In briefMonte Carlo Estimation for KV Cache Eviction (LORE-KV)
SummaryThis paper addresses which key-value cache entries a language model should keep under a fixed memory budget during decoding. It proposes LORE-KV, a training-free method that samples short continuations from the frozen model and uses their response-side query states to score prompt-token utility. Reported results show gains on LongBench and RULER at a budget of 128, with gains shrinking at larger budgets, some task-level regressions, and a compression overhead of about 1.46 to 2.77 times the baseline's per-sample time.
Ahsan Bilal, Muhammad Ahmed Mohsin, Muhammad Umer et al.
In briefPreliminary study on automatically synthesizing interpretable chess tactics
SummaryThe paper proposes a symbolic sub-policy model for chess that draws on chess tactics to improve interpretability. It adapts patterns learned by the inductive logic programming system PAL to build the model. The authors introduce a divergence metric to compare the model against a random baseline. They find tactics that suggest moves of playing strength similar to a human beginner.
In briefLearning Explainable Representations of Complex Game Strategies
SummaryThis paper proposes training reinforcement learning agents in a way similar to human cognition, synthesizing learned strategies and policies as executable procedures built from sequences of gameplay actions. The authors present methods to automatically learn such programs for playing chess and solving tasks in a grid-based environment. They report that the learned strategies produce effective actions and can be learned from gameplay data.
Abhijeet Krishnan, Colin M. Potts, Arnav Jhala et al.
In briefAsymptotic analysis of risk minimization on heavy-tailed random data
SummaryMany real datasets contain extreme values far more often than Gaussian models predict. This paper studies linear regression with empirical risk minimization on entry-wise i.i.d. symmetric α-stable data. Using the replica method, it fully characterizes the generalization error when sample size and feature dimension grow at a fixed ratio. The analysis also yields a heavy-tail universality law, scaling laws, and the Bayes-optimal prediction error.
In briefMeasuring climate backlash in Twitter and Reddit archives using lexical definitions and participant turnover
SummaryThe paper analyzes four Twitter and Reddit archives, processing all files without sampling and using transparent lexical rules to link climate-backlash frames with platform-specific response measures. It separates changes among returning authors from participant turnover, finding both contribute to a rise in protest-language shares around the 2019 global climate strike. The authors show that operational definitions, platform response fields and observation boundaries shape what can be claimed, and the study does not measure individual opposition or persuasion effects.
In briefCombining supervised and self-supervised graph representations for out-of-distribution learning
SummaryThis paper studies whether self-supervised graph representations can complement supervised learning to improve node classification when test graphs differ from training graphs. The authors propose two backbone-agnostic methods: Co-Train, which jointly learns and adaptively integrates both representation types during training, and Dual-Space Retrieval, a non-parametric approach that fuses predictions from two representation spaces at inference. Across four benchmarks with temporal and cross-domain shifts, Co-Train consistently outperforms strong supervised baselines.
In briefLearning to Go Beyond a Fixed Theory: Experimental Discovery Outside the Initial Hypothesis Space
SummaryThe paper argues that scientific discovery systems usually optimize experiments within a fixed hypothesis space, which fails when all candidate models omit the same mechanism. The authors propose experimental model-class revision, where a policy jointly proposes a structural edit and a diagnostic experiment testing whether the edit is needed, using anytime-valid sequential evidence to trigger revision. On 400 held-out controlled dynamical environments, the method reaches 89.5% exact recovery with 32 real experiments, 10.0 percentage points above the strongest matched baseline.
WhyJointly learns structural model revision and diagnostic experiments as one sequential decision, with anytime-valid evidence; validated on 400 environments and multiple transfer tests.
In briefA new sentence stress detection method that uses auxiliary word stress modeling and a loss regularizer
SummaryThis paper proposes a method for detecting sentence stress, the emphasis on semantically important words, by jointly modeling it with auxiliary word stress detection. It also introduces a word-span stress regularizer that concentrates token-level stress probabilities within each stressed word. Experiments on the TinyStress-15K benchmark show the method outperforms strong baselines, with the full configuration achieving the best sentence stress detection result.
In briefStateless Language Agents: Scaling Long-Horizon Automated Research
SummaryThis paper proposes Stateless Language Agents (SLAs), a design for long-running automated research in which agents do not carry conversation history across calls. Instead, a harness stores the research state, such as candidate solutions and measured outcomes, and rebuilds a fresh, role-specific context for each invocation. Evaluated on software engineering, kernel optimization, and algorithm design with budgets up to one billion tokens, SLA achieved the best final result on every task and matched the strongest kernel baseline using over 84% fewer tokens.
WhyProposes stateless agents with harness-owned research state, validated against three frameworks at billion-token budgets; impact mostly limited to automated research systems.
In briefExplicit tighter bounds on sequential forecast calibration error beyond T^{2/3}
SummaryThe paper studies sequential forecasting of binary outcomes, where predicted probabilities should match observed event frequencies. It introduces a two-phase recursive labeling strategy for the sign-preservation-with-reuse game and refines the reduction from sign preservation to calibration. Combining these improvements yields an explicit bound of O(T^{0.662942288}), the first explicit exponent below 2/3 for sequential calibration.
WhyGives a new two-phase labeling strategy and sharper reduction, yielding the first explicit exponent below 2/3 with rigorous analysis, though impact is confined to calibration theory.
In briefExplore-then-commit: discovering scientific laws with fewer measurements using language models
SummaryThis paper proposes an explore-then-commit protocol in which a language model suggests hypotheses, a programmatic planner collects measurements, and a fresh prompt derives the final law from fixed observations. Across 576 NewtonBench trials, interpreter-enabled planners used far fewer measurements on medium-difficulty tasks and reduced error. However, symbolic-accuracy gains were inconsistent across modules, and the authors note that causal components and generalization beyond noiseless direct-equation tasks remain unresolved.
In briefAFA-BANDIT: Near-optimal online feature selection for classification under a budget
SummaryThe paper studies online active feature acquisition, where an agent decides which costly features to obtain before predicting each sample's label. The authors frame this as a combinatorial Bandits with Knapsacks problem and derive an improved regret upper bound. They also propose LP-Chain, a variant that searches a cost-aware chain of feature subsets whose size grows linearly with the number of features. On synthetic data, LP-Chain outperforms HEDGE-based BwK and deep reinforcement learning baselines.
In briefBioStudyBench: Testing AI agents on biomedical studies published after model cutoffs
SummaryThis paper introduces BioStudyBench, a benchmark of 25 long-horizon analysis tasks based on biomedical studies published after the evaluated models' knowledge cutoffs. Agents receive only a neutral research question and must find public data and search literature restricted to pre-cutoff records. Each task is run with and without data and tools to measure gains over prior knowledge. Access to data and tools raised pass rates by 47 percentage points on average, with the best closed-weight model passing 94.7% of tasks and the best open-weight model 81.3%.
In briefLearning where a crane grapple should grasp logs from point clouds
SummaryThis paper presents a learned policy that uses unsegmented point clouds to choose where and how to orient a grapple for clearing dense log piles with a trailer-mounted hydraulic forestry crane. The method combines behavior cloning with reinforcement learning and is compared against a geometric heuristic in simulation and in twelve field trials. In the field, the learned policies deposited a larger share of logs than the heuristic, but the stability gain seen in simulation did not carry over to the crane.
George Sideris, Lucas Bessai, Heshan Fernando et al.
In briefHub-and-Spoke Latent Geometry for Dual-Mismatched Semi-Supervised Learning
SummaryThis paper addresses semi-supervised learning when unlabeled data differs from labeled data in both class distribution and label space, containing imbalanced classes and unknown-class samples. The authors propose a hub-spoke latent geometry where known classes are evenly arranged around a central hub with compact prototype-based clusters, and the hub anchors a low-evidence region for high-uncertainty unknown-class samples. Combined with an evidence-based classifier, the approach improves feature discriminability and uncertainty separation. Experiments report that it outperforms state-of-the-art methods, with a maximum improvement of 3.25% across settings.
In briefPhysLDM: Latent Diffusion for Accurate Simulation of Deformable Objects
SummaryThis paper introduces PhysLDM, a latent diffusion framework for one-shot simulation of deformable bodies. Its core is a spatiotemporal VAE that compresses high-resolution 3D volumetric meshes, reaching about 2.48 mm reconstruction precision on meter-scale scenes with up to 78x token compression. The authors compare regression and diffusion, finding that diffusion better captures the distribution of chaotic deformable dynamics, where regression tends to average out physical behavior. Trained purely kinematically, the model generalizes zero-shot to unseen datasets.
In briefLinear fitness subspace in protein language models for sample-efficient directed evolution
SummaryThe paper proposes the Linear Fitness Subspace hypothesis: within mutation-induced changes in protein language model representations, a small assay-specific set of directions lets fitness variation be captured linearly from few labeled variants. Building on this, the authors introduce Subspace-Guided Evolutionary Search (SGES), which performs surrogate modeling, uncertainty estimation, and acquisition within the learned subspace. Across ProteinGym assays and a budgeted search evaluation, SGES reportedly improves fitness prediction and search efficiency over zero-shot protein language models and recent ML-guided optimization baselines.
In briefVALSE: A Method That Skips Layers Adaptively per Input to Make Large Language Model Inference More Efficient
SummaryThe paper develops a theoretical framework for skipping layers adaptively in a vertical direction, including a closed-form expected FLOPs formula and results showing skip-layer models are strictly contained in but well approximate the full-layer function space. It also draws a structural duality between this approach and Mixture-of-Experts architectures. Building on the theory, the authors propose VALSE, which scores each input with a lightweight difficulty estimator and skips redundant layers, including non-contiguous middle layers, and its feasibility is preliminarily assessed at prototype scale.
In briefEmoception: Efficient fine-tuning of video transformers to detect player arousal changes from gameplay footage
SummaryThis paper proposes SALFT, a method that fine-tunes only selected layers of a Video Vision Transformer to recognize player arousal changes from gameplay videos. Layers are chosen by measuring the L2-norm change in their parameters after brief adaptation. On the Arousal Video Game AnnotatIoN dataset, SALFT performs comparably to full fine-tuning across all games without statistically significant degradation while updating about 8% of parameters. The authors also introduce a method for tracing attention patterns to improve model interpretability.
In briefA neural JKO method for gradient flows in the Hellinger-Kantorovich geometry, using paired transport and growth maps
SummaryThe paper proposes a mesh-free neural JKO scheme for advection-reaction-diffusion equations viewed as gradient flows in the Hellinger-Kantorovich geometry of unbalanced optimal transport. Each step uses a spatial map and a mass-change factor, so transport and local mass creation or loss are handled in one variational update. The authors establish existence and mass bounds for JKO minimizers, and under extra assumptions prove positivity, regularity and a discrete Euler-Lagrange equation. Numerical experiments check agreement with the PDE and energy dissipation.
In briefBeyond Scalar IoU: Structured Verification for Video Temporal Grounding
SummaryThis paper introduces SUTURE, a reward verifier for reinforcement learning on video temporal grounding that scores generated intervals jointly across a rollout group rather than independently. It uses disagreement across rollouts and per-position coverage to reweight and redistribute reward. Across five benchmarks, SUTURE improves grounding performance at every reported IoU threshold and reduces video-start anchoring in reasoning traces.
In briefRobust Decision-Making in World Model Latent Spaces via Latent Disturbance Modeling
SummaryThis paper studies robust decision-making in the latent space of world models by modeling latent disturbances as perturbations to learned latent dynamics. The authors build a set of plausible pessimistic transitions using a dynamics-aware similarity metric and out-of-distribution detection, calibrated with conformal prediction. Robust actions and worst-case disturbances are jointly optimized via game-theoretic methods and applied to safety filtering and policy steering. Franka manipulator experiments report 70% fewer failures in safety filtering and 54% fewer in sampling-based policy steering.
WhyIntroduces latent-space robust optimization for world models using conformal calibration and game-theoretic optimization; hardware tests cut failures 70% and 54%.
In briefRobot Descriptions Versus Robots: GaugeBench Tests Policy Robustness to Equivalent Re-Descriptions
SummaryThe paper introduces GaugeBench, which keeps a robot's physical mechanism fixed while rewriting its description using physically equivalent conventions such as joint-axis direction and joint-angle zero. The same policy weights score far lower on re-described robots than on genuinely new robots, with axis reversal alone reproducing most of the drop. The authors report that exact transport between descriptions and training across equivalent axis conventions can recover performance, and they argue cross-embodiment evaluation should test representation robustness as well as mechanism robustness.
WhyProposes description-only convention changes with fixed mechanism, showing policy collapse across models and frameworks, separating mechanism from representation robustness with supporting ablations.
Rahath Malladi, Arshia Sangwan, Rajesh K. Gupta et al.
In briefBiGym 2.0: Benchmark for Humanoid Household Tasks Using Learned and Agent-Written Robot Control
SummaryThis paper presents BiGym 2.0, a benchmark adapting BiGym to the Unitree G1 humanoid across 20 household tasks, with a shared whole-body controller and 60 human VR demonstrations per task. It compares vision-language-action fine-tuning, imitation learning, demonstration-driven reinforcement learning, and cold-start coding agents. Vision-language-action fine-tuning achieves the highest nine-task mean, and agent-developed programs beat the demonstration-driven RL baselines on that mean. Cross-workspace stacking and multi-object transport remain difficult. Environments, demonstrations, and evaluation traces are open-sourced.
WhyAdapts BiGym to Unitree G1 with unified controller, comparing multiple methods across 20 tasks; useful for the subfield.
In briefLSC-DPO: A Preference Training Method That Controls the Learning Signal
SummaryThis paper addresses a limitation of Direct Preference Optimization (DPO), where the training loss becomes less responsive as the preference margin grows. The authors treat the sigmoid factor as a learning signal and propose LSC-DPO, which dynamically regulates this signal near a target regime. Experiments on AlpacaEval 2, MT-Bench, and Anthropic-HH report consistent gains over DPO and strong baselines. The paper also introduces a signal-budget compensation rule that reduces performance variation across different coefficient initializations.
In briefRecurrent Looped Transformer: a model whose computation path grows with sequence length
SummaryThe paper introduces the Recurrent Looped Transformer (RLT), which splits its layers between a parallel causal encoder and a recurrent decoder that merges encoder output with the previous token's decoder state. On six algorithmic tasks, RLT generalizes far beyond training lengths on parity, permutation tracking, and modular arithmetic, where a standard Transformer performs much worse. Ablations show these gains depend on the feedback connection.
In briefARMD-Predict: multi-site machine learning to predict antibiotic resistance before culture results
SummaryClinicians must choose empiric antibiotics before culture susceptibility results are available. The authors trained ARMD-Predict on 7.5 million culture-antibiotic observations from four academic medical centers, using models for individual antibiotic susceptibility and a hierarchical model that predicts the organism before resistance phenotypes. Median AUROC was 0.77 for inpatient urine and 0.70 for blood cultures. The authors state that independent calibration, threshold selection and prospective validation are needed before clinical use.
WhyMulti-site hierarchical models predicting antibiotic resistance; evaluated across four centers with patient-grouped and chronological tests, supporting practical empiric prescribing relevance.
Fateme Nateghi Haredasht, Nicholas P. Marshall, Wenyuan Chen et al.
In briefPersonal-agent-mediated recommendation using cross-platform user history
SummaryThis paper proposes a paradigm in which a user's personal LLM agent adjusts a platform's recommendation ranking using user-authorized cross-platform history to produce the final top-K slate. The authors introduce MediateRec, a benchmark with proxy and real cross-platform tests, and PAMO, a training method that balances beneficial rescues against harmful overrides. Experiments show meaningful platform corrections but non-negligible harmful overrides even for strong proprietary LLMs, with PAMO outperforming matched outcome-only RL.
In briefREViT-v2: A hierarchical vision transformer that stays equivariant to rotations and reflections
SummaryThe paper proposes a scalable vision transformer that is equivariant to rotations and reflections, built on windowed group-convolutional self-attention and a hierarchical feature design. The authors show the approach scales to models with millions of parameters and to large datasets with practical image sizes such as ImageNet. Code and pretrained weights are publicly released.
In briefLooking Inside GraphCast: How It Internally Represents Atmospheric Rivers
SummaryThis paper uses sparse autoencoders to examine the internal workings of GraphCast, an AI weather model, focusing on atmospheric rivers. It finds that the model internally computes atmospheric river intensity, measured by integrated vapor transport, even though this quantity is neither an input nor a target. The Matryoshka variant ranks learned concepts by importance and reveals their relations, and direct interventions confirm that these internal variables are causal.
WhyApplies sparse autoencoders to GraphCast, finding a stable internal IVT representation of atmospheric rivers, with interventions confirming causality.
Madelyn Mathai, Timothy B. Higgins, Kevin M. Grise et al.
In briefMeasurement bias and resolution effects in 3D mitochondria shape analysis
SummaryThis paper examines how three-dimensional mitochondria measurements can differ systematically depending on the processing method. Occupancy-based volumes exceed reference mesh volumes by 3.665% on average, and removing a depth offset in a controlled pipeline reduces this inflation by about 45%. A frozen regression lowers median absolute percentage error on unused objects, but its error bound coverage differs between overall and low-occupancy subgroups. Coarser in-plane spacing also changes surface area and sphericity in rat-cortex objects.
In briefLOGIC: Benchmark for Intent-Based Change Impact in Aerospace Electrical Systems
SummaryThis paper presents LOGIC, a controlled benchmark for testing whether language models can select only the design changes an engineering request actually authorizes in aerospace electrical revisions. Selected changes are then propagated through a typed electrical traceability graph, which separates selection errors from propagation errors. The results show that structured evidence methods perform best on explicitly anchored requests, while performance drops for paraphrased requests and larger candidate lists. The authors conclude that literal evidence and model reasoning should be combined with engineering review when intent cannot be reliably established..
Muhammad Faraz Shoaib, Muhammad Qasim, Raisulhaq Mohammed Rizwan et al.
In briefCooperating with Future Collaborators: Multi-agent RL when agents join at different times
SummaryThis paper studies cooperative multi-agent reinforcement learning where agents participate at staggered times, so information left by earlier-acting agents can help later ones. The authors propose Staggered Participation Learning (SPL), a training-time method that gives earlier agents prospective information-acquisition supervision and later agents outcome-supervised learning for using that information. Across 60 MPE/RWARE backbone setting comparisons, SPL achieves higher observed mean task completion in every case, with an average difference of 14.1%. Gains are also reported for eight-agent teams and a physics-based UAV-UGV environment in Isaac Lab.
In briefCETUS: How Well Earth-Trained Image Models Transfer to Titan's Cassini Radar Images
SummaryThe study uses Cassini synthetic aperture radar images of Saturn's moon Titan to test whether image representations learned from Earth imagery can classify its terrain. It compares pretrained encoders (DINOv2, DOFA, CROMA) with classical image features and an untrained vision transformer, evaluating on geographically separate regions. Pretrained encoders achieved higher mean macro F1 under logistic regression, but rankings shifted with training settings, and further Titan training helped DINOv2 while hurting DOFA. Because the labels partly derive from the same radar data, scores reflect agreement with expert interpretation.
In briefTwo Vectors Replace In-Context Demos: Structured Task Adaptation via Embeddings (STAVE)
SummaryThis paper addresses the high cost of in-context learning in large multimodal models, where demonstration images must be re-encoded for every query. It proposes STAVE, which replaces demos with two task-specific vectors added to existing input embeddings: a readout vector for answer-producing tokens and a context vector for other structural token groups. Experiments on six multimodal models and five language models show STAVE matches or beats prior methods with fewer task parameters, and outperforms 15-shot in-context learning on 18 text tasks at zero-shot inference cost.|
In briefUnanimously Wrong: Certified Abstention from Medical LLM Consensus
SummaryThe paper shows that agreement among multiple answers from a medical question-answering system does not guarantee correctness, since every sample can share the same wrong answer. It proposes ProbeGuard, which decides whether to abstain by examining how consensus formed, using agreement trajectories, minority persistence, retrieval saturation, rationale coherence, and an active counter-evidence probe. Evaluated on three medical QA benchmarks, it raises discrimination between correct and incorrect unanimous consensus to 0.696 AUROC.
In briefOpenSplatGraph: Building Persistent Object-Level Scene Graphs from Dense Semantic Maps for Robots
SummaryThis paper presents OpenSplatGraph, a framework that builds persistent 3D scene graphs directly from an online Gaussian-based open-vocabulary semantic map. It adds a reliability-aware semantic field that tracks observation statistics to extract query-conditioned objects, which are then linked to graph nodes and updated over time. The authors report competitive results on standard 3D scene understanding benchmarks and in real-world robot experiments for online open-vocabulary perception and downstream tasks.
Binh Long Nguyen, Kien Nguyen, Clinton Fookes et al.
In briefAIMS: scalable novel-view rendering by fusing nearby views into anchor views
SummaryThis paper presents AIMS, a framework that makes feed-forward novel view synthesis scale to large numbers of input views. It selects spatially spread anchor views with farthest point sampling and fuses nearby observations into each anchor using a lightweight learnable integrator, so extra observations still contribute while the global view budget stays fixed. On RealEstate10K and ScanNet it reaches 29.41 dB and 17.73 dB PSNR, with rendering averaging 7.24 ms per view.
In briefComplementary Feature Domains: Information Preservation Is Not Predictive-Contribution Preservation
SummaryThe paper develops Complementary Feature Domains (CFD) theory, which treats predictive value as a context-dependent contribution system induced by representations and their realization families. It shows that an invertible representation change can keep target information intact while altering predictive contributions under a restricted decision family, and defines a CFD contribution defect to measure this change. In a controlled ECG experiment, a nonlinear bijective recoding preserved information in a frozen time-frequency representation but changed accuracy under a fixed affine learner, and applying the exact inverse restored all tested coalition accuracies.
In briefHouseholdBench: Testing whether AI language models can predict household economic decisions
SummaryThis paper introduces HouseholdBench, an evaluation that combines 6 U.S. household surveys and 32 prediction tasks covering consumption, income, labor, expectations, and housing. It tests whether large language models can predict household behavior, including responses to policy changes, using past behavior, demographics, and macroeconomic conditions. Most models beat a no-change baseline, with the best lowering numeric error by 12.2%, but gradient-boosted trees rank first on most tasks and open-weight models lag behind. The authors also report that a 4-billion-parameter open-weight model can match proprietary models through fine-tuning and aggregating 16 predictions per observation.
WhyIntroduces a new 32-task household benchmark, comparing 13 LLMs to baselines; fine-tuning and aggregation let a small model match proprietary ones.
Jin Huang, Diego Ferreras Garrucho, Yutong Xie et al.
In briefLearning a Mixture of GFlowNets: training several GFlowNet samplers together to explore discrete distributions better
SummaryThis paper studies using an ensemble of GFlowNets to sample from discrete target distributions, aiming for better state space exploration and convergence. The authors propose a general theoretical framework for mixtures, split into continuously and discretely indexed variants. They also introduce Stratum-Conditioned GFlowNets, which partition the state space and train components in parallel, and report faster convergence and better mode coverage without significant extra computation.
In briefLARK: A Low-Cost Surgical Tracking System Using Cameras and Kalman Filtering
SummaryThis paper presents LARK, an optical tracking system for image-guided surgery that uses multiple commodity RGB cameras and fuses their views to track instruments. Two methods, multi-view monocular pose fusion and multi-view triangulation, are tested under different occlusion levels and compared with a commercial stereo infrared tracker. With five cameras and adaptive Kalman filtering, median target registration errors were 0.64 mm for point localization and 0.73 mm for trajectory tracking on a precision grid. The tracking hardware costs under $1,000 USD.
George Sideris, Justin Cree, Andrew Stirling et al.
In briefSearching a synthesis-route latent space to design makeable molecules
SummaryThis paper presents RouteFlow, which reframes synthesizable molecular design as a search over a continuous route latent space where each latent decodes to a complete synthesis route. It uses reward-guided flow matching for sampling and adds a cycle-consistency mechanism to stabilize fine-tuning. Across 16 optimization tasks from Therapeutic Data Commons, the authors report the best sample efficiency among synthesizability-aware baselines, along with the best synthetic accessibility and highest retrosynthesis success rate.
In briefTAFFY: A Task-Adaptive Tabular Foundation Model with In-Context Diversity
SummaryThis paper introduces TAFFY, a tabular foundation model that uses an In-Context Diversity Prior to build synthetic pretraining contexts from multiple related environments created by controlled interventions and distribution shifts. It also uses a Task-Conditioned Looped Transformer that selectively reapplies shared Transformer blocks, with a task-conditioned gate, to refine contextual representations. The authors report that TAFFY attains the lowest average rank across six classification and five regression benchmark datasets.
In briefSeeing Invisible Hazards: Physics-Guided Visual Prompts for Safer Robot Navigation
SummaryThe paper proposes Physics-Guided Visual Prompting (PG-VP), a plug-and-play module that adds virtual obstacles to camera views so a frozen vision-language-action navigation model avoids invisible radiation or thermal hazards without retraining. The virtual obstacle moves across frames and is only rendered when a hazard is detected. On R2R-CE and RxR-CE val-unseen splits, PG-VP steered the policy toward low-risk actions in 84.9% and 83.2% of cases, with navigation success dropping by 6.8 and 7.9 points. Real-robot tests reported improved worst-10% trajectory safety of 63.45% against thermal sources and 32.59% against radiation sources.
In briefDecoupled Multi-Agent Orchestration (DeOrch): A learned method that separates task planning from worker selection
SummaryThe paper introduces DeOrch, which separates worker-agnostic planning from concrete worker selection. A two-stage planner first decomposes a task without worker information, then chooses collaboration operations using compact matchability feedback, while a lightweight matcher adapts online with a contextual bandit. According to the authors, DeOrch outperforms prior automatic multi-agent orchestration methods with fewer worker calls and transfers to an unseen worker pool without retraining.
In briefGlobal transport couplings to improve classifier-free guided flow models
SummaryThe paper introduces Global Transport (GT), a class-agnostic optimal-transport coupling computed without class labels, for training conditional flow models. GT hurts performance without guidance but consistently improves generation when combined with classifier-free guidance (CFG) across domains, model scales, and sampling budgets. The authors argue that couplings for conditional flows should be evaluated under the guided flow used at inference. They present coupling design as a training-time lever that requires no changes to architectures, samplers, or guidance mechanisms.
Katarina Petrovi\'c, Zander W. Blasingame, Danyal Rehman et al.
In briefGradient admission for small language model finetuning via LoRA
SummaryThe paper identifies three problems in LoRA fine-tuning of small language models: conflicting gradients, static data selection, and subspace saturation that overwrites useful updates. It proposes GRADE, which combines a state-aware selector that admits samples aligned with the evolving multi-task gradient field and a step-level gate that rejects likely destructive updates near saturation. Across three current backbones and seven heterogeneous instruction datasets, GRADE outperforms data-selection and PEFT-stabilization baselines and is the only method that consistently improves over standard LoRA on every architecture.
In briefDoes gradient EM for high-dimensional Gaussian mixtures really need a √d separation?
SummaryThe paper studies learning Gaussian mixture models with the EM algorithm and its gradient-based variants. Prior work required the minimum separation between true components to scale with the square root of the dimension d. The authors consider a hybrid EM algorithm and prove that when the separation is of order d^(0.5-ε), population gradient EM with random initialization cannot guarantee global convergence in sub-exponential time, even in the over-parameterized regime.
In briefFoundation-model-aided multi-agent reinforcement learning for optimizing wireless random access networks
SummaryThis paper addresses the high training overhead of multi-agent reinforcement learning used to optimize random access wireless networks. The authors propose a foundation-model-aided actor-critic algorithm within a consensus-based decentralized architecture. They provide a convergence analysis under local reward exchanges and nonlinear value function approximation, showing the same convergence order as conventional methods. Numerical results indicate the approach significantly speeds up learning for random access network optimization.
Myeung Suk Oh, Zhiyao Zhang, Alvaro Velasquez et al.
In briefSelf-correcting offline speech translation on a low-power edge device
SummaryThis paper presents a fully offline speech-to-speech translation system running on a Jetson Nano (4 GB) that corrects weak translations without retraining. It uses multilingual BERT cosine similarity as a quality estimation gate that triggers a secondary pass when confidence is low. Among three correction methods tested on English-Spanish FLORES-200 sentences, Minimum Bayes-Risk decoding (M2) gave significant gains, QE reranking (M1) gave none, and constrained beam search (M3) was significantly worse than baseline.
In briefA systematic study of bias in LLMs judging advertising relevance
SummaryThis paper systematically examines whether large language models fairly judge how well advertisements match search queries. Using a counterfactual framework, the authors test how advertiser identity, possible popularity, input language, and demographic wording affect relevance judgments, using GPT-4o and a Qwen-7B model trained for relevance prediction. Changing the advertiser or input language can alter the assessment, and some gender and occupation comparisons follow common stereotypes. The authors also study mitigation methods, finding their effectiveness depends on whether advertiser information is relevant to the query and on label distribution in training data.
In briefKeeping Unstable Modes in JEPA World Models via Inverse Dynamics
SummaryRobots often have unstable modes that grow unless corrected by feedback, so controlling them from camera images requires representations that keep these modes. The paper shows that next-step prediction with anti-collapse regularization alone can lose controllable unstable modes, making stabilization impossible from the learned representation. The authors add an action reconstruction (inverse dynamics) loss, prove its effect on linear systems, and test the approach on nonlinear visual control tasks.
Leonardo F. Toso, Yann LeCun, James Anderson et al.
In briefSeparating model effects from persona effects in mixed-model LLM social simulations
SummaryThis paper simulates a social network in which agents are powered by several different base language models. It finds that the engagement an agent receives depends more on its base model than on its assigned persona. As more models are added, the attraction or repulsion effects of each base model grow markedly, suggesting that network dynamics may converge to base model effects at scale. Content-mediating analyses examine how predictable models are across contexts and how their lexical patterns relate to engagement-maximizing style.
In briefSkillFormer: Skill-Decomposed Adapters for Audio Language Models
SummaryAudio language models must handle many distinct skills, but training them jointly on all skills causes interference between tasks. SkillFormer splits audio understanding into skill-specific low-rank adapters and uses a learned router to combine them according to each question at inference time. An alternating schedule updates each adapter on its own skill cluster before jointly calibrating the router. It adds fewer than 4% of the base model's parameters and leaves the audio encoder and language backbone unchanged, raising average accuracy across three models on MMSU, MMAU-Pro, and MMAR.
In briefTargeted search reveals random-device testing underestimates worst-case error in a simulated wave-based neural operator
SummaryThis paper studies how large errors can grow in wave-based optical processors used as Fourier layers in neural operators, given fabrication and alignment variation. In a simulated case study, the authors compare randomly sampled devices with a searched device and find the searched device's error often exceeds the maximum over many random devices. The authors conclude that random testing estimates how often errors exceed a threshold, while worst-device search gives a lower bound on how large errors can get.
Samrendra Roy, Jason Yoo, Souvik Chakraborty et al.
In briefActivation Denoising: Parallel LLM Quantization via a Robustness View
SummaryThis paper addresses how quantization errors compound across layers when compressing large language models after training. It proposes parallel quantization with activation denoising, which treats upstream errors as noise and applies a robustness regularization through preprocessing and metric-weighted rounding. The authors report that this recovers much of the benefit of slower sequential quantization in a single parallel pass, at a fraction of the time, and that it complements orthogonal rotation methods.
Yan Scholten, Rachel Lawrence, James Hensman et al.
In briefGrounding What Shapes the Plan: Rethinking Groundedness for Driving
SummaryThis paper argues that driving models that ground reasoning in causal, spatial, and perceptual evidence still need a clear link to the final action. It introduces GroundAct, which treats physical entities as the unit of grounding, uses a lightweight reference token to keep their continuous states addressable, and corrects the plan only through interactions between referenced entities and the evolving proposal, a approach called grounded planning. The method is evaluated in open-loop and closed-loop simulation settings.
In briefFrom Local Evidence to Safety Verdicts: Causal Tracing in Vision-Language Models
SummaryThis paper studies how vision-language models combine an image and a prompt to reach a joint safety verdict. The authors build SSU-Bench, a dataset of matched safe and unsafe image-text pairs, and swap internal states between paired inputs in three models. They find that intervening at changed input positions works in earlier layers, while intervening at the final token works in later layers.
Faezeh Dehghan Tarzjani, Mevan Wijewardena, Alexander Romanus et al.
In briefMobileVISTA: Generative Data Augmentation for Robust Robot Pose Generalization
SummaryThe paper introduces MobileVISTA, a framework that turns demonstrations collected at one canonical robot pose into pose-perturbed training data by augmenting egocentric images and retargeting actions to match base pose changes. Tested in simulated humanoid and bimanual tasks and on a real Galaxea R1 Pro, policies trained on the augmented data were more robust to out-of-distribution poses at test time. The benefit was largest for humanoids, where the camera moves with the robot's kinematic chain and the robot fills much of the view.
Suzannah Wistreich, Stephen Tian, Isabella Huang et al.
In briefOpen-ended information seeking by AI agents that gather information through questioning
SummaryThis paper studies information elicitation, an open-ended problem where an agent must keep deciding which information to pursue as new information emerges. It compares how 11 large language models of different families and sizes judge information value and how those judgments shape sequential information seeking. Using a controlled simulation where models share the same information space and selection rule, the authors characterize breadth-depth behavior and test robustness through sensitivity analyses and ablations.
In briefJarvis: A Proactive Speech Agent for Multi-Party Conversations
SummaryMost speech agents only respond when addressed and interact with one person at a time. This paper presents Jarvis, a real-time agent that joins multi-person conversations and speaks only when it can help. Grounded in a document shared beforehand, it intervenes when the group misses or misstates a fact and does not correct itself within a few turns. The authors also release CHI-180-proactive, a synthetic multi-party dataset with known gaps, errors, and self-corrections. Jarvis is correct on most events it addresses and stays silent 97% of the time when the group resolves issues itself; a live study with 23 participants confirms these trends.
Seunghyun Oh, Hirotaka Hiraki, Shuyue Stella Li et al.
In briefMARS: A Multi-Resolution Adaptive Routing Method for Sequential Recommendation
SummaryThe paper observes that long user histories mix short-term intent, medium-term interests, and long-term preferences, but single cached memories retain these scales unevenly, recovering recent and mid-range content much worse than long-range content. The authors propose MARS, which writes the full history into recurrent state tracks tied to different half-lives and uses sparse routing to pick relevant temporal resolutions for each seed memory. Experiments on three public datasets show MARS outperforms strong baselines, with gains growing as history length increases.
In briefTwo-sample tests for random graphs without vertex matching
SummaryThe paper studies how many graphs are needed to tell apart two groups of random graphs when their vertices have no known correspondence, and which graph statistics detect which differences. For an Erdős–Rényi null and a planted two-block difference that keeps expected degrees unchanged, signed triangle counts reach the stated sample-size rate. Without vertex alignment the required sample size grows by a factor of order t^{-2}.
WhyProves tight sample complexity m≍t^{-3} for unaligned graph two-sample testing, shows tree statistics fail, and simulations match predicted exponents.
In briefAutomating physician queries to fill gaps in ambient clinical documentation
SummaryThis paper studies whether an LLM can automate the Draft, Ask, Update (DAU) query loop that documentation specialists use to clarify missing information in clinical notes. The authors audited 3,000 real visits to identify sources of missing documentation and built five transcript-degradation benchmarks from public data. Analysis of 21k clarification turns found that useful-question predictors differ by task, and that about 9% of turns hurt performance.
Joseph Paul Cohen, Raj Shah, Han-Chin Shing et al.
In briefSelective test-time adaptation for multimodal wearable sensors using source-learned reliance
SummaryThis paper addresses reliability of multimodal wearable systems when sensor streams become noisy or unavailable. The proposed method, CARAT, separates model reliance on each modality from runtime corruption detection, estimating reliance during source training. At deployment, a lightweight one-class detector flags suspect streams, and the system either replaces them with a missingness symbol or attenuates their representations. Across four wearable datasets, five corruption types, three backbones and eight baselines, CARAT achieves the highest overall macro-F1 and best mean rank, while using fewer GFLOPs and updating fewer parameters than EATA.
In briefDoes Muon Need Fine-Grained Spectral Shaping?
SummaryThis paper asks whether the Muon optimizer needs per-direction spectral weighting in its matrix updates. It proposes BulkBoost, which splits singular modes into a bulk below a noise edge and spikes above it, and boosts the bulk with one shared gain. Across 30 continued-pretraining settings, two-band reweighting is competitive with the fine-grained Freon profile and outperforms Spectra.
In briefGeometry-aware diffusion sampling for reconstructing CT images from sparse or limited angles
SummaryThis paper addresses ill-posed image reconstruction in sparse-view and limited-angle CT, where missing angular information makes results ambiguous. The authors propose a geometry-aware diffusion-guided stochastic framework built on a regularized noise-weighted pullback metric derived from the CT forward operator and noise covariance. The metric adapts updates and exploration to how sensitive each direction is to measurements. Experiments report competitive reconstruction quality, strong measurement consistency, and spatially resolved uncertainty estimates.
Honglei Brinkmann, Carola-Bibiane Sch\"onlieb, Ander Biguri
In briefWho Pays for Shared Limits? Learning Responsibility Shares in Multi-Agent RL
SummaryWhen several agents share one cost budget, a common Lagrange multiplier enforces the total limit but does not say how the penalty should be split among agents. The paper proposes Lagrangian Responsibility Allocation (LiRA), which learns each agent's share of the common multiplier by maximizing social welfare over a finite training horizon. Tested on CityLearn, MABIM, Harvest, and MetaDrive with 3 to 400 agents, LiRA improves average social welfare by up to 29% over uniform and agent-specific multiplier baselines while keeping costs within budget.
In briefRobust importance sampling for rare-event probabilities using constrained Gaussian mixtures
SummaryThis paper estimates rare-event probabilities for Gaussian variables passed through a general function, using importance sampling. The method separates the problem into a coverage stage and a fitting stage, and constrains the final Gaussian mixture proposal so its importance-sampling variance is finite. Experiments report lower variance and more stable results than several baselines, especially in high-dimensional and multimodal settings.
Pawe{\l} Lorek, Rafa{\l} Nowak, Rafa{\l} Topolnicki et al.
In briefSpecBraM: Which target should EEG foundation models predict? Band-power versus waveform reconstruction
SummaryThis paper proposes masked band-power prediction (MBP), a self-supervised EEG pretraining method that predicts fixed narrow-band log spectral energy for masked patches instead of reconstructing waveforms. On sleep staging with ISRUC and HMC, MBP outperforms waveform-based baselines under a strict linear probe, with larger gains when only 1% of labels are used. The gains do not extend to all tasks, such as motor imagery, depression screening, and vigilance regression.
In briefCombining Classic Document Automation with Generative AI
SummaryThis paper examines how legal document automation services built on expert-system and other symbolic approaches could enhance and be enhanced by current generative AI. It discusses possible benefits and challenges of combining the two. The authors also report preliminary experiments using large language models to identify and fix issues in texts written by laypeople.
In briefAdapting RL policies to changing agent behavior using finite-depth sensitivity estimates
SummaryThe paper addresses how a reinforcement learning policy can adapt to changes in another agent's behavior without large amounts of new interaction data. It proposes a finite-depth framework that estimates policy sensitivity by approximating the policy Hessian and mixed derivative from a reference environment, with an adjustable propagation depth. Truncation-error bounds are derived, and in a belief-driven pursuit-evasion game the method lowers derivative-estimation error as depth increases and improves zero-shot return over direct transfer.
In briefPsyCIDRA: A Dual-Agent AI Framework for Psychiatric Interviews and Diagnostic Reasoning
SummaryThis paper presents PsyCIDRA, a two-agent system in which an interviewer agent conducts free-form psychiatric interviews using tools for notes, expert skills, and ICD-11 references. A separate diagnostic reasoning agent reviews the transcript, reporting hypotheses with supporting, conflicting, and missing evidence, and withholds a diagnosis when evidence is insufficient. In simulated cases it outperformed direct prompting on diagnostic agreement and rank-1 accuracy (60.5% vs. 51.9%), and in a blinded study with 101 participants it matched psychologists on whether to propose a hypothesis in 79.6% of cases versus 65.4% for direct prompting.
Milad Mohammadi, Fatemeh Akrami Shamsabadi, Zahra Mohseni et al.
In briefStructure, Not Belief: Correlated Thompson Sampling Guided by LLM-Derived Covariance
SummaryThis paper studies combinatorial Thompson sampling where an LLM partitions the arms once, and the partition is converted into a correlation matrix that shapes how posterior samples are drawn, while Beta posteriors are updated only from real rewards. The authors provide a Bayesian regret bound for an idealized Gaussian sampler and report regret reductions of 19% on 16 synthetic Bernoulli families at T=2,500 and 41% on the MIND-small news benchmark compared with CTS. An LLM-free ablation indicates that gains on unstructured instances come from the kernel shape rather than injected beliefs.
In briefCOMPASS: Locating Where Reasoning Happens in Language Models
SummaryThis paper introduces COMPASS, an inference-time method that steers language models toward reasoning by identifying the attention heads that carry it. The authors use the correctness of the model's own direct answers as a signal to find a latent direction that elicits reasoning, and show that only a small subset of heads can be effectively intervened on. Across three model families and several math benchmarks, COMPASS outperforms the activation-steering baselines compared, raising GSM8K accuracy by 16 points on average while generating 20-70% fewer tokens than CoT.
In briefEfficient multimodal inference by adaptively choosing and fusing inputs
SummaryMultimodal systems usually encode every available input, even when a subset would be enough for prediction. This paper introduces SemARC, which pairs a Sequential Modality Aggregator (SeMA) with an Adaptive Runtime Controller (ARC) that uses acquired evidence to choose the next modality to encode or to stop. Across six multimodal classification datasets, SemARC achieves on average 3.2% higher macro-F1 and 61.4% lower total inference GFLOPs than each dataset's most accurate baseline.
In briefProtective noise that survives image resizing against malicious AI edits
SummaryThe paper addresses instruction-guided diffusion editors that can forge personal photos, and notes that existing protective perturbations weaken once photos are downscaled by an unknown factor. The authors propose SRIM, which uses anchor scales to cover the range of likely downscaling factors and optimizes at the scale where protection is currently weakest. Experiments on full-resolution photos and popular editors show SRIM gives the strongest protection, even with one quarter of the baselines' ℓ∞ budget, and remains strongest under input purification.
In briefElasticFit: Fit-Aware 3D Object Insertion Using VLM Reasoning and Generative Adaptation
SummaryThis paper presents ElasticFit, a framework for inserting objects into existing 3D scenes so they fit local geometry. A vision-language model reads a language instruction and scene renderings to infer placement, volume, orientation, and an adaptation mode such as rigid placement, uniform scaling, or elastic fitting. The object is then generated, reconstructed in 3D, and refined under collision and contact constraints. Against fixed-asset baselines, spatial relation success rose from 50.8% to 69.7% and support success from 48.3% to 91.7%.
In briefAuditable claims about AI agents: what must be defined and recorded to verify them
SummaryThe paper argues that a claim about an AI agent can only be checked if its policy, scope, settling records, and record writers are named first. It defines a claim as auditable when these elements and a decision rule are fixed and the records are obtainable. The authors prove that support is impossible without independent record coverage, per-action authorization, and completeness. A claim-check table and a worked five-state case illustrate the method.
In briefInterpretable hypergraph learning using neural additive models for transparent predictions
SummaryThis paper introduces the hypergraph neural additive network (HGNAN), a learning framework for hypergraph-structured data that is interpretable by design. It extends neural additive models to higher-order relational data by combining feature-wise nonlinear decomposition with hypergraph-aware structural aggregation. Experiments on benchmark datasets show performance comparable to state-of-the-art hypergraph learning methods, along with intrinsic interpretability.
In briefAlignQuant: Tile-aligned mixed-precision quantization for faster LLM generation
SummaryAlignQuant is a post-training quantization method that uses GPU-compatible two-dimensional weight tiles as the shared unit for precision allocation, storage, and computation. Precision is assigned according to sensitivity, using joint prefill and decode calibration under a model-wide storage budget. Across four LLMs from 3B to 14B parameters, it reports up to 2.50x generation speedup over BF16 while preserving model quality, with evaluations on three GPUs and contexts up to 64K tokens.
In briefSelecting Fewer Survey Items Over Time to Reduce Participant Burden While Predicting Outcomes
SummaryThis paper addresses longitudinal psychology studies, where collecting many variables improves prediction but burdens participants. It proposes distilling neural-network-based Longitudinal Active Feature Acquisition policies into interpretable decision trees that choose which items to collect at each timepoint. The method was tested on a simulation and on an ecological momentary assessment dataset forecasting daily alcohol consumption. The authors report that the number of items per occasion can be meaningfully reduced with minimal loss in accuracy.
In briefFork-and-Flush: Helping Research Agents Escape Stuck Solution Areas
SummaryIndependent runs of the same autoresearch agent often stall at very different scores, and the gaps persist even with more compute. The authors call these localized regions of the solution space "idea basins" and propose fork-and-flush, which splits the agent into parallel trajectories that keep the workspace but start with fresh chat context, then continues from the best one. Across 13 long-horizon research and engineering tasks under an equal compute budget, it improved the min-max normalized average score by 66.0% over a single run and 44.4% over best-of-N.
Ziyang Cai, Christos Ziakas, Vasilis Kontonis et al.
In briefSeparating what from where: how a small GUI grounding model should receive the action type
SummaryThe paper studies how a small GUI grounding model should receive the action type, such as tap or swipe, when deciding where to act on a screen. The authors fine-tune Qwen2-VL-2B with LoRA on the Android in the Wild dataset and compare five conditioning methods against a flat baseline. The auxiliary loss, additive embedding, and prompt word each improve hit rate by roughly five to seven points, though much of the gain traces to a preprocessing choice rather than a spatial prior.
In briefArtifact removal cleans electrodermal signals but does not improve classification in a VR balance task
SummaryThis paper tests whether removing artifacts from electrodermal activity signals improves classification in a virtual-reality balance-disturbance task. The artifact-removal network detected artifacts well and reduced waveform error on a benchmark, but classifier accuracy on the VR recordings did not improve over raw input. The authors conclude that the waveform benefit is lost before the decision stage, so preprocessing should be judged by the decisions it supports against an unprocessed baseline.
In briefStaFIR: convex learning of causal filters that balance stationarity and input similarity
SummaryThe paper proposes StaFIR, a causal finite-impulse-response filter that learns a nonnegative mixture of exponential lag profiles to reduce time-series nonstationarity. Its convex objective balances empirical stationarity with similarity to the input. Tests on ARFIMA–GARCH simulated data and rolling financial series, including realized-volatility forecasting, show that it adjusts filtering strength to persistence. In downstream forecasting it shows no clear accuracy difference from fixed half-order differencing, but achieves higher similarity to the raw signal.
In briefAdaptive Gait Biofeedback for Chronic Ankle Instability with Participant-Specific Model Updating
SummaryThe study evaluates a temporal convolutional classifier that labels gait cycles as GOOD or BAD using angle thresholds, tested with leave-one-subject-out validation in 20 participants. Seven intervention participants completed nine sessions over three weeks, and models updated after failed sessions were compared with their parent models. Frontal-plane ankle angle was compared against 10 control participants, showing a Post-intervention association in a nonrandomized design. The findings do not establish independent clinical gait classification or a causal benefit of model updating.
Jaeyoon Kim, Veronika Lebisova, Jeniya Sultana et al.
In briefAccentCL: A robust English accent classifier that can add new accent categories over time
SummaryThis paper presents AccentCL, a class-incremental learning framework for classifying English accents that stays robust to class imbalance and differences between recording corpora. It uses multi-layer features from a frozen Whisper-Large-v3 encoder, trained with an imbalance-aware loss and a domain mean alignment loss, and expands its label set through replay-based continual learning. On a five-class task it reports 77.1% balanced accuracy and 76.9% macro F1. When Spanish-accented and Chinese-accented English were added incrementally, the new classes reached 83.3% and 61.8% F1 while base-class balanced accuracy stayed around 77%.
Mu-Ruei Tseng, Waris Quamer, Ghady Nasrallah et al.
In brief2D-FET-Bench: From Spatial Reasoning to FET Design on Flakes
SummaryThe paper introduces 2D-FET-Bench V2, a benchmark of 128 layout tasks for field-effect transistors on exfoliated 2D flakes, built from microscopy-derived contours. Agents generate layouts rendered to GDSII, which a deterministic verifier checks for geometric and structural requirements. The best configuration, GPT5.6-Luna with ReAct-3, passes 62.3% of attempts and solves 80.5% of tasks at least once.
In briefBenchmarking online updates with revealed labels for EEG brain-computer interface decoding
SummaryEEG brain signals drift over time, so brain-computer interface models trained once can lose accuracy in practice. The paper benchmarks online adaptation for two common pipeline families, Common Spatial Patterns and Riemannian covariance methods, using time-ordered test-then-train evaluation. Label-revealed online updates improved 13 of 14 model-dataset pairs on the two largest streams, with relative accuracy gains up to about 18% over a frozen model.
Bogdan Kozyrskiy, Artem Grachev, Abraham I. Camelo Guerrero
In briefLearnable Spectral Activations: Letting Neural Networks Learn How to Combine Frequencies
SummaryThis paper proposes learnable spectral activations (LSA), which replace fixed neuron nonlinearities with a residual truncated Fourier series whose harmonic amplitudes are learned during training. Linear weights select features while activation coefficients shape the spectrum, and the two are updated by separate gradients. The authors report that this factorization concentrates more target-signal energy in the leading eigenmodes of the neural tangent kernel and improves reconstruction quality across audio, image, neural radiance field, and neural acoustic field tasks.
In briefBayesian optimization over function spaces using sparse kernel-based methods
SummaryThis paper addresses Bayesian optimization where the variable being optimized is an infinite-dimensional function rather than a finite vector. The authors propose L0 Manifold Optimization (L0MO), which searches functions in a Reproducing Kernel Hilbert Space that have a sparse representation, optimizing both kernel locations and coefficients. They also create new benchmark functions that adapt standard finite-dimensional test functions to infinite dimensions. Experiments reportedly show the method performs better across a wide range of benchmarks.
Davide Sartor, Meghan E. Huber, Donghyun Kim et al.
In briefEvaluating Feature-Acquisition Policies with Pretrained Tabular Models
SummaryThis paper studies how to learn and evaluate policies that sequentially acquire features to gain information about a target, using prior-data fitted networks (PFNs) trained without task-specific fitting. The authors show that with imbalanced offline data coverage, using total predictive entropy as a reward biases evaluation against sparsely observed features, because it mixes epistemic and aleatoric uncertainty. They instead target the posterior expected aleatoric entropy. Experiments on synthetic and real datasets show reduced value estimation bias and credible intervals with strong empirical coverage.
In briefWhat pass@k Misses: Diversity and Capability Retention After RL Post-Training
SummaryThe paper argues that pass@k, the standard metric for judging reinforcement-learning post-training, depends only on each problem's probability of a correct sample and cannot capture how correct outputs are distributed, so it cannot measure diversity. Training Qwen2.5-1.5B-Instruct on GSM8K with GRPO and rejection-sampling fine-tuning moves diversity measures in opposite directions, while pass@8 and pass@32 show no consistent winner. The authors present this as a concrete case of a standard evaluation protocol missing a property it is routinely used to certify.
In briefLayered defenses for LLMs: testing memory gates against sycophancy from activations and memory
SummaryThis paper studies a problem with long-term memory in language models: stored user history can make models favor a user's beliefs over objective evidence. The authors propose a 2×2 framework that separates internal activation steering from external memory handling, and evaluate four open-weight models on MemSyco-Bench. Selective filtering with a Router Gate preserves more average accuracy than removing all memory. Mild inverse steering reduces sycophancy in one setting, but the effect is not statistically significant, so the authors conclude that external memory filtering is the part that holds up.
In briefFree Detection of Fabricated Citations for Dissertation Assessment
SummaryThe paper introduces the Fabrication Severity Stack (FSS), a three-tier framework that classifies citation fabrications by how hard they are to detect, from non-existent references to subtle content misattribution. It proposes a staged heuristic verification workflow that can be run through mainstream LLM interfaces, aimed at computer science examiners assessing dissertations. In a pilot on 100 citations with 20% seeded fabrications, the workflow helped LLMs flag 9 to 10 of 10 non-existent references and partly detect metadata corruption.
In briefFed-BRDECS: A Privacy-Preserving, Heterogeneity-Aware Federated Deep Clustering Method
SummaryThis paper addresses federated clustering of unlabeled data spread across clients while keeping client data private. It proposes Fed-BRDECS, which replaces the global soft-assignment objective of Deep Embedded Clustering with a locally computable sample-stability loss, so clients do not share sensitive assignment distributions. To handle non-IID data, it adds prediction-balanced sampling and centroid-level restarting, and experiments on image and text benchmarks show gains over representative baselines. The authors also report that the method improves reconstruction-based federated time-series anomaly detection without extra inference-time cost.
Haemin Park, Diego Klabjan, Martin W. Braun et al.
In briefA perspective on approximating complex simulation likelihoods with chained aggregated normalizing flows
SummaryThe paper proposes a new perspective on likelihood approximation in simulation-based inference, using a chain of aggregated normalizing flows built from sequential sets of bijective transformations. Parameters are estimated stage by stage, and then updated and aggregated using empirical likelihood under moment restrictions and a sequential decision-making framework. The resulting surrogate model can generate samples and support hypothesis testing, uncertainty quantification, and Bayesian inference.
In briefInference and learning in sparse autoencoders as natural gradient flow
SummaryThis paper unifies feature inference and dictionary learning in sparse autoencoders as natural-gradient flows on a shared variational free energy. It introduces BeFOND, an encoder-free sparse coding model with closed-form dynamics. On synthetic data and language-model activations, it reports better dictionary recovery, rare-feature detection, concept detection and intervention, with quality improving as dictionary width grows.
WhyUnifies inference and dictionary learning as natural-gradient flows with closed-form dynamics; supported by comparisons on synthetic data and language-model activations against baselines.
