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 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 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 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 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.
In briefNovoCRF: Globally Normalized De Novo Peptide Sequencing on a Prefix-Mass Lattice
SummaryThis paper presents NovoCRF, a conditional random field whose states are prefix masses, for de novo peptide sequencing from mass spectrometry data. It is trained with exact normalization over all mass-feasible peptides and uses exact k-best decoding plus two BiGRU rerankers. On seven held-out 2026 datasets, it reports higher peptide and amino-acid recall than Casanovo v5 and much higher throughput.
In brief"Going Deeper With Memories": a multiscale byte-level language model with memory blocks
SummaryThe paper introduces DeepMem, a causal multiscale architecture for byte-level language modeling that uses a learned invertible subband transform to compress byte sequences into shorter representations while preserving byte identity through a lossless carrier and detail sidebands. It adds fast weight memory blocks in the encoder, iterative shared state computation at the compressed bottleneck, and an attention-based super-resolution decoder that restores byte-level representations. The authors report stronger language modeling than a matched byte-level Transformer on FineWeb-Edu and PG19, and the best macro average across evaluated FLORES languages.
In briefLogarithmic regret bound for alternating mirror descent in zero-sum games
SummaryThis paper studies alternating online mirror descent (AltOMD), a learning rule for zero-sum games. The authors prove that, for two-player zero-sum bilinear games with compact convex strategy sets and mild regularity assumptions on the regularizers, the sum of both players' external regrets is O(log T), which is nearly constant. These assumptions hold for the negative entropy regularizer, so the alternating multiplicative weights update algorithm also attains logarithmic regret. The bound further implies an O(log T / T) duality gap for the averaged strategies.
WhyThe authors resolve an open question, proving an O(log T) regret bound for alternating OMD in bilinear zero-sum games and verifying it for the negative entropy regularizer.
In briefShrinkage Priors for Implicit Generative Models
SummaryThis paper proposes a sample-efficient training method for implicit generative models built on deep neural networks. It combines regularized inference with shrinkage priors and the generalized Bayes framework, using a scoring rule posterior based on a strictly proper scoring rule. The authors introduce a block SGMCMC within Gibbs sampling to explore high-dimensional parameter spaces and identify sparse representations of these models. Performance is demonstrated on linear regression with p ≫ n and on an application to neural generative models.
Shreya Sinha-Roy, Christian P. Robert, Ritabrata Dutta
In briefPost-stratified importance sampling: a method that improves importance sampling estimates by reducing them to one-dimensional modeling.
SummaryImportance sampling estimates can have very high or infinite variance, and their empirical standard errors are often unreliable. Inspired by multilevel regression and poststratification in survey sampling, the authors propose post-stratified importance sampling (PIS), which turns high-dimensional weighting into two one-dimensional problems: regression for the conditional mean of the importance ratio and density estimation for its marginal law. Ordinary importance sampling is a special case. A model-based implementation uses one-dimensional Gaussian process regression and heavy-tailed density estimation, and simulations and cross-validation applications show gains in estimation accuracy and uncertainty quantification.
In briefARLoop: Fully Open Pretrained Looped Transformers Using Attention Residuals
SummaryThis paper proposes ARLoop, a looped transformer that repeatedly applies shared layers and selectively reuses earlier layer outputs through attention over depth. The authors find that sliding-window depth attention matches or beats full-history attention at larger recurrent depths. Scaled to 1.2B parameters and pretrained on 0.5T tokens with three recurrences, the model outperforms Huginn-3.5B on all evaluated tasks, with especially strong gains in math and coding.
In briefDiagnosis by Composition: Finding and Locating Unseen Diseases from Combined Radiological Findings
SummaryThis paper introduces diagnosis by composition (DbC), a learning framework that mirrors how radiologists move from visual findings to disease through guidelines. The model learns a vocabulary of radiological findings, and each disease is defined by its required findings, combined without learned parameters into a diagnosis map. On multi-phase liver CT, DbC detects and localizes unseen diseases, rare diseases from another institution, and external cohorts better than all comparators.
In briefRegret-based characterization, information limits, and auditing sample complexity for adaptive policies
SummaryThe paper studies whether a policy's performance comes from its average action placement or from reacting to information in a finite-horizon MDP. It shows that the gap between a policy and its open-loop marginal equals a sum of covariance terms, and bounds this gap by the information the policy uses. The authors also give the number of rounds needed to estimate this gap from bandit-feedback logs, and provide an importance-sampling auditor for episodic MDPs.
In briefPost-Inference Model Chaining: A Problem Statement and a Model-based Bootstrap for Training End-to-End Digital Twins
SummaryThis paper examines how separately fitted stage models are chained after inference to build end-to-end digital twins, and how a common Monte Carlo simulation approach behaves when predictions are forwarded through such chains. Analytical examples and simulations of multi-stage linear-quadratic chains show interval and coverage inflation, along with input-dependent bias. The authors propose a model-based bootstrap, which in the tested scenarios substantially reduces relative mean bias and brings coverage closer to the nominal 90%.
In briefLearning curves in natural and artificial systems: a unified mathematical model
SummaryThe paper proposes a general model of experience formation, defining the learning level as the probability that the next uncertainty value has been mastered and not forgotten. It derives that the learning curve shape is mainly determined by the distribution of situation occurrence, giving exponential curves for uniform distributions and power-law curves for power-law distributions. The authors relate these mechanisms to scaling laws in deep learning and address choosing a training distribution that differs from the deployment distribution.
In briefPLANTEUR: A Crop-Type Segmentation Benchmark for Tropical Regions
SummaryThe paper introduces PLANTEUR, a benchmark for mapping crop types from satellite images in French overseas territories. It provides aligned Sentinel-2, Sentinel-1, and Landsat-8 image sequences with parcel-based labels covering six tropical regions and eight classes. The authors define a standard test split and a leave-one-island-out protocol to measure generalization across territories, and they evaluate fully supervised models, linear probes, and fine-tuned foundation models. The code and data are open source on GitHub and Hugging Face.
Piper Wolters, Gabriel Tseng, Favyen Bastani et al.
In briefAUC-based fairness measures for clinical prediction AI
SummaryThis paper examines eight AUC-based fairness metrics used to compare clinical prediction models for binary outcomes across subgroups. It explains that AUC is non-collapsible, so subgroup and overall AUC values cannot be directly compared, and analyses what each metric estimates. Simulations based on the EuroSCORE II in-hospital mortality model illustrate their behaviour. The authors recommend reporting an AUC matrix and, when needed, within-group AUC parity.
Jo\~ao Matos, Ben Van Calster, Richard D. Riley et al.
In briefPredicting flood peak depth from high-water marks using terrain, satellite embeddings and weather data
SummaryThe paper trains a fusion network on 5,689 USGS high-water mark surveys from 39 flood events to predict flood depth at surveyed sites six hours before the peak. It reports a mean absolute error of 0.39 m, compared with 0.93 m from differencing measured water levels against the same elevation model. On eleven held-out events, however, the model does not resolve depth at the marks (R2 = 0.03).
In briefPhysics-informed reduced models for multiscale systems using Mori-Zwanzig methods
SummaryThis paper applies the Mori-Zwanzig formalism, a model reduction approach, to bases discovered by physics-informed machine learning rather than standard functions like Fourier modes. The fluctuation-dissipation theorem is used to close the resulting non-Markovian equations for the learned basis coefficients. The authors also introduce a complementary generative closure model. Both closures substantially improve predictions of filtered states compared with Markovian dynamics alone.
In briefDifferentially private agnostic learning of halfspaces with no extra sample cost
SummaryThis paper presents the first computationally efficient differentially private algorithms for agnostically learning halfspaces. It covers three settings: improper learning under Gaussian and log-concave marginals, proper learning under Gaussian marginals, and semi-agnostic learning under Gaussian marginals. In all three, the extra sample complexity caused by privacy matches the corresponding non-private sample complexity.
WhyThe authors report the first efficient differentially private agnostic halfspace learners, with privacy adding no sample complexity beyond non-private bounds.
In briefMulti-objective generative design via expected hypervolume fine-tuning
SummaryThis paper introduces Hypervolume Fine-Tuning (HVFT), a method that adapts a pretrained generative model so its samples better balance several competing properties. It optimizes the expected hypervolume of generated samples, rewarding both candidate quality and collective trade-off coverage, while keeping the model close to its pretrained distribution via a KL constraint. The authors report state-of-the-art expected hypervolume and trade-off coverage on therapeutic-peptide and molecular design tasks using discrete diffusion and flow-matching models.
Juan Guevara, Riccardo De Santi, Andreas Krause et al.
In briefUltraDiff: Self-Supervised Ultrasound Video Restoration Using Diffusion Priors
SummaryThis paper presents UltraDiff, a method that jointly increases the resolution and removes speckle noise from ultrasound videos without paired training data or network retraining. It uses a frozen pre-trained diffusion model as a plug-in prior and optimizes each frame's sampler seed under a range-constrained maximum a posteriori formulation, with a speckle-aware fidelity term and a total-variation regularizer. On EchoNet-LVH videos, it improves PSNR by 1.24 dB over the strongest baseline, and on real echocardiography videos it reduces the mean absolute error of zero-shot ejection fraction prediction from 7.31 to 6.37 points.
In briefTAQ-MOTR: Tracking-Aware Detection Queries for End-to-End Multi-Object Tracking
SummaryThis paper proposes TAQ-MOTR, a framework that forms detection queries in end-to-end multi-object tracking with awareness of trajectory states. A pre-decoder Hybrid Query Generator fuses current-frame image features, track memory, and learnable memory to help discover untracked objects and complement tracked ones. A training-only auxiliary detection branch adds extra detection supervision. The method reports improvements on DanceTrack, SportsMOT, and BFT, reaching 72.0 HOTA, 63.1 AssA, and 76.6 IDF1 on the DanceTrack test set.
In briefEmpirical Bayes flow matching for continuous cryo-EM structural heterogeneity
SummaryThis paper studies how to learn generative priors over high-dimensional latent variables from indirect, noisy observations, focusing on continuous heterogeneity in cryo-electron microscopy. It proposes Empirical Bayes Flow Matching (EB-FM), which learns a flow matching prior directly from observations by alternating guided flow-based posterior sampling with model updates. The method is validated on two-moons and MNIST inverse problems and on simulated cryo-EM data with known poses, where it recovers continuous distributions of protein conformations in atomic-coordinate space.
In briefRobust Reconstruction of Power Lines from Incomplete Mobile and Aerial LiDAR Data
SummaryThis paper presents an automated method for reconstructing individual power conductors from mobile and aerial LiDAR point clouds, even when returns are sparse, incomplete, or occluded by vegetation. It combines scaled density-based clustering, robust polynomial fitting, and iterative fragment merging, applying the same pipeline to both data types without modality-specific tuning. On the aerial dataset it reports 96.87% precision, 96.66% recall, and a 0.127 m geometric RMSE; on the mobile dataset, a 0.045 m point-to-model RMSE and 95.27% coverage. Source code and data are publicly available.
WhyThe authors apply one modality-agnostic pipeline to mobile and aerial LiDAR power-line reconstruction, reporting 96.87% precision and 96.66% recall on aerial data with degradation and sensitivity expe
Mohamed Gaha, Damos Ayobo Abongo, Philippe Massicotte et al.
In briefEnsembles of random partition models under covariate shift
SummaryThis paper studies covariate shift, where the input distribution changes between training and deployment, and why ensembles tend to be more robust than single models. The authors extend covariate shift severity theory from single models to ensembles and introduce new discrepancy measures to control the risk of partition-based kernel estimators. They show that with sufficiently large ensemble size, these models can reach minimax-optimal risk convergence rates under covariate shift. The results suggest that larger ensembles are needed as covariate shift becomes more severe.
In briefFactorized MLLM embeddings that model intended edits for composed image retrieval (FactEmbed)
SummaryThis paper addresses composed image retrieval, which finds a target image from a reference image plus a modification text. It proposes FactEmbed, which separates understanding the composed query from building discriminative embeddings, using factorized latents that imagine the intended visual changes. These latents are trained with a factor-level contrastive loss and distilled into a deeper attention layer of the MLLM. The authors report that FactEmbed outperforms prior methods across CIR benchmarks.
In briefFrom Natural to Medical Images: Whether Diffusion Models Transfer to Medical Image Generation
SummaryThis paper examines whether diffusion denoisers pretrained on natural images transfer to medical image generation. Experiments span latent and pixel-space paradigms, model sizes from 131M to 952M parameters, and two radiology datasets, RadImageNet and BraTS. Natural-image initialisation consistently improves generation quality, and the benefit tracks spatial structure rather than semantics. The authors conclude that natural-image denoisers transfer positively to medical ones in their setup.
WhyThe authors systematically study natural-image transfer to medical diffusion denoisers, identify spatial structure as the key channel, and report 21-34% gFID gains on RadImageNet and BraTS.
Efthymios Georgiou, Ioannis Kakogeorgiou, Sarah Br\"uningk
In briefAdjusting learning rates for cross-entropy training using local quadratic estimates
SummaryThe paper presents two methods that adjust an existing learning rate schedule for stochastic gradient descent with cross-entropy loss, each using one extra forward pass to estimate loss derivatives. Both methods yield very similar step sizes, which are used to slightly raise or lower the scheduled learning rate. Experiments training ResNets on Cifar10 and Cifar100 show faster loss decrease and greater monotonicity than the perturbed schedule. Test accuracy improves slightly on Cifar10 and stays the same on Cifar100.
In briefProvably Learning Mixtures of Linear Experts
SummaryThis paper studies whether gradient-based training of Mixture-of-Experts models converges to good solutions. It analyzes a gated mixture with linear experts and a gate defined by Gaussian clusters in the input space, and proves that gradient EM on the likelihood jointly trains the gate and experts to any prescribed accuracy. Under logarithmic overparameterization and a separation condition on gate centers, convergence occurs in polynomially many iterations from random gate initialization and least-squares-initialized experts.
WhyThe authors give a global convergence guarantee for joint gradient EM training of gated linear-expert mixtures, with experiments on the predicted phase transition.
In briefWhat the Model Sees: Input Choice Validity in AI Video Analysis
SummaryThis paper examines how social scientists using multimodal large language models to analyze video should choose among text, audio, video, and combined inputs. The authors compare these input representations across six open-weight models and three human-annotated video datasets. They find that no single representation performs best consistently, combining inputs rarely helps, and higher classification accuracy does not reliably yield estimates closer to human codes. The authors conclude that input representations must be validated both for the construct measured and for the estimated quantities derived from the labels.
In briefOperAItor: An agent framework for fault-tolerant operation of chemical plants
SummaryOperAItor is an open-source framework for building and testing large language model agents that manage faults in chemical processes. It is demonstrated on a simulated reactor-distillation plant producing acetone, using 20 scenarios with single and multiple faults. Without agent intervention, 18 scenarios caused plant stops; with validated agent configurations, three of the four tested models cut stops to 5-8 and recovered $265,000 of the loss.
In briefPadLoc: Training-free sensor localization on floor plans using sparse geometric matching
SummaryPadLoc is a training-free localization back-end that places a moving sensor within a known floor plan. It keeps a swappable depth front-end but replaces the dense pose grid with sparse matching: dominant wall orientations generate a small set of pose candidates, which are scored against free space in the plan and filtered with a Gaussian-mixture Bayes filter. The authors report that it matches or beats competing methods on Gibson trajectories and on Bonn LiLocBench and Hilti-Trimble datasets.
In briefAgentAM: An AI Agent for Optimizing Laser Powder Bed Fusion Processes with Spatter and Defect Awareness
SummaryThis paper presents AgentAM, a large language model agent that lets process engineers query CFD-trained surrogate models of spatter and melt pool geometry for SS316L via natural language. The polynomial surrogates, trained on 279 single-track simulations, predict melt pool dimensions and spatter volume, and geometric criteria assign defect regimes. The agent selects tools for prediction, process-window scans and constrained optimization, and was evaluated on a 145-prompt benchmark.
In briefLearning Behavior-Related Brain Dynamics from Multimodal Neural Recordings
SummaryThis paper presents BREM-NET, a nonlinear multimodal dynamical model that fuses spiking activity, local field potentials, and behavior. A structured learning framework separates behaviorally relevant neural dynamics from residual dynamics. On a public non-human primate dataset with simultaneous spiking and LFP recordings during a 2D reaching task, the model shows better behavior decoding, neural prediction, and latent alignment with behavior. It maintains these results when neural modalities are asynchronous or have dropouts.
Parastoo Azizeddin, Eray Erturk, Maryam M. Shanechi
In briefCalibrated Answer Repair for Financial QA: When Self-Checking Backfires
SummaryThis paper studies financial question answering, where a model may pull a plausible number from the evidence but attach it to the wrong entity, period, unit, or accounting concept. It proposes Calibrated Answer Repair (CAR), which first estimates whether a draft is likely to benefit from repair, then runs a targeted check, localizes the relevant evidence or draft span, and limits the permitted edit. Across FinRank, Fin-RATE, and FinQA with three open-weight models, CAR outperforms generic verification in all settings, recovering 14.6% of initially missed answer requirements while damaging 2.1% of initially correct ones.
In briefBasin2Vec: Learning Basin-Centric Embeddings for Scalable Hydrological Modeling
SummaryThis paper presents Basin2Vec, a self-supervised framework that learns compact basin representations from hydroclimatic and physiographic data without requiring observed streamflow. Across several forecasting architectures, it outperforms predefined attributes and general-purpose geospatial embeddings in temporal generalization and zero-shot spatial transfer. The authors report average relative gains of up to 8.3% temporally and median predictive performance 3.5 times that of the strongest baselines on unseen basins.
In briefRoboTryOn: Creating Robot Demonstrations from Human Videos via Virtual Try-On
SummaryRoboTryOn converts human videos into videos of a target robot specified by a reference image, keeping the original task, objects, and background. It translates keyframes autoregressively for visual consistency, then synthesizes video conditioned on the human hand skeleton. An inverse dynamics model using depth features infers action labels from the synthesized videos for policy learning. Policies trained on this synthetic data raised average success on novel tasks from 2% to 60%, at about a third of teleoperation collection cost.
WhyThe authors treat human-to-robot translation as virtual try-on, synthesizing robot videos for policy learning; the abstract reports bimanual success rising from 2% to 60%.
In briefPerturb, Repair, Verify: Self-Play Vision-Language Verifiers for Compositional Understanding
SummaryThis paper introduces PROVE, a self-play framework that trains vision-language verifiers to judge whether images match their text descriptions. Starting from a correct image-caption pair, a Mutator perturbs objects, attributes, counts, or relations, the verifier judges the altered caption and gives feedback, and a Repairer restores a faithful caption for re-verification. Experiments report improved fine-grained compositional verification across model scales and families, with limited and benchmark-dependent changes on broader multimodal tasks.
In briefMacrocanonical Generator Networks: Fast, data-efficient neural stand-ins for physics simulations
SummaryThis paper introduces macrocanonical generator networks, a method for training fast neural surrogates that reproduce the multiscale statistics of stationary physical processes. The generator is trained to satisfy these statistical constraints in expectation, so realistic sample-to-sample fluctuations are preserved, and each new sample then takes a single forward pass. Experiments on cosmology and fluid simulations show realistic samples can be recovered from very small training sets, with generation far faster than the reference simulator.
In briefAsynchronous Robot Policies: Running Reasoning and Action Generation at Separate Rates
SummaryLarge vision-language models improve general robot manipulation, but their computational latency slows real-time physical reactions. The paper proposes Asynchronous Robot Policy (ARP), which decouples semantic reasoning from low-level action generation: a high-capacity vision-language model updates a semantic memory at low frequency, while a lightweight vision-action expert generates actions at a much higher control frequency. On Franka manipulation tasks, the authors report higher success rates and shorter execution times than baselines. They also state that the vision-language model can run at about 10 times lower than its nominal frequency with no significant performance loss.
In briefBoosting Interactive HTML Generation with High-Fidelity Reinforcement Learning
SummaryThis paper addresses generating interactive HTML pages from natural-language specifications, noting that current models often fail to ensure functionality and visual quality during real interaction. Using mini-games as a test setting, the authors propose WebCatalyst, a browser-interaction reinforcement learning framework. It combines a game-level GUI assessor, Visual Anchor to mitigate VLM reward hacking, and Action-Trace Credit for runtime-grounded credit assignment.
In briefCreating Power Flexibility in Datacenters via Query Routing
SummaryLarge language model inference is raising datacenter power demand, so facilities need ways to adjust consumption while still serving interactive users. This paper proposes query routing, which spreads queries across model instances without deferring them, and presents an analytical framework linking arrival rates and sequence lengths to power consumption. Validated on four GPU platforms, the framework was tested in a trace-driven simulation of a 10,000 H100 GPU datacenter, showing flexibility up to 16% of total power, with 90% of the desired reduction reached in about 10 seconds.
Shivam Patel, Bingqing Chen, Jared Fernandez et al.
In briefGM League Arena: A Benchmark for Long-Term Sports Team Management by AI Models
SummaryThis paper introduces GM Bench, a benchmark that tests large language models acting as general managers of professional sports teams across multiple seasons under resource limits and competition. It builds simulators for four sports, adding financial constraints, player demands, and market mechanics. Over five NBA seasons, Claude Opus 5.5 ranked first among 29 evaluated LLMs.
In briefCell crops sharing a source image inflate cervical cytology benchmark accuracy
SummaryPublic cervical cytology benchmarks cut many cell crops from far fewer microscope fields or slides and often split data at the crop level, so test cells share acquisition sources with training cells. In a preregistered study, crop-level splitting raised accuracy on SIPaKMeD and CRIC. The authors conclude that decodability and saliency analyses cannot replace a parent-disjoint split.
Snehasish Ghosh, Soham Mitra, Antareep Majumdar et al.
In briefBeyond Foundation Models: Using PFN Priors to Tune Ordinary Tabular Models
SummaryThis paper proposes PriorTune, a method that uses synthetic data priors originally built for pretraining tabular foundation models to select hyperparameter settings for conventional tabular models. Offline, it picks the subset of configurations that wins most often against a random baseline on synthetic datasets, and online it only tries those selected settings on a new dataset. On the BeyondArena benchmark, LightGBM with TabPrep features and these configurations ranks above TabPFN-3 and RealMLP, while being more efficient at inference.
In briefMeta²: Precomputed Reusable Knowledge for Adapting Agent Harnesses Online
SummaryThis paper introduces Meta², a method for improving long-horizon AI agents by optimizing the harness, which consists of prompts, tools, memory, and hooks around a frozen model. Offline, it learns a reusable textual Meta-Skill by contrasting training runs with and without an evolver, guiding when to change the harness and which edits to avoid. Online, an evolver uses this guidance and emerging run evidence to either keep the harness unchanged or make task-local edits. On Terminal-Bench 2.1 and DeepSWE 1.1 with GPT-5.6-Terra, Meta² achieves the highest reported performance among evaluated methods, at 81.00% and 44.44%.
In briefRandom matrix analysis of symmetric tensor unfolding, a method for recovering hidden signals from noisy high-order data
SummaryThis paper studies recovering a hidden unit vector from a spiked tensor corrupted by symmetric Gaussian noise. Instead of replacing the noise with an i.i.d. tensor as earlier work did, the authors analyze the symmetric model directly using random matrix theory. They show that square unfolding undergoes a sharp phase transition at signal strength around n^((p-2)/4), and derive precise expressions for the spectral distribution, threshold, and overlap.
WhyThe authors analyze symmetric tensor unfolding directly via random matrix theory, showing a sharp phase transition at λ≍n^{(p-2)/4} and validating an open conjecture.
Tom Szwagier, Xiaoyi Mai, Jos\'e Henrique de Morais Goulart
In briefFrom Opinions to Commitments: Unstable Assignments in AI-Drafted Proposals
SummaryThis paper examines whether large language models that turn citizens' opinions into collective proposals produce stable task assignments when the same opinions are processed again. Testing four models on opinions from DeliberationBank and Polis, the authors found that 17% to 20% of five-run sets showed assignments in some runs but not others. Variation persisted under safeguards such as temperature zero, and prompts seeking agreement covered minority positions less than prompts asking to describe differences.
Sachit Mahajan, Javier Argota S{\'a}nchez-Vaquerizo, Arthur Capozzi et al.
In briefPAC learning where each example has several correct answers
SummaryThis paper studies PAC learning when each example has at most ℓ correct labels, and a prediction counts as correct if it lies in that set. Training may reveal the full set, only whether a proposed label is correct, or a single correct label chosen by a demonstrator. The authors show that learning from demonstrations can be much harder than learning from full sets or bandit feedback, and they give dimension-based characterizations of learnability for finite and countable label spaces.
WhyThe authors characterize learnability from full sets, bandit feedback and demonstrations, and report a tight ε^{-ℓ} sample bound with proofs for finite and countable label spaces.
Steve Hanneke, Farnam Mansouri, Alireza F. Pour et al.
In briefTraining-free mapping of solar panel modules from multi-view drone images
SummaryThis paper presents MV-PVMapper, a training-free framework that maps individual photovoltaic modules by combining vision foundation models with drone photogrammetry. It generates segmentation candidates with SAM 3, projects them into object space with slope correction, and selects instances across views using module-size checks, then feeds accepted results back for iterative refinement. On surveys of three rooftop PV plants, it recovered 7,710 of 7,711 reference modules with no false positives, reaching an aggregated Jaccard index of 97.55%.
In briefSolving the nonlinear Westervelt equation for ultrasound using physics-informed neural networks
SummaryThis paper solves the one-dimensional Westervelt equation, which describes finite-amplitude ultrasound propagation in biological tissue, using physics-informed neural networks (PINNs). It tests three synthetic cases: linear, shock-dominated, and dissipation-dominated. A sinusoidal activation function is used to capture oscillatory behavior. The results indicate that the attenuation coefficient controls the rate of amplitude decay, while the nonlinearity coefficient governs frequency-dependent phase speeds.
In briefDo ray-traced city models beat standard channel priors for unseen base stations? A pre-registered simulation study and a reversed test on measured data
SummaryThe paper asks whether a channel prior pooled from many ray-traced city scenes serves a newly deployed base station better than standard 3GPP CDL models. In simulation, the pooled prior was about 0.6 dB behind a site-trained estimator, roughly matching CDL after 50 measurements, and learned models added little. On measured outdoor channels from the DICHASUS dataset, the result reversed, with the ray-traced prior 3.2 to 4.1 dB worse than CDL.
In briefHaving the Right Source Is Not Enough: Why Old Audio Distorts Speech AI Answers
SummaryThis paper studies why irrelevant earlier audio in a multi-turn conversation changes a speech large language model's answers, even when the queried event is still represented. Controlled experiments on MOSS-Audio and MiMo-Audio show that historical values steer errors toward specific competing answers. Causal interventions find that in open-ended question answering most such errors vanish when later computation cannot access the irrelevant source, while many errors persist in multiple-choice answering. The authors conclude that reliable conversational memory requires controlling how retained information enters answer computation, not just retaining and retrieving it.
In briefOptimal control theory for early-exit neural networks: turnpikes in ResNets
SummaryThis paper analyzes jointly optimizing losses at the output and intermediate layers of neural networks using optimal control theory. It proves that ResNets trained this way exhibit a turnpike phenomenon, where the layers needed to reach a set loss level are bounded regardless of total depth. The authors connect this to overthinking and explain early-exit criteria based on softmax probabilities, with experiments confirming the effects.
Jens P\"uttschneider, Simon Heilig, Asja Fischer et al.
In briefDistributed bandit learning for freshness-aware data preprocessing in edge-enabled digital twins
SummaryThis paper studies freshness-aware preprocessing of IoT data in digital twin networks supported by multi-access edge computing (MEC) servers, which reduces transmission volume and DT-server load. CPU-allocation selection is formulated as a fixed-confidence distributed best-arm identification problem in linear bandits, solved with the DistLinGapE collaborative learning method. Numerical results indicate the method identifies the optimal feasible strategy and reduces preprocessing delay under stationary task assignments and workloads. In the evaluated settings, collaboration gives an identification speedup nearly proportional to the number of participating MEC servers.
In briefLearning intersections of degree-2 polynomial threshold functions beyond multi-index models
SummaryThe paper presents a PAC learning algorithm for intersections of k degree-2 polynomial threshold functions over R^d under a Gaussian distribution. It achieves error ε in time and samples d^{O(k)}·2^{O_k(log^3(1/ε))}, improving on the d^{O(k^2/ε^2)} running time of polynomial regression. An SQ lower bound shows the d^{Ω(k)} dimension dependence is unavoidable.
WhyThe authors report a learner whose dimension exponent is independent of accuracy, with a matching SQ lower bound for degree-2 PTF intersections.
Ilias Diakonikolas, Daniel M. Kane, Mingchen Ma et al.
In briefFrom AutoResearch to Continual AutoResearch: Do AI Agents Learn Across Tasks?
SummaryThis paper introduces CONTINUAL AUTORESEARCH, an environment that compares a persistent AI agent against a reset version of the same agent on matched task streams, data, and time budgets. It covers nine task families, including synthetic tasks, symbolic regression, time-series forecasting, and tabular AutoML, and tests five frontier agents. Persistent state improves research efficiency for all five agents, helping them find stronger solutions sooner, but these gains do not consistently yield better final solutions.
Jeremy Qin, Ruizhi Deng, Mohamed Osama Ahmed et al.
In briefLabel-preserving Fourier domain adaptation using generative-AI images for 3D mitral valve segmentation
SummaryThis paper addresses the difficulty of segmenting the mitral valve in 3D transesophageal echocardiography, where expert-annotated public datasets are scarce. The authors test Fourier Domain Adaptation using images from general generative-AI platforms as annotation-free style sources, comparing them against real-image style donors and other training conditions on the MVSeg2023 dataset. Generative-AI-based adaptation significantly reduced boundary and surface errors, but its effect on Generalized Dice Score depended on the generator, with only MAI showing a significant gain over the baseline. The non-AI control performed comparably to ChatGPT, so the authors do not conclude that the generative approach is superior to using real-image donors.
Javeria Javed, Ahmed Salim, Muhammad Mahboob Ur Rahman
In briefOptical circuit switching set once per deployment to speed up large language model inference decoding
SummaryThis paper addresses how communication latency during LLM inference decoding limits per-user interactivity in scale-up fabrics built from electronic packet switches. It proposes workload-static optical circuit switching, which configures circuits once per deployment and co-designs them with model parallelism and topology-aware collective schedules. Evaluated with the GXNET emulator on a 576-accelerator domain serving 5T-parameter MoE models, the all-optical fabrics improve interactivity by up to 65% and energy efficiency by up to 119% over 4:1 oversubscribed packet-switched fabrics.
Maximilian Th\"urmer, Nicolas Scotto di Perto, Eloise Stein et al.
In briefMolecular Prior Networks: Predicting Drug-Target Affinity in a Single Forward Pass
SummaryThis paper proposes a new approach to modelling drug-target affinity, in which a transformer predicts the labels of query molecules from labelled support molecules in one forward pass. The model is trained on thousands of existing affinity datasets, so it can learn general patterns of molecular activity across targets, chemical series and dataset sizes. For the smallest support sets, it matches the state of the art on the regression variant of the FS-Mol benchmark and runs faster because it requires no task-specific training.
In briefMeasuring vertical continuity of forest fuel layers using Gaussian decomposition of laser scan point clouds
SummaryThis study applies Gaussian decomposition to vertical height profiles from terrestrial laser scanning point clouds across 25 permanent forest plots. It compares geometric, probabilistic, and overlap metrics at multiple spatial scales to characterize vertical fuel continuity. The authors report that 3 to 6 m cells offer a useful compromise, and suggest the approach can improve ladder fuel characterization in wildfire risk assessment.
Jes\'us Torralba, Luis \'A. Ruiz, Pablo Crespo-Peremarch et al.
In briefDual-VaDE: Stable, Robust Medical Image Prototype Learning Using Dual-Domain Priors
SummaryDual-VaDE is an unsupervised generative framework that learns interpretable anatomical prototypes from unlabeled medical images. It couples a spatial anatomy prior with a phase-preserving Fourier spectral prior in a dual-domain Gaussian mixture, optimized end-to-end with an alternating EM formulation. Evaluated on OASIS-3 brain MRI, it yields sharper prototypes, cluster separation improvements of up to 94% over spatial-only baselines, and low inference overhead of 0.28 ms per image.
In briefThe Platonic Brain: Convergence of Representations in EEG Foundation Models
SummaryThis study tests whether the Platonic Representation Hypothesis applies to human brain representations using EEG foundation models. Recordings from six participants watching a television series, paired with aligned visual and language content, were used to examine whether model representations converge within and across modalities as decoding of EEG markers improves. EEG models with greater decodability showed stronger alignment across model families in both local and global structure, not consistently increasing with parameter count. Cross-modal convergence depended on modality and metric: higher EEG decodability was linked to stronger global alignment with language models, but not consistently with vision models, and local alignment showed no consistent association.
Nicolas Waehner, Enzo Tagliazucchi, Jacobo Sitt et al.
