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 briefJev with Foresight: Safety-Focused Language-Guided Self-Driving Using a Language World Model
SummaryThis paper presents Jev-AD, which pairs Jev, a decision model that answers predefined questions with probabilities, with a language world model that predicts how nearby vehicles move over the next two seconds for each candidate action. The predicted consequences are compared by the decision model rather than inferred from raw coordinates. In 500 simulated highway merging episodes, the collision rate dropped from 26.0% to 2.4% at 0.25 seconds per decision. In 500 replayed real-world NGSIM episodes, it dropped from 48.2% to 2.8%.
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 briefLIBRA: Automatically Building Executable Tool Libraries for LLM Agents
SummaryThis paper introduces LIBRA, a framework that turns a flat catalog of tools into an executable tool library. An LLM proposes candidate invocations, reusable outputs, state changes, and success conditions for each tool, and LIBRA tests these compositions in a sandbox, keeping only relations supported at runtime. The library is built before user queries and needs no dependency annotations or task-specific training. On AppWorld, BFCL, and SciToolEval, LIBRA improves one-shot Recall and NDCG over the same dense retriever by 7.9 and 3.7 percentage points on average.
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 briefSurgTransfer: Multi-Step Latent Action Modeling for Surgical Action Transfer
SummaryThis paper proposes a multi-step action learning framework that learns context-invariant latent actions in a surgical world model, so they can transfer across different surgical contexts. The method learns from unlabeled videos and is evaluated on the public SurgToolLoc-2022 and PitVis-2023 datasets using quantitative metrics and qualitative assessment by surgical experts. The code and implementation details are published on GitHub.
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.
In briefLatentTriad: Pre-action risk representations for detecting malicious agent skills
SummaryLarge language model agents are increasingly extended with reusable skills, but untrusted skill content can trigger operations beyond the user's task. The authors find that hidden states formed during prefill contain readable information about whether a task and its operations are consistent. They introduce LatentTriad, which combines contextual boundary features, group-adjusted operation directions, and departures from benign state relationships. Detection runs on a single shared prefill without generating output tokens, invoking tools, or executing skill code.
In briefSparse Anchors for Aligning Repeated Robot Recordings in Changing Environments
SummaryThe paper addresses aligning a new robot recording of an environment to a previously built reference map, even when the scene has changed. The method gathers absolute pose estimates via place recognition, feature matching and geometric checks, keeps the consistent ones as anchors, and combines them with within-session relative poses in a robust pose-graph optimization. On 15 cross-session experiments from the HILTI26 construction-site dataset, it reaches 94.4% recall within 1 m and 100.0% within 5 m, outperforming independent image localization and direct registration of reconstructed geometry.
Yunus Talha Erzurumlu, Niklas Trekel, Saurabh Gupta et al.
In briefImage matching method using deep feature manifold embedding and neural adjoint maps
SummaryThis paper addresses finding semantic correspondences between different object instances of similar categories in real-world images. It adapts functional maps, used for 3D shape matching, by deriving a discrete Laplace-Beltrami operator from a manifold embedding of deep image features via Galerkin's method, giving image grids a meaningful geometric structure. A neural adjoint map is added to represent non-linear mappings. The authors report state-of-the-art results among zero-shot image correspondence methods on multiple dense correspondence benchmarks.
In briefCRONOS: A benchmark for reinforcement learning on robot manipulation without episode resets
SummaryThis paper introduces CRONOS, a simulation benchmark for fine-tuning vision-language-action (VLA) robot models with reinforcement learning when scenes are not reset between episodes. The authors find that naively fine-tuning pre-trained policies fails in this setting because of shifts in object configurations and end-effector poses. They report that targeted reset allocation and addressing biases in pre-trained models can mitigate these problems. Non-episodic training is also found to improve long-horizon manipulation and generalization to unseen object configurations and task sequences.
Po-Yi Wu, Djengo Cyun-Jyun Fang, Dian Cheng et al.
In briefRollout-free world model for scoring self-driving trajectories
SummaryThis paper proposes RFWorld, a rollout-free world model for evaluating candidate trajectories in end-to-end autonomous driving. Instead of predicting full future scenes for each candidate, it builds a shared scene memory that candidate trajectories query directly through sparse spatial readout. The authors report that this reduces inference cost and avoids accumulated prediction errors, and that it achieves state-of-the-art results on NAVSIM and Bench2Drive.
In briefVIGOR: Visual Gain Ordering to Reduce Hallucination in Diffusion-Based Multimodal Models
SummaryMultimodal discrete diffusion language models generate answers by iteratively predicting masked tokens, but they can still hallucinate content that fits language patterns yet lacks support in the image. The paper proposes VIGOR, a training-free, plug-and-play inference method that reorders which masked positions to fill next, based on how much each prediction depends on visual evidence. It compares token confidence with the original image against a pass where the visual input is ablated. The authors report that VIGOR reduces hallucination and improves multimodal reasoning on LLaDA-V and Lumina-DiMOO.
In briefFloMo: 3D Scene Flow as an Intermediate for Learning Robot Actions from Human Video
SummaryThis paper presents FloMo, a 5B video-generation model fine-tuned to predict 3D scene flow and action tokens. By rendering scene flow as RGB and encoding it in the same latent space as video, a pretrained video backbone adapts efficiently to motion prediction. Co-trained on action-free human video and robot teleoperation data, the model generalizes to unseen objects, motion primitives, and tasks.
In briefDo Agentic Policies Need to Look at Every Step? A Geometric Abstraction for Reasoning and Action
SummaryThis paper proposes replacing repeated raw-image reasoning in robot control with a 3D geometric abstraction written as text and symbols. Typed, evidence-aware references to object parts are grounded from calibrated RGB-D data, and after each action the system returns compact textual checkpoints with metric residuals. Across offline policy synthesis and online execution, reasoning over this text and symbolic state improved multi-step manipulation completion while substantially reducing repeated multimodal queries.
In briefAblation study of protein ensemble generator design choices across structural types
SummaryThis paper examines design choices for generative models of protein conformational ensembles. Using a minimal Cα flow-matching model, the authors ablate architecture, sequence conditioning, rigid-body alignment and training data source. Variants are evaluated against experimental SAXS and NMR chemical-shift data across ordered, disordered, multi-domain and transitional proteins. The training data, rather than the architecture, is identified as the main constraint.
In briefLearning to Recognize Footsteps Across Unseen Footwear
SummaryThis paper studies identifying people from footstep patterns recorded on pressure- or force-sensitive walkways, where enrollment uses one footwear type and recognition uses unseen footwear. The authors propose BiStepNet, which splits spatial and temporal identity cues into two independently trainable streams and outperforms compact 3D spatiotemporal baselines on the StepUP-P150 dataset. Results show cross-footwear performance improves systematically with the volume and diversity of footwear in training, and auxiliary gait datasets with partial modality overlap can further improve robustness.
In briefHow correlations between emitting partons affect direct-photon yields and elliptic flow in heavy-ion collisions
SummaryThis paper studies how correlations between partons that emit photons affect the yield and elliptic flow of direct photons in heavy-ion collisions. The authors show that the local 2→2 photon rate depends on the joint distribution of incoming partons, so a correlation can change photon emission without altering single-particle distributions. Folded over the evolving medium in Au+Au and Pb+Pb collisions, the correlation raises the photon yield toward fitted PHENIX and ALICE data and increases v2 mainly at intermediate transverse momentum.
In briefPositional encoding to relax equivariance constraints during training, more efficiently and scalably
SummaryThe paper addresses optimization difficulties in strictly equivariant neural networks, which can struggle when data only approximately follow a symmetry. Prior work relaxed equivariance during training with non-equivariant dense layers, but at significant computational cost. The authors propose injecting a positional encoding with learnable scale to break the symmetry, gradually annealing it to zero so the original equivariant model is recovered by the end of training. They report significant gains for translational and rotational equivariant models with only marginal compute and memory overhead.
In briefALBI: LLM-guided inductive logic programming through iterative generate-test-constrain refinement
SummaryThe paper introduces ALBI, a framework that replaces the ASP-based hypothesis generator in Learning from Failures with a large language model while keeping SWI-Prolog as a formal verifier. It runs an iterative loop in which the LLM proposes candidate hypotheses, Prolog tests them, and diagnostic feedback guides refinement. On 30 standard ILP benchmarks, three frontier LLMs each reached 100% success, matching Popper with about 1.4 to 1.7 iterations on average. Under semantic anonymization, all models still achieved 100% success, with a moderate increase in iterations.
Aboubaker Seddiq Benterki, J\"uri Vain, Gert Kanter et al.
In briefToolSpace: Tool-centric environment synthesis for training long-horizon AI agents
SummaryToolSpace proposes a framework that builds training environments around tools rather than data, mining dependencies between adjacent tool calls from real MCP service execution traces using deterministic rules. It clusters the resulting tool-composition graph into environments and generates tasks via state-aware random walks, keeping only those that replay successfully. On seven long-horizon tool-use benchmarks, the method substantially raises the average scores of Qwen3.6-35B-A3B and GLM-5.2.
WhyThe authors propose tool-centric environment synthesis mined from execution traces, reporting gains across seven long-horizon tool-use benchmarks.
In briefFederated Transfer Learning for Medical Image Classification under Mixed Skew
SummaryThis paper addresses federated learning on multicentric medical data, which exhibits mixed skew across feature, label, concept, and quantity shifts between centers. It proposes PAFP, a personalized federated learning framework built on a pre-trained backbone that inserts client-specific Feature Alignment modules before fine-tuned blocks to align hidden features toward a server-aggregated target. On the FedISIC benchmark, PAFP achieves the highest unweighted balanced accuracy and reduces inter-client performance variance roughly six-fold relative to the strongest baseline.
Alexandre G. Leclercq, S\'ebastien Bougleux, Aur\'elien Corroyer-Dulmont et al.
In briefWhose Effect Is It? How Objectives Inflate Learned Subspace Interventions
SummaryThe paper examines activation patching, a method that copies part of a language model's internal state between inputs to locate concepts by measuring output change. The authors prove that, to first order, the effects of orthogonal parts sum to the whole state's effect, so a learned part exceeds the whole only when the remaining components push the output back. Objectives that reward larger output effects select such parts, and changing the objective to target the network's own output removes most of the excess.
In briefComparing representations and fitting methods for learning port-Hamiltonian models
SummaryPort-Hamiltonian models describe physical systems by separating stored energy, dissipation and external inputs. The paper compares neural networks, analytic expressions and combinations of the two for representing these components, and compares gradient-based optimisation with sparse regression for fitting. Tests on simulated mechanical and electrical systems, with and without noise, report prediction accuracy and computational cost for each approach.
In briefDRAP: A Replay-Free Method for Continual Learning in Physical AI
SummaryThis paper studies continual learning for physical AI, where small per-step policy errors can accumulate along a trajectory and cause closed-loop task failure. It proposes DRAP, a replay-free method that estimates each adaptable tensor's historical importance from action gradients weighted by local trajectory-amplification risk, and uses this to reduce plasticity in components that supported earlier behavior. Applied to a fixed subnetwork of OpenVLA and tested on the LIBERO Spatial, Object, and Goal suites, DRAP is reported to retain previously learned behaviors substantially better than standard AdamW fine-tuning and EWC while still acquiring new suites well.
In briefHaptic Shared Control for Multi-Robot Systems: Letting Operators Steer Robot Groups at Multiple Scales Through Touch Feedback
SummaryThe paper modifies a global connectivity maintenance algorithm so robots can use measurements from their local frames, taken with sensors that have limited fields of view. This algorithm is combined with three adaptive supervision strategies that let an operator control a single robot, pin constraints on several robots, or act on the group's centroid. These methods keep inter-agent constraints satisfied while giving the user haptic feedback about the system's connectivity state. A user study compared performance with and without haptic feedback across scenarios designed for each control mode.
In briefDeep learning for on-phone RGB photo enhancement: Mobile AI 2026 challenge report
SummaryThis report describes the Mobile AI 2026 challenge, which asked participants to design fast models that enhance RGB photos while running on smartphone GPUs from Qualcomm and Arm. Participants used the DPED dataset of image pairs captured by an old iPhone 3GS and a professional Canon 70D DSLR. Models were timed on Snapdragon 8 Elite Gen 5 and Dimensity 9500 chipsets, and the top solutions enhanced Full HD images in under 300 milliseconds on both platforms.
Andrey Ignatov, Georgy Perevozchikov, Radu Timofte et al.
In briefHow finite spectral depth affects early rank selection and bypassing the first gradient-flow saddle in matrix factorization training
SummaryThis paper studies training dynamics of matrix factorization, a simple model of neural network training, under small initialization. It analyzes continuous-time dynamics that applies a finite number of Newton-Schulz iterations to the normalized matrix gradient (Muon). The authors show that the number of iterations can change the rank selected early in training and let trajectories bypass the first nonzero stationary point that gradient flow approaches. Under an additional residual-inertia condition, they prove bypass occurs for all sufficiently large finite depths, give an explicit sufficient threshold, and construct an example where the threshold is four.
In briefMatched but Not Better: Testing Quantum Machine Learning Against Classical Models on Six Clinical Datasets
SummaryThe paper compares a quantum kernel SVM and a variational quantum classifier with five tuned classical models on six public clinical datasets, using identical inputs and 25 cross-validation splits. Paired AUROC differences were assessed with equivalence tests, showing the quantum models approached but did not significantly exceed classical performance. The authors also propose QML-Health-READY, an eight-domain framework defining reporting requirements for credible quantum machine learning claims.
