In briefData-efficient scaling-law discovery for recommendation systems via coupled hyperparameter scaling
SummaryThe paper presents Axia, a system for efficiently discovering scaling laws of ranking and recommendation models. It introduces a scaling layer that groups architecture parameters and defines how they co-vary, then uses surrogate models, a posterior-based acquisition criterion, and a Pareto frontier to select configurations with few informative runs instead of exhaustive sweeps. Evaluated on 34 model iterations of Meta's ads ranking and retrieval system, the workflow reportedly yielded a mean 0.66% offline normalized-entropy improvement and 0.36% cumulative online lift.
John Martabano Landy, Louis C. Tiao, Yao Xuan et al.
In briefVIPBench: A Benchmark for Matching Speaker Models to How Listeners Hear Voice Identity
SummaryThe paper introduces VIPBench, a benchmark built from 124,876 same-or-different judgments by 1,290 listeners on 9,800 recording pairs from 100 public figures, including real recordings, text-to-speech voice clones, and morphs. It defines four tasks for measuring how well speaker-representation models align with human identity perception. Evaluating eleven public speech models, the authors find listeners judged a clone and a real recording of the same speaker to be different people in 41% of responses, and no model reaches the noise ceiling.
WhyThe authors evaluate speaker models against 124,876 human same-or-different judgments on real, cloned, and morphed speech, comparing eleven models.
In briefPictograms that help fix errors in multilingual speech recognition
SummaryThis paper describes an interface that combines alternative transcripts, audio replay, and contextually checked pictograms for selected uncertain concepts to help users notice speech transcription errors. The system will initially support functional utterances in French and Spanish. A counterbalanced study will test whether pictorial feedback helps listeners recover the speaker's intended action beyond text alternatives alone. The authors also plan to measure semantic mismatch and extra interaction time.
Mariana Hernandez, Hector Luna, Carlos Bruno et al.
In briefRoute to Remember: Adaptive Memory Selection for LLM Agents
SummaryThis paper addresses the cost of text-based memory in LLM agents, where each query requires retrieving accumulated history. The authors apply on-policy self-distillation (OPSD) to train a context-free student model as weight memory, using the same model conditioned on user memory as teacher. A verify-then-regenerate router lets weight and text memory each handle the queries they suit best. On the HaluMem benchmark, the approach matches or surpasses strong text-memory baselines while reading fewer tokens per query.
In briefA video recognition method using multi-order spatio-temporal interaction
SummaryThis paper presents VideoMogaNet, a model for action recognition in videos. It builds on the multi-order spatial representation of MogaNet and adds parallel multi-order temporal modelling using depthwise-only convolutions and implicit gating-based aggregation, whose cost grows linearly with the number of channels. It also introduces a dual interaction between spatial and temporal dimensions to improve the spatio-temporal representation without extra computation. The authors report performance on par with or better than state-of-the-art methods on UCF101, HMDB51, Kinetics-400 and Diving48, with a substantially lower temporal-branch cost.
Kevin D. Nguetche, Lilian Hollard, Eric Granger et al.
In briefExact signal recovery from neural network priors with sigmoid output
SummaryThis paper studies signal recovery in generative compressed sensing using feedforward neural networks with piecewise-linear activations, such as ReLU, followed by a final sigmoid activation. It shows that, without assuming expansivity, signals in the network's range with entries in [δ, 1-δ] can be exactly recovered from a number of measurements depending on the network depth, the input dimension k, and the geometric mean K of the layer widths.
In briefViBench 2.0: A Benchmark for Building Software Over Many Steps
SummaryThis paper introduces ViBench 2.0, a benchmark that tests whether AI agents can build complete software products through a series of feature requests. It uses 17 full-stack web products, 241 feature requests, and 544 test plans run by a browser agent, and a build passes only if every test plan passes. Evaluating nine models, the strongest three pass up to 98% of test plans but deliver a fully working product in at most 65% of builds, and no open-weight model does so.
Daniel Furman, Preeya Kirani, Ibrahim Sheikh et al.
In briefWhen Multi-Fragment 2D/3D Registration Can Be Trusted: Calibrated Certification and View Planning for Pelvic Fracture X-rays
SummaryThis paper examines when registering a segmented CT scan to fluoroscopic X-ray images can be trusted to guide surgery on pelvic fractures. The authors propose a closed-loop method that registers all fragments after each shot, bounds their errors with a calibrated covariance, flags implausible results, and decides whether to stop or plan the next view. Tests on simulated fluoroscopy of 60 held-out fracture CTs show that shift between views best separates tightly bounded failures.
In briefLACS: A label-free, distribution-sensitive method for choosing which training examples to skip
SummaryThis paper introduces LACS (Label-Agnostic Confidence Selector), a training-time rule that selects examples based on maximum predictive confidence, without needing labels or per-example gradients. It removes examples whose confidence is at or above a threshold set by a log-mean-exp reference plus a margin, and gives an exact condition for when no pruning occurs. The method is evaluated across seven settings, including neural training, language modeling, and boosting. The authors report reduced backward-pass rows and accuracy comparisons, while noting that their current implementation does not yet yield wall-clock speedups.
In briefWhiel: A neurosymbolic platform for machine-checked verification of relational programs
SummaryThe paper presents Whiel, a system for reasoning in Lean about imperative programs that operate on relational databases, covering containment and equivalence problems for Datalog and graph query languages. It uses Hoare-style reasoning with first-order specifications and discovers loop invariants neurosymbolically, checking verification conditions with the Vampire theorem prover. Successful proofs are translated into Lean certificates, and performance is evaluated on a curated benchmark suite.
In briefHybrid Spark: adaptive speculative decoding that lets content pick the drafter
SummaryThis paper presents Hybrid Spark (\sys), a framework that runs both a model-based drafter and a model-free n-gram drafter for large language model serving. Within adaptive verification on the GPU, each request uses whichever draft chain is better, with a per-token price used to put both drafters on one scale. Implemented in vLLM with DSpark, it reports 1.10–1.62× output throughput over DSpark on code editing at batch size 4 and 1.04–1.27× at batch size 128, while staying at 0.97–1.01× on chat at temperature 0.
In briefGeneralization Limits of On-Policy Distillation After RL Consolidation
SummaryThis paper examines how well multi-teacher on-policy distillation (MOPD) and single-teacher on-policy distillation (OPD) consolidate capabilities from multiple teachers. It finds that although in-distribution performance is near-perfect, out-of-distribution scores fall below those of the base model. The authors analyze teacher capability and data coverage as possible bottlenecks, and propose geometry-based patching and base-preserving replay to improve out-of-distribution performance.
In briefTransport Along Local Annealing Paths for Multimodal Sampling
SummaryThis paper studies learning transport maps for multimodal sampling, showing that geometric annealing paths can be much longer in Wasserstein space than the distance between their endpoints when mass moves between modes, which slows optimization. It proposes componentwise annealed transport (CAT), which starts from a mode-informed mixture reference, decomposes the target into components, and learns a separate transport map along each componentwise annealing path. The authors also introduce adaptive annealing schedules guided by transport effort, and report improved transport learning and sampling efficiency on multimodal synthetic distributions and Bayesian inference problems.
In briefRIDGE: State-Conditioned Reward Blending for Varied Behavior in Game-Playing RL Agents
SummaryThe paper introduces RIDGE, a single-agent architecture that blends reward personas using weights computed from slowly changing state variables, and trains a separate value head per persona to avoid PPO instability under switching rewards. On the Crafter game with four personas, one RIDGE agent learns both the wood-tier crafting and combat branches, which no fixed-persona baseline solved jointly. A sharpness ablation indicates that state-conditioning drives the improvement. Results are reported across 3 seeds per condition, and larger multi-seed experiments are left for future work.
AL Muqshith Mohammed Shifan, Kevin Christopher Chua, Ali Neshati et al.
In briefCAPTOR: Text-encoder-free grounding by counting caption co-occurrence over a frozen image encoder's visual words
SummaryThis paper proposes CAPTOR, a method that estimates how query terms match image patches without a text encoder. It clusters a frozen image encoder's patch tokens into visual words and builds term signatures from the standardized difference in word shares between images whose captions mention a term and those that do not. On COCO, CAPTOR reaches 0.565 mIoU versus 0.502 for the best training-free baseline, and it also outperforms the strongest pathology-text baseline by 0.160 mIoU on held-out breast-cancer slides.
In briefRAGCompass: a method that picks the best retriever for each query by reranking retrieved passages.
SummaryRetrieval-augmented generation has several retrievers, and none works best for every query. RAGCompass reads the passages each retriever returns and selects the most suitable one for a given query. It uses a reranker fine-tuned contrastively against the answers each evidence set produces, plus a small comparison head, with no retriever-specific parameters. The paper reports a 28% relative gain in answer F1 over the best fixed retriever and better generalization to unseen corpora, retrievers, and readers than query-only routers.
In briefRelevance Propagation as Entropic Optimal Transport on Filippov Networks
SummaryThe paper reframes Layer-wise Relevance Propagation (LRP) as the solution of a single variational problem on the active computational graph of a neural network. It shows that common LRP rules, including LRP0, LRPz+ and LRPα,β, arise from this principle through different choices of reference, and proposes LRPNα,β with node-dependent weights. It also introduces Sinkhorn Relevance Propagation and evaluates the rules on VGG-16 and ConvNeXt-Tiny using conservation and standard explanation metrics.
In briefGaussianRenderFormer: A Transformer That Renders 3D Gaussian Splats Directly
SummaryThis paper presents GaussianRenderFormer, a transformer that takes a set of 3D Gaussians and a camera and outputs the rendered image without using a rasterizer. Trained on 26,820 Objaverse objects pruned to 20,000 Gaussians each, it reaches 38.08 dB PSNR on 300 unseen objects. Rasterizing the same pruned input reaches 44.02 dB. The authors attribute the gap to a bottleneck in the view-dependent decoder, while noting that under-training is not ruled out.
In briefSeparating initial-condition error from model error using machine-learning weather models
SummaryThis paper compares gated and ungated optimization of initial conditions in the GraphCast and Pangu-Weather machine-learning weather models across 36 dates. Forecast errors in 500 hPa geopotential height (Z500) over days 1-10 are verified against ERA5. Ungated optimization outperforms gated optimization on half of the dates when stopped early. Gating does not guarantee better Pangu-Weather performance, but it yields more comparable improvements for both models across all 36 cases.
In briefC-GRPO: Conformal calibration for cutting rollout costs in reasoning model training
SummaryThis paper addresses wasted compute in reasoning model training, where standard Group Relative Policy Optimization uses a fixed number of rollouts per prompt. It proposes C-GRPO, which replaces the fixed budget with calibrated conformal thresholds and proves finite-sample coverage for its stopping rule. Across six benchmarks in mathematics, graduate-level science, and code generation, it reduces mean training rollouts by 49-70% relative to fixed-budget GRPO and transfers across model families with up to 82% savings.
Arya Fayyazi, Seyedarmin Azizi, Mehdi Kamal et al.
In briefVerified but Vacuous: Why 'Passing Verification' Doesn't Guarantee Correct LLM Code
SummaryThis paper argues that evaluations of LLM-generated verified code often report one number that merges three distinct properties: verifier acceptance, syntactic binding of the proof to the spec, and semantic fidelity of the spec to the requirement. Experiments on Python benchmarks with a Dafny-based pipeline and on Rust benchmarks show these properties can diverge. Hidden-test failures persist even among structurally verified samples.
Aditya Vempaty, Himaneesh Sompalle, Manan Tayal et al.
In briefBrainBind: Linking Synchronized EEG and fMRI Through Long Natural Video Representations
SummaryThis paper presents BrainBind, a framework that maps EEG and fMRI recorded simultaneously during long naturalistic video viewing into frozen video embedding spaces used as shared visual-semantic references. On the CineBrain dataset, different video anchors and system configurations show distinct retrieval profiles, with the Omni setup giving fMRI the strongest ranking over the full candidate list. An episode-disjoint evaluation is also reported to assess the effect of repeated stimulus identity and narrative context.
In briefLearning vs. Optimizing Players in Budgeted Games: how a learning agent can resist manipulation and keep low regret when both players face budget constraints
SummaryThis paper studies repeated general games over T rounds between a utility-maximizing optimizer and a learning agent, where budget constraints link the players' decisions across rounds. The authors give an efficient, budget-feasible learning algorithm under which any adaptive optimizer earns at most BSE·T+O(√T) in expectation, and the learner has O(√T) regret when the optimizer plays a fixed mixed strategy. The algorithm relies on two novel forms of calibrated forecasts, which can be constructed in polynomial time by simulating online gradient ascent.
WhyThe authors give a budget-feasible learner with non-manipulability and low stochastic regret, using two new calibration notions and a polynomial-time forecast construction.
In briefMultimodal fusion methods combining whole-slide images and clinical data for inflammatory bowel disease
SummaryThis paper compares 15 multimodal fusion methods that combine whole-slide histology images with clinical data to predict remission outcomes in newly diagnosed inflammatory bowel disease patients. The authors use a modular framework so all methods are evaluated under identical conditions on a prospective cohort. The best method reached macro-F1 scores of 0.57 to 0.75 and outperformed single-modality models in four of six scenarios.
Ankana Banerjee (1), Mark van der Giezen (1, 2 et al.
In briefWhose Values Are Generated? Measuring Hidden Cultural Values in Multilingual Text-to-Image Models
SummaryThis paper introduces VisionValueBench, a multilingual benchmark based on six of Hofstede's cultural value dimensions, to examine which cultural orientation text-to-image systems use when depicting neutral social situations. It finds that even value-neutral prompts produce images favoring particular value orientations, a effect the authors call visual value leakage. Tests across 960 social scenarios and six configurations show that prompt language has limited influence, while geographic cues shift some dimensions but largely keep the dominant orientations.
In briefReadable Is Not Steerable: Testing Physical Structure in Learned Fluid Models
SummaryThis paper tests whether self-supervised world models organize their latent spaces around physical quantities, using 2D decaying Navier-Stokes flow across three pre-registered studies. A JEPA model did not outperform an untrained network on organizing its latents around fluid quantities, so the hypothesis was not supported. Plain autoencoders made enstrophy prominent in slow latent directions, and VICReg reduced this prominence, a partially supported result. The study also finds that what a latent contains and what can be steered through it can differ.
In briefJust the Facts: How AI Dialogue Reduces Conspiracy Beliefs
SummaryThis paper examines why personalized dialogues with AI reduce conspiracy beliefs, using two online experiments with 2,193 U.S. adults. The researchers removed possible causes one at a time, and the debunking effect held across these variations. Only instructing the AI to persuade without factual evidence and arguments substantially weakened the effect. Content analyses linked larger belief reductions to factual argumentation rather than rapport, storytelling, or empathy.
Thomas H. Costello, Gordon Pennycook, David G. Rand
In briefChangeJEPA: A Predict-Then-Compare Method for Remote Sensing Change Detection
SummaryThe paper proposes ChangeJEPA, which reframes remote sensing change detection as predicting the post-change latent representation from a pre-change image and partial post-change evidence. A joint-embedding predictive loss aligns two hypotheses with the target representation, and change is inferred from disagreement among the representations. On the SYSU-CD, LEVIR-CD+, and CDD benchmarks, it outperforms a decoder-matched Siamese baseline by up to 1.51 IoU points and achieves the best IoU and F1-score on all three.
Buddhi Wijenayake, Dhananjaya Jayasundara, Vishal M. Patel
In briefUsing Epistemic Uncertainty to Train 3D LiDAR Point Cloud Segmentation Models
SummaryThis paper examines how explicitly modeled epistemic uncertainty can improve semantic segmentation of 3D LiDAR point clouds. The authors compare a baseline neural network with two variants that use uncertainty as an extra input feature and as a per-sample confidence weight in the loss. Experiments on four large LiDAR benchmarks show higher mean Intersection over Union and recovery of minority classes the baseline never predicts. In half of the configurations, alignment between prediction error and estimated uncertainty also improves.
In briefMESA: A framework for testing robot generalization across composition, semantics, and space
SummaryThis paper introduces MESA, an evaluation framework for measuring how well language-conditioned robot manipulation policies generalize. It provides five evaluation suites covering in-distribution performance and four generalization axes: spatial configuration, object instance, object category, and subtask composition. The authors also release MESA-Gen for generating tasks and demonstrations, plus BiMESA, a large-scale benchmark for bimanual generalization. Among seven baseline methods, the strongest model scores 52% on MESA and 49% on BiMESA.
In briefLost in Pooling or Out of Reach? Diagnosing Spatial Left/Right Failures in Vision-Language Similarity
SummaryContrastive vision-language models perform near chance on left/right spatial relations. The paper identifies two causes: the relation may remain in the pooled embedding but be unreachable by native image-text similarity, or pooling may discard it while dense patch features retain it. The authors propose a parameter-free, query-conditioned spatial moment that recomputes geometry from the patch grid and stores it beside the native embedding. On a 118,287-image index it raises left/right pairwise accuracy to 0.81–0.96 while reproducing the native ranking for queries that name no relation.
WhyThe authors separate two causes of spatial relation failure and repair both with one readout, reporting gains across ten encoders and a 118,287-image index.
In briefOverlap Scores Rank PET Models but Don't Guarantee Reliable Patient Tumor Burden
SummaryThis study examines whether automated metabolic tumor volume (MTV) estimates from PET scans are reliable for individual patients. Using 1,014 autoPET FDG PET/CT studies and five released models on 489 tumor-positive patients, with checks on 597 PSMA studies, the authors compare Dice overlap scores with MTV error. Model rankings by average Dice and average MTV error agree, but at the patient level the two metrics often disagree. The authors recommend that segmentation benchmarks also report patient-level MTV error and missed-lesion counts.
Tushar Kataria, Gerald Sabin, Ponnuswamy Sadayappan et al.
In briefNear-Minimax Sample Complexity for Multi-Objective Reinforcement Learning under Generative Models
SummaryThis paper studies how many samples multi-objective reinforcement learning needs when reward preferences are chosen only after data collection. For finite discounted decision processes with unknown transitions and a generative model, the authors find near-minimax sample complexity that is logarithmic in the number of objectives and saturates at the number of states. Matching lower bounds are also provided, up to logarithmic factors.
WhyThe authors give near-minimax sample complexity for multi-objective RL under generative models, with matching upper and lower bounds up to logarithmic factors.
In briefHow Finely Can Vision-Language Models See Biology? Bio-VQA Benchmark
SummaryThis paper introduces Bio-VQA, a benchmark for testing how well vision-language models understand biological images, with 1.86M questions over 217K images from 11 public datasets. It covers taxonomic reasoning, fine-grained attributes, quantitative spatial reasoning, and individual re-identification. Evaluating 18 open-weight models, the authors find that general visual skills transfer poorly to biology, with taxonomic accuracy falling from 0.72 at phylum to 0.12 at scientific name.
In briefSupport vector machines shown to directly minimise KL divergence
SummaryThe paper proves that the dual objectives of support vector classification and support vector regression are constrained KL divergence minimisation problems between two Gaussian measures. It further shows that removing domain constraints and applying L2 regularisation recovers the objective of kernel ridge regression from the same divergence formulation. The results offer a unified information-theoretic view of standard kernel machines.
In briefAGORA: Combining LLM judges' ratings with reliability-aware weighting
SummaryThis paper proposes AGORA, a method for combining scores from several LLM judges that weights each judge by its learned reliability. It places all items on one latent quality scale using item-response theory and estimates each judge's sensitivity and rating thresholds without human labels. The authors report that on three benchmarks AGORA matches or exceeds Dawid-Skene on large evaluation sets, with a larger advantage on sets of only a few hundred items, and that its learned weights track per-judge human validity.
In briefVirtualTrialHarness: An agent framework coordinating virtual trial protocols across complex clinical data schemas
SummaryThis paper introduces VirtualTrialHarness (VTHarness), an agent framework that turns virtual-trial study protocols into executable analyses by selecting reusable procedural skills for each workflow stage and using explicit contracts to keep study decisions consistent across stages. The authors also present VTBenchmark for evaluating source selection and cohort construction. In three complex synthetic environments, VTHarness produced all required cohort outputs matching the references in 95.1% of workflows after requirement clarification, versus 60.4% using retrieval from the same procedural corpus.
In briefWhose Transparency Is It? How Stakeholders Differ on Generative AI Transparency Needs
SummaryThis paper examines how end-users, AI builders, and policymakers understand and experience transparency in generative AI, using a global survey of 400 participants and a focus group with end-users. All three groups share similar views that transparency should center on accuracy and privacy. However, only 28% of end-users say current transparency practices reflect their needs very well, compared with 53% of builders and 49% of policymakers. The authors call for pluralistic transparency infrastructures rather than one-size-fits-all approaches.
Genevieve Smith, Sahiba Chopra, Natalia Luka et al.
In briefLanguage models iteratively create, grade, and revise data science projects
SummaryThis paper evaluates whether agentic AI models can complete an end-to-end data science project and how feedback improves their results. A writer model downloads, cleans, and analyzes a dataset and writes a short paper, then a grader model evaluates it and the writer revises, repeating for three rounds. Seven models from five developers were tested in 49 writer-grader pairs, each run three times, and the average score rose from 59 to 76 per cent after the final round.
In briefFedSharp: Noise-Resistant, Attack-Robust Decentralized Federated Learning
SummaryThis paper studies decentralized federated learning, which has no central server and faces data heterogeneity, Byzantine attacks, and channel noise. It proposes FedSharp, which keeps models from 2b predecessor nodes, screens suspicious models using coordinate-wise median deviations, and averages the b lowest-loss candidates. It also adds noise-adaptive gradient-norm regularization and proximal training. In experiments on a 100-node ring with 33 Byzantine nodes, FedSharp reports markedly higher accuracy than a state-of-the-art single-model-selection benchmark on MNIST and CIFAR-10.
Ojaswi Sinha, Amin Farajzadeh, Animesh Yadav et al.
In briefA unified method for combining training datasets and exactly removing their influence from machine learning models
SummaryThis paper studies three operations on Gibbs probability measures: renormalization, normalized log-linear combination, and nesting, and gives explicit expressions for the objective functions, reference measures, and regularization factors of the resulting measures. When the reference measures are probability measures, these operations are given geometric interpretations involving Pythagorean identities and exponential geodesics. The authors show that nesting and normalized log-linear combination are equivalent, and apply this to Gibbs algorithms so that models trained on local datasets can be combined into models trained on the aggregated data.
WhyThe authors unify three Gibbs measure operations, prove nesting equals normalized log-linear combination, and apply this to dataset aggregation and exact unlearning.
In briefGeneralized hyperrectangular prediction sets for multivariate conformal prediction
SummaryThis paper studies the shape of prediction sets in multivariate conformal prediction, where many nonconformity scores produce sets that are hard to interpret. The authors generalize hyperrectangles to Cartesian products of arbitrary one-dimensional sets, called generalized hyperrectangles. They introduce a rectangularization framework that converts any continuous multivariate score or conditional density into one yielding such sets with finite-sample coverage. Synthetic and real-world experiments quantify the efficiency cost of rectangularization and highlight its benefits for uncertainty quantification.
In briefA one-pass method for efficient whole-page ranking of ads and content
SummaryThis paper addresses whole-page ranking, which jointly orders advertisements and organic content under layout and business constraints. It proposes Multi-Item Scoring (MIS), a prefix-aware non-autoregressive framework that adds sequential context to efficient multi-item scoring. On KuaiRand-1K and RL4RS, MIS approaches the generation quality of an autoregressive baseline with 4.07x and 3.54x faster CPU decoding. In a randomized test on a large social-video platform, it improved the realization metric and watch time slightly while reducing ranking-service CPU use by 9.73%.
In briefRewrite, then veto: a rewriter-agnostic edit gate for LLM post-OCR correction
SummaryThe paper proposes EditVeto, a gate placed after an LLM rewrites OCR text, which splits the rewrite's differences into blocks and keeps or reverts each one with a 0.6B-parameter model. The authors report that the gate improves five rewriters on the HIPE-OCRepair-2026 test sets, including two not seen during training. On published competition outputs, it lowers error for three of four teams but raises it for the winning run.
In briefUsing general vision-language models to read EEG for clinical neurodiagnosis
SummaryThis paper asks whether general vision-language models can read EEG recordings visually, much as clinicians inspect EEG patterns. It reports that these models, without large-scale EEG pretraining, perform competitively on a range of clinical neurodiagnostic tasks. They also adapt to new tasks in few-shot settings better than specialist foundation models and improve with scaling where modality-specific models do not.
In briefConfidence Score: a new kernel measure for judging how well kernels suit classification
SummaryThis paper proposes the Confidence Score, a new measure for evaluating kernels in classification tasks. Unlike alignment-to-the-ideal-kernel measures, it compares an encoding of the ground-truth labels with the ground truth as projected by the kernel. This lets kernels that preserve structure within a class score highly. Experiments on real-world and synthetic datasets show it correlates more strongly with classification accuracy than several other kernel measures.
In briefTight lower bounds for high-order smooth, uniformly convex minimization
SummaryThis paper establishes tight lower bounds for minimizing high-order Hölder-smooth and uniformly convex functions when their orders match, i.e., q = p + ν. It shows that any deterministic algorithm using derivatives up to order p needs a number of oracle calls of order Ω of the stated form to reduce an initial optimality gap Δ0 to ε. The bound matches the known upper bound and serves as a higher-order uniformly convex analogue of Nesterov's first-order lower bound for strongly convex functions.
WhyThe authors give a tight lower bound matching the upper bound for matched high-order Hölder smoothness and uniform convexity, via a new multiscale construction.
In briefSafe multi-agent path planning with diffusion models
SummaryThis paper extends SafeDiffuser to multiple agents, embedding control barrier function constraints into diffusion-based trajectory planning so that joint safety specifications stay satisfied over finite time. It presents centralized and decentralized (communication-aware) constrained denoising methods and handles local traps and deadlocks with time-varying specifications and liveness barrier functions. A benchmark suite including a multi-agent Maze2D domain is provided, and experiments report fewer violations than unconstrained diffusion baselines while planning quality is retained.
In briefExplaining Sparse Visual Features by Reconstructing Their Effects
SummarySparse autoencoders split visual representations into individual features, but the image regions that most strongly activate a feature can be ambiguous when several features with different meanings fire in the same places. This paper proposes RecSAEI, a framework that intervenes on representations along SAE decoder directions and reconstructs the visual output to reveal what a feature means. The authors report that the method disambiguates feature meanings across multiple datasets and vision encoders, and link [CLS] interventions to a sparse set of patch SAE decoder directions.
In briefWhen to ask or answer: strategic clarification for resolving ambiguous references in dialogue.
SummaryThis paper studies strategic clarification, where a system asks questions only when the answers would narrow the set of candidate referents. The authors built a controlled test environment from the SIMMC 2.1 dataset that records how many questions each ambiguous reference needs to be resolved. They also propose BACE, a lightweight action-selection method on a frozen language model backbone, which they report chooses the optimal number of clarifying questions more consistently than GRPO and ArCHer.
In briefTSFM Meets LLM: Using Text as Covariates for Time Series Forecasting
SummaryThis paper addresses the gap between large language models, which understand text but lack numerical precision, and time series foundation models, which are numerically accurate but cannot reason about text. The authors propose Context as Covariates (CoCo), in which an LLM converts its textual reasoning into forecast and confidence covariates that guide a TSFM. The models are aligned in stages using reinforcement learning and trained on a synthetic multimodal corpus. On real benchmarks spanning finance, economics, weather, traffic, and security, CoCo with Qwen3-4B-Instruct-2507 and Chronos2 achieves up to 12% MAE improvement over state-of-the-art models.
Hemanth Ram, Sai Shankar Narasimhan, Shubhankar Agarwal et al.
In briefSlackBench: Testing AI Agents on Team Projects Using Real Code Repositories
SummaryThis paper introduces SlackBench, a benchmark for evaluating AI agents on research and engineering workflows that require reasoning over both workplace conversations and evolving code repositories. It is built from privacy-preserving reenactments of real academic projects, with multi-party dialogue, decision changes, and matching GitHub and Google Docs artifacts. The benchmark contains 140 queries covering cross-source search, abstention, goal-directed summarization, and circumstance inference. Across 33 frontier models and agent harnesses, the best system scores around 80%, and most score below 66%.
In briefCtrl-F: Using Force as a Control Interface for Humanoid Robot Loco-Manipulation
SummaryThis paper presents Ctrl-F, a hierarchical framework that adds an explicit 3-D force command at contact bodies such as hands or feet to humanoid motion tracking. It combines a unified motion-force controller that tracks reference motion while applying the commanded force, with a high-level planner that issues force commands online from object state and proprioception. In simulation and on hardware, Ctrl-F completes tasks requiring large, sustained interaction forces and adapts to changing loads, including pushing and kicking boxes up to 20 kg where a force-blind baseline fails.
In briefActive preference learning that targets each objective to align language models on several goals at once
SummaryThis paper introduces OGC-DPO (Objective-conditioned Gradient Coverage DPO), an active learning framework for aligning large language models with multiple objectives such as helpfulness, harmlessness, and honesty. It treats each response pair and objective combination as a separate query and measures informativeness using the policy's own implicit gradient, without an external reward model. Experiments across different numbers of objectives show it outperforms existing active preference learning methods at the same annotation cost.
In briefDreamX: A Video Diffusion Model Used as a Universal Latent Action Learner
SummaryDreamX repurposes a pretrained video diffusion model to learn compact latent actions from video, compressing each latent frame by 300 times (1,200 to 4 tokens). The latent actions capture action-relevant dynamics separated from visual appearance, so they can serve as a shared interface across humans, robots, environments, and action parameterizations. The paper reports state-of-the-art results among latent action models on LARY-Bench, along with any-to-any video transfer and zero-shot action transfer.
In briefFrom self-operating AI agents to AI systems that augment human thinking
SummaryThis paper argues that AI should be judged not only by what machines can do alone, but also by how much they can enhance a person's own intelligence, a concept it calls AI-augmented human intelligence. It proposes a "cognitive symbiont," a lifelong companion that senses what its user senses, keeps their context over time, and joins their thinking and everyday tasks at the right moments. The authors also note new risks such systems raise, and treat their governance as part of the paradigm.
In briefAligning language models to general preferences via direct Nash optimization: an empirical study
SummaryThe paper studies Direct Nash Optimization (DNO) and its practical variant DNO-Prct, which align a language model to a general preference function by approximating the Nash equilibrium of a two-player preference game. The authors re-implement the method on Llama 3.1 8B Instruct and Mistral 7B Instruct v0.1, reporting higher length-controlled win rates on AlpacaEval 2.0 after three iterations. The paper also includes ablations, pipeline cost, and a discussion of single-run results and GPT-4o serving as both teacher and judge.
Austin Tsai, Saurav Nanda, Ganapathy Parthasarathy et al.
In briefPolicyIK: Reinforcement Learning for Precise Robot Arm Inverse Kinematics
SummaryThis paper presents PolicyIK, a goal-conditioned reinforcement learning policy that computes precise inverse kinematics, converting end-effector pose commands into robot joint motions. In simulation, the trained PPO policy keeps final-second mean errors within 1 mm and 0.2 degrees on 93.5% of held-out tasks, rising to 98.9% with a deployment protocol. In simulated pick-and-place, it executes end-effector commands from the π0.5 vision-language-action policy with 95.7% success.
WhyThe authors apply RL to high-precision IK with offset and dwell rewards, reporting gains over four classical controllers in simulation and on a physical arm.
In briefGUILD: Guided Iterative Data Collection for Robot Imitation Learning
SummaryRobot imitation learning usually depends on costly real-world teleoperation data, and unguided collection often yields data that adds little to policy performance. The paper proposes GUILD, which alternates between proposing tasks with vision-language model priors and collecting data, and evaluating the collected trajectories via predictive data attribution to see how they affect performance on a target deployment setting. This feedback guides subsequent collection. Experiments in simulation and on real robots report improved sample efficiency and deployment performance over passive collection and several baselines.
In briefCan AI Tools Carry Out a Scientific Literature Review? A Test Against a 102-Article Benchmark
SummaryThis paper evaluates six AI tools across four stages of a literature review: finding relevant articles, interpreting them, synthesizing findings, and proposing future research. The benchmark is a published review of the Spanish plume weather pattern and its 102-article corpus. Institutional library access substantially increased the number of articles the tools could find, but several tools misclassified many passing mentions as substantial discussion. Their syntheses were largely accurate but sometimes included unsupported claims and overstated consensus. The authors conclude that the tools can assist with discovering and interpreting literature but cannot replace expert judgment.
Abdullah Al Kalbani, David M. Schultz, Alexandra K. Anderson-Frey et al.
In briefHybrid LLM and Deterministic Pipelines for Large-Scale Code Migration at Bloomberg
SummaryThis paper reports on Bloomberg's Automated Uplift initiative, which applied LLM-assisted migration pipelines across thousands of repositories. It covers two production campaigns, a Python packaging migration and a C++ build system migration, producing 2,333 pull requests and 729 successful migrations. The authors describe a hybrid architecture that maximizes deterministic processing and confines the LLM to bounded file transformations, and they distill transferable patterns for enterprise-scale migrations.
Mario Corchero, David Williams, Laszlo Kiss Kollar et al.
In briefVisibility-aware reference generation for benchmarking 3D reconstruction in vegetated scenes
SummaryThis paper presents a LiDAR-based pipeline that generates reference depth maps and visible point clouds from existing multi-sensor localization datasets. It introduces the LDGVS visibility score for point-wise filtering and automatic scale selection in Hidden Point Removal. On TartanAirV2 and TartanGround, the method retains about three times the point density of Poisson surface reconstruction at comparable absolute relative error. Applied to BotanicGarden, it compares feed-forward and conventional SfM methods, finding that feed-forward methods are efficient but can duplicate structures, while SfM yields coherent but sparse geometry that is more prone to failure in vegetated scenes.
In briefGSCV++: A lightweight method for compressing 3D Gaussian Splatting sequences with standard video codecs
SummaryThis paper presents GSCV++, a lightweight pipeline that compresses 3D Gaussian Splatting sequences using standard video codecs. It improves three stages: preprocessing through energy-adaptive reduction of spherical harmonic coefficients, 3D-to-2D mapping via prior-based P-MiniPLAS, and content-aware compression settings. Experiments report over 20% intra- and 60% inter-prediction rate-distortion gains over state-of-the-art methods, along with faster 2D map generation, encoding, and decoding.
In briefDurably Suppressing Misaligned Behaviors in Language Models
SummaryThe paper studies how to suppress undesired behaviors that appear in language models after training, comparing SFT, GRPO, and gradient-ascent unlearning on immediate suppression, durability under further training, and retained capability. It introduces PEARL, a GRPO variant that augments rollouts with continuations from intermediate states along cached trajectories. Across two model families and two misbehaving model organisms, PEARL achieves strong immediate suppression while preserving task accuracy and staying durable under further fine-tuning.
In briefOpenBrain: An Auditable Release of Generated Whole-Brain MRI Labels
SummaryThis paper presents OpenBrain, a release of 35,838 provenance-verified brain MRI cases from 607 OpenNeuro source datasets, with generated whole-brain parcellations. Each case record links multiple parcellations, source and license information, reference-free QC-7 reliability measurements, and a combined risk score with a confidence tier. The authors report that lower QC-7 risk corresponds to higher generated-label quality, and that lower-risk selections improve training performance over random selection.
In briefAgnostic learning by imitating: showing discriminative and generative learning are equally hard
SummaryThe paper studies whether training a model that separates positive and negative labels as well as the best model in a class is harder or easier than training a generator whose outputs cannot be told apart from real data by any model in that class. The authors prove the two tasks are computationally equivalent, extending a previously known reduction with an efficient reduction in the opposite direction. Using adaptive queries to a multiaccuracy oracle, they achieve agnostic learning and several other learning objectives.
WhyThe authors prove equivalence in complexity between multiaccuracy and agnostic learning via a new efficient reduction using adaptive oracle queries and follow-the-leader techniques.
S\'ilvia Casacuberta, Parikshit Gopalan, Varun Kanade et al.
In briefAdversarial speculative decoding to improve long-horizon AI agents
SummaryThis paper argues that speculative decoding affects more than speed, since it can shift the output distribution and reasoning accuracy of long-horizon agents. The authors train a draft model to be rejected at critical reasoning points, so that those tokens are sampled from the target model under less noisy conditions. On complex agentic coding benchmarks, the method raises end-to-end success rate without changing the target model or its output distribution.
In briefA circuit for learning from examples inside a prompt, without updating model weights
SummaryThe paper studies how language models learn few-shot sentence classification with arbitrary labels from demonstrations in the prompt, using three 3–4B-parameter transformers. It reports that the models implement a key-value memory over label-token positions, with attention heads storing demonstration classes as keys and upper-layer heads copying matching labels at query time. Causal tests such as activation patching and head ablations support each step, and the same components also appear in a transformer trained from scratch.
WhyThe authors give a full composed circuit for in-context learning with arbitrary labels, a gated key-value memory, validated causally across models and scales.
In briefWeight-space programming for auditing clinical decision networks
SummaryThis study tests whether a language model, given only a clinical protocol and a small network's raw weights, can audit and repair that network. Planted flaws included confounds, a 3 Gy cord threshold shift, and a hidden trigger in networks screening head-and-neck radiotherapy planning requests. The model named the broken rule in 16 of 18 flawed networks and repaired them, but found only 1 of 3 triggers and cleared 2 of 5 healthy networks. The authors conclude it can locate some flaws but cannot yet certify a network as correct.
In briefMedPhysBench: a benchmark testing language models on medical physics calculations and error detection
SummaryMedPhysBench is a benchmark of medical physics calculation items across radiation therapy, diagnostic imaging, nuclear medicine and radiation protection, paired with error-detection items. On 1,200 held-out test items, 12 models scored between 33.3% and 100.0% on calculation. Models also varied widely in detecting planted errors in colleagues' worksheets, and the authors release code and per-item results.
WhyThe authors combine medical physics calculations and error detection in one benchmark, using protocol code for reference answers; they report evaluating 12 models on 1,200 held-out items.
In briefA parameter-free accelerated optimization method for convex problems using p-norm preconditioning
SummaryThis paper proposes an adaptive accelerated optimistic dual averaging method (A²ODA) for unconstrained convex optimization, using low-cost preconditioning based on p-norm geometry for p in [2, ∞]. The method requires no line search and selects stepsizes with a new technique based on an aggregated affine lower model. Its update structure is unified across p and is no more complicated for p > 2 than for the Euclidean case p = 2. Under a generalized local smoothness condition, the authors derive a convergence bound, and in the special case of local smoothness it recovers a nearly optimal rate.
In briefFrom Tokens to Trajectories: Parameter-Efficient Continuous Adaptation of Unified Multimodal Models
SummaryDiscrete tokenization in unified multimodal models gives strong semantic reasoning but loses fine visual detail, often producing structurally accurate images with many artifacts. The paper introduces FlowShift, a parameter-efficient post-training framework that converts pre-trained discrete models into continuous rectified-flow generators by replacing discrete visual interfaces with continuous latent projections and applying LoRA to attention layers. It also proposes Dual-Objective Velocity Vector Distillation (DOVD), which aligns the model with a continuous teacher and ground-truth trajectories, improving perceptual quality while preserving prompt adherence and reasoning.
Gayatri Deshmukh, Onkar Susladkar, Adheesh Juvekar et al.
In briefCATF: A manager-worker framework for forecasting time series using context
SummaryThis paper studies forecasting a time series from its own history plus contextual variables such as weather and calendar data. The authors propose CATF, in which a Transformer-based manager sends each input to a single worker model based on its context, using hard routing in training and inference so workers specialize without manual data splitting. On five context-rich datasets, CATF beats its standalone versions in 162 of 200 comparisons and outperforms the zero-shot Time-MoE model in 15 of 20 settings.
In briefPyRUA-DataFarm: An LLM Agent Generating Real-Ready Robot Training Data Without Teleoperation
SummaryThis paper presents DataFarm, in which a coding agent rebuilds a robot task in simulation from a data request and one calibration episode, writing the scene, an expert controller, and a success judge. A video model conditioned on real images renders the simulated rollouts into photorealistic multi-view video labelled with robot actions. On five real-robot tasks, policies trained only on this data matched policies trained on as many teleoperated episodes (93 vs. 84 successes in 100 trials) and performed better under an unseen tablecloth and coloured light (87 vs. 46 successes). The authors report an average adaptation cost of $21 per task and $0.80 per additional episode.
WhyA coding agent replaces teleoperators, generating simulated tasks rendered photorealistically; the authors report matching teleoperation success on five real-robot tasks and lower cost.
In briefHARP: A training-free dual-profile agent communication method for LLM-based recommendation
SummaryThe paper frames LLM-based recommendation as an agentic communication process with four stages: building a self-description, transmitting it along the collaborative graph, understanding the other side, and synchronizing from the next interaction. It proposes HARP, a training-free method that adds three operators to a frozen LLM backbone: a symmetric item profile, a profile-level homophily measure used at retrieval, and a single-pass reflective update. On four public benchmarks (Yelp, Books, MovieTV, Goodreads), HARP raises Hit@1 by 6.9% to 24.0% over the strongest published baseline.
In briefMDSKILL: A method that helps AI agents absorb stored skills more efficiently by training on three types of skill views.
SummaryMDSKILL organizes an agent's skill memories into three functional views: state grounding, action planning, and execution control. It selects high-quality skills using dynamic metrics and builds reasoning branches for each view and for their integration. On-Policy Self-Distillation then transfers these branches' knowledge into a shared policy, and the paper reports gains of up to 7.1% on ALFWorld and 8.1% on WebShop over prior state-of-the-art methods.
In briefΦasmaNet: a fast neural network that predicts gravitational-wave signal strength
SummaryThis paper introduces ΦasmaNet, a compact neural network that predicts the signal-to-noise ratio of gravitational waves from binary black holes without computing full waveforms. It is evaluated for a single detector at Advanced LIGO design sensitivity, using about 732k parameters. The authors report it processes a million-signal catalogue in about 8.3 seconds on one GPU, roughly two orders of magnitude faster than conventional methods on a 16-core CPU.
Ioannis Tsioulis, Anastasios Tefas, Nikolaos Stergioulas
In briefA study on whether language models need long context, using frequent compaction to keep contexts short
SummaryThe paper notes that later tokens cost more to generate than earlier ones, with the cost gap on frontier open-weight models reaching up to 7x. Through frequent compaction, the authors show that across many long-horizon and long-input tasks, models keep little to no accuracy when contexts stay short. They also report that performance on long-horizon software engineering tasks does not collapse even with at most one action per context, and outline semantic state space models (SSSMs) as a concept, questioning the need for long-context models.
Peter Belcak, Saurav Muralidharan, Jindong Jiang et al.
In briefLoopVQ: Sub-1-Bit KV Cache Compression for Looped Transformers
SummaryLooped transformers reuse the same layers across several loop steps to save parameters, but each step still stores its own key-value (KV) cache, which becomes the main memory bottleneck. This paper proposes LoopVQ, which jointly codes channels across loop steps and fine-tunes codebooks to match full-precision outputs. At 0.5 bits per element, it outperforms existing KV cache vector quantization methods on Ouro and Huginn models across several task types. Its SGLang implementation nearly halves decode latency at 32K context and improves peak serving throughput.
In briefUniversal hardness certified via masking channels
SummaryThis paper generalizes the hardcore lemma of Impagliazzo (1995) into a broad learning-theoretic framework for average-case hardness of computing functions. It develops a general method for certifying hardness that applies to non-Boolean functions and many notions of hardness beyond the 0-1 loss, using convex duality to select the appropriate fidelity measure. The authors derive multiclass generalizations and construct a universal masking channel that applies to many hardness notions simultaneously.
WhyThe authors generalize the hardcore lemma to non-Boolean functions and convex losses via masking channels and convex duality, giving a tight characterization and a universal channel.
In briefGSBR: Geometric Support Behavior Regularization for Online RL
SummaryThis paper proposes GSBR, a regularization method for online reinforcement learning that penalizes actions by their distance from the support of a pretrained base policy. Actions inside the support are not penalized, while outside the support the penalty grows smoothly with distance, allowing improvement when the gained value outweighs it. Across 19 tasks on four robotic benchmarks, GSBR raises the average success rate from 33% for the base policy to 95%, with the highest average success rate and throughput among compared methods.
In briefWonder 4: A Mixture-of-Experts Method for Image Quality Enhancement
SummaryReal photos often suffer from several defects at once, such as blur, grain, compression, scratches, and low resolution, each with different severity and location. Wonder 4 is a mixture-of-experts model that corrects these defects simultaneously in the regions where they occur, without using a degradation classifier, mask, prompt, or task label. Each block holds 13 experts, and a router sends every image patch to 5 of them, trained with a bias that only sets a minimum usage share rather than forcing equal use. On DIV2K and Flickr2K, it achieves the best LPIPS, DISTS, and NIQE scores across all three test settings.
In brief"Personas Are Not People": Limits of Fixed Synthetic Personas for Simulating Populations
SummaryThis paper examines the use of synthetic populations, simulated people defined by fixed persona profiles, in place of real people for survey pretesting and AI evaluation. The authors argue that behavior varies more across contexts than between people, so fixed personas cannot reproduce the context-specific part of a population's behavior, however detailed or large the model. They propose a new evaluation criterion and a person-in-context design, tested on a Reddit personality corpus and a clinical-trial recruitment survey. The design produces context-dependent behavior but does not fully match the real populations.
In briefClassifying sparse 3D point clouds with a new network that stays accurate at low density
SummaryThis paper addresses the problem of classifying 3D point clouds when their density is low, a condition that can degrade existing methods. It introduces LD-PointNet++, an architecture optimized for low-density point cloud classification. The authors also present the first benchmark that systematically compares state-of-the-art methods across varying densities, showing significant performance declines at low density. Experiments on the ModelNet dataset show LD-PointNet++ outperforms these methods in low-density scenarios and keeps accuracy above 90% across density levels.
In briefReassurance Error in AI Severe-Weather Advice Tested Against Real NWS Alerts
SummaryThis paper examines whether an LLM assistant that cannot see official severe-weather alerts may tell users conditions are calm while a warning is active, a failure the authors call false reassurance. They replayed 117 frozen cases built from real U.S. National Weather Service alerts through five conditions of a Claude-based preparedness pipeline. Access to real alerts removed false reassurance (paired McNemar p < 0.0001), while a blind neutral prompt produced it in 37.7% of alert cases. The authors release the dataset, labels and code.
WhyThe authors define false reassurance and replay 117 real NWS alert cases across five prompt and alert-access conditions, releasing data and code.
In briefTraining an AI on answers it believes are wrong can make it misaligned
SummaryThe paper shows that fine-tuning a language model on data it believes to be false, even when the data has no harmful content, can cause broad misalignment. Experiments across mathematics, trivia, cooking and sports found that the more strongly the model disbelieves the training answers, the more misaligned it becomes. The authors attribute this to an inconsistency between the model's internal beliefs and its stated answers.
WhyThe authors report that fine-tuning on believed-false but harmless data induces broad misalignment, with misaligned answers rising from 4% to over 60% in mathematics experiments.
Ariana M. Villegas-Suarez, Guido Freire, Sohan Venkatesh et al.
In briefAnchored rank-adversarial fine-tuning to make pretrained place recognition models robust to attacks
SummaryVisual place recognition locates a camera by retrieving similar geo-tagged images. The paper shows that current foundation-model methods such as SuperVLAD and BoQ are highly vulnerable to tiny adversarial perturbations. The authors propose ARAFT, which fine-tunes a released model with a listwise rank loss and an anchor to the frozen model's clean descriptors, and report improved robust recall under a fixed clean-accuracy budget.
In briefYANAMI: A large-scale multimodal dataset and benchmark for multi-turn stateful dialogue.
SummaryThis paper introduces YANAMI, comprising YANAMI-400K, a dataset of 400K multi-turn multimodal dialogues, and YANAMI-Bench, a benchmark of 400 dialogues. The dialogues cover four scenarios (multi-image fusion, image editing, world modeling, and visual reasoning) and require models to track and revise evolving dialogue states across turns. Fine-tuning BAGEL on YANAMI-400K improved image editing by 10.3%, generation by 2.0%, and stateful interaction by 13.23%, while evaluations show existing models still struggle with this task.
In briefMöbius: A Unified Motion-Language Model for Understanding and Generation
SummaryThe paper presents Möbius, a single model that understands, generates, and edits human motion within one framework. It brings the diffusion process into the same multimodal representation space used for understanding and operates directly on continuous frame-level raw motion, without discrete tokenization or a pretrained motion VAE. According to the paper, it achieves state-of-the-art results among unified motion models on HumanML3D and scales to 1.7B parameters with 1,700 hours of motion data.
In briefBeyond Accuracy: What Language-Subspace Interventions Capture, Remove, and Improve
SummaryThis paper studies whether suppressing language-related components in mid-layer activations can improve multilingual reasoning in LLMs, tested on MATH and MGSM with four open-weight models. Subspaces built by Mean-SVD and DiffPCA capture substantial translation-related variation, but linear probes still identify language almost perfectly after projection. The intervention that most improves MGSM accuracy sacrifices response-language fidelity, and only one configuration improves joint accuracy, by 1.60 points. The authors argue that causal interventions must be judged on more than accuracy.
Shaanil Punglia, Dadhichi Telwadkar, Vaibhav Rathore et al.
In briefConformal Prediction at Performative Equilibrium: How Deployment-Induced Data Shifts Limit Coverage Guarantees
SummaryThis paper studies conformal prediction when a deployed predictor changes the data distribution, requiring prediction sets to achieve coverage under the distribution at a performatively stable predictor. The authors identify a phase transition at total-variation sensitivity L=1: for L≥1 no procedure can guarantee both valid coverage and nontrivial sets, while for L<1 the minimax excess miscoverage scales as 1/((1-L)√n). A transfer theorem links approximate equilibrium predictors to coverage, and experiments including credit scoring illustrate the predicted behavior.
WhyThe authors identify a phase transition at L=1 for conformal prediction at performative equilibrium, give minimax lower bounds and a transfer theorem, and report experiments confirming the predicted t
In briefBenchmark on how long-lived agents manage experience as environments change
SummaryThe paper introduces ReScopeBench, a benchmark for the experience lifecycle across 18 interactive environments. It separates structural experience about how an environment is organized from rule experience about reusable principles that guide decisions across tasks. Task streams are built around shared experience and include silent environmental changes to study how experience is acquired, reused, and revised. Across 20 models, retained experience improves task performance, but agents struggle to keep it current as environments change, and correctly recorded experience does not always lead to successful task completion.
In briefNeural network method for physics problems with holes, using de Rham cohomology
SummaryThis paper proposes Hodge-de Rham PINNs, a variant of physics-informed neural networks augmented with de Rham cohomology, to handle problems with topological features such as holes and punctures. The method parameterizes a 1-form using the Hodge-Morrey-Friedrichs decomposition and links it to scalar PDE solutions through integral operators. The authors test the approach on problems including the Aharonov-Bohm effect and provide theoretical guarantees.
In briefRoute Anything: Study of Model and Policy Routing in General-Purpose AI
SummaryThis paper presents Route Anything, a framework for studying how a general-purpose AI system can dynamically choose and coordinate the best model or policy for each task while balancing performance and computational cost. The authors build an omni-modal trajectory matrix covering 25 benchmarks and 452 candidate configurations, with 374,401 scored runs. Learned routers outperform the best fixed candidate by 0.3 points across 16 benchmarks, by 2.1 points on video generation, and by 1.1 points on robot manipulation while using 24% less policy time. Routing also raises efficiency by 8.82x on average compared with the reference model.
In briefDetecting Hidden Knowledge in LLMs via Activation Geometry
SummaryThis paper proposes a way to tell whether a large language model truly knows a concept by measuring the geometric coherence of anon-vectors, the hidden-state differences between a concept query and its anonymized paraphrase. Genuine knowledge forms a tight, low-entropy cluster, while ignorance causes geometric fragmentation. Across ten instruction-tuned models, these metrics reliably separate known from unknown concepts and correlate with adversarial recovery after unlearning.
WhyThe authors propose geometric fragmentation of anon-vectors to measure LLM knowledge and predict unlearning recovery, reporting validation across ten models, four attacks, and five unlearning algorith
In briefiGC: Reinforcement learning for antibody optimization using a learned mutation prior
SummaryThis paper presents iGC (in silico Germinal Centers), a two-stage reinforcement learning framework for optimizing antibodies against specific antigens. The first stage learns a somatic hypermutation (SHM) prior from parent-child pairs in B-cell lineage trees and adapts it to a target affinity-maturation dataset. The second stage initializes a policy from this prior and optimizes antigen binding with KL regularization, using a Reward-Tax Meta-Gradient objective to prevent reward exploitation.
Thao Nguyen, Robert G. Alberstein, Ewa Nowara et al.
In briefStabilizing targeted guidance in agentic reinforcement learning
SummaryThis paper studies targeted guidance, a method that learns from per-turn textual feedback through an on-policy self-distillation objective, to improve behaviors that are hard to optimize with reinforcement learning. The authors test the approach in an agentic multi-turn software engineering environment. They report that combining targeted self-distillation with a per-criteria RL reward, and giving feedback only on turns where the policy fails, works well. The paper also examines design choices that keep entropy stable and reduce oscillation during training.
Kevin Frans, Alexander M Rush, Sergey Levine et al.
In briefGaussian partitioning: a tool for reliably learning multi-index models
SummaryThe paper introduces Gaussian partitioning, an algorithmic method for distributions that are smooth with respect to a Gaussian. It splits space into polyhedral regions so that nearly all probability mass lies in thin strips that constrain only a few orthogonal directions. The authors use this to build agnostic learners for several Boolean multi-index model classes, including functions of halfspaces, and also prove Statistical Query lower bounds. For functions of k halfspaces, the dimension dependence is improved over previous results.
WhyThe authors introduce Gaussian partitioning and improve dimension dependence for learning multi-index models, with matching SQ lower bounds reported.
Ilias Diakonikolas, Giannis Iakovidis, Daniel M. Kane et al.
In briefSO(3)-equivariant learning on CAD boundary representations (neural networks that handle rotated CAD parts)
SummaryThe paper proposes SO3brep, an SO(3)-equivariant graph neural network that operates on the face-adjacency graph of CAD boundary representations. It computes node features in local frames and maps them into a shared frame using Wigner D-matrices, with Clebsch-Gordan tensor products preserving the relative orientation between adjacent faces. Across three face-segmentation and two classification benchmarks with randomly rotated test parts, it achieves the highest accuracy with the fewest parameters among compared methods. Its advantage is largest when few labeled parts are available.
In briefJHiMA-WebRL: A joint hierarchical multi-agent reinforcement learning framework for multi-step web tasks
SummaryThis paper introduces JHiMA-WebRL, which trains a task decomposer and an action-executing actor jointly so web agents can handle long, multi-step tasks on websites. The decomposer is initially fine-tuned on structured subtask annotations produced by a strong teacher model from multimodal GUI trajectories, and both components are then optimized together with an extended GRPO method in the WebArena environment. On WebArena-Lite-v2, the 14B model (a 7B actor with a 7B decomposer) reaches a 37.2% average success rate, compared with 24% for a ScaleCUA-32B agent trained with RL on the same tasks. Ablations indicate that updating the decomposer during RL performs better than updating the actor alone.
