-abdomen-pelvis CT scans. It was trained on a multicenter retrospective dataset from North America, Europe, and Asia, mostly from tumor lesion examinations. On disease detection, survival prediction, and image retrieval tasks, CAPTOR outperformed state-of-the-art CT foundation models within its target clinical context. The authors also release their code and model weights.
WhyA clinically aligned vision-language CT model, outperforming baselines across multiple oncologic downstream tasks with multicenter data and significance testing.
S\'el\'ena Jean-Mactoux, Sophie Loizillon, Ghina Jardali et al.
In briefHarness^2: Recursive Agent Harnessing for Open-World Tasks
SummaryThis paper introduces Harness^2, a plug-and-play framework that recursively improves the execution harness of LLM agents for open-world tasks. Task-level recursion refines each task's harness using a contrastive Propose-Probe-Compose step that works without ground-truth verifiers, while harnessing-level recursion extracts experience from these edits to support future tasks. The authors report that a single step improves GLM-5 and Gemini 3.5 Flash across three professional benchmarks, with further gains as the task-level recursion budget grows.
In briefLearnable low-rank polynomial sketch for more accurate linear attention in Transformers
SummaryStandard softmax attention in Transformers scales quadratically with sequence length, which limits long-context use. This paper analyzes existing linear attention methods through polynomial kernel approximation and proposes LLoPS, a framework that learns a low-rank polynomial sketch to approximate polynomial kernels with linear attention. The authors claim their method provably achieves a strictly smaller approximation error than prior approaches. They report that LLoPS performs best across extensive benchmarks against various linear attention baselines.
In briefCortex: Measurement-Gated Evolution and Pruning of Coding-Agent Skills and Rules
SummaryThe paper introduces Cortex, a tool that maintains the text files coding agents like Claude Code read, such as skills and rules. It saves developer corrections as replayable tasks, finds recurring fixes, and adds one skill or rule at a time, keeping it only if repeated runs show a net gain. In the authors' tests, the rules Cortex built made the agent follow project conventions far more often, and unused text was not kept.
In briefEmbedding Graph Kernel Entropy (EGKE): Embedding-Based Fine-Grained Hallucination Detection for LLMs
SummaryThis paper proposes EGKE, a method for detecting hallucinations in large language models. It modifies Kernel Language Entropy by replacing the NLI-derived semantic graph with an RBF kernel applied to BGE sentence embeddings, avoiding separately loaded NLI classifiers. The authors evaluate it across 30 experimental setups using AUROC and AUARC, reporting competitive performance against state-of-the-art methods.
Dilrangi Sankalpana, Pragatheeswaran Vipulanandan, Uthayasanker Thayasivam et al.
In briefModeling when meaningful events occur in language using neural point processes
SummaryThe paper proposes using neural temporal point processes to model when key information-bearing events, namely new words and high-surprisal tokens, arrive in text. The model takes event interval histories and contextual representations from LLMs as input to predict future arrival times. Results show LLMs produce new words at significantly higher rates than humans, while high-surprisal tokens show no such difference, and new-word intensity decays over time for both humans and LLMs but not for high-surprisal tokens.
In briefComparing Cross-Epoch Agreement and Softmax for Spatial Disambiguation
SummaryThis paper studies how to assess model confidence when several objects of the same class appear in a 3D scene on the ScanRefer dataset. On the full validation set of 9,508 queries, it compares calibration error for object-level softmax and cross-epoch agreement. It also breaks down errors on Multiple queries, finding most are wrong-instance errors. The authors suggest calibrating abstention on Multiple queries and reporting wrong-instance rates when only agreement was logged.
Micha{\l} Zakrzewski, Mariia Shpir, Taras Rumezhak et al.
In briefShould Small Models Drive? Local Models as Interpreters in Cloud-Written Agent Programs
SummaryThe paper examines an intermediate design where a cloud-written agent program hands bounded questions about user data to a small local model via an ordinary function call, so the data stays on the device. On AppWorld, this raises the success rate of a code-only agent in all six combinations of backbone model and test partition evaluated, recovering part of the gap to a cloud-hosted ReAct loop while keeping user data local.
In briefBudget-limited object retrieval for 3D scene VLMs: category names vs. pairwise geometry
SummaryThis paper studies which objects and relations a 3D scene-grounded vision-language model should see first under a fixed context budget. On ScanNet referring benchmarks, matching queries to object category names beats pairwise geometry scoring at every reported K, with the largest gaps at high recall. A downstream 3DGraphLLM evaluation under a matched object budget also finds that semantic retrieval outperforms hand-designed geometry.
Micha{\l} Zakrzewski, Mariia Shpir, and Taras Rumezhak
In briefRefined Grid Sampling: Faster Training of Implicit Neural Representations for Large Scientific Data
SummaryThis paper addresses the inefficient training of implicit neural representations (INRs), which must evaluate dense coordinate grids. It proposes Refined Grid Sampling (RGS), combining stratified grid sampling, error-based refinement, and dynamic grid-resolution adjustment to reduce spatial redundancy without evaluating errors over the full grid. On high-resolution Navier-Stokes fields, RGS achieves higher reconstruction accuracy within a fixed wall-clock budget and reaches full-grid training accuracy up to 6 times faster than the evaluated baselines.
In briefClothCode: Generative Modeling for Clothing State Estimation and Motion Prediction
SummaryThis paper presents ClothCode, a generative framework that handles both garment state estimation and action-conditioned dynamics prediction in one shared discrete configuration space. It encodes garment configurations as residual vector-quantized patch tokens and predicts them with masked generative transformers using parallel, confidence-guided refinement. Across three garment templates, it reduces mean displacement error by 26.6% for state estimation and 37.0% for dynamics prediction compared with the strongest baselines. When integrated into a model-predictive controller, it raises real-world folding and flattening success from 56.7% to 78.9%.
In briefSimple linear fixes to recover search quality after compressing late-interaction document vectors
SummaryLate-interaction models store many token-level vectors per document, which is costly, and common pooling or pruning compression hurts retrieval quality at aggressive settings. The authors learn a global linear correction to the scoring function by distilling the uncompressed model's score distributions, keeping the encoder and compressor frozen. Tested with ColBERTv2, Jina v4, and ColPali2 on MS MARCO, BEIR, and ViDoRe v3, the correction improved over baseline compression in 154 of 160 in-domain BEIR settings. Transfer from MS MARCO was more variable across models and datasets.
In briefTracing which training examples and image regions shape vision-language model outputs
SummaryThe paper introduces MOSAIC, a method for explaining vision-language model predictions by attributing training-example influence to specific text phrases and image regions. It learns a shared sparse image-text latent space and combines it with EK-FAC to decompose influence through these concepts. Experiments on Qwen and Gemma across captioning, chest X-ray understanding, and visual question answering report more localized, coherent features without degraded task performance.
