In briefWhen Instruction Files Go Stale: Documentation Debt and Hidden Cost in Coding Agents
SummaryThis study examines how outdated repository instruction files, such as AGENTS.md and CLAUDE.md, affect coding agents like Claude Code. The authors tested three real stale instructions from open-source projects across 22 maintenance tasks, comparing stale, corrected, and removed versions. Agents often followed stale guidance when relevant, which added extra tool calls, yet task success alone hid this cost. None of the agents repaired the instruction files themselves.
Tural Mehtiyev, Biruk Tadesse, Lawrence Arkoh et al.
In briefParallel Agent Teams: A Better Way to Scale Test-Time Compute
SummaryThis paper proposes parallel agent teams, a method that samples many small agent teams in parallel at inference time and keeps the result of the best team. Tested on 12 long-horizon tasks, including GPU kernels and formal proofs, with a 27B model and a budget of 100 agents, parallel teams of ten beat parallel individual agents on 10 of 12 tasks and beat a single team of 100 on all four tasks where such teams were run. At equal dollar cost, they beat a frontier model on half the tasks.
In briefAR-Edit: Training-free, inversion-free autoregressive video editing for real-time streams
SummaryThe paper introduces AR-Edit, a method for editing video streams in real time without requiring future frames, additional training, or inversion. By analyzing self-attention in DMD-distilled autoregressive video diffusion models, the authors find that queries and keys stay stable while values encode the evolving content. AR-Edit caches source values from a single forward pass and selectively injects them during denoising to preserve the source video.
Hovhannes Margaryan, Quentin Bammey, Christian Sandor et al.
In briefSeeing What Matters: A Benchmark for Task-Directed Spatial Understanding in Robot Manipulation
SummaryThis paper introduces Seeing What Matters, a benchmark testing whether robots choose camera views that gather the visual evidence a manipulation command needs. It covers coffee workstation, drawer/cabinet, and microwave/appliance tabletop scenarios, with 860 frozen evaluation cases. A diagnostic reference pipeline reaches about 98.8% closed-loop success, while seven vision-language models reach only 26-43%.
In briefDeciding When to Step In: Causal Evaluation of AI Agents
SummaryThis paper proposes a framework for estimating the causal effects of randomized interventions, such as verification or screening, within AI agent trajectories. It uses a relatively small dataset of trajectories collected under randomized interventions to estimate dynamic causal effects of contextual intervention policies. Applied to tau-bench Retail, ToolSandbox, and ImpossibleBench, the authors find that intervention effects vary substantially across agent actions and states.
In briefStudies whether retrieving similar problems helps language models solve competitive programming tasks.
SummaryThis paper proposes a retrieval-augmented generation method that uses embedding-based retrieval and similarity search with a dynamic threshold to find analogous problems in a knowledge database. Experiments on the CodeContests dataset compare problems given only their descriptions, with random training examples, and with retrieved similar examples, and add a Min-K membership inference analysis to check for memorization. Performance is measured by pass rates and CodeForces Elo ratings, and the results indicate that similarity-based retrieval substantially improves model performance on competitive programming.
Gabriel Souza Baggio, Gabriel Machado Lunardi, Guilherme Medeiros Machado et al.
In briefDo LLMs Follow Prompt Word Meanings or Their Own Knowledge? A Translation Conflict Study
SummaryThe study builds a controlled Mandarin-to-English translation task in which familiar source words are given new meanings in the prompt, to examine how large language models resolve conflicts between these supplied meanings and their prior lexical knowledge. It evaluates the effects of word frequency, sentence context, and prompt framing across nine models. Reliance on the supplied meanings is greater for less frequent words and in contexts that support them, and larger models within a family use them more than smaller ones.
In briefFast approximate planning for a weaker class of partially observable decision problems
SummaryThis paper studies planning in finite-horizon partially observable Markov decision processes (POMDPs) with finite state, action, and observation spaces. For ρ-decodable POMDPs, it gives quasipolynomial-time approximate planning algorithms for every fixed ρ>0. The authors show that decodability follows from both ℓ1- and ℓ2-observability and is strictly weaker than observability, and that the optimal decodability constant can be computed in polynomial time.
WhyThe authors introduce decodability, extend quasipolynomial-time approximate planning to it, and report polynomial-time computability of the decodability constant.
In briefROSE: Risk-aware method for keeping multimodal AI models accurate through continuous knowledge edits
SummaryMultimodal large language models need ongoing factual updates, but sequential edits can erode existing abilities and cause earlier updates to be forgotten. ROSE restricts parameter changes to the orthogonal complement of previously edited subspaces and uses a dynamically computed risk mask to protect factual parameters important for stability. At inference, it applies multi-granularity integration to activate edited knowledge only for relevant queries. On two multimodal benchmarks, ROSE is competitive with state-of-the-art methods on five key metrics and remains stable over 1,000 sequential updates.
In briefAnimWeb-Bench: A Benchmark for Generating Animated Webpage Code from Videos
SummaryThis paper introduces AnimWeb-Bench, the first benchmark for generating animated webpage code from videos, built from 124 real-world webpages across 7 application domains and 33 animation types. It proposes a two-dimensional evaluation protocol that separately measures static appearance fidelity and animation fidelity against a reference video. Testing 14 multimodal large language models and three coding agents, the authors find that reproducing animations is much harder than reproducing static appearance, and that frame inspection and render feedback improve results.
In briefLanguage as Trait, Politeness as Strategy: Controlled Tests of Linguistic Effects in LLM Social-Deduction Agents
SummaryThis paper uses the LURK Werewolf framework to test whether varying an LLM agent's linguistic range and politeness changes how other agents judge and act. Only the wording of messages is rewritten while game content is kept, and each game day is replayed from saved checkpoints against a re-sampled control. Neither manipulation changed the trust other agents placed in the speaker, and no effect on votes or night actions survived multiple-comparison correction.
In briefTurning PAC bounds on generalization error into PAC bounds on parameter estimation error
SummaryThis paper uses oracle bounds derived from PAC bounds to obtain finite-sample bounds on parameter estimation error for dynamical systems. As an illustration, it gives such bounds for linear and bilinear systems with bounded inputs and noise, applicable to identification algorithms that minimize a regularized empirical loss. The approach relies on an oracle inequality, combined with the convexity of the true loss or system-theoretic properties of the loss function. The authors also compare the convergence rates of these bounds with classical asymptotic results.
In briefA tight hierarchy theorem for chain-of-thought computation in transformers
SummaryThis paper studies the computational power of transformers that use chain-of-thought (CoT) reasoning. It proves the first hierarchy theorem for CoT-based computation, showing that a larger number of CoT steps strictly increases the class of solvable decision problems. The result uses known simulations between CoT-augmented transformers and Turing machines, together with a padding technique adapted from complexity theory.
WhyThe authors establish the first hierarchy theorem for CoT computation, adapting padding to obtain a tight hierarchy within transformer models.
In briefDiffusion-free score matching for sampling exact conditional distributions.
SummaryThis paper introduces DISCO, a method that avoids the time-dependent score fields used in diffusion models by estimating a single score of a slightly perturbed data distribution. The exact conditional score for this distribution is readily available. Experiments show DISCO matches state-of-the-art diffusion models in sample quality, better represents the underlying data distribution, and samples from arbitrary conditionals more accurately than standard heuristic samplers.
In briefA lightweight neural network for detecting targets on edge devices
SummaryThis paper proposes a lightweight convolutional neural network for edge devices that recognizes and classifies threats or targets of interest. It includes a real-time hotspot detection algorithm, reporting inference speeds of up to 20 fps and detection ranges over 400 meters with typical thermal sensors. The authors report high accuracy, sensitivity, and a low false positive rate after testing.
In briefWavelet-based feature extraction for laser spectroscopy classification of human plasma
SummaryThis study applies continuous wavelet transform to extract spectral peak features from laser-induced breakdown spectroscopy of human plasma samples, distinguishing SARS-CoV-2-positive from negative cases. Plasma from 181 individuals was dried on silicon substrates and classified with a support vector machine, reaching 93.00% accuracy in leave-one-out evaluation. The authors note these estimates are provisional and require independent validation.
Md Mahmudur Rahman, Atqiya Munawara Mahi, Nazim Ahmed Belabbaci et al.
In briefA new measure of structural similarity between two-dimensional point sets
SummaryThis paper introduces the Multi-scale Closing Index (MCI), a dissimilarity measure bounded between 0 and 1 that uses multiscale mathematical morphology to assess how well two 2D point sets align, with smaller values indicating greater structural agreement. The authors tested MCI on synthetic point clouds generated from Perlin-noise spatial structures, applying perturbations, point removal, translation, rotation, and impulsive noise. MCI increased consistently with perturbation severity and clearly separated clouds from the same structure from those of different structures. Comparisons with Hausdorff-based measures and Sliced Wasserstein distance suggest MCI is a useful complementary metric for evaluating 2D point-set registration.
Nicol\`o Zava, Angel Franco-Alvarez, Samy Blusseau et al.
In briefZero-shot benchmark of unified AI models for judging and repairing shallow-water drone photos
SummaryThis paper systematically tests unified multimodal large language models on shallow-water drone images, checking whether they can both assess water-induced distortions and restore the images without task-specific training. The benchmark has 1,000 image pairs and evaluates five open-weight models with several prompting strategies. Assessment performance varies widely across models and prompts, restoration gains are small, and some outputs alter scene structure or hallucinate content. The authors conclude that recognizing degradation does not reliably lead to physical restoration and offer a reproducible baseline for future work.
In briefA privacy-aware deep learning benchmark for recognizing facial emotions
SummaryThis paper proposes a lightweight facial emotion recognition pipeline that uses only geometric features, 478 facial landmarks and 52 ARKit blendshape coefficients, avoiding raw video storage. Six neural networks trained from scratch on the fused CREMA-D and RAVDESS datasets were compared with an ImageNet-pretrained ResNet50 baseline under 5-fold subject-independent cross-validation. The geometric models outperformed the pixel baseline on unseen subjects with 12 to 124 times fewer parameters, and performed better on subtle expressions such as sadness.
Larissa Palhares, Maria Machado, Hudson Bruno et al.
In briefCoordinated mixed-precision memory compression for LLM multi-agent systems
SummaryThis paper presents CoMixKV, a method that compresses the key-value caches of multiple LLM agents running at the same time by assigning different bit-widths to each agent and layer under a shared memory budget. It estimates sensitivity online from recent queries, keys, and values, without calibration data or extra model passes. Tested on four benchmarks and three LLM backbones at average 2, 3, and 4 bits, it reports strong accuracy and, on Qwen3-32B, a 32.08% reduction in peak KV-cache memory and a 1.38x speedup over FullKV.
In briefWhat to Delegate, What to Keep: Human Preferences for AI Agents in Daily Life and Work
SummaryThis paper reports a scenario-based survey of 289 U.S. adults with prior AI experience, covering 45 use cases across nine life domains, to examine how people want AI agents to handle tasks on their behalf. Greater trust in the agent, desired AI leadership, prior task-specific AI use, and anticipated financial savings were associated with higher automation preference. Concern that automation would reduce personal fulfillment predicted lower preference. The authors argue that delegation is best understood as a configuration of execution, collaboration, leadership, and authority.
Sarah Wheeler, Veronica Rammouz, Yaoyao Qian et al.
In briefOne Run Is Not Enough: A Monte Carlo Method for Testing How Reliably AI Models Handle Data Science Tasks
SummaryThe paper introduces DS-QA, a benchmark of 100 data science and statistics questions drawn from UK university courses, spanning 16 datasets and three difficulty levels. The authors propose a three-layer evaluation framework covering general response properties, coarse-grained metrics, and task-specific metrics. They test OpenAI models with 100 repeated runs and find substantial variability in responses to data visualisation and supervised learning tasks, especially at medium and hard levels. The study emphasises that consistency should be assessed alongside response quality.
In briefBudget-limited adaptive data buying for robust off-policy evaluation across several logging policies
SummaryThis paper studies off-policy evaluation when extra data can be bought from several existing logging policies, each with a different cost per observation, under a limited budget. It proposes Adaptive Greedy, which sequentially buys data from the affordable logger giving the largest cost-normalized variance reduction across target policies. The authors show the empirical allocation converges almost surely to the optimal design and that the pooled estimator attains the semiparametric efficiency bound. Simulations and a semi-synthetic AIDS Clinical Trials Group Study 175 application show it outperforms standard allocation strategies.
In briefContrastive tuning to reduce stain and scanner bias in pathology foundation models
SummaryThis paper presents SPECTRA, a model-agnostic contrastive post-training framework that encourages histopathology foundation models to learn representations robust to scanner and stain differences. It uses LoRA adapters trained on matched whole-slide images from the PLISM dataset, aligning spatially corresponding regions across acquisition conditions. SPECTRA improved robustness, gene-expression prediction, slide-level classification, and retrieval across models, and overall THUNDER tile-level classification, with exceptions for segmentation and calibration.
In briefTriNetra: A surveillance system that weighs thermal images by reliability for RGB-thermal detection and tracking
SummaryTriNetra is a surveillance framework for combining visible and thermal camera images, which can each fail at night or under saturation or blur. It computes a frame-level Thermal Reliability Coefficient from five label-free image cues and uses it to gate thermal features in a dual-branch YOLOv8m detector and a ByteTrack-based tracker. On the M3FD benchmark, the adaptive gate matches a fixed gate on clean data but performs better when the thermal stream is corrupted.
Sadique Khatib, Aditya Ahirrao, Vedant Sonawane et al.
In briefAssessing ACL Injury Risk with a Graph Convolutional Autoencoder
SummaryThis paper presents an unsupervised learning model built on an autoencoder with multiple graph convolutional network layers, which uses pose estimation data to determine a person's anterior cruciate ligament (ACL) injury risk. The model runs entirely on a smartphone and is applied to two tests, the Drop Vertical Jump (DVJ) and the Single Leg Hop (SLH). The authors report that it accurately identifies athletes at high risk on both tests, supporting the development of a widely accessible smartphone tool for real-time ACL injury risk screening.
Athanasios Babouras, Paul A. Martineau, Thomas Fevens
In briefAccelerating fixed-point iterations for optimization over probability measures on neural Wasserstein manifolds
SummaryThis paper proposes a fixed-point acceleration method for optimizing over probability measures, which are represented as pushforwards of a reference distribution through a neural network. The parameter space is equipped with the induced pullback Wasserstein metric, so natural gradient descent becomes a fixed-point iteration on a Riemannian manifold. The authors accelerate this iteration with Riemannian Anderson mixing, which uses historical residuals instead of Hessians. Tested on multimodal sampling, PDE-based inverse problems, aggregation dynamics, Bayesian logistic regression, and generative modeling, the method reduces natural-gradient iterations and is competitive with SGD and Adam in loss-gradient evaluations and wall-clock time.
In briefFOSCAR: a faster decode-attention method for 2- and 4-bit KV caches on GPUs.
SummaryThis paper presents FOSCAR, a GPU kernel library that speeds up LLM decode attention over 2-bit and 4-bit quantized key-value caches. It uses affine folding to move scale and offset operations onto the other matrix operand or the output, cutting the number of such operations sharply. Kernels are built for H100 and B200 GPUs, and on H100 they achieve geo-mean speedups of 3.36x (INT2) and 2.83x (INT4) over FlashInfer BF16. Integrated into SGLang, it raises peak long-context serving throughput by up to 170% over 2-bit baselines and 341% over 16-bit baselines.
WhyThe authors apply affine folding to move scale and offset out of matrix products, reporting speedups on H100 and B200 versus several baselines and SGLang serving gains.
Donglin Zhuang, Pragaash Ponnusamy, Xiaoxia Wu et al.
In briefEvent-camera spacecraft pose estimation: direct regression versus geometric recovery
SummaryThis paper compares two approaches to estimating the 6-DoF pose of spacecraft from event-camera data. One is a modular pipeline combining learned keypoint detection, geometry-based pose recovery and temporal smoothing; the other directly regresses translation and rotation. Both were evaluated under the same SPARK 2026 hidden-test protocol on real event data from SPADES. Direct regression lowered orientation error, while the modular pipeline lowered relative translation error, showing complementary error profiles.
In briefIntent-aware caching to speed up LLM serving
SummaryLarge language model serving systems cache key-value states from earlier conversation turns so later requests can skip redundant computation, but cache hits depend on whether and when a conversation continues. This paper finds that user intent signals different continuation probabilities and turn delays, and proposes Intent-Aware Caching (IAC), an online eviction policy that classifies requests by intent and uses learned reuse statistics. On real workloads (WildChat and ShareChat), IAC achieves a 34.6% cache hit rate, above a prior learning-based method (18.8%) and LRU (12.2%), and reduces average serving latency by up to 46.1%.
Kazi Hasan Ibn Arif, JinYi Yoon, Dimitrios S. Nikolopoulos et al.
In briefAdds coronary ostium markers and centerline graphs to the ImageCAS heart-scan dataset, plus open localization models.
SummaryThe paper adds coronary ostium annotations for all 1000 patients in the public ImageCAS dataset and builds centerline graphs rooted at these new ostia. Six 3D localization models trained on ImageCAS, CAT08 and ASOCA were benchmarked on 58 held-out volumes. SwinUNETR-V2 and nnResUNet localized every ostium with median errors of 0.70 to 0.87 mm.
Matteo Leccardi, Andrea Moglia, Marco Morandini et al.
In briefAdapting frozen promptable segmenters to custom class labels with few-shot learning
SummaryThis paper presents SNAP, an annotation pipeline that trains a per-project classification head, SnapHead, on top of a frozen SAM2.1 model so that it learns each project's own class boundaries from the labels annotators already draw. On eight look-alike cell lines, SnapHead agrees with the true labels far more closely than SAM 3 does. SNAP also outperforms SAM 3 on five microscopy projects and on five held-out datasets.
Khayrul Islam, Christopher Frederic Miller (Computational Engineering Division, Lawrence Livermore National Laboratory et al.
In briefICare-DT: a digital twin for continuous forecasting of intensive care patients
SummaryThis paper proposes ICare-DT, a Transformer-based digital twin that updates a latent patient state hourly from new measurements without retraining, to forecast vital signs in intensive care patients. Trained on 140,473 eICU-CRD stays, its later forecasts outperformed its own hour-6 forecasts of the same windows in all 12 test settings. The authors also build a benchmark that replaces real treatments with simulated doses and injects known-size responses to test treatment-response prediction.
WhyThe authors propose an hourly-updating Transformer digital twin and a benchmark injecting known treatment effects, validated on eICU and MIMIC-IV.
In briefRidgeFill: Training-free late-layer cache filling to speed up prefill
SummaryRidgeFill is a training-free method that speeds up prompt prefill by computing only the lower layers exactly. A per-layer ridge regression map, fitted once from calibration text, predicts the inputs of the later layers, which then write their caches from that prediction. On Qwen3.8-27B and Qwen3.8-Flash-Next, the paper reports 1.25 to 1.60 times faster time to first token, with small changes on RULER and LongBench v2.
In briefWhere Should Learning Live? Regret Analysis for Text and Parameter Adaptation
SummaryThis paper studies how language-model systems repeatedly choose among prebuilt prompts and parameter adapters during deployment, using feedback from earlier choices. It measures excess cumulative loss against the best fixed prompt-parameter pair and splits it into selection regret within the candidate family, the gap from restricting that family, and the effect of the actual deployment history. Experiments show that fewer switches can either lower or raise deployed loss depending on whether inherited histories help or hurt. On Spider with Qwen3.5-9B, text-only selection has lower within-family regret than joint selection but produces more failures because its candidates exclude stronger configurations.
In briefAttention Sparsity Method That Adapts Each Head's Selection Resolution
SummaryThis paper proposes HASA, a training-free sparse attention framework that adapts selection resolution for each pretrained attention head based on its query-key structure. It uses the singular spectrum of each head's query-key operator to set an effective rank in the feature dimension and merges neighboring context regions by directional coherence along the sequence dimension. According to the paper, HASA captures 88.1% of dense-attention mass, improves accuracy over UNIQUE by up to 3.47 points on RULER and 2.57 points on LongBench-Pro, and reaches up to 14.07x speedup over dense attention.
In briefVerifier Noise as Temperature: Testing Tilt Predictions in LLM-Guided Search
SummaryThis paper studies LLM-guided search loops that propose a modified candidate, score it, and keep it if it scores better. Under specific noise assumptions, the authors derive that the stationary distribution is the mutation equilibrium tilted by an exponential factor, and test this prediction on tasks such as bin packing. The measured gains did not match the predicted values, and runs from different starting points did not reach a common level, so whether the tilt describes the stationary law of LLM editing remains open.
In briefPoseRieFlow: Riemannian Flow Matching for Text-Driven Whole-Body Pose Generation
SummaryPoseRieFlow is a compact flow-matching model that generates whole-body 3D human poses from text descriptions. It works directly on the rotation manifold of SMPL-H joints, moving samples from a Gaussian prior to the data distribution along Riemannian geodesics. A part-tokenized Diffusion Transformer groups the 52 joints into eight body-part tokens, with dedicated tokens for each hand. On PoseScript, it achieves the lowest whole-body kinematic error among methods that produce hands, using 21.6M trainable parameters.
Jan Spindler, Bardia Ardakanian, Patrick Stotko et al.
In briefA deep learning method that removes speckle noise from radar images across different polarization types.
SummarySpeckle noise in synthetic aperture radar (SAR) images complicates interpretation. The paper proposes one deep learning framework that uses a single trained convolutional neural network to despeckle single-, dual-, and full-polarization SAR data. It uses a common nine-channel polarimetric covariance matrix representation and a polarization-aware masking strategy, without changing the network architecture. Comparisons on three scenes from different sensors show competitive performance.
Gianpaolo Passarello, Sergio Vitale, Hossein Aghababei et al.
In briefMETA-IR: A Meta-Learning Framework for Adaptive Safety Alignment in LLMs
SummaryThe paper proposes META-IR, a framework that helps large language models resist jailbreak attacks and harmful prompts by combining introspective reasoning with meta-learning. An inner loop performs structured risk analysis and self-correction using SI-MCTS, while an outer loop optimizes the introspection policy across diverse attack types. Experiments on Llama-3.1-8B-Instruct and Qwen-2-7B-Instruct report improved adaptation speed, robustness, and reasoning stability over baselines.
Mahabubur Rahman Miraj, Hongyu Huang, Nankun Mu et al.
In briefSURE: A Spatiotemporal Uncertainty Method for Data-Efficient Robot Vision-Language-Action Learning
SummaryThis paper presents SURE, a training framework for vision-language-action (VLA) robot control models that explicitly models spatial and temporal uncertainty in actions. It uses a lightweight transposed-convolutional probabilistic action decoder to capture correlations among neighboring action bins, and Soft-DTW alignment supervision to handle timing differences between predicted and demonstrated action sequences. The authors report that SURE matches or surpasses baseline VLAs on LIBERO, RoboTwin 2.0, and real-world manipulation tasks while using only 10% of the demonstration data or 20% of the training steps.
In briefMultiscale Single-Index Model: Mean-Squared-Error Analysis of Online SGD Dynamics
SummaryThis paper studies the Multiscale Single-Index Model, a simplified model of hierarchical feature learning in which each layer extracts a single-index feature at one scale and passes it on. For the case of two hidden layers, it analyzes joint online spherical stochastic gradient descent under mean-squared-error loss. The authors show weak recovery from random initialization with O(d log² d) samples, followed by strong recovery to accuracy ε with O(d log(1/ε)) additional samples.
In briefAutonomous particle image velocimetry with zero-shot sim-to-real reinforcement learning
SummaryThis paper studies how to place the camera in particle image velocimetry, a method that reconstructs flow fields from tracer particle motion. A reinforcement learning agent controls a six-degree-of-freedom manipulator carrying a camera to locate and frame a hydrofoil wake in a water tunnel. The agent is trained entirely in a photorealistic simulation and transferred zero-shot to the real water tunnel. It reliably frames wakes of randomly placed hydrofoils, resolving them at higher resolution and with lower flow-statistics error than a fixed camera view.
Chang Rajani, Luke Sebastien Marsden, Stephan Kunz et al.
In briefSame Pixels, Different Documents: Measuring Part-Level Editability in Vectorized Images
SummaryImage vectorization is usually judged by how faithfully it renders, yet documents that render identically can support different edits. The paper introduces Pass@k, which checks whether a prescribed edit can succeed using objects selectable within k clicks under a specified interaction model. The authors evaluate twelve open and commercial methods on 211 references, with 635 edit tasks on 93 of them, using eight document diagnostics. Parts composed of multiple shapes mostly fail at Pass@1, and higher fidelity does not guarantee higher Pass@1.
In briefGoodhart in the Planner: how robot search can exploit reward models learned from human preferences
SummaryThis paper studies what happens when robots choose actions by searching against a reward model learned from human preferences, using model-predictive control or best-of-N selection. Across several simulated tasks, harder search reduced true return by up to 76%. The authors propose a calibrated χ²-regularized MPPI rule, which they found to be the safest fix among those tested.
WhyThe authors propose a chi-squared regularized MPPI temperature rule, prove worst-case optimality, and report true-return results across several robot tasks.
In briefTransferring rib effects to new sheet-forming shapes with a graph Transformer
SummaryThis paper studies surrogate models for local features such as side-wall ribs in sheet metal forming. The authors propose R-BiLoGo, a baseline-conditioned residual surrogate, which learns rib effects on one global geometry and transfers them to others. Across 22 target geometries, transfer learning often matched or outperformed training from scratch, and its benefit decreased as geometric differences grew.
In briefOnline causal configuration for networked systems using doubly robust steady-state learning
SummaryThe paper proposes CAUSAL-UCB, an algorithm that selects network configurations using optimistic bounds on target causal values obtained during measurement exploration. It uses an interference-aware clipped doubly robust evaluator that issues predecision certificates separating statistical variation from identification, transport, and temporal-dependence errors. The authors prove lower bounds on minimax inference error and measurement-induced deployment loss, and compare the method with certified IPS-UCB in a high-mismatch benchmark.
In briefDual-architecture fusion for vehicle lidar 3D object detection
SummaryThis paper proposes SpconvTransformer, a 3D backbone network that fuses a convolutional branch and a Transformer branch for lidar-based 3D object detection. The convolutional branch uses sparse convolution to learn local voxel features, while the Transformer branch uses self-attention to model long-range relationships between voxels. Combining both feature types addresses the limited receptive field of 3D convolutional networks. The authors report average accuracies of 81.45% for 3D detection and 88.09% for bird's-eye view detection, and an average orientation similarity of 91.10%.
In briefHow Much Chat History Does an AI Agent Need? A Study of Memory Budgets and Scoring
SummaryThe study tests five fixed passage budgets on 1,100 BEAM questions while holding retrieval ranking and the reader model constant. Mean scores rose as the passage budget grew, with smaller gains per extra token at higher budgets. It also compares retaining user messages with token-matched recent messages in a controlled task, and notes that the correctness of BEAM answers remains unresolved after a scoring audit.
In briefVideo2Asset: Building Persistent 3D Street Assets Online from Video
SummaryThis paper presents Video2Asset, an online framework that builds a continually updated 3D street scene from multi-view video, with each object stored as an independently editable asset. It uses occupancy as an intermediate representation and links current observations to accumulated identity, geometric, and appearance evidence to keep object identities consistent. Experiments on several datasets show improved temporal identity consistency and asset quality, including a 15.4% reduction in asset fragmentation on nuScenes with SGE.
In briefDecoupled Mode Connectivity for Generalizing from Base to Novel Classes in Vision-Language Models
SummaryThis paper addresses prompt learning for CLIP, a vision-language model, where specializing to base classes and staying close to zero-shot features conflict when both act on one prompt. The proposed Decoupled Mode Connectivity (DMC) trains a separate prompt for each objective and merges their text features with a fixed coefficient at test time. Across 11 datasets, DMC improves the average harmonic mean of base and novel accuracy over MergeTune for both CoOp and KgCoOp.
Imad Eddine Marouf, Khalid Oublal, Enzo Tartaglione et al.
In briefTensorMutation: Learning Weight-Space Mutations of Deep Learning Models on Neural Graphs
SummaryThis paper presents TensorMutation, which represents trained deep learning models as neural graphs and uses a mutation-conditioned graph neural network to predict parameter changes, generating executable mutants without retraining. The method was evaluated on a corpus of 25,252 original-mutant pairs built from 71 multilayer perceptron classifiers. On 12 real-fault subjects, it achieved the highest median behavioral similarity among the compared approaches, with a statistically significant improvement over DeepMutation++.
In briefA separation principle that splits multi-agent learning into single-agent learning plus optimal transport
SummaryThis paper studies cooperative multi-agent reinforcement learning, where agents follow separate noisy dynamics but share a population-level objective, such as a mobility operator routing a vehicle fleet. The authors prove a separation principle showing the population cost is upper bounded by an optimal transport problem whose cost is a target-conditioned single-agent cost. They introduce SALT, which learns a single-agent policy and coordinates agents via optimal transport, scaling to large fleets and transferring to unseen fleet sizes. In a Manhattan vehicle-delivery experiment, SALT improves travel time by 15-21% over a shortest-path heuristic.
Lucia Pezzetti, Nicolas Lanzetti, Antonio Terpin et al.
In briefNAVE: A Conformer Model for Detecting Antarctic Whale Calls in Noisy Recordings
SummaryThis paper presents NAVE, a Conformer-based detector that finds whale call boundaries in long underwater recordings and classifies them by call type. It combines normalized input features that suppress site-specific ambient sound and adapt to changing noise, frequency-adaptive convolutions, and a convolution kernel matched to call duration. On BioDCASE 2026 Task 2 with leave-one-site-out validation, a single NAVE model outperforms the WhaleVAD-BPN baseline, and an ablation study shows each component helps recover the hard-to-detect D downsweep class.
In briefSteerable multi-objective training of language models using jointly learned low-rank adapters (LambdaMix)
SummaryThe paper introduces LambdaMix, a parameter-efficient method that jointly trains one low-rank adapter per objective on a frozen base model. A preference vector selects the adapter mixture at inference, so preferences can change without further optimization. Experiments on helpfulness and harmlessness and on instruction-following math problems report higher hypervolume than all evaluated baselines, with fewer training completions and stored adapter parameters than GRPO specialist policies.
In briefStoring memory in world model parameters at test time for reinforcement learning
SummaryThis paper introduces Test-Time Model-Based Reinforcement Learning (TT-MBRL), which uses gradient descent at test time to encode an episode's observations into world model parameters. Each episode thus yields its own world model, aiming to improve long-term memory over methods that rely only on context. The method is first validated on a simplified 2D Memory Maze, then achieves new state-of-the-art results on the 3D Memory Maze benchmark. It also maintains strong performance on DeepMind Control and Atari 100k.
In briefBiLoGo: A neural simulator for sheet forming that keeps local thinning sharp
SummaryThis paper presents BiLoGo, a learned simulator for sheet metal forming that combines local node-element message passing with a Transformer-based neural operator for long-range information exchange. It uses three modules: a bidirectional GraphSAGE model, a dual-stream Transformer neural operator, and a gated residual fusion module. Evaluated on a dome-shaped cold forming case and two corner-shaped hot stamping cases, BiLoGo reports the lowest positional and thinning errors compared with the baseline models tested.
In briefCRTrack-Cap: A New Benchmark for Describing Objects Across Multiple Camera Views
SummaryThe paper introduces cross-view dense video object captioning, which links per-camera tracks of the same person into scene-level trajectories and generates one caption for each. It builds the CRTrack-Cap benchmark from CRTrack, containing 425 global trajectories, 838 local tracklets across 13 scenes, and canonical captions for 335 identities. The authors also propose the CV-CHOTA metric and a baseline framework combining View-Exclusive Cross-Camera Association and Conflict-Evidence Gating.
In briefTrace-square-root potential method for contextual search and bilateral trade.
SummaryThe paper presents a unified potential method for contextual search and bilateral trade with one-bit feedback, yielding order-tight algorithms. In the Bayesian setting, contextual search has total loss bounded by O(Tr√Σ), and bilateral trade has regret bounded by O(Tr√Σ_S + Tr√Σ_B). Adapting this potential to the worst-case setting, both problems achieve tight O(√d) regret, improving on previous guarantees.
WhyThe authors introduce a unified trace-square-root potential, obtaining tight O(√d) regret for contextual search and bilateral trade, improving prior bounds.
Romain Cosson, Federico Fusco, Anupam Gupta et al.
In briefKnow What You Need: Efficient Activation Checkpointing — a method to cut memory use when training large language models
SummaryThis paper proposes an approximate rematerialization method for activation checkpointing, a technique that saves memory during LLM training by recomputing discarded activations during backpropagation. The method uses information from the initial forward pass to recompute only the most important layer components, selected greedily or via randomized sketching, aiming for unbiased gradient estimates. On a Llama-488M model, recomputing about two-thirds of layer components increased perplexity by 1.7%, 1.2%, and 2.5% depending on whether feed-forward layers, attention layers, or both were targeted.
In briefLearning Analytic Operators and Parametric PDEs with Kernel Methods
SummaryThis paper studies learning analytic operators from finite samples using kernel methods, examining how the target's regularity should guide kernel choice. It shows that analytic operators embed into vector-valued reproducing kernel Hilbert spaces and gives finite-sample, high-probability guarantees for uncertainty from random sampling. The theory is applied to the solution operator of a parametric elliptic PDE, yielding error bounds that separate sampling and discretization contributions and specify how mesh size should scale with sample size. Numerical experiments illustrate the results.
Giovanni S. Alberti, Markus Holzleitner, Lorenzo Rosasco et al.
In briefNot Your Diary, Just Your Topic: Testing Personalization Controls
SummaryThis paper examines a common personalization test that swaps a user's history with another user's, noting that the swap also changes topics, so accuracy drops may reflect topic mismatch. Using a frozen language model on LaMP-2 news categorization, the authors find the user's own history adds 2.3 points over no history, and category labels alone account for 86% of the gap to a random user's history. They argue personalization should be evaluated on unseen users, with similar-content histories and no-history models as comparisons.
Karina Romanova, Igor Kiselev, Vladimir Platonov et al.
In briefCoresets for capacitated clustering via dual concentration
SummaryThis paper studies coresets, small weighted data summaries that approximately preserve clustering cost, for the k-median and k-means problems with capacity constraints on cluster sizes. For finite metric spaces, the authors obtain optimal-size coresets that match the best known bounds for unconstrained clustering. In Euclidean space they give improved bounds, including the first coresets whose size depends nearly linearly on k and is independent of dimension and input size. The key technical tool is concentration bounds for dual solutions of min-cost flow problems, which simplify and tighten the analysis of the sampling distribution.
WhyThe authors give optimal-size coresets for capacitated clustering on finite metrics, matching unconstrained bounds, and first near-linear-in-k Euclidean bounds via min-cost flow dual concentration.
Milind Prabhu, Chris Schwiegelshohn, Sudarshan Shyam et al.
In briefClassifying kidney tumors from CT scans using regions around the tumor and slice aggregation
SummaryThis paper presents a region-aware model for classifying clear cell kidney cancer from CT images. It pools frozen DINOv3 features separately from the tumor and a 0 to 10 mm peritumoral context region, aggregates slices, applies training-only PCA, and fuses the result with radiomics. On KiTS23 and 111 external cases, internal and external AUCs are 0.8078 and 0.9094.
Yuan Liang, Sourav Bhattacharjee, and Abraham Campbell
In briefTransformer-FSQ: Discrete codes for compressing a model's working memory
SummaryThis paper proposes Transformer-FSQ, which replaces standard key and value projections with a finite scalar quantisation bottleneck so the KV cache stores packed discrete integer codes for each token. On models from 0.6B to 8B parameters, it is the most accurate of the five baseline families evaluated on the RULER long-context retrieval benchmark at compression ratios from 12x to 97.5x. A fused Triton kernel decodes directly from the integer codes, reaching about 1.01x the throughput of native attention on uncompressed Qwen3-8B-Base at 32x compression and about 1.54x at 97.5x.
Andras Szecsenyi, Thomas L. Lee, Vaisakh Shaj et al.
In briefRuACT: Turning Inactive Scoring Rubrics into Adaptive Critics and Teachers for Reinforcement Learning
SummaryThis paper addresses a problem in rubric-based reinforcement learning, where rubrics that every response fails or passes provide no learning signal. The proposed RuACT method uses unreachable rubrics as privileged information for reward-gated self-distillation, and splits saturated rubrics into child rubrics that separate response quality. RuACT outperforms GRPO and other baselines on four medical and scientific benchmarks.
WhySplits inactive rubrics into unreachable and saturated cases, using them as teachers and critics; the authors report gains over GRPO on four medical and scientific benchmarks.
In briefCoordinatewise epsilon settings for Adam under heavy-tailed noise
SummaryThe paper studies how Adam's practice of adding one shared scalar epsilon to every coordinate's denominator affects convergence when noise and curvature differ across coordinates. Under heavy-tailed noise with finite p-th moments, choosing a separate epsilon for each coordinate attains the benchmark rate O(T^{-(p-1)/(3p-2)}) for l1 stationarity. For 1<p<2, Adam with a shared epsilon has a lower bound of order Omega(T^{-(p-1)/(2p)}), which holds for any predictable scalar choice of the step-size and momentum parameters and epsilon.
WhyThe authors propose coordinatewise epsilon choice, matching the optimal time exponent, and prove a lower bound for shared-scalar epsilon Adam.
In briefMSR-ABC: A Method Combining Complementary Features to Classify Kidney Tumours
SummaryThe paper introduces MSR-ABC, a framework that fuses MedVAE, MedSigLIP, and radiomic features to classify clear cell renal cell carcinoma from CT scans. It uses cross-representation attention, auxiliary supervision, and gated residual fusion so the features complement each other. On an internal set of 396 KiTS23 cases and an external set of 111 TCGA/TCIA cases, it achieves mean AUROCs of 0.806 and 0.879.
Yuan Liang, Sourav Bhattacharjee, and Abraham Campbell
In briefHalf-precision (fp16) running of AI weather models to cut compute cost with little loss in accuracy
SummaryThe paper runs five public AI weather prediction models on a single 16 GB GPU with their networks in half precision while keeping the physical state in 32-bit, and verifies forecasts against ERA5. The global RMSE changes by at most 1.4%, and about 0.5% or less for four of the five models. A 10-day forecast in fp16 is 1.3 to 3.5 times faster and uses 1.4 to 3.7 times less energy than in fp32.
In briefStitching videos from several freely moving cameras
SummaryThis paper tackles aligning videos of the same scene recorded by multiple freely moving cameras. The proposed method, MViS, is a lightweight test-time optimization framework that jointly processes frames from all views in a short time window and predicts per-view, per-time homographies. It optimizes in two stages, geometry-guided alignment followed by appearance-based refinement, enforcing multi-view and temporal consistency. MViS matches or surpasses prior methods on a standard two-view benchmark, and the authors introduce an annotated dataset with more than two moving cameras.
Shira Ifergane, Roy Amoyal, Shahar Benishay et al.
In briefAlignment by Confident Debate: Learning Preferences from Multi-Agent Debate
SummaryThis paper introduces Alignment by Confident Debate (AbCD), a method that turns multi-agent debate traces into preference data without gold answer labels. A sampling-diverse committee debates each problem, and final answers are clustered by task-level semantic equivalence. Experiments on verifiable tasks such as math reasoning, multiple-choice quantitative QA, and code generation across several model families show gains in consensus accuracy over inference-only ensembling, debate, and SFT on the same traces.
Nikhil Verma, Dingxian Wang, Mohammad Reza Rezaei et al.
In briefTurtLES: A Large-Scale Benchmark for Turbulent 3D Neural PDE Surrogates
SummaryThis paper introduces TurtLES, a 14 TB dataset of three-dimensional turbulent cylinder wakes in the shear-layer transition regime, computed with high-fidelity implicit large-eddy simulations. It contains 400 trajectories of 400 time steps each, with 3 to 11 million points per frame, covering variations in geometry, Reynolds number, and angle of attack. The authors benchmark state-of-the-art neural operators on autoregressive rollouts and identify failure modes such as temporal instability, loss of small-scale energy, and difficulty learning on large irregular domains.
Jarno Platenburg, Armand Kassa\"i Koupa\"i, Paola Cinnella et al.
In briefSolver-free decision-focused learning using neural optimization layers for edge selection
SummaryThis paper proposes a decision-focused learning framework that avoids solver calls for NP-hard edge-selection problems with uncertain parameters. A neural optimization layer is first trained without solution labels using problem objectives and constraint penalties, then frozen while a prediction model is trained through its continuous output. Experiments on the asymmetric traveling salesperson and endpoint-free orienteering problems report better decision quality than prediction-focused learning, less training time than evaluated decision-focused baselines, and faster deployment than solver-based methods.
In briefAchieving Optimal Regret in Two-Sided Profit Maximization for Trade
SummaryThis paper studies repeated bilateral trade, where an intermediary sets a purchase price for a seller and a sale price for a buyer, with valuations drawn from an unknown joint distribution. The intermediary observes only whether trade occurs, not the valuations themselves. The authors design an algorithm with regret Õ(T^{2/3}) against the best fixed price pair, matching the classical one-dimensional dynamic pricing lower bound up to logarithmic factors. The result shows the one-dimensional regret rate persists in two dimensions.
WhyThe authors report regret Õ(T^{2/3}) for bilateral trade, matching the one-dimensional dynamic pricing lower bound up to log factors.
In briefQuery-guided RoPE adjustment to reduce mid-context information loss in long-context LLMs
SummaryLong-context language models often overlook information in the middle of their input, a problem called "lost-in-the-middle." This paper proposes QGPM, a training-free method that adjusts rotary position embeddings using query relevance computed from the model's own attention heads. Low-relevance tokens are shifted on fast-rotating planes to pull relevant evidence closer to the query. Experiments on NoLiMa and RULER show gains over base models and strong baselines, largest on retrieval tasks.
Peiyu Zhang, Heng Ping, Shinibali Bhattacharyya et al.
