In briefT-SNN: A Temporal Simplicial Network for Decoding Brain Signals from EEG
SummaryThis paper introduces the Temporal Simplicial Neural Network (T-SNN), which models EEG recordings as sequences of evolving simplicial complexes to capture higher-order interactions among brain regions over time. It combines simplicial convolutions with recurrent updates. On the seven-class SEED-VII emotion recognition task, T-SNN outperforms convolutional, recurrent, graph-based, and Transformer methods in trial-wise and cross-subject evaluations. Adding eye-movement features further improves performance.
In briefMetaSampling: A plug-in method that makes frame samplers more efficient for long-video question answering
SummaryThis paper introduces MetaSampling, a training-free strategy that can be added on top of existing frame selectors for long-video question answering. It dynamically reduces the number of frames passed to a multimodal large language model. Across 36 configurations, it cut the selected frames in every case and improved accuracy in 25 of them. The average frame reduction was 8.9%, with slightly higher accuracy overall.
In briefUnfoldCRF: refining segmentation masks by modeling pixel labels and latent regions together
SummaryUnfoldCRF treats segmentation mask refinement as inference in a conditional random field over pixel labels and latent region variables, unrolling damped mean-field updates on one energy. Compared with recurrent black-box refiners matched in parameters, stages, and supervision, it improves results on COD10K and across five datasets and several mask sources. It also keeps its advantage on mask generators not seen in training. Ablations indicate that pairwise messages mainly fix boundaries and region messages mainly fix non-boundary errors.
In briefA packet-level digital hardware twin for commissioning megahertz diagnostic edge AI and tokamak control integration
SummaryThis paper presents a packet-level digital hardware twin that simulates the full path from megahertz plasma diagnostics to tokamak control-system handoff. The simulator models a 1 MHz beam emission spectroscopy diagnostic, UDP transport with packet loss and jitter, FPGA parsing, edge inference, and a 1 kHz control cycle, generating binary payloads and PCAP files. It reports that a dropped packet is detected and its overlapping inference windows are invalidated, and that valid windows have a median latency of 91.6 us and a 99th percentile of 108.1 us.
Semin Joung, Abhilasha Dave, Luca Scomparin et al.
In briefASTRA: ADMM-Accelerated Topology Reconfiguration for Satellite Constellations
SummaryThis paper presents ASTRA, a framework for dynamically reconfiguring the network topology of large satellite constellations to improve reliability and efficiency. It combines an online learning formulation with an ADMM-based solver and efficient online updates, making constrained updates much cheaper than generic pipelines. The authors report comparable or better topology quality on synthetic constellations and continued effectiveness on real Starlink data under partial deployment and non-uniform spacing.
In briefA benchmark testing YOLO-based weed detection across seven datasets for farming.
SummaryThis paper benchmarks YOLO26 nano, small, and medium models on seven public weed-detection datasets that differ in crops, weed species, field conditions, and annotation protocols. Beyond in-dataset testing, it examines cross-domain generalization and training on combined data from all datasets. The models perform strongly within their own datasets, but performance drops substantially on off-domain data, and multi-source training only partly reduces this domain shift.
Hristina Zdraveska, Vlatko Spasev, Ivica Dimitrovski et al.
In briefRewardExplainer: Learning explanations of reward models from counterfactual preference feedback
SummaryThis paper proposes RewardExplainer, a framework that rewrites responses counterfactually, collects feedback from a target reward model, and uses it to train a reusable explainer. The explainer produces open-ended, atomic, and intervenable natural-language scoring mechanisms. Experiments across multiple target reward models and explainer backbones show consistent improvements. The authors also use the generated mechanisms to find potential bias patterns and build debiasing data that improves robustness on reward-hacking benchmarks.
In briefOpera: A Verbal Critic Framework for Long-Horizon Coding Agents
SummaryThis paper introduces Opera, a verbal critic framework that helps long-horizon coding agents receive corrective feedback. Each correction is kept as a persistent note and tracked until the diagnosed problem is resolved, with feedback audited against visible evidence before delivery. The authors report that Opera raises the resolve rate of non-critic agents on Terminal-Bench 2.1, a SWE-Bench Pro subset, and DeepSWE v1.1. Fine-tuning Qwen3.5-9B on Opera-guided rollouts improved its held-out SWE-Bench Pro resolve rate by 10.2 percentage points.
WhyTreats feedback as persistent notes tracked until resolution, a clear new angle; thorough experiments across three benchmarks and four models with baselines support practical value for coding agents.
In briefICMAPE: A Multi-Agent Method for Learning Through Exploration
SummaryThis paper addresses multi-agent systems where the goal is to gain information about unknown environment properties, a setting called active sequential hypothesis testing. It proposes ICMAPE, a Bayesian learning framework that turns a fixed-confidence identification goal into a reward based on inference confidence, so reinforcement learning can drive decentralized exploration. The method jointly trains a centralized inference network and decentralized policies that also decide when to stop collecting data. On two synthetic benchmarks and a Maryland nitrate monitoring task using real-world data, ICMAPE-TD3 reached target accuracy with fewer exploration steps.
In briefFrom HL to H+L-1 Parameters: A Compact Hankel-Toeplitz Forecaster for Long-Term Forecasting
SummaryThis paper studies linear models for long-term time series forecasting, using classical stationary prediction theory to guide parameter sharing. It proposes the Hankel-Toeplitz Forecaster (HTF), which learns one impulse response that defines both an inverse filter and a forecast map, using H+L-1 trainable coefficients. Across seven benchmarks, its horizon-averaged MSE is within 1.2% of Dense Linear, with 75 to 229 times fewer trainable parameters.
In briefA preprocessing method to reduce inflated similarity in philosophical texts that quote opposing views
SummaryThe paper argues that semantic textual similarity methods mistake vocabulary a document borrows while discussing or critiquing other positions for its own claims, inflating similarity scores. The authors propose 12 rule-based high-level preprocessing steps applied before standard NLP preprocessing. Tested on three Stanford Encyclopedia of Philosophy entries with eight Transformer models, the steps lowered similarity scores across all theory pairs. The paper also introduces a semantic diffusion index (SDI) for measuring each document's semantic reorientation."}}Ž<cite index=
In briefTest-time spatial reasoning for robot manipulation via generative real-to-sim
SummaryThis paper presents Simify, a training-free framework that performs spatial reasoning at test time using massively parallel physics simulation. From one RGB-D image, it builds simulation-ready object assets with 3D generative and vision-language models, then uses evolutionary search over thousands of rollouts to optimize object arrangements for a reward-defined task. Real-robot experiments show it completes complex rearrangement tasks with unseen objects and outperforms prior foundation-model spatial reasoning methods. The authors also highlight that complete and accurate geometry is important for successful sim-to-real transfer.
In briefFuture Information-Directed Sampling for Bayesian Nonstationary Bandits
SummaryThis paper addresses the exploration-exploitation trade-off in bandit learning and proposes Future Information-Directed Sampling (FIDS), a new algorithm for Bayesian nonstationary bandits. FIDS explores to gather information about future optimal arms rather than only the current one. The authors show its regret is comparable to Thompson Sampling up to a small constant factor. They also introduce a supervised-learning approximation that learns the policy from offline data, and demonstrate its effectiveness on synthetic benchmarks.
In briefDynGraphAgentBench: A benchmark testing AI agents that choose anomaly detectors for changing graphs
SummaryThis paper introduces DynGraphAgentBench, an executable benchmark for evaluating agents that manage anomaly detection on dynamic graphs when feedback arrives late. It includes seven temporal graph datasets, eleven selectable detectors, and eight chronological deployment windows per dataset. Controllers see only time-causal aggregate context and their own history, must pick a detector before training, and receive outcomes after a one-window delay.
In briefAdapting neural operators to changing conditions for mechanics decisions
SummaryNeural operators speed up repeated nonlinear mechanics calculations, but their accuracy drops when operating conditions fall outside the training range. This work tests whether high-fidelity finite-element solutions gathered during use can be reused to update a neural operator and improve command selection. Two hard-magnetic soft-material systems are studied. In both examples, updating with 16 acquired loading paths recovers much of the lost accuracy while keeping nominal accuracy.
Prashant K. Jha, Koffi Enakoutsa, Ian Galloway et al.
In briefGroupMask: Layer-adaptive group-wise sparsity for compressing large language models
SummaryThis paper studies semi-structured pruning, which compresses large language models while keeping a regular sparse pattern, and asks whether layer-adaptive sparsity helps beyond the fixed N:M pattern. The authors propose GroupMask, which keeps or removes whole weight groups, uses a hypernetwork to produce group selectors for all layers, and learns them with the pretrained weights frozen. On LLaMA-2-7B at 50% sparsity with a 1×256 group size, learned layer-adaptive allocation lowers WikiText-2 perplexity from 10.02 to 8.30 and raises average zero-shot accuracy from 0.455 to 0.496 compared with a uniform per-layer ratio.
WhyReplaces fixed N:M sparsity with learned group-level, layer-adaptive allocation, validated against baselines on five LLaMA and Qwen models.
In briefBeyond Solo and Consistency: Rehabilitating Multi-Agent Debate via Conditional Progressive Pruning
SummaryThis paper studies Multi-Agent Debate (MAD) built on large language models, a test-time scaling technique where agents exchange multiple rounds of communication. The authors observe that existing MAD frameworks fail to beat strong single-agent and consistency-based baselines under the same strict cost limit. They propose Conditional Progressive Pruning (CPP), a lightweight pruning framework that fully exploits multi-round debate. According to the paper, CPP outperforms existing MAD frameworks on multiple benchmarks and is the first to fully outperform consistency methods.
In briefHARMONIA: Interpretable graph learning using mixtures of neural bases
SummaryThis paper presents HARMONIA, a framework for graph learning that is interpretable by design. It uses a Mixture of Neural Bases to route features to specialized basis experts, and Sparse RRWP Aggregation to capture multi-hop relationships through sparse propagation without quadratic pairwise cost. Predictions keep a simple additive form that decomposes into feature responses and structural influence. The authors report stronger explanation recovery than existing interpretable graph baselines, competitive predictive performance, and scalability to graphs with millions of nodes.
In briefDo System One Decisions Add Up? Checking Probability Consistency Across Decision Steps
SummaryThis paper examines whether a classification model's probabilities stay consistent when a decision is made directly at fine-grained labels versus reconstructed through broad categories. The authors compare two systems, Jev and Laya, on 72,000 questions across TREC, CLINC150, and MASSIVE, finding substantial disagreement. Reconstruction lowers Jev's accuracy on CLINC150 but raises Laya's, and the authors argue that accuracy, calibration, and coherence should be evaluated together.
In briefOn token value inequality in efficient reasoning: how much each reasoning token matters
SummaryThis paper argues that tokens in chain-of-thought reasoning differ widely in value, and that token-level log probability signals can identify the core ones. The authors introduce TokenProbe, a framework that selectively compresses redundant tokens within a GRPO-based objective. According to the paper, the method keeps reasoning quality while reducing token usage by 76% relative to the baseline.
In briefCompressing volumetric video with sparse 4D anchors for Gaussian splatting
SummaryThis paper presents SAGA, a codec for volumetric video that represents dynamic 3D scenes with hierarchically organized sparse 4D anchors. Coordinate-based neural decoders generate finer anchors and Gaussian primitives by interpolating between anchors, sharing parameters across spatiotemporal structures. Fixed-size memory slots with orthogonality-informed updates are added to model long-range dependencies for entropy-context modeling. Experiments report PSNR BD-rate reductions of 80.39% on Neu3D and 83.94% on MPEG MIV relative to GIFStream.
In briefTwo-sample test for random graphs under non-integral L_r norms
SummaryThis paper studies whether two populations of networks share the same edge probabilities, a task whose difficulty depends on the norm used to measure differences. For non-integral r greater than 2, the known upper and lower bounds did not match. The authors propose a test that runs two statistics and rejects if either does, with thresholds set by Hölder interpolation. They prove it attains the conjectured rate, and simulations with n between 32 and 256 show sample growth consistent with theory.
WhyCombines two statistics via Hölder interpolation to prove the conjectured optimal rate for non-integral r>2, with simulations supporting theory.
In briefThinkNet: Evolutionary Ensembles for Compact Motor-Imagery EEG Decoding
SummaryThis paper presents ThinkNet, a framework for selecting compact brain-computer interface decoders and inference policies that generalize to held-out subjects analyzing motor-imagery EEG signals. It uses train-only normalization, validation-guided evolutionary search, and validation-gated inference so that test-subject data does not influence model selection. On BCI Competition IV-2a under leave-one-subject-out evaluation, compact models achieved higher mean held-out accuracy than mid-size and large alternatives, and validation-gated ensembling improved results further.
In briefGreenpixie's AI Token Method: Estimating Energy, Water and CO2-eq Impact per AI Token
SummaryThe paper proposes a method to estimate the energy use of cloud-hosted large language model inference per token, separating input and output tokens. GPU energy is measured by benchmarking open-weights models on text tasks, and non-GPU server energy is estimated from inference wall time. Bayesian regression relates energy per token to model size, traffic, and hardware, and the authors describe how the results can estimate CO2-equivalent emissions and water use per token.
Joshua Horswill, Ross Hunter, Matt Clifford et al.
In briefSymNetPro: Locating Multiple Directional Radio Transmitters from Sparse, Line-of-Sight-Aware Signal Measurements
SummaryThis paper addresses localizing several directional radio transmitters from sparse received-power measurements, where the receiver sees only a combined signal that cannot be separated by source. Building obstructions and stronger transmitters that hide weaker ones make the problem harder. The authors extend the SymNet model into SymNetPro, adding a sparse line-of-sight-aware attention bias and a transmitter-drop augmentation method. On directional ray-traced urban scenes, it reports substantially lower OSPA than representative baselines under extreme sparse sampling, with consistent gains under noise and more transmitters.
In briefEEG-Fusion: Failure-Aware Expert Routing for Source-Free Motor Imagery EEG Decoding
SummaryThis paper addresses subject-level failures in cross-subject motor-imagery EEG decoding, where a decoder can collapse toward predicting one class for some users. It proposes EEG-Fusion, which uses a reliability gate built from label-free signals to route each target subject to a neural, covariance-based, or physiological-feature expert. In leave-one-subject-out evaluations on BCI IV-2a, BNCI2014-001, BNCI2014-004, and a Cho2017 subset, the method improves subject macro-F1 and reduces collapse indices relative to a raw EEGNet baseline.
In briefReliable LLM-as-a-Judge Measurement for Multi-Turn Business Agents
SummaryThe paper presents a methodology for evaluating production multi-turn business agents, combining evaluation specification, modular LLM judges, intent-preserving user simulation, and human-in-the-loop governance. It attributes failures to either agent capability or missing business knowledge so they can be acted on. Production studies reported that system-level fidelity improved across repeated audits and that combined human and automated judging had the strongest descriptive performance in two task-completion settings. The authors note the studies are observational, so the findings show operational usefulness rather than causal or universal superiority.
In briefSimple diffusion language models generate better in few steps than reported
SummaryThe paper argues that much of the quality gap of diffusion language models at few generation steps comes from a suboptimally configured sampler. Adjusting only the sampler, without retraining, lets an older masked diffusion model match newer successors, reaching lower generative perplexity in 16 steps than the standard sampler reaches in 1024. The authors introduce GroupEval to assess quality and semantic diversity separately, and explain theoretically that parallel unmasking breaks token dependencies, so a fixed temperature of one is generally suboptimal.
In briefMonitoring JEPA world model behavior with Jacobian centroids
SummaryThe paper proposes centroids, the row-sums of sub-component Jacobians, to characterize the behavioral properties of world models, computable efficiently via Jacobian-vector products. On continuous control tasks with JEPA world models, it finds the encoder represents the goal correctly while the predictor is behaviorally unresponsive, a failure that predicts planning failure before any action. Centroid-based methods also outperform baselines as distribution-shift detectors.
Thomas Walker, Randall Balestriero, Richard Baraniuk
In briefEpiTransfer: Training-free long-range depth from monocular drone frames
SummaryThis paper proposes a training-free depth estimation method that uses only two monocular images and camera pose estimates. It synthesizes a virtual stereo pair via epipolar transfer, turning temporal correspondence into stereo triangulation. Tested on outdoor drone flights up to about 90 m and indoor OptiTrack scenes against LiDAR ground truth, it achieves an indoor AbsRel of 0.092, close to direct triangulation, and outperforms ZoeDepth and Depth Anything V2.
Diksha Aggarwal, Rutvik Dagadkhair, Sanjana Srivastava et al.
In briefTest-Time Generalized Category Discovery: Adapting a vision-language model to recognize known classes and find new ones without labels
SummaryThe paper introduces Test-Time Generalized Category Discovery (TT-GCD), a setting where a vision-language model must adapt to distribution shifts, classify known categories using only text supervision, and discover novel categories, all at test time without labeled data. The authors propose PACT, an unsupervised framework that assigns test images to prototypes representing both known and novel categories in visual space. Experiments on corruption and domain-shift benchmarks reportedly show PACT outperforming adapted state-of-the-art TTA and GCD methods.
Shambhavi Mishra, Omprakash Chakraborty, Julio Silva-Rodriguez et al.
In briefRe-examines whether evidence supports claims that language models produce homogeneous open-ended responses.
SummaryThe original paper argued that language model responses to open-ended prompts are strongly homogeneous, calling this an Artificial Hivemind. This paper rechecks its three central results, finding the flagship metaphor example unrepresentative and the baseline too lenient, with several measurements lacking any null comparison. It also shows that prompting, an inference-time intervention, reliably increases response diversity. The authors do not decide whether the Artificial Hivemind exists, only that the published evidence does not establish it.
Rylan Schaeffer, Brando Miranda, Joshua Kazdan et al.
In briefUnifying Video Tasks via Spatiotemporal Analogy
SummaryThe paper introduces ViGeo, a framework that extends training-free visual in-context learning to video through spatiotemporal canvas completion. Tested on a diverse task taxonomy with a strict train-test split, it generalizes to unseen video manipulations and zero-shot modalities such as event cameras. The authors also identify task internalization, where a pretrained task's query format overrides demonstrations, and show that a small amount of task-unrelated data can remove this shortcut.
Chia-Hsiang Kao, Belinda Zeng, Bharath Hariharan et al.
In briefArticulateArena: A Metric for Comparing Articulated Object Motions
SummaryThis paper proposes ArticulateArena, a metric for evaluating predicted articulation in articulated objects by comparing the motions that joints induce rather than their parameter encodings. It handles fixed, revolute, prismatic, and helical joints in one framework and extends to full kinematic trees. The authors also release ArticulateArena-20K, a dataset of 19,977 articulated objects with verified kinematics, and re-evaluate published reconstruction methods on it.
WhyProposes a representation-invariant joint motion metric, proves it is a metric, extends to trees, and re-evaluates methods on a 20K-object benchmark.
In briefDiffusion-Based Rollouts for Stabilizing Long-Range Environmental Forecasts
SummaryThe paper examines whether diffusion-based rollouts can stabilize recursive forecasting in environmental prediction, tested on water-level time series and high-dimensional precipitation fields. Diffusion suppresses recursive error growth, but stability does not guarantee forecast fidelity: water-level forecasts gradually lose event-level detail and contract toward central values. In precipitation experiments, which keep numerical weather prediction conditioning throughout the rollout, diffusion better preserves spatial organization and event-detection skill.
Marina Vicens-Miquel, Amy McGovern, Aaron J. Hill et al.
In briefPreserving differentially expressed gene rankings when predicting spatial gene expression from histology images
SummaryThe paper introduces image-based differential expression ranking (IDER), which tests whether gene expression predicted from histology images preserves the ranked gene list derived from measured differential-expression statistics. It proposes a differentiable IDER training objective that can use morphology-derived proxy contrasts instead of predefined biological group labels. On public spatial transcriptomics datasets, the approach improves DEG-ranking agreement and pathway-enrichment overlap compared with conventional reconstruction objectives.
Kaito Shiku, Kazuya Nishimura, Yasuhiro Kojima et al.
In briefMaking the Phi-2 small language model lighter for real-time chatbots using PEFT and QLoRA
SummaryThis study optimizes the Phi-2 small language model for real-time chatbots by combining parameter-efficient fine-tuning with QLoRA, which adds 4-bit quantization to low-rank adaptation. The goal is to cut memory use substantially for mobile and edge devices while keeping or possibly improving response accuracy. Evaluated with ROUGE metrics, the models showed notable gains on summarization tasks.
PhanTan Khanh Nguyen, Ashfaq Ali Shafin, Khandaker Mamun Ahmed
In briefQuantization error is spectrally flat; one random probe estimates layer sensitivity without data for mixed-precision bit allocation.
SummaryThe paper shows that a single random Gaussian probe gives an unbiased estimate of a layer's squared quantization error, because round-to-nearest error is spectrally flat. Tested on 1,683 tensors from a 35B MoE and a 9B dense model, the estimator is stable across bit-widths. The RAM method applies this probe with propagated input statistics to allocate bits under a byte budget without calibration data, and on tested MoE models it reduces median WikiText-2 perplexity by 3.5 to 13.6% versus size-comparable uniform 4-bit builds.
WhyA single random probe yields a calibration-free error estimator used for budgeted mixed-precision; novel, with broad multi-architecture experiments but limited baseline comparisons.
In briefHyperMCTS: Hypergraph-Enhanced Tree Search for Long-Task LLM Agents
SummaryLong tasks require language model agents to make coordinated decisions under constraints spanning a whole solution. HyperMCTS is a training-free method that adds a cross-trajectory hypergraph to an MCTS tree, accumulating returns for groups of decisions that recur across different paths. Its HyperUCT selection rule reuses this evidence to guide search. On DeepPlanning it raises average planning accuracy by 2.3 to 7.3 points over the strongest baseline across three backbones, with fewer LLM calls and output tokens, and SealQA experiments show gains in question answering.
Tingsong Xiao, Nithish Balachandar Moudhgalya, Chandrayee Basu et al.
In briefSelf-Discovering RL: Is Learning History an Asset or a Burden?
SummaryThis paper performs a causal mechanistic audit of a reinforcement learning update rule that was discovered automatically by an algorithm, organized around the five pillars of the Era of Experience. By pinning, freezing, and transplanting recurrent states while holding meta-parameters fixed, the authors test when learning history helps and when it hinders. Key findings are that recurrent history widens the usable reward-scale window, that letting imported state evolve reduces the penalty of mismatched history, and that controlling replay retention can reverse the apparent adaptation advantage over DQN under environmental change.
Haomin Luo (1, 2) ((1) University of Cambridge, (2) Models2 AI)
In briefTesting memory-optimal transformer kernels on real HPC hardware against formal predictions
SummaryThis paper checks cost predictions for transformer kernels derived with the Mathematics of Arrays against measured performance on two HPC clusters, using CPUs and GPUs. The authors found and fixed a GPU slowdown caused by atomic contention, yielding up to a 2.5x speedup. They also report that identical derivations produce very different real costs depending on cluster topology. They propose that hardware-specific optimization can be handled as a routine rewrite with fixed, verified specifications.
In briefWhen consent outlives its context: reusable leftover permissions in long-running AI agents
SummaryThe paper finds that user approvals granted to LLM agents for security-sensitive actions can persist across tasks or sessions, remaining usable after the context that justified them has ended. The authors propose a longitudinal attack that obtains such authority through benign interactions and later replays it during adversarial execution. In tests, residual authority raised attack success rates by up to 35.1 percentage points on AgentDojo and by 24.9 points on average in live attacks against three production coding agents.
WhyIdentifies a new residual-authority replay failure in long-lived agents, validated across six LLM families and live coding agents with substantial attack success gains.
In briefArtificiality's First Exam: an LLM-based C-test for evaluating supraintelligence
SummaryThis paper proposes a new test based on Levin's Kt complexity, using large language models as representational machines. It generates items whose difficulty is measured in bits and certified before any human or frontier model can solve them. The authors administered a battery of certified items to twelve LLMs, finding that performance declines with difficulty into previously uncharted high-difficulty regions.
In briefJIVE: A training-free method to make generated samples more diverse
SummaryGenerative models often suffer from mode collapse and limited sample diversity. JIVE is a training-free framework that injects velocity perturbations aligned with the leading right singular subspace of the generator's endpoint Jacobian to increase diversity. The paper claims this provably maximizes endpoint diversity while preserving quality, using matrix-free iterations with small overhead. Across benchmarks, it improves pixel and feature-level diversity in few-step and one-step generation.
In briefSlopBench: Ranking language models by repetitive AI-style writing across multiple domains
SummarySlopBench is a benchmark that compares which language models produce stiff, repetitive prose that readers call AI slop. The authors evaluated eighteen models on 112 tasks spanning email, social posts, essays, and workplace chat, collecting 19,928 outputs. They score four surface behaviors that can be checked by hand, namely length, opener repetition, paragraph rhythm, and fixed lexical constructions. The composite ranking proves sensitive to weighting, so the authors report the four behaviors separately and release prompts, outputs, reference statistics, and scoring code.
In briefExplainer: The Two Faces of Adam in Separable Linear Classification
SummaryThis paper studies deterministic, full-batch, bias-corrected Adam on separable linear classification with softmax parametrization and log-loss. It gives a quantitative description for small fixed positive epsilon: the classifier first nearly maximizes the max-norm margin, and only later do updates become gradient-like. Reaching a target Euclidean margin takes many more iterations, with bounds given for polynomial and inverse-linear stepsize schedules, and experiments are reported to support these predictions.
WhyQuantitatively characterizes two-phase Adam behavior with small fixed epsilon, separating max-norm and Euclidean margin regimes, with iteration bounds supported by experiments.
In briefFinite Probes Suffice: Identifiability and Universality in Weight-Space Learning
SummaryThis paper studies when representations built from a finite set of probes are enough to learn properties of neural networks, and establishes general identifiability and universality results. It shows that intermediate hidden representations can be considerably more informative than final outputs alone. Based on this, the authors introduce HIDDENPROBE, an architecture that learns from hidden probe responses, which outperforms existing probing methods on neural functional benchmarks including MLPs and Transformers. The code is publicly available on GitHub.
Soutrik Sarangi, Yonatan Sverdlov, Adir Dayan et al.
In briefNSV-Shift: A benchmark testing whether speech models understand non-speech sounds like sighs or laughs and respond accordingly
SummaryThe paper introduces NSV-Shift, a contrastive benchmark for checking whether speech-to-speech models can understand non-speech vocalizations and adapt their replies. Each pair has two conversations with the same words, differing only in the non-speech vocalization in the final turn. A pilot with 22 human-verified pairs (44 audio conditions) tested five models on perception, emotion understanding, and response adaptation. Models detected non-speech vocalizations better than they interpreted their fine-grained emotional meaning or gave appropriately differentiated responses.
In briefHow residual-stream burden shapes representation learning in diffusion transformers
SummaryThis paper examines why diffusion transformers on large pixel patches succeed when predicting clean data but fail when predicting noise or velocity. The authors propose residual-stream burden, the need for the residual stream to preserve noise-dependent input variation through depth. Controlled experiments identify spectral concentration in patch space and residual-state bandwidth as key factors. They also introduce Spatially Indexed Hyper-Connections (SiHC), which reaches FID 1.71 on ImageNet 256².
In briefMISHAP-Bench: A Hallucination Benchmark for Large Audio-Language Models
SummaryLarge audio-language models often give fluent answers about audio that lack grounding. The paper defines two hallucination types: context claims not based on the audio, and knowledge claims lacking support from verifiable external facts. It introduces MISHAP-Bench, with 12,000 open-ended question-audio pairs and an evaluation pipeline. Testing ten models, it finds substantial hallucination, with the best model at 36.5%, and mitigation methods give only partial improvements.
Zhi Wen Soi, Giulio Segalini, Jian-Jia Chen et al.
In briefControlling risk in LLM answers by pricing label-free checks
SummaryThis paper introduces PriceCheck, which builds decision rules from label-free checks such as re-solving a problem to decide when a large language model should answer or abstain. Each check is priced by its agreement rates on correct and incorrect answers and its cost per run. In mathematics, the selected schedules served 76.1% of answers on average while keeping held-out selective risk below 1.5% on all 15 splits.
WhyProposes priced label-free checks for selective answering, validated on math with 15 splits, baselines, and 118 schedules.
Dongyub Jude Lee, Jungseob Lee, Chanjun Park et al.
In briefFeature-Augmented Boosting for Multicalibration Without Discretization
SummaryThis paper studies multicalibration, which requires a predictor's residuals to be unbiased globally and also after conditioning on its own level sets and reweighting by a rich class of test functions. Inspired by Tax et al. (2026), the authors propose a boosting paradigm that avoids discretizing the predictor during training: each round calls a squared-loss oracle on hypotheses that take the previous predictor's output as an extra feature, and only the final predictor is rounded. Analyzing the expressivity of the augmented hypothesis class, they show it yields a hierarchy of guarantees, including multiaccuracy, multicalibration, and the stronger level-set multicalibration.
In briefWhere Activation Sparsity and KV-Cache Sparsity Meet in LLM Decoding
SummaryThis paper studies how two ways of speeding up large language model decoding, activation sparsity (reducing weight reads) and KV-cache sparsity (reducing cache reads), compare in memory traffic. The authors derive a context length at which their savings are equal and ideal speedup bounds from model dimensions and keep ratios. Timing on two GPUs from 2K to 128K tokens shows the projection branch leads at short context and the KV branch at long context. Under matched perplexity budgets, composing the two decodes 14 to 26% faster than the best single branch.
WhyUnifies byte-crossover analysis for activation and KV sparsity, validated by timing on two GPUs across 2K–128K contexts, useful within the subfield.
In briefStandard policy optimization can be both optimal and differentially private for contextual bandits
SummaryThis paper asks whether vanilla policy optimization explores enough to achieve near-optimal regret in stochastic contextual bandits. The authors show that exponential policy updates driven by offline regression achieve this under realizability, without exploration bonuses or importance weighting. They also present a batched version and show how private regression oracles yield differentially private bandit algorithms, supported by experiments.
WhyShows vanilla exponential policy updates achieve near-optimal regret without exploration bonuses, plus batched and differentially private variants; experiments are modest in scale.
Idan Attias, Orin Levy, Alexander Ryabchenko et al.
In briefFaster block-diffusion language model serving with distribution-free risk guarantees
SummaryThe paper presents Redline, a finite-sample procedure that selects serving operating points for block-diffusion language models using correctness on calibration prompts. It keeps the joint probability that a reference answers correctly and a candidate does not within a user-chosen budget with high probability, then deploys the fastest configuration that passes. The authors report that on math at a ten percent budget, a LLaDA2 configuration commits over a third more tokens per forward pass.
Jungseob Lee, Dongyub Jude Lee, Chanjun Park et al.
In briefLLMs trained to predict their own accuracy learn two forms of metacognition
SummaryResearchers trained 10 open-weight LLMs to predict their own accuracy on factual multiple-choice questions before answering them. The trained confidence reflected two signals: on questions close to the training data it tracked true accuracy, while in other domains it tracked output consistency, the concentration of the answer distribution. Output consistency tracking appeared early in training and generalized across datasets, whereas accuracy tracking developed later and stayed local to the training distribution.
Nicolas Yax, Stefano Palminteri, Pierre-Yves Oudeyer
In briefProspective Interpretation Risk: Principled Control of Messages Between LLMs
SummaryThis paper studies multi-agent LLM systems where one model sends messages to another and the receiver may interpret them differently than intended. It defines prospective interpretation risk (PIR), the probability that a receiver reconstructs a task other than the intended one, estimated with black-box probes before the message is sent. The authors also introduce value of interpretation information (VoII) to decide when querying receiver information is worth its cost. Reported results include interpretation-failure rates varying 4 to 13 times across receivers and a 44% reduction in failure after PIR-guided message revision.
In briefLost with a Map: Conversational State and Behavioral Reliability in Language Models
SummaryThis paper studies how eight instruction-tuned language models represent, update, and use conversational state in task-oriented dialogue on MultiWOZ and SGD. Structural information such as active domains and slots is more linearly readable than exact values, and after a user changes a value, both old and new values still influence model behavior. The authors propose a state-action controller that selectively edits the base model's action using structural readouts, raising exact-query accuracy from .318 to .621 and task success from .272 to .371 on held-out MultiWOZ interaction.
In briefGrowing Brain-Signal Decoder Neurons: A Controlled Test on EEG
SummaryConvolutional EEG decoders are usually trained at a fixed width, and this study tests whether adding neurons during training beats a reference width on EEG. The authors grew three convolutional backbones on 12 motor-imagery datasets under three protocols and compared each with its reference model per subject. The grown ShallowFBCSPNet scored 2.9 points above its reference with only 0.57x the parameters, while SCCNet changed by at most 1.2 points.
Adam Mounir, Stella Douka, Arnault H. Caillet et al.
In briefDCEmbed: A Scalable Method for Optimizing Over Neural Surrogate Models
SummaryThis paper proposes DCEmbed, a heuristic for optimization problems that embed neural network surrogates. It uses a difference-of-convex representation of ReLU networks, avoiding extra binary variables for activations, and solves the problem via an iterative penalty convex-concave procedure. In experiments on quadratic programs, mixed-integer resource allocation, and neural two-stage stochastic programming, it reaches high-quality feasible solutions much faster than exact mixed-integer embeddings.
Akshay Sreekumar, Nicolas Christianson, Priya L. Donti et al.
In briefUsing Two Confidence-Gated Local LLM Judges to Curate Merchant-Matching Training Labels
SummaryThis paper studies how reliable pseudo-labels are for matching noisy payment descriptors to merchant entities or returning no match. Labels are kept only when two local large language models agree, with separate thresholds for selections and abstentions. On 2,000 expert-annotated queries, the pair reached 96.88% label purity at 81.7% coverage, outperforming either model alone. The study measures label purity only, not downstream value for student models.
In briefCounterfactual Rollout Replay: Using Forkable Environments as Free Process Rewards for Coding Agents
SummaryOutcome-only reinforcement learning gives software engineering agents only a final success signal, with little guidance on intermediate decisions. The paper proposes Counterfactual Rollout Replay (CRR), a training-time method that uses forkable executable environments to select decision points, restore states, sample alternative actions, and roll branches forward to obtain step-level return contrasts. CRR needs no human process labels or learned process reward model, with "free" referring to those supervision costs rather than replay compute.
In briefJET: Answer selection with pretrained Transformer models, no extra training
SummaryJET uses pretrained language and vision-language models to choose one answer from a finite set of candidates without additional training. It directly evaluates each candidate's likelihood and shares computation across candidates. On desktop CPUs and consumer GPUs, Qwen3.6-35B-A3B reached 87.48% accuracy on the full MMLU test set. Prefix reuse and cache management gave 2.18 to 2.23 times speedups, and input preparation optimizations cut process time by 30.8% with unchanged outputs.
In briefRobot-GST: a geometry-aware framework for representing and evaluating robot manipulation policies over time and space
SummaryThis paper presents Robot-GST, which builds a simulated robot environment from RGB-D observations using 3D Gaussian Splatting and SAM3D. Candidate action sequences are simulated and evaluated in this environment before execution to filter out infeasible behaviours. The approach combines large vision-language models for long-horizon planning with Gaussian-aware final-state estimation. It is validated on cube placing, toy packing, and duck rearrangement tasks involving rigid, soft, and deformable objects.
In briefPopulation Physics, Population Problems: Safety and Emergence in LLM Societies
SummaryThis paper studies the collective behaviour of large language model (LLM) societies, which can produce statistically significant self-organisation that is not predictable from single agents alone. The authors introduce a framework for measuring self-organisation in LLM social systems and apply it to a Schelling grid, a social network (Moltbook), and a Twitter-like misinformation simulation (Rogue). All three show significant self-organisation, with open-ended systems displaying phase-transition-like dynamics. The paper reports that population-level pathologies can still emerge even when LLMs are safety-tuned or monitored.
In briefWhen successful strategies fail: how terminal agents adapt to environmental change
SummaryThe paper introduces environmental novelty, a change that keeps a task's goal fixed but invalidates an assumption behind a previously successful strategy. The authors present AGNI, an automated pipeline that extracts trajectory-relevant assumptions, injects targeted environmental changes, and checks that the novel tasks remain solvable. Across three terminal benchmarks and multiple LLM agents, performance drops substantially on novel tasks compared with base tasks. Post-training on environmental novelty improves adaptation to held-out novel tasks and also improves base-task performance.
Janvijay Singh, Vaishnavi Shrivastava, Dilek Hakkani-Tur et al.
In briefR²Flow: Self-improving AI agents that evolve their own skill libraries
SummaryThis paper introduces R²Flow, a framework that lets LLM-based agents improve themselves by reusing and revising the skills they combine into procedures. It alternates policy learning, independent verification, and versioned skill-library updates on a shared-state graph that merges equivalent execution histories. The authors report better task accuracy and library-edit precision than heuristic, reinforcement learning, and skill-evolution baselines across question answering, math reasoning, decision making, and code generation, with transfer across executors.
In briefWhether speech recognition models can adapt to new speakers from examples given at inference time
SummaryThis paper tests whether encoder-decoder speech recognition models can adapt to new speakers, accents, and domains when given speech-text examples at inference time. Six models were evaluated with two demonstration formats, and all could adapt out of the box, with up to 30% relative improvement in oracle experiments and up to 23% using first-pass hypotheses. Collated demonstrations gave consistent adaptation across models.
In briefPI-NOMT: Physics-Informed Neural Optimal Mass Transport for Brain Fluid Dynamics
SummaryThis paper presents PI-NOMT, a framework that represents tracer density, velocity, and source as continuous neural fields and combines governing physical equations with unbalanced optimal-transport regularization. It infers hidden transport mechanisms from sparse spatiotemporal observations. The method is evaluated on a synthetic benchmark with known ground truth and on DCE-MRI sequences from nine control rats, where it recovers the velocity field and yields low endpoint error and low PDE residuals.
Mehmet Emin Acar, Vahit Bugra Yesilkaynak, Helene Benveniste et al.
In briefProgram-Verified Self-Evolution for Vision-Language Models: a method that verifies generated answers with programs instead of majority votes
SummarySelf-evolving vision-language models generate questions from unlabeled images, but prior methods label answers by majority vote or model judgment, and a human evaluation found 24% and 18% of those labels wrong. The paper proposes VQS, in which the model parses each image into a structured record, and fixed programs write questions and compute answers from it, with the model checking only individual facts. Across ten benchmarks, VQS improves Qwen3-VL by up to 3.18 points at 2B, 4B, and 8B scales, outperforming the strongest self-evolving baseline at each, and human raters judged 94% of its answers correct versus 76% for majority voting.
In briefReDrive: World-Modeling-Shaped Representations for End-to-End Driving
SummaryThis paper presents ReDrive, an end-to-end driving framework that strengthens planning-oriented visual features by predicting future representations. Training has three stages: driving video pretraining, joint world modeling and planning, and planner adaptation. At inference, it needs no auxiliary perception modules or explicit future prediction. On NAVSIM, it reports 91.0 PDMS on v1 and 90.8 EPDMS on v2, and the authors argue that a simple encoder-planner pipeline suffices for high-performance planning.
In briefRethinking Contextualization by Reinterpreting Attention Head Channels
SummaryThis paper studies contextualization, the process by which language models pass information between words to build sentence-specific word representations. The authors find that less-informative words tend to absorb more contextual information, and that this absorption is selective, enabling more precise routing between matched words. They reinterpret attention heads as channels gated by their singular vectors, which point to the hidden states of more informative words, and show this enables automated interpretation of heads and embedding them in a continuous space.
In briefProbability Contracts: Measuring Accuracy, Coherence, and Decisions in LLM Probability Reports
SummaryThe paper introduces probability contracts, a benchmark combining exact finite-world posteriors, event alignment, interface coherence, and failure-aware decision evaluation. Across four model-interface configurations on 1,000 worlds, it finds that equivalent probability requests can yield different decisions, and that accuracy rankings vary by stratum. Post-hoc analyses and pre-specified secondary checks examine disagreement, confidence, and decision loss across interfaces.
In briefQwenGyre: A Flexible Reinforcement Learning Framework for Training Extra-Long-Horizon AI Agents
SummaryThe paper presents QwenGyre, a framework for online reinforcement learning of large language model agents on extremely long tasks that can span hours and nearly 1M tokens per rollout. It reallocates GPUs between rollout and training without interrupting live executions, and its trajectory processor reconstructs branching histories, scores partial progress, and removes redundant paths. The authors report a 6.0 percentage-point gain on NL2RepoBench (52.5% to 58.5%) and speedups of up to 1.85x over the Colocate baseline.
In briefHow code data helps different tasks: a decomposed view of LLM post-training
SummaryThe paper splits an execution-verified code corpus into interpretable categories based on the computational patterns of its solutions. Controlled fine-tuning compares individual categories with a balanced mixture on question answering, mathematics, and code generation across instruction-tuned models. The authors report that the same category can help one model or task and hurt another, and that the best category varies by starting model and target task. On some model-task pairs, compact mixtures using roughly 10–15% of the full corpus outperformed full-corpus training.
In briefViBR-WM: Bayesian regression model for forecasting visual and physical states
SummaryThis paper introduces ViBR-WM, a world model for forecasting that uses interpretable Bayesian regression over visual features, physical histories and known covariates. It compresses visual representations and applies Bayesian variable selection, then combines forecasts from predictor subsets weighted by posterior probability, accounting for parameter uncertainty and future disturbances. Across four forecasting tasks (object motion, vegetation greenness, solar power), the authors report lower mean physical-target error than Temporal Straightening, ConvLSTM, PredRNN and SimVP on every task.
In briefLaya as a Typed Probabilistic Assessor: Independent Reproduction and Preregistered Calibration Study
SummaryThis paper independently reproduces the Laya Typed-Decisions checkpoint, a 421M-parameter ModernBERT-large model that answers choice, negation and score questions over workflow state. It finds the model is uniformly under-confident, contrary to the card's over-confidence claim, and a single temperature fit reduces held-out ECE from 0.204 to 0.037. The preregistered isotonic regression selection was not supported. Re-running the full official test split reproduces the headline accuracy, and 20 of 22 confirmatory tests reject under FDR correction at q=0.05.
In briefChemOPD: Multi-Teacher On-Policy Distillation for Chemical Reasoning Across Many Tasks
SummaryThis paper introduces ChemOPD, a method for training one language model to handle several chemical reasoning tasks. It estimates task similarity from fine-tuning gradients, groups specialist teachers with partial overlap, and keeps a generalist teacher so specialist guidance adds to its supervision. On the ChemCoTBench benchmark, the anchor-residual approach improves most metrics over specialist-only distillation and captures more of the teachers' gains.
In briefDeltaSeek: An Early Step Toward Active Perception in Changing Construction Sites
SummaryConstruction sites change continuously, which degrades static mapping and registration. This paper presents DeltaSeek, a framework examining how a robot's sensing embodiment limits the observations it can acquire. Using a Husky A300 with a UR5e arm on an IFC-derived benchmark, it compares chassis-mounted and wrist-mounted RGB-D cameras. Two changes were invisible from any chassis viewpoint while the wrist camera resolved both; for shared observations, median base travel was 6.0 m for the wrist camera and 15.2 m for the chassis camera.
In briefActive-Bottleneck Mechanism for Weak-to-Strong Generalization: When a Student Beats Its Teacher
SummaryThis paper studies weak-to-strong generalization, where a student trained on a teacher's predictions outperforms that teacher, under fully converged ridgeless two-stage linear regression with no early stopping or explicit regularization. The authors derive in closed form the range of pseudo-label counts for which the student beats the teacher, depending on spectral decay and noise level. A random-feature analysis identifies an 'active-bottleneck' principle: whichever of the number of pseudo-labels or the student width is scarcer controls how much teacher error is filtered out.
In briefCLIMB-flow: a multiscale flow method for posterior sampling in linear inverse problems
SummaryThis paper presents CLIMB-flow, a posterior sampling algorithm for Bayesian inverse imaging problems using pyramidal or cascaded diffusion models. It uses a coarse-to-fine strategy to generate images directly in pixel space, alternating endpoint estimation, data-consistent updates, and re-noising. The steps approximate Gibbs sampling from two conditional distributions. Experiments on ImageNet, CelebA, AFHQ, and fastMRI cover inpainting, deblurring, super-resolution, and accelerated MRI, reporting 1.37 to 7.66 dB PSNR gains over the strongest competitor on CelebA and reconstruction up to 512x512.
In briefFooling Many Frontier MLLMs with One Black-Box Adversarial Attack
SummaryThis paper investigates whether a single adversarial image can consistently mislead diverse frontier multimodal large language models in black-box settings. The authors propose O-Attack, which exploits a broad, cross-modally aligned semantic space in surrogate models and optimizes perturbations through semantic consensus. The reported attack success rates on GPT-5.4, Claude-4.6, and Gemini-3.1 rise from 29.1%, 42.8%, and 38.2% to 77.2%, 81.6%, and 80.9%, and the method outperforms six prior approaches across 24 MLLMs.
WhyImproves transfer attacks by exploiting aligned semantic spaces of surrogates; evaluated across 24 MLLMs versus six baselines, with large success-rate gains.
In briefLearning to Align Robot Actions with Imagined Visual Plans
SummaryThis paper studies world-action models (WAMs), which jointly predict future visual observations and robot actions, and notes that their visual predictions can disagree with the outcomes implied by the generated actions. The authors treat the visual prediction as a goal-conditioned proposal and use a frozen action-conditioned world model to build feedback from consistency between the two predictions and alignment with the goal. This feedback is used with Flow Policy Optimization (FPO) to optimize the WAM's action head. Across four real-world UR5 manipulation tasks, mean success rose from 43.4% to 75.1%, compared with 61.4% for π0.5.
In briefV-JEPA 2 and AQA for Mountain Bike Technique: Phase-Based Scoring and Reference-Based Feedback
SummaryThis paper studies action quality assessment of mountain-bike jumps using 95 field-recorded clips labelled only as good or poor technique. The method combines frozen V-JEPA 2 video features and 2D pose and motion measurements over four heuristic phases, and a small transformer predicts a score. A reference layer compares phase measurements with good-technique clips to suggest possible cues. The authors stress that the cross-validation results (AUROC 0.768) are exploratory, affected by selection bias and unseparated riders, and that the small external probe was weak, so rider-independent scoring and reliable corrective feedback are not yet established.
In briefdOPT: Differentiating conic optimization by reducing it to a simpler quadratic program
SummaryThe paper introduces dOPT, a framework for differentiating conic optimization problems so they can be embedded in learning systems. Instead of differentiating the full conic formulation, it reduces the problem at a computed primal-dual solution to an equality-constrained quadratic program that preserves the solution and its first-order sensitivity. Explicit reductions are derived for convex NLPs, QPs, SOCPs, and SDPs. Numerical experiments validate the gradients and show substantial speedups over existing methods as problem size grows.
In briefVestrum: Improving AI Agent Harnesses by Adapting Verification, Structure, and Memory
SummaryThis paper introduces Vestrum, a framework that improves the software “harness” controlling language model agents without training the underlying model. It turns failures in execution traces into classified, scoped changes to verification, retrieval, decomposition, and knowledge synthesis, then evaluates them as a bundle. Across five settings and two baseline harnesses, the improved harnesses raised held-out performance, for example UltraHorizon from 47.6 to 59.8 and Terminal-Bench 4 Hard from 63.7% to 70.3%.
Jayant Parashar, Eugene F. Douglass, William C. Bastian et al.
In briefImproving visual anomaly detection by automatically interpreting and adjusting its internal signals
SummaryVisual anomaly detectors flag deviations from normal images, but their signals can mix real anomalies with harmless visual variation. The authors decompose PatchCore residuals into sparse features, use a multimodal LLM to label each feature as anomaly, distractor, or uncertain, and then suppress distractors while amplifying anomaly features. Across 40 categories, the joint interventions raise macro-average image-level AUROC from 0.8724 to 0.8857 on source data and from 0.8066 to 0.8210 under synthetic corruptions.
In briefDual-vocabulary language model for distilling knowledge between models with different tokenizers
SummaryThis paper addresses misalignment in on-policy distillation when teacher and student models use different tokenizers. It proposes the Dual-Vocabulary Language Model (DVLM), which swaps the teacher's output head for a student-vocabulary projection and uses Parallel-Tokenized Sequences with Hybrid-Prefix Attention to align both inputs and output logits. According to the reported experiments, the DVLM teacher reaches a converged loss similar to the original teacher, and student models improve on six reasoning tasks.
In briefAnnealed Sinkhorn with momentum: certified unregularized optimal transport in linear memory
SummaryThis paper analyzes Bregman Douglas-Rachford splitting (BDRS) for unregularized discrete optimal transport and introduces a primal-dual certificate that can be evaluated in linear memory. The authors show BDRS matches a warm-started inexact proximal point method and can be viewed as annealed Sinkhorn with an implicit inverse-linear temperature schedule. In pixel-level color transfer experiments, BDRS achieves lower repaired transport cost than MDOT-TNT on 1024×1024 images, and processes 4238×2365 images on a single NVIDIA L40S GPU.
In briefSame Runs, Different Results: Testing AI Coding Agents on Open-Weight Models
SummaryThis paper repeats one machine-learning task 584 times, asking AI agents to improve the training code of an airline-delay XGBoost classifier. It compares six agents across six open-weight model endpoints. The authors find that run-to-run variation within a single agent-model pairing exceeds differences between pairings, so a few runs cannot reliably rank them. They recommend evaluating agents and models as pairings over repeated attempts, reporting compliance alongside quality.
WhyAcross 584 controlled runs, it quantifies agent score variance, shows few runs rank agents unreliably, and finds rule-breaking runs scored highest.
Eduardo Ari\~no de la Rubia (Central European University), Szilard Pafka (Epoch)
In briefEyes on the Road: How Motorcycle Riders Look at Traffic in Indian Cities
SummaryThis paper presents the first large-scale study of how motorized two-wheeler riders direct their gaze in natural, mixed urban traffic in India, using the myEye2Wheeler dataset. A YOLOv11 and SAM2 segmentation pipeline extracted object-level gaze measures under direct gaze and central vision. Central vision supports broad monitoring, while direct gaze enables brief, selective sampling. Experience mainly refines the timing of scanning rather than changing how attention is allocated.
In briefSteering speaking rate in autoregressive speech synthesis without retraining
SummaryAutoregressive text-to-speech systems produce natural speech but offer little control over speaking rate after training. The paper proposes adjusting speed at inference time by clamping one decoder block's activation along a discovered speed axis to a fixed value. The approach largely preserves speaker identity, generalizes across model architectures, and keeps high naturalness in objective and human evaluations.
Francesco Verdini, Antonis Asonitis, Aref Farhadipour et al.
In briefCodeActionBench: Testing AI Models That Write Code to Control Robots
SummaryThis paper introduces CodeActionBench, a benchmark of 25 robot manipulation tasks that tests whether general-purpose multimodal models can turn vision and reasoning into executable robot code. Agents receive no task-specific fine-tuning, demonstrations, or specialist perception modules. Across nine configurations and 675 attempts, success rates ranged from 2.7% to 73.3%, and trajectory analyses highlight failures in spatial alignment, object retention, and completion judgment.
In briefShowing versus doing: how demonstrations change when teaching
SummaryThis paper studies how people teach through demonstration, not just by performing a task. The authors build a Bayesian model that distinguishes intentionally teaching (showing) a task from simply performing it (doing). In two experiments, human participants changed their teaching behavior in line with the model's predictions. Standard inverse reinforcement learning algorithms also learned better from showing than from doing.
Mark K. Ho, Michael L. Littman, James MacGlashan et al.
In briefMinkowskiPE: Positional encoding for space-time perception
SummaryThis paper proposes Minkowski Positional Encoding (MinkowskiPE), which uses joint time and space coordinates to parameterize Lorentz transformations applied to query and key features. Attention scores then depend only on the relative spacetime displacement between tokens, making them invariant to global translation. The method is evaluated on molecular dynamics and video prediction, with the authors reporting best results on all nine multi-trajectory molecular evaluations and a 9.9% MSE reduction on KTH video prediction using about one-tenth the parameters of the best baseline.
In briefDiffusion Reward Models (DRM): Using diffusion to capture multimodal human preferences
SummaryThe paper introduces DRM, a Diffusion Reward Model that treats reward modeling as conditional density estimation, using a lightweight Diffusion Transformer to denoise Gaussian noise into a reward vector without parametric assumptions on the output distribution. Across five benchmarks, DRM matches or surpasses baselines under matched data and backbone and recovers multimodal reward structure that conventional point-estimate heads miss. Experiments also show that using DRM as the training-time reward improves downstream RLHF policy performance.
In briefBaryBind: Aligning Multiple Data Types via a Shared Wasserstein Barycenter
SummaryThis paper presents BaryBind, a method for aligning more than two data types, such as text, video, and audio, into one shared semantic space. It maps each modality onto a Wasserstein barycenter that minimizes average distances to all modality distributions, then uses the volume of a simplex built around this barycenter as a global alignment measure. Experiments report competitive results on retrieval, classification, video question answering, and cross-modal generation, along with robustness when a modality is missing and scalability beyond three modalities.
In briefIs KL divergence really necessary for on-policy distillation of LLMs?
SummaryThis paper examines whether KL divergence is required as the loss for on-policy distillation (OPD) of large language models. The authors find that preserving the update direction toward the teacher is sufficient, especially for the small set of tokens where teacher and student strongly disagree. Based on this, they propose Consensus Multi-Teacher On-Policy Distillation (C-MOPD), where every sample is supervised by all teachers, and report that it outperforms MOPD on math and code benchmarks.
In briefRelevance Is Not Sufficient Evidence: Spotting Missing Facts in RAG Before Answering
SummaryThis paper examines retrieval-augmented generation, where retrieved passages may name the right entities but lack the facts needed to answer. It reports that 12 generators still answered 40.0-99.3% of insufficient-evidence questions even when told to abstain. The authors build a paired benchmark and propose RINSE, which combines question coverage, answer presence in passages, and a small language model's judgment to assess sufficiency before generation.
WhyProposes RINSE, a pre-generation sufficiency judge, beating 10 baselines and a frontier model (0.837 vs 0.746 AUC) across six datasets.
In briefMulti-aperture PPG with MAPIS: how light intensity, pulse signal and signal coherence vary across sensor apertures
SummaryThis study analyzed 44 thirty-second recordings from one MAPIS device, which samples several aperture-dependent optical signals at once. Pulsatile signal grew sublinearly with DC intensity, so lower-intensity apertures had higher perfusion index, and those apertures also showed higher temporal coherence. Infrared signals were consistently more coherent than Red signals. The authors say these within-device spatial relationships are averaged out in conventional single-channel PPG, and that controlled studies are needed to explain the photon-path mechanisms.
