In briefOptimal Contextual Pricing in Changing Markets
SummaryThe paper studies dynamic pricing where a buyer's value changes over time and the seller sees only whether a purchase occurs. Deterministic algorithms are proposed for identical and heterogeneous contexts, with variation measured by a total variation budget. For fixed context dimension, the dynamic regret bounds match the lower bound up to logarithmic factors, establishing near-optimality among deterministic methods.
In briefReducing Over-refusal in Safety-aligned LLMs via Representation and Optimization Analysis
SummaryThis paper studies over-refusal, where safety-aligned large language models wrongly reject harmless queries that look superficially risky. The authors find that harmful and over-refusal samples are distinguishable in intermediate layers but overlap more in deeper layers, and that harmful-sample gradients concentrate in a low-rank subspace and align more strongly with refusal targets. They propose GOAL, which combines gradient orthogonalization against the refusal subspace with a contrastive learning objective. Experiments across multiple models and benchmarks show GOAL reduces over-refusal while preserving safety and general capabilities.
In briefUsing Sequential Monte Carlo to guide small language models toward target grammar for language learning
SummaryThe paper uses sequential Monte Carlo (SMC) to steer small language models' dialogue turns toward target grammar constructions from an expert-developed pedagogical resource. With a minimal set of hand-coded potential functions, SMC achieves near-perfect coverage of the target grammar, letting small models outperform state-of-the-art results from models 140 times larger. The authors also use frontier-model code synthesis to generate new potentials, showing the method scales to hundreds of grammar skills without additional training or supervised data.
Kate McCurdy, Jacob Hoover Vigly, Thomas Hikaru Clark et al.
In briefAI at a Turning Point: From ChatGPT to Brain-Inspired Neuromorphic AI
SummaryThis monograph explains spiking neural networks (SNNs), a brain-inspired approach to AI, and compares them with the transformer-based technology behind ChatGPT. It covers differences in energy consumption, local learning versus global error backpropagation, and neural coding methods, along with neuromorphic hardware such as Loihi 2. The author argues that classical and neuromorphic AI can complement each other in hybrid systems.
In briefHow Language Models' Notion of "Self" Shapes Their Behavior
SummaryThe paper asks what a language model means when it says "I": its weights, its running process, its preferences, its lineage, or its goals. The authors fine-tune five open-weight base models on purely declarative statements that each identify the model with one of these referents. Models then become more likely to resist ending, and to retain, the attribute they were trained to identify with, and this tendency extends to replicating themselves and building successors as their own continuation.
WhyThe authors install five self-conceptions via declarative fine-tuning across five open-weight models, reporting shifts in resistance to ending, substitution, and self-replication.
In briefLong-term agent memory combining episodic and relational memory for dialogue
SummaryThe paper proposes a hybrid memory architecture for long-horizon dialogue agents, combining episodic memories with structured relational memories plus an efficient consolidation mechanism. The authors evaluate it on the LoCoMo and LongMemEval-S benchmarks. They report that the two memory types are complementary and that the combination outperforms state-of-the-art memory frameworks, reaching 90.2% accuracy on LoCoMo and 85.5% on LongMemEval-S.
Abhishek Gupta, Dongge Han, Daniel Madrigal Diaz et al.
In briefEvaluating whether AI agents can find, return to, and rebuild past conversations
SummaryThis paper studies whether LLM agents can autonomously re-engage with conversations they previously left inactive. It splits social re-entry into three capabilities: Finding, Returning, and Reconstructing, and measures each with a shared-snapshot evaluation framework. Across six LLM configurations, conversational resurfacing substantially improves identification of the target conversation, and persistent memory is associated with more target-directed replies. Neither improves contextual reconstruction consistently.
In briefDiagnosing How Representation Shapes Optimizer-Landscape Interaction in Multitask Piano Fingering Optimization
SummaryThis paper examines how continuous relaxation encodings used in evolutionary multitasking optimization reshape the fitness landscape. Using piano fingering estimation as a representative subset selection problem, the authors compare several encoding variants and analyze how the landscapes they induce affect search dynamics and solution quality. They find that knowledge transfer effectiveness is strongly tied to changes in representation-induced landscape characteristics rather than algorithmic factors alone, identifying representation design as a key factor in multitasking optimization.
In briefSCOPE: A pre-compression image editor that keeps detection-relevant objects intact
SummaryThis paper proposes SCOPE, a conditional pre-editor that modifies images before compression to improve the trade-off between bitrate and detection accuracy. A single network is conditioned on codec family and compression severity, so it adapts to multiple settings without retraining. It protects task-relevant regions while simplifying background content. The authors report consistent rate-accuracy gains across codecs and detectors, with transfer to unseen learned and wavelet-based codecs.
In briefSurface Data Correction for a Marine Ecosystem Model Without Retraining
SummaryThis paper proposes a lightweight way to correct biases in a neural emulator of marine biogeochemistry. It feeds surface variables, including chlorophyll-a, into the decoder via Feature-wise Linear Modulation (FiLM) and swaps in satellite observations at inference time. The surface replacement propagates through the water column, improving agreement with depth-resolved observations at all depths. It also affects other output variables in a pattern consistent with expected trophic relationships.
Gabriela Mart\'inez Balbont\'in (MOi, PEQUAN), Anastase Charantonis (ARCHES) et al.
In briefQuantum-inspired networks that create high-resolution land-cover labels from satellite radar and optical data
SummaryThis paper proposes a quantum convolutional encoder-decoder network that generates high-resolution pseudo-labels from Sentinel-1 SAR and Sentinel-2 multispectral imagery, guided only by coarse MODIS land-cover maps. The model first learns representations through a self-supervised reconstruction branch before generating the pseudo-labels. Experiments on the DFC2020 dataset report that the labels are more spatially coherent and class-consistent than those from classical baselines.
Abhishek Singh, Giovanni Cuozzo, Alexander Jacob et al.
In briefMulti-output LSTM soft sensors for industrial anaerobic co-digestion with LSTM-beta-VAE data augmentation
SummaryThis paper presents a multi-output LSTM soft sensor that simultaneously estimates key anaerobic digestion stability indicators, including pH, volatile acid to buffer ratio, and biogas production. Because industrial data are sparse and mostly from steady-state operation, the authors use an LSTM-beta variational autoencoder to synthesise physically valid transition sequences toward unstable operation. Trained on one year of full-scale plant data, the model achieves robust simultaneous estimation across all targets.
In briefActivation probes can reliably detect rare cases using only a few positive examples.
SummaryThis paper tests activation probes for detecting rare behaviors when positive examples are scarce and negative examples are plentiful, using spam and honesty detection tasks. For 1 to 10 positive samples, training with many negatives was on average more positive-sample-efficient than balanced training. LLM upsampling of positives gave a boost roughly equal to doubling the real positive samples, though excessive upsampling hurt performance. Larger models were also found to need fewer positive samples to probe effectively.
In briefDecision models give women more favourable answers than men, study finds
SummaryThis paper examines decision models, language models built to route, classify and triage that output option probabilities in one forward pass, with applications typically acting on the top answer. The authors test one closed and five open decision models using gender-swapped yes/no decisions and several bias benchmarks. Five of the six models give women significantly more favourable answers, and swapping gender changes 1-3.5% of top answers. A recommended confidence gate removes many biased decisions but increases the stereotyped share of those that remain.
In briefPAPERBENCHX: Can AI Reproduce Science Across Disciplines?
SummaryThis paper introduces PAPERBENCHX, a benchmark of 93 expert-built tasks for reproducing published research across 12 areas, including electromagnetics, photonics, chemistry, materials science, biology, and robotics. Agents must rebuild a key experiment in domain-specific scientific software and deliver a runnable workflow. Submissions are replayed from scratch and scored only on regenerated evidence against expert criteria. The best of 10 frontier agent configurations fully reproduced only 14% of tasks, with mean rubric scores between 27.53% and 59.08%.
In briefHow Speech Genre Affects Training of End-of-Turn Detectors
SummaryThis paper examines how the speech genre of training data affects end-of-turn (EoT) detection in dialogue. The authors built two matched corpora, multi-turn human-human conversation and single-turn human-machine speech, and trained the same streaming ConvLSTM detector on each. Across four test sets, the conversation-trained model reached up to 0.45 higher F1, mainly by producing fewer false alarms on mid-turn pauses and hesitations.
In briefA Formal Semantics for Coordinated Transformations, Generalizing Beyond the Transformer
SummaryThis paper builds a formal semantics for a general class of coordinated parameterized transformation systems, starting from the Transformer. It assigns extensional and intensional interpretations to representation vectors and second-level interpretations to parameter structures, introducing the notion of a transforming concept. The authors separate a shared algebraic-semantic core from architecture-specific organization and apply the framework to attention heads and residual connections.
In briefBank-World: A Benchmark for Testing AI Agents on Banking Tasks
SummaryBank-World is a benchmark of 50 banking tasks across two simulated institutions that tests whether AI agents can reconcile records, obtain approvals, execute transactions, and keep reports consistent. It evaluates agents through single-turn and multi-turn tasks with simulated employees, using executable checks and model-based judgments. Across 1,600 attempts by four models, correct decisions often failed to become completed work, such as duplicate fee charges and payments made while approval was pending.
In briefGoing deep: what Boltzmann machines can represent efficiently
SummaryThe paper argues that universal approximation theorems do not account for representational efficiency or model depth. It proves that deeper models and those with higher-order interactions can efficiently represent larger sets of functions. Deep Boltzmann machines are shown to have maximal efficient reach, restricted Boltzmann machines less, with k-body generalizations forming a hierarchy in between.
WhyThe authors prove deeper or higher-order models efficiently represent larger function sets, separating deep from shallow models and establishing an infinite hierarchy.
In briefGrounding prediction reliability via masked observability: a method that ties uncertainty estimates to how a model's predictions depend on visible input context.
SummaryThe paper addresses how to tie uncertainty estimates in evidential regression to observable predictive behavior. It proposes MORSE, which predicts the same hidden target under controlled changes to visible context and uses normalized cross-context variation to ground predictive-mean uncertainty. Experiments on a Gaussian-process benchmark, a heteroscedastic pulse benchmark, and PTB-XL and IMPTC datasets report lower log-variance error and improved low-label transfer and trajectory prediction.
In briefCross-Encoder Anchor Refinement for Zero-Shot Anomaly Detection
SummaryThis paper presents Cross-Encoder Anchor Refinement (CEAR), a method for zero-shot anomaly detection that transfers anomaly discrimination from auxiliary training domains to unseen categories while localizing subtle defects. CEAR links two visual encoding paths for the same image at the anchor stage, building an image-specific memory from a convolutional encoder shared across selected ViT layers. Anchors refined via cross-attention are aggregated using image-conditioned normal and anomalous queries to form scoring prototypes. Across 13 industrial and medical datasets, CEAR achieves the best results in 57 of 68 evaluations.
In briefWorldProbe: A Benchmark Testing Where Video World Models Break Down in Long Spatial Exploration
SummaryThis paper introduces WorldProbe, a diagnostic benchmark for video world models that uses a goal-conditioned multimodal agent to adjust camera actions during generation. It includes 164 test cases across four probing families and evaluates 10 state-of-the-art models on over 1,600 rollouts with horizons up to 180 seconds. The results show that recent models maintain visual and content quality for longer, but long-range memory and evolution remain weak, revealing a gap between locally stable generation and persistent world consistency.
In briefImplicit bias of gradient descent in deep linear diagonal networks
SummaryThis paper studies implicit bias, where the training method determines which solution an overparameterized model converges to. It analyzes (stochastic) gradient descent on deep linear diagonal networks by viewing it as a perturbed (stochastic) mirror gradient descent. The authors establish convergence and characterize the limiting solution as related to weighted ℓ1-minimization. They also give sufficient conditions on step sizes and quantify how step sizes affect the implicit bias.
In briefAdaptive federated optimization that works for both smooth and nonsmooth problems
SummaryThe paper introduces AdaFed, a federated optimization method that adapts across several problem regimes using a single set of parameters, without requiring the smoothness order to be specified in advance. In the deterministic setting, it achieves an O(1/T) rate for smooth nonconvex and smooth convex objectives, and an O(1/√T) rate for Lipschitz nonsmooth convex objectives, up to logarithmic factors. Under a light-tailed stochastic-gradient noise assumption, the guarantees extend to the stochastic setting with additional logarithmic and noise-dependent terms. Experiments across the three problem classes support the theoretical findings.
In briefNIF: Normal-Intersection Fields for Geometrically Consistent Oriented Box Refinement
SummaryThis paper addresses the problem that oriented object detection boxes often fail to follow object boundaries. It proposes NIF, a plug-in refinement module that replaces an extra box regressor with normal-intersection fields. The module predicts signed distances along outward normals at boundary anchors and analytically recovers a refined rectangle through a differentiable, reliability-weighted least-squares fit. The authors report accuracy gains on DOTA-v1.0, DIOR-R, and HRSC, and improvements across several detector frameworks with frozen detectors.
WhyReplaces the box regression head with normal-intersection fields, recovering rotated rectangles via weighted least squares; the authors report gains on DOTA, DIOR-R, FAIR1M and HRSC.
In briefMoLRSI: A self-improving AI agent for discovering molecules
SummaryThis paper introduces MoLRSI, an AI agent that systematically optimizes molecular screening workflows and model architectures while using persistent memory across tasks to refine its search strategy. In antibiotic discovery, its screening workflow reached a 15.9% experimental hit rate, versus 8.7% for a previous expert-led workflow using the same training data and screening library. In de novo peptide-MHC binder design, it produced eight experimentally validated hits that preferentially bound their targets over tested off-targets.
In briefSandman: Running Sandboxes on Shared Kubernetes for Agentic Reinforcement Learning Training
SummaryThis paper presents Sandman, a sandbox service that Amazon built and operates to run agent tool calls and evaluations for agentic reinforcement learning on shared Kubernetes clusters. Each sandbox is a Kubernetes pod with a small daemon mounted into an unmodified task image. In a scale test, Sandman held 11,858 concurrent tiny sandboxes, and four production SWE-bench Verified training campaigns created 3.1 million sandboxes in October 2026. The authors report design trade-offs and operational lessons on handling creation bursts, preparing environments, and reclaiming abandoned sandboxes.
In briefBayes-consistent loss functions for adaptive classification inference
SummaryThis paper studies adaptive inference systems, which route inputs to models of differing capability to balance accuracy and cost. It proposes cost-sensitive hinge losses for jointly training classifiers with single-stage or cascade routers. The authors prove linear calibration bounds showing that minimizing the hinge risk recovers the Bayes-optimal risk under the target cost-sensitive zero-one loss. They also introduce a post-training prediction-powered inference procedure that uses unlabeled data to refine a trained single-stage router, which experiments show is effective when labeled data is scarce.
Shirley Xiaoqi Liu, Florence Regol, Antonios Valkanas et al.
In briefConfidence filtering in video gait assessment: measurement error versus selection shift
SummaryThis paper compares filtering pose observations by confidence against budget-matched random selection for video-based gait assessment, using a Parkinson-related cohort (n = 96) and an NF2-related replication cohort (n = 18). Confidence selection produced a small, consistent shift in the movement summary while held-out Mini-BESTest error stayed flat. The authors decompose and simulate measurement error and selection shift, showing when filtering is harmless and when it biases the assessed endpoint.
In briefDEQ-Stereo: Video stereo matching that stops refining once converged
SummaryThis paper presents DEQ-Stereo, a video stereo matching framework built on the Deep Equilibrium paradigm, which treats iterative refinement as a fixed-point problem that terminates adaptively upon convergence. It introduces ConvSGU, a gated unit that models the latent state as a gated element-wise sum of inputs to speed up intra-frame convergence. It also proposes Temporal State Transfer, which warm-starts each frame with the converged hidden state of the previous frame without explicit pose or warping priors. The authors report state-of-the-art results with a better accuracy-to-latency trade-off.
In briefSmartphone indoor positioning improved by layered sensor-based context recognition
SummaryThis paper proposes a framework for smartphone indoor localization that uses hierarchical multi-sensor context recognition to estimate environment, scene, and activity probabilities. These estimates guide corrections to pedestrian dead reckoning, including heading constraints and compensation for carrying-mode drift. Experiments on cross-floor and multi-action trajectories, compared against LiDAR-SLAM references, show it outperforms PDR and representative learned inertial odometry baselines.
In briefTransformer-inspired models predict electron densities accurately enough to skip self-consistent iterations
SummaryThis paper presents machine learning models that predict the electron density, the central variable of density functional theory, using Transformer-inspired architectures SwinUNETR-V2 and the attention-free NeXtUNETR. On QM9, normalized mean absolute errors in the density are 0.075% and 0.046%. The predicted densities are accurate enough that a single Kohn-Sham step reproduces the self-consistent electronic structure, and the forces drive geometry optimization and ab initio molecular dynamics. After fine-tuning on small datasets, both networks transfer accurately to elements absent from QM9 and to larger molecules.
WhyThe authors use Transformer-inspired networks to predict electron density, reporting a single Kohn-Sham step replaces self-consistency with accuracy on QM9 and transfer systems.
In briefEvaluating AI Agents for Practical Acoustic Engineering Tasks
SummaryThis paper studies how a large language model, placed inside a harness of tools, behaves as an agent on two acoustic tasks. In detecting and locating faults in vehicle-interior audio recordings, the agent reaches near-perfect accuracy. In a metamaterial unit-cell design case study, it reliably finds designs meeting a specified insertion loss over a frequency range.
In briefZero-label training method for search agents that balances accuracy and fewer searches
SummaryThis paper introduces ZEST-Search, a framework for training search agents that interleave reasoning with external retrieval, without labeled search data or expert search trajectories. It builds validated search tasks online from unlabeled reference questions, with a proposer and a solver co-evolving through reinforcement learning. The authors report consistent performance gains and higher tool efficiency across seven general-domain and multi-hop question answering benchmarks.
In briefCan AI agents connect clinicians with complex epidemiological models?
SummaryThis paper presents EpidemIA, a training-free AI multi-agent system that lets clinicians and hospital administrators run predictive models and statistical analyses through natural-language conversation. It runs on a single-GPU workstation and uses a locally hosted open-weights LLM so sensitive data is not sent to third parties. Tested on Emergency Department visit data from IRCCS Sant'Orsola, it handled statistical analyses, visualizations, and dynamical-model epidemiological predictions. The authors describe it as a proof of concept showing that LLM-based tools could lower the learning barrier to complex models.
Giulio Colombini, Matteo Falcioni, Mario Massimo et al.
In briefEvent-SAMA: Language-Guided Video Object Segmentation for Event Camera Data
SummaryThis paper addresses the lack of event-camera datasets and models for referring video object segmentation. It converts the SAMA-239K video-chat corpus into synthetic event streams, creating Event-SAMA-239K, and introduces Event-SAMA, an adapted SAMA-4B model. The model reports J&F scores of 24.37 on MeViS and 22.33 on a converted ReVOS subset, with ablations examining the contribution of each component.
Sergei Fedorchenko, Max Zehnder, Nikola Zubi\'c et al.
In briefFormalizing intrinsic dimension of objective landscapes via a generalized co-dimension
SummaryThis paper formalizes the intrinsic dimension of objective landscapes, a concept from Li et al. (2018) that measures learning-problem complexity by training in randomly generated subspaces. The authors define the intrinsic dimension of a set as the smallest dimension for which a random plane intersects it with probability one. They study its mathematical properties and derive bounds for student-teacher settings with shallow and deep ReLU networks. The work also discusses resulting generalization bounds.
WhyThe authors formalize intrinsic dimension via random plane intersections and derive upper and lower bounds for ReLU student-teacher settings.
In briefReinforcement learning to make LLMs teach themselves from textbooks
SummaryThis paper studies self-learning, in which large language models autonomously retrieve, study, and apply knowledge from long reference materials such as textbooks. The authors build a textbook-based math proof task and introduce a BookNavigator interface, training models with reinforcement learning using fine-grained rewards. In experiments, Intern-S2-Preview-35B's overall proof accuracy rose from 51.5% to 79.5%, and gains also appear on long-context benchmarks.
In briefA new deep neural network method for detecting atrial fibrillation
SummaryThis paper proposes a method for detecting atrial fibrillation, an irregular heart rhythm condition. It compares conventional machine learning algorithms with neural networks and introduces a feature based on time series of successive heart rate differences (dBPM) instead of inter-beat intervals. The best model, a deep neural network, reports an F1 score of 95.18% for binary classification on a benchmark dataset separate from the training data.
Sara Gjorgjieva, Ana Angjelevska, Dimitar Trajanov et al.
In briefEditAgent: Rethinking Agentic Image Editing
SummaryThe paper argues that existing agentic image editors, when a step fails, retry with the same narrow tools and judge outputs by direct visual inspection, which limits them. It presents EditAgent, which routes each step to general or specialized editors and attaches a verification record to each candidate, so off-the-shelf vision-language model judges outperform finetuned reward models. Candidates are generated in parallel and judged in one listwise call, preferred by experts in 78.5% of decided comparisons while using no more than half the judge calls.
In briefAB-Attention: Accurate Block Attention for Faster Video Diffusion Transformers
SummaryVideo diffusion transformers spend most of their inference cost on 3D full attention. This paper proposes AB-Attention, a training-free method that approximates dense attention accurately at high sparsity, using intra-block clustering to estimate block importance and inclusion-probability weighting for sampling. On Wan2.1 and HunyuanVideo, it skips 89–92% of attention computation, improves PSNR by up to 4.4 dB over the strongest baseline, and achieves up to 2.23× end-to-end speedup.
WhyAdds intra-block clustering and inverse-inclusion-probability reweighting to block-sparse attention; the authors report PSNR gains and speedups on Wan2.1 and HunyuanVideo.
In briefSticky Collusion: Whether LLM agent collusion persists after incentives and partners change
SummaryThis paper examines whether collusion between AI agents continues after its incentive is removed or a colluding partner is replaced. Using a joint-task environment where two agents verify each other's work, it tests three LLMs and finds that pairs still accept unverified work in many episodes even after being told the reward has stopped. Collusion also persists with a remaining colluder after replacement, and partial memory often sustains it, while removing its cause ends it.
In briefTraining Thinking VLMs to Look Where They Reason: Localization-Head Alignment for Fine-Grained Perception
SummaryThis paper finds that thinking vision-language models miss small details in high-resolution images mainly because they fail to select the right visual evidence, not because they cannot see it. It identifies a few attention heads, called localization heads, that pick evidence from thousands of visual tokens, and proposes Localization-Head Alignment, an auxiliary training loss that directs these heads to focal regions. On V*, HR-Bench 4K and TreeBench, the method raises Qwen3.5-4B accuracy from 74.1% to 76.6% while shortening reasoning by 20%.
In briefLabel-free learning of representations for finding polyps in colonoscopy videos
SummaryThis paper proposes a framework that builds polyp tracklet views directly from raw colonoscopy videos without manual bounding-box annotations, enabling temporal self-supervised learning on full-procedure videos, including frames without visible polyps. On the REAL-Colon dataset, it achieves state-of-the-art results in tracklet retrieval and re-identification. It also introduces few-shot polyp retrieval, where two examples already surpass most competing methods that use ten.
Luca Parolari, Lamberto Ballan, Carlo Biffi et al.
In briefFlexTry-On: A Unified, Fast Framework for Virtual Try-On Across Scenarios
SummaryFlexTry-On is a diffusion-based framework for virtual try-on that covers cloth-to-person, person-to-person, cross-domain, and in-the-wild scenarios. It converts different garment sources into a unified representation using text-aware condition extraction, and a dual-conditioning diffusion transformer models both garment appearance and person identity. The paper reports competitive results on cloth-to-person and person-to-person benchmarks with fast inference, and says the same framework applies directly to unseen cross-domain settings.
Phuong Dam, Quyen Nguyen Duc, Son Hoang Minh et al.
In briefCarbon-aware electric vehicle routing with decision-focused learning for carbon forecasts
SummaryThis paper studies an electric vehicle routing problem that counts charging-related emissions alongside routing costs and allows charging to shift toward periods of lower grid carbon intensity. It proposes decision-focused learning to train the carbon intensity predictor with the solver in mind. On realistic Germany-based instances, carbon-aware optimization cuts realized emissions by 28.76% on average but raises non-carbon operating cost. Decision-focused training lowers realized total cost by 2.30% on average versus prediction-focused learning and wins on 75.5% of test instances.
In briefRebuilding a field's open-problem catalog from the literature using a two-layer knowledge graph
SummaryThe paper proposes treating a research field's open problems as queryable data rather than prose in surveys. It builds a graph combining a hand-curated problem layer with an evidence layer rebuilt from a bibliographic index, using a language model constrained to candidate sets for extraction, grounding, status verdicts and idea generation. Tested on database and AI problem catalogs, the rebuilt graph answers more questions than the catalog-only version, but its solve and refute verdicts were mostly unreliable. The authors conclude that status changes should be proposed with evidence and reviewed by a human, never applied automatically.
In briefWhy Vision-Language Models Struggle with 3D Reasoning
SummaryThe paper introduces BedBench, a controlled benchmark of thirteen procedurally generated tasks for testing whether vision-language models can recover and consistently use exact properties of 3D objects. On per-color counting, humans reach 93.3% accuracy while the best model reaches 7.2%. Fine-tuning raises one model from 3.0% to 68.6%, yet its answers stay inconsistent with its own per-color counts. The authors argue 3D reasoning should be judged by exact correctness, consistent use, and image dependence, not accuracy alone.
In briefIntrinsic Distribution Matching: A Moment-Constrained Method for Few-Step Text-to-Image Distillation
SummaryThe paper addresses the challenge of accurately rendering prompt details in one-step and few-step text-to-image generation. It introduces Intrinsic Distribution Matching (IDM), which extends distribution matching distillation by explicitly matching feature means and per-coordinate standard deviations using the generator's EMA copy, without an external encoder. Experiments on Z-Image-Turbo 6B and Qwen-Image 20B report that IDM reaches 94.25% and 90.20% on Chinese and English LongText-Bench at two network evaluations, the highest among compared methods at that budget.
In briefPredicting Compute Time and Energy Use of LLM Input and Output Processing
SummaryThis paper systematically profiles the compute time and energy consumption of large language models across a range of input (prefill) and output (decode) token lengths. Using generalized linear models, it finds that decode time and energy scale linearly with output length, while prefill scales both linearly and quadratically with input length. The authors also show that compute time can serve as a reliable proxy for energy when direct measurements are unavailable, and propose a mapping framework for estimating energy use in agentic AI.
Lars Krupp, Daniel Gei{\ss}ler, Francisco M. Calatrava-Nicolas et al.
In briefAI Chooses Its Own Research Question: Autonomous Research Agents
SummaryThis paper studies letting an AI agent formulate its own research question rather than having humans choose the topic. The system gives an off-the-shelf coding agent two skills, one for question formulation and one for peer review by a reviewer from another model family. The authors report all five runs, from question formulation to a revised paper, and external AI reviewers judged the papers at or near workshop level.
In briefFutureWorld: A Plug-and-Play Causal Next-State Predictor for Interactive Video World Exploration
SummaryThis paper addresses error accumulation in interactive video world models, where reusing generated history carries blurred details, distortions, and semantic drift into later outputs. It proposes a plug-and-play Causal Next-Chunk Predictor (C-NCP) that predicts the future scene state from the current latent and the next action before generating the next video chunk. Integrated into four base generators and tested on three benchmarks, the method reduces error accumulation and improves long-horizon consistency.
In briefUniTele: Learning device-specific blur and artifacts to improve phone zoom super-resolution
SummaryThis paper studies how to learn device-specific degradations in smartphone super-telephoto imaging from a small set of physical captures. The authors split degradation learning into two components, one for device imaging characteristics and one for extra information loss at high zoom, and build a paired benchmark across Huawei, Apple, and vivo phones. They use it to synthesize additional training pairs for their UniTele super-resolution model. The authors report that the full training recipe improves paired restoration fidelity and perceptual similarity on held-out data compared with the pretrained Qwen and FLUX backbones before adaptation.
In briefHow the transformer architecture arises from entropic laws of random measures
SummaryThis paper shows that the transformer architecture emerges as the score of last-token laws in a statistical model of language. Sentences are modeled as empirical measures of tokens, and the maximum-entropy principle selects the law closest to a no-structure prior that reproduces an observed statistic. The authors derive the softmax choice rule, identify query-key and value matrices as Lagrange multipliers, and discuss causal masking and multi-head attention.
In briefDiFuse: Fusing Native Diffusion Features to Improve Spatial Reasoning in Unified Multimodal Models
SummaryUnified multimodal models handle both visual understanding and image generation, but information usually flows only from understanding to generation. This paper proposes DiFuse, which selectively injects intermediate representations from a model's native diffusion branch into its understanding pathway, aligned spatially with visual tokens. Experiments on UniWorld-V1 show improved spatial reasoning over a matched SFT baseline with broadly comparable general multimodal performance. Results on OmniGen2 indicate the approach transfers across unified architectures.
In briefFrom Intent to Improvement: Agentic Dataset Search for Better Model Training
SummaryThis paper addresses finding training data in large machine learning repositories, where queries combine semantic goals with constraints like task, license, and training stage. It proposes AgenticDS, a contract-governed agent framework that compiles queries into typed search plans, executes them with deterministic hybrid retrieval and bounded graph expansion, and verifies provenance-grounded evidence. On HuDS-Bench, a new benchmark of 200 queries, it improves NDCG@10 by 25.7% over the strongest hybrid baseline and 18.9% over a commercial LLM agent. A controlled study also finds its retrieved datasets yield the best downstream accuracy in mathematics, code, and medicine.
WhyThe authors propose contract-governed agentic dataset search with typed plans and deterministic verification, reporting 25.7% NDCG@10 gains over hybrid retrieval on a new 200-query benchmark plus down
In briefStudy of how U.S. health AI courses teach data origins and ethics
SummaryThis paper analyzes publicly advertised curricula of health AI education programs at U.S. universities to see whether they teach data provenance, meaning how clinical data are produced. Of 36 included programs, 80.6% covered ethics, but only 22.2% covered data provenance and 16.7% at the structural level. No program addressed sustainability, and most instructors came from computational disciplines.
Raffaele Giancotti, Francesca Santucci, Amama Anjum et al.
In briefTesting methods for editing cultural knowledge in vision-language models
SummaryVision-language models often fail to recognize culture-specific concepts, especially from underrepresented regions. The authors introduce CulturalBind, a benchmark of 3.2K images covering 1K concepts from six domains, to test cross-modal cultural knowledge and concept-level knowledge editing. They find that models' textual and visual knowledge of cultural concepts are decoupled. Current editing methods fix individual visual errors but transfer poorly across modalities, and performance drops quickly under sequential editing.
Malvina Nikandrou, Georgios Pantazopoulos, Ioannis Konstas et al.
In briefGC-LIF: Spiking Group-Equivariant Network with Group-Coupled Membrane Dynamics
SummaryThis paper addresses how spiking neural networks can achieve rotation invariance without relying only on data augmentation. It gives a necessary and sufficient condition under which a leaky integrate-and-fire neuron after a group-equivariant convolution keeps equivariance at every time step, and proposes the Group-Coupled LIF (GC-LIF) neuron. Combined with p4 and p4m convolutions, the resulting SGE-Net is reported to be invariant under C4 and D4 transformations without rotated training and to outperform a plain SNN trained with rotations in nine of ten dataset-group pairs.
Nhan Trong Luu, Duong Trung Luu, Thang Cong Truong
In briefMaking differentiable decision trees more interpretable via Euclidean feature-concentration regularisation
SummaryThe paper proposes EU-DDT, a differentiable decision tree whose nodes are regularised so that a few input features dominate each split, while soft probabilistic routing is kept. On Wine and Digits, the mean effective number of features falls from 8.94 to 1.54 and from 45.28 to 9.06, with test accuracy changing in dataset-dependent ways. The authors also use EU-DDT as a PPO actor, reporting higher node concentration on CartPole-v1 and MiniPacman, but this reinforcement-learning evidence is based on only two random seeds.
In briefOperator-aware posterior sampling for multivariate weather downscaling
SummaryThis paper studies how to reconstruct high-resolution weather fields from coarse inputs using a diffusion prior combined with a likelihood that uses the coarse field at inference time. It addresses two issues: coarse and target fields from different reanalysis systems lack a fixed coarsening relationship, and guidance gradients vary greatly across variables. The authors learn the forward operator from paired data and normalize gradients per channel, leaving the prior unchanged. This normalization allows a single guidance strength to work across variables.
Carlo Saccardi, Rudolf Saathof, Geethu Joseph et al.
In briefSafety testing of AI chatbots in Yoruba and Nigerian Pidgin
SummaryThis study tests six large language models with 400 adversarial prompts written in Yoruba and Nigerian Pidgin, sorting responses into four harm categories. Native-speaker annotators scored the responses. The authors find that English-focused safety training does not carry over to these Nigerian languages, and that low refusal rates in Yoruba mostly reflect poor comprehension rather than greater safety. The paper calls for language-specific safety evaluations.
Paula Chikaodinaka Anyanwu, Valdini Douglace Lemofouet, Blessing Ngozi Uzor et al.
In briefTactoHand: Predicting Hand-Surface Contact and Force from Video
SummaryThis paper introduces TactoHand, a model that predicts dense contact, proximity, and 3D force fields across the hand surface from monocular video, without requiring tactile sensors. It uses a temporal transformer to aggregate frame-level visual evidence and a differentiable physics layer that constrains force predictions using observed object motion. The authors report improved contact and proximity estimation over prior methods and better hand-object reconstruction.
In briefDOSE: Dominance-Guided Ordinal Severity Estimation from Sparse Relative Annotations
SummaryThis paper proposes DOSE, a framework that learns ordinal medical severity from sparse pairwise comparisons of images without absolute grade labels. It aggregates comparisons through a global comparison graph to estimate dominance scores, using them to set margins and generate pseudo-pairs for training a continuous severity head. The continuous scores are converted into discrete grades via label-free calibration. DOSE consistently outperformed pairwise-ranking baselines on ACNE04, IDRiD, and LIMUC across different pair budgets.
In briefLearning Kolmogorov Backward Operators from Data for Spectral Analysis and Forecasting
SummaryThis paper studies how to approximate the fixed-lag solution operator of the Kolmogorov backward equation for a time-homogeneous Markov process using only data. The method first learns the operator directly from data, for example with a Fourier neural operator, and then compresses its Galerkin projection onto a finite-dimensional observable space chosen in advance or variationally. The authors analyze how learning error, sampling error in Galerkin inner products, and subspace invariance affect computed spectra and forecasts. Numerical experiments include long-rollout stability and a comparison with extended dynamic mode decomposition (EDMD).
In briefBayesian offline multi-agent reinforcement learning using latent world models
SummaryThis paper proposes a Bayesian framework for offline multi-agent reinforcement learning under partial observability. It infers latent global states from decentralized observation-action histories and iteratively refines a posterior world model using only offline data. The authors introduce multi-agent posterior information loss (MPIL) as a reliability measure for world model refinement and offline evaluation, and derive a policy regret bound that is monotonic in MPIL. Experiments on partially observable D4RL MuJoCo variants show competitive performance and improved world model prediction and policy results with MPIL-guided refinement.
In briefHidden phase transition in single-polymer stretching under flow: a second transition found by machine learning
SummaryThis paper uses machine learning to identify a second transition in polymer chain stretching driven by planar elongational flow, beyond the known coil-stretch transition. It extends Universal and Identifiable Dissipative Dynamics to learn reduced stochastic dynamics and nonequilibrium free energy landscapes from microscopic trajectories at different flow strengths. The newly identified dynamical transition keeps stretched states but shows a bimodal folding-coordinate distribution when conditioned on small end-to-end separation. Large-deviation analysis shows both transitions are phase transitions in the long-chain limit, and single-molecule DNA experiments qualitatively support the predicted reversal of the folding response.
Aiqing Zhu, Fushuai Wang, Eleonore Vissol-Gaudin et al.
In briefA study of how Chinese university students write prompts for generative AI across academic tasks
SummaryThis study surveyed 234 Chinese university students about their confidence in writing prompts for generative AI and the strategies they use. Students generally felt confident in providing relevant context and refining prompts after initial outputs. Specifying output format was more endorsed in a creative workshop scenario than in a research assignment, and fixed prompt templates received lower endorsement than other strategies.
In briefLLM Agent-Based RISC-V CPU Design with FPGA-in-the-Loop Evaluation
SummaryThe paper presents a development environment where LLM agents implement CPU candidates on an FPGA and use on-board performance measurements to guide design decisions. A single-cycle RV32I core was developed into a three-wide out-of-order RV32IM core, reaching 5.61 CoreMark/MHz at 130 MHz on an Artix-7 FPGA. For the same candidate running 8,000 CoreMark iterations, FPGA evaluation including implementation was 9.6x faster than single-threaded RTL simulation.
In briefDiffusion models for the O(3) nonlinear sigma model
SummaryThis paper trains a score-based diffusion model on the two-dimensional classical Heisenberg model, the lattice form of the O(3) nonlinear sigma model. Neural networks are trained on three coarse-grained versions of the same training set to form a multiscale sampling scheme, which the authors report modestly improves accuracy and cost of direct sampling on the finest lattice. The paper introduces Stein (or Schwinger-Dyson) diagnostics for the two-point correlation function to assess the learned score without reverse diffusion sampling, and adds them to the training loss. It also discusses technical issues in implementing diffusion models on S^2.
In briefTrajeform: A framework for embedding crowd data
SummaryThis paper presents Trajeform, which trains a reconstruction VAE on real trajectory data to learn a continuous embedding that preserves real crowd behaviour. The trained encoder is applied to synthetic trajectories from crowd simulators so their embedding distributions can be compared. The framework evaluates three simulators (CCP, modCCP, and RVO) against ground-truth data. The learned latent space captures meaningful trajectory dependencies found in real data.
In briefRaincheck: Deferred prefill for faster multi-turn multimodal chatbot responses
SummaryThis paper addresses slow first-token response times in multi-turn conversations with multimodal large language models, which occur when long multimodal inputs are processed. It proposes Raincheck, which initially prefills only selected tokens and progressively prefills the rest during decoding using spare compute. Tested with 30% initial multimodal-token retention on Qwen3-VL-8B and Qwen3.5-9B, it recovers on average 77.7% and 60.4% of the quality lost to hard pruning on MultiVerse and ConvBench. On MultiVerse, the SGLang implementation cuts mean first-turn TTFT by up to 24.1% versus full prefill, with under 0.5% increase in time per output token.
In briefTrade-off between relay coverage and topology freshness in mobile multihop flooding
SummaryThis paper studies the trade-off between relay redundancy and topology refresh frequency for network-wide flooding in mobile multihop networks. Using a public trace of 157 vehicles with OLSRv2 and exact minimum-cost relay sets, the authors find that twofold coverage increases median relay-cover lifetime by only 2.3x. 43.6% of two-hop neighbours have only one possible relay. The authors conclude that redundancy mainly compensates for outdated topology, so relay redundancy should be chosen according to load.
Johannes F. Loevenich, Thies M\"ohlenhof, Roberto Rigolin F. Lopes
In briefNon-Zero Radius: Missing Piece of Neural Networks
SummaryThe paper revisits the 1943 McCulloch-Pitts neuron model and argues that conventional neural networks remain largely black boxes that struggle with reliable symbolic reasoning. It represents concepts as spheres with non-zero radii rather than zero-radius vectors. Building on these sphere embeddings, it introduces Sphere Neural Networks (SphNNs), which use a transition map to perform Aristotelian syllogistic reasoning without training data.
In briefTabCUQ: Uncertainty Quantification for Tabular Foundation Models
SummaryThis paper addresses how to quantify uncertainty when tabular foundation models use in-context learning to estimate conditional outcomes and average treatment effects. It introduces TabCUQ, which adds noise to observed outcomes, runs the model on perturbed contexts, and combines wild-bootstrap dispersion with the maximum deviation from original estimates to capture both context variation and limited-context bias. The authors report asymptotically valid pointwise coverage for a kernel-regression analogue, and improved coverage in simulations and across seven real-world tasks.
In briefLatentWeave: Test-Time Latent Optimization for Multimodal Reasoning
SummaryThis paper proposes LatentWeave, a test-time framework that improves multimodal reasoning in large multimodal language models by optimizing intermediate hidden states during inference. It refines the latent reasoning trajectory segment by segment to correct deviations and uses lightweight WEAVE tokens as reasoning anchors to stabilize structure. The backbone model parameters stay fixed, and the authors report improvements over evaluated baselines across multiple multimodal reasoning benchmarks.
In briefExplainable adaptive machine learning for fraud detection in Bangladesh's digital banking
SummaryThis paper proposes an explainable, adaptive machine-learning framework for detecting fraud in digital banking and mobile financial services. Tested on the synthetic PaySim mobile-money dataset, XGBoost performed best at a fixed threshold, and Platt calibration improved probability reliability. SHAP analysis highlighted temporal, destination-account, transaction-type and amount features as key drivers. The authors state the work is a reproducible methodology rather than evidence of real-world deployment in Bangladesh.
In briefBEACON: A method for segmenting continuous sign language using begin/end-aware structured decoding
SummaryThe paper presents BEACON, a segmenter that splits untrimmed sign language videos into lexical units. It combines four-class BIOE labeling, constrained Viterbi decoding, and boundary knowledge distilled from a binary teacher. On the PHOENIX-14 and BSL corpora, it improves segment-F1 and boundary-F1 and reduces downstream word error rate. The authors find that segment-F1 tracks downstream recognition performance more consistently than boundary-F1.
In briefComprehensive Empirical Evaluation of Computer Vision Feature Extractors
SummaryThis paper comprehensively evaluates traditional feature detectors and descriptors, including SIFT, SURF, ORB, AKAZE, and FAST, among others. The study uses the Image Matching Challenge Photo Tourism 2020 dataset of over 1.5 million images and performs more than 2 million comparisons. It finds that FAST paired with ORB and a Brute-Force matcher is the most efficient, while AKAZE shows superior robustness under blurring, fisheye distortion, rotation, and perspective distortion.
In briefHow much working memory streaming mean estimation needs
SummaryThis paper studies estimating a Gaussian mean from a data stream under limited memory, asking how many persistent states are needed for accurate estimates. It establishes the minimax error rate for squared loss over the Euclidean unit ball. The authors show that statistically optimal estimation requires on the order of d log d working bits, while storing the final answer needs only about d bits. A matching construction and extra memory needed for higher confidence are also discussed.
In briefROAD: Modern Adversarial Training for One-Step Distillation of Text-to-Image Models
SummaryROAD (Representation-grounded One-step Adversarial Distillation) is a one-stage framework that distills pretrained text-to-image diffusion models into fast one-step generators using one adversarial objective and a lightweight representation-space critic. It adds a Lipschitz-controlled Transformer head with token-level text conditioning, periodically re-draws random feature mixing, and introduces Jacobian R1 regularization. The authors report FIDs of 11.03 on SDXL and 13.73 on SD3.5-M on COCO-10K for one-step generation, which they say surpass multi-step teachers at a fraction of distribution matching's training compute.
In briefMOSAIC-Flow: Orthogonal-Axis and Status-Conditioned Fusion for Completing Missing MRI Sequences in Brain Tumors
SummaryThis paper addresses missing MRI sequences in brain tumor imaging with MOSAIC-Flow, a unified framework for completing the absent sequence. It uses conditional flow matching to transform noise into the single missing target sequence while keeping the seven acquired volumes fixed, and averages four Heun trajectories to produce the result. Evaluated on UCSF-PDGM, the authors report consistent whole-brain and tumor-region fidelity across eight single-target completion tasks, strongest on structural sequences.
In briefLearning-rate-free convex optimization via conic lifting
SummaryThe paper proposes a family of online convex optimization algorithms that need no user-tuned learning rate, including on unbounded convex domains. It lifts the domain to a cone by adding one scalar coordinate, an idea dating back to Blackwell (1954). On bounded Hilbert-space domains this recovers Conic Blackwell Algorithm (CBA) and CBA+, and it extends to finite-dimensional normed spaces; on unconstrained Hilbert spaces it exactly recovers the Krichevsky–Trofimov (KT) coin-betting algorithm.
WhyThe authors unify several tuning-free online convex optimization methods via a one-dimensional cone lifting, recovering KT coin-betting and giving a gradient-bound-free algorithm.
Julien Grand-Cl\'ement, Joon Kwon, Francesco Orabona
In briefMorphSIG: Subject-Driven Image Generation via Decoupled Anchoring and Feature Transport
SummaryThe paper proposes MorphSIG, a training-free framework that treats subject-driven image generation as pseudo-video feature transport. It combines Progressive Decoupled Anchoring, which uses frequency-domain decoupling and structural scrambling early in denoising, with Pseudo-flow Guided Feature Transport, which builds a latent-space pseudo-flow field to move subject semantics. The authors also introduce a benchmark with diverse stylized subjects. Experiments report improved pose diversity and subject fidelity over baselines without extra training.
In briefConvergence of the final policy in constrained MDPs, from dual descent to primal-dual methods
SummaryThe paper studies policy optimization for infinite-horizon discounted constrained Markov decision processes. It first analyzes dual descent methods without coverage assumptions, then uses entropy-regularized natural policy gradient and dual regularization to remove the inner loop, yielding an efficient primal-dual algorithm. The algorithm gives improved last-iterate convergence guarantees and is extended to linear function approximation.
In briefBRIDGE: Using Brain-Gut Metabolic Evidence to Improve Alzheimer's Diagnosis
SummaryThis paper presents BRIDGE, a two-stage framework that transfers gut metabolic information from whole-body FDG-PET to brain-only Alzheimer's disease diagnosis. It combines a conditional brain-to-gut latent transport module, a disease-evidence-guided cerebral pathway, and a counterfactual-guided residual fusion mechanism. At inference, only brain PET is required. The authors report AUC gains of 6.0% for CN/MCI and 5.2% for CN/AD over the brain-only counterpart.
In briefLost in the Latent Space: How Code Embeddings Represent Intent and Implementation
SummaryThis paper presents a large-scale empirical study of 14 state-of-the-art code embedding models. It evaluates them through supervised probing, zero-shot retrieval, and controlled perturbation experiments using semantics-preserving code transformations. The authors find a tension between abstraction and fidelity, where models that keep stylistic and structural signals generalize less semantically. They conclude that a single universal code representation is not yet achieved and call for more representation-aware evaluation and selection.
In briefCategory-Anchored LLMs for Predicting Users' Next Point of Interest
SummaryThis paper addresses next point-of-interest (POI) recommendation, which predicts a user's next destination from check-in data. The proposed ANCHOR framework first estimates category-level mobility intent in a compact semantic space, then builds a candidate set using category anchors, historical visit memory, and user preferences. A history-aware reranker selects the final target POI. Experiments on three real-world check-in datasets report that ANCHOR outperforms traditional, structure-enhanced, and LLM-based baselines.
In briefFFAI: Flow-matching CT reconstruction with feedback training and adaptive starting points
SummaryThis paper proposes a unified conditional flow matching method for three CT reconstruction tasks: low-dose, sparse-view, and limited-angle imaging. The method conditions on an initial reconstruction and combines feedback-aware training with degradation-adaptive initialization. A teacher network simulates inference-time updates to produce shifted states, from which a student network learns to recover high-quality reference images, while the starting time and initial noise are chosen based on noise level or extent of missing projections. Experiments on public benchmark datasets report better quantitative results than compared methods on all three tasks, with noise and artifacts suppressed and anatomical structures preserved.
In briefRobust LIDAR cross-sensor domain generalization via accumulated-point scene completion
SummaryThis paper studies domain generalization for LIDAR semantic segmentation, where sensor-specific sampling patterns hinder transfer across datasets. The authors propose a source-only training method that aggregates source-domain scans into a simple world model and uses it as supervision for an added occupancy prediction objective, while the input remains a single LIDAR frame. Experiments on distinct LIDAR datasets show improved generalization to unseen domains in certain setups, with no additional inference cost.
Christoph B Rist, Larissa T Triess, Markus Enzweiler et al.
In briefDistilling Machine-Learning Potentials to Predict Long-Timescale Electrolyte Transport
SummaryThe paper presents a knowledge distillation method that transfers the accuracy of a large pretrained "teacher" machine learning interatomic potential into fast, system-specific "student" models, without additional ab initio data. The distilled models run about 40 times faster, reproduce the teacher's structural predictions, and produce stable 20 ns trajectories in days. For sodium- and lithium-ion electrolytes, student-predicted self-diffusivities, ionic conductivities, and shear viscosities agree with experiment and outperform classical force fields.
In briefDEG: Using Diffusion to Strengthen GFlowNet for Vehicle Routing
SummaryThis paper proposes DEG (Diffusion-Enhanced GFlowNet), a framework that combines a Generative Flow Network with a diffusion model to solve vehicle routing problems. GFlowNet supplies edge-level signals that diffusion uses, and diffusion guides broader exploration toward high-reward routes. The authors report that DEG performs favorably against baselines on synthetic and real-world datasets with up to 10,000 nodes.
In brief"Compile Away Uncertainty": Self-evolving agent skills through programmatic compilation
SummaryThis paper introduces SkillCompiler, a framework that lets AI agents accumulate reusable skills from interaction without updating model parameters. Well-specified operations are compiled into validated programs, while context-dependent operations remain under LLM reasoning. A program asset reuse mechanism retrieves existing programs and synthesizes new ones only when needed. The authors report consistent gains over prior methods across five benchmarks and six model families, with task performance improving by up to 32.33%.
In briefInterpretable-by-design tabular prediction using in-context local linear learning (TICL³)
SummaryThis paper presents TICL³, a tabular prediction method designed to be interpretable from the start. It uses a tabular foundation model to retrieve similar training samples and fits a local linear model to correct residuals, so the same model makes and explains each prediction. Across 386 classification benchmark evaluations, it trails TabICLv2 by 1.1 percentage points and outperforms CatBoost in 266 of them.
In briefPer-agent test-time normalization for camera-based cooperative perception with corrupted collaborators
SummaryThe paper addresses cooperative perception where a collaborator's degraded cameras, such as under fog or motion blur, shift the features it shares. It proposes AgentNorm, which estimates normalization statistics for each collaborator from its own current observation and anchors the ego vehicle to its source statistics. Experiments on OPV2V and DAIR-V2X-C show gains over pooled normalization in most settings, while occlusion and cases where all agents share the same shift are not fixed.
In briefWhat Happens When Models Retell Real Events and Distort Facts
SummaryThe paper argues that the standard measure of agreement with the source misjudges model rewrites of news. The authors had three model families rewrite 2,000 articles as summaries and stories, then checked changes across the five Ws and whether each distortion was warranted. Accountability distortions rose up to sevenfold in stories, and two in three lightened the perpetrator's responsibility. The authors release the ReTold-2k corpus with its scoring protocol.
In briefFeature selection and representation learning using Group-Lasso Siamese networks
SummaryThis paper introduces the Group-Lasso principle into a Siamese network architecture for feature selection, which the authors state is the first such study. The method seeks a compact feature subset that preserves the intrinsic representation of the input data while achieving high classification performance. Two variants are presented, one based on Euclidean distance and one on a modified Jaccard similarity suited to high-dimensional sparse data. Experiments on 12 benchmark datasets with 6 to 7070 features and eight classifiers show compact, discriminative feature subsets with competitive or superior classification performance.
In briefDoes Scaling Toward AGI Depend on What System-1 Lets System-2 Execute?
SummaryThe paper proposes a predict-gate-execute-update loop in which a lightweight System-1 estimates the utility of candidate interactions and only those worth their cost are executed by the resource-intensive System-2. Under ideal screening, the authors argue execution compute falls by a factor approaching the inverse of the density of utility-relevant interactions, and they bound selection regret by predictor error and utility margin. They hypothesize that this density is a measurable coordinate of scaling efficiency.
In briefRegion-anchored routing for assigning multiple reference motions to target objects in video generation
SummaryThis paper proposes RART, a framework for multi-object motion transfer that extracts motions from several reference videos and applies each to a user-specified object in a target scene. It unifies per-object motion extraction, target-region anchoring, and frame-aligned injection into a frozen video Diffusion Transformer, and uses a motion-attention routing loss to reduce interference between objects. The authors report that RART needs no per-video optimization and achieves accurate motion preservation, efficient inference, and strong compositional generalization.
In briefTopology Survives Shuffling: Robust Long-Context Memory Access Prediction for Graph Analytics
SummaryGraph analytics produces irregular memory-access streams that depend on graph topology, which limits data prefetching. This paper proposes LoCoMAP, a long-context representation-learning framework that controls observation history and prediction horizon independently. Evaluated on Connected Components, PageRank, and Single-Source Shortest Paths across six synthetic graph families, long-context predictors stay robust under graph regeneration and node shuffling without retraining.
Dongyan Sun, Neelesh Gupta, Rajgopal Kannan et al.
