In briefGeometry-guided method for 3D instance segmentation that transfers 2D semantics into 3D
SummaryThis paper addresses 3D instance segmentation, where models need both geometric structure and strong semantics. It treats transferring 2D semantic priors into 3D as a masked reconstruction task: part of the lifted background priors is masked, and the 3D encoder is trained to recover them. It also introduces geometry-modulated semantic injection, which uses raw 3D geometry to modulate the lifted 2D priors before sparse encoding. The authors report state-of-the-art results in full-data and few-shot settings on ScanNetV2 and ScanNet200, and the code is public.
In briefA Shared-Datastore Approach to Scaling LLM Multi-Agent Systems
SummaryThis paper addresses poor scaling in LLM-based multi-agent systems and proposes ScaleMAS, where asynchronous agents coordinate through a shared datastore of linked messages. Agents can publish, search by content with a bounded result size, and visit messages by address, with results ranked by relevance and citation count. The authors report that across four agentic tasks, ScaleMAS outperforms baselines at lower inference compute, scales to 128 agents, and continues working when agents fail midway.
In briefElephant in the Fridge: fixed-memory frame packing for long video understanding
SummaryLong videos make vision-language models use visual tokens that grow linearly with length, causing high memory use and latency. This paper proposes FramePack-U, a training-free method that plugs into existing models and compresses any video into a fixed-size visual representation. It ranks frames by text relevance and diversity, then packs the selected frames at varying resolutions into a fixed context window. The authors report gains over full-resolution baselines and prior methods on LongVideoBench, MLVU, and Video-MME.
In briefMulti-agent AI teamwork when colleagues give unreliable information
SummaryThe paper argues that current LLM-based multi-agent systems assume collaborators always give correct, complete, and prompt information, which is unrealistic. It introduces UnCoBench, a benchmark organizing failures into Misaligned, Incomplete, and Missed settings, along with an automated, verifiable scenario generation pipeline. Tests across ten models show that similar reliable-setting performance can lead to different robustness to unreliable evidence, and that models often omit corrective follow-up queries.
In briefForecasting Solar Output by Modeling Regional Generation Fields
SummaryThis paper addresses ultra-short-term forecasting for densely distributed photovoltaic systems by converting distributed generation measurements into image-like regional generation fields and modeling their spatiotemporal evolution. The SatPV-Field framework combines historical PV fields, satellite-derived future shortwave radiation, and meteorological forecasts to predict regional generation states, which then guide individual station forecasts. Experiments on 496 distributed PV systems show a station-level MAE reduction of about 10.0% over the best single-station temporal baseline and about 2.5% over a graph-based regional method.
In briefLessons from Luck: Whether LLM-written lessons pass chance outcomes into an agent's next choice
SummaryThis pre-registered study examines whether self-improving agents, which write lessons from their run outcomes, carry chance results into later decisions. Across 256 synthetic worlds with two tools of recorded reliability, models wrote lessons that then guided the next tool choice. After a chance failure of the more reliable tool, the lessons carried that chance into later decisions. After a lucky success of the less reliable tool, the effect was bounded below 0.10, which is a bound rather than evidence of no harm.
In briefGUIDE: A Computer-Assisted Consensus System for Trustworthy Clinical Guidelines
SummaryThis paper presents GUIDE, a multi-agent copilot system that helps humans and AI collaboratively write and update clinical practice guidelines. Specialized AI agents handle labor-intensive tasks while experts retain judgment and final decisions. On 12 NICE ophthalmology datasets, GUIDE reached 88.85% literature recall and 91.67% recommendation agreement, using about 162 minutes per dataset with 28.68% human intervention. A prospective real-world validation reported outcomes comparable to human-only updating, with less time and effort.
In briefSilence and Noise: Fixing Suppression and Overvaluation in Multimodal Entity Alignment
SummaryThis paper studies multimodal entity alignment across knowledge graphs, arguing that missing modality features and structural differences between graphs interact and harm fusion. It proposes ALMEA, which uses cross-modality imputation and latent semantic calibration in a bidirectional loop, plus controlled iterative selection for pseudo-labeled pairs under sparse supervision. The authors report consistent gains over state-of-the-art baselines, with up to 5.20% MRR and 6.40% Hits@1 improvements at a 20% supervision rate.
In briefEmbedding-space noise for better exploration in LLM reinforcement learning
SummaryThis paper addresses a problem in RLVR for LLMs: with small rollout groups, samples from the same policy tend to be similar, weakening learning signals. The proposed method, EDEN, replaces a small share of standard rollouts with ones generated from the prompt plus one perturbed embedding position, using Gaussian noise and a shared anchor. Across model scales and math reasoning benchmarks, it outperforms GRPO and other exploration baselines, with the largest gains in sampled solution coverage.
In briefSelectTTRL: Cheaper test-time reinforcement learning by selecting which problems to train on
SummaryTest-time reinforcement learning lets language models adapt on unlabeled test problems using pseudo-labels from sampled responses, but applies the same costly process to every problem. The authors find that a cosine-similarity score between a model's response representation and a correctness-related direction correlates with problem solvability across models and datasets. They propose SelectTTRL, which ranks test problems by this score and applies standard TTRL only to a chosen percentile range. Experiments on multiple reasoning benchmarks reportedly cut test-time computation while keeping or improving standard TTRL performance.
In briefPolanyi Meets Galois: Tacit Knowing as Algebraic Deterministic Operation for Re-rooting AI Science
SummaryThe paper argues that tacit knowing is better modelled as an operating regime of a memory than as content, using VaCoAl, an XOR-and-shift hyperdimensional memory over GF(2). Repairing collisions gives exact auditability, while tolerating them yields a path confidence, CR2, formalizing Polanyi's claim that tacit and explicit knowledge are not sharply divided. It also argues that AI-solved results lacking follow-ups are rootless explicit knowledge, and proposes a division of roles: LLMs generate, VaCoAl/PyVaCoAl keeps an auditable trace, and humans value the results.
In briefCloneOT: neural inverse optimal transport that learns clone-specific perturbation responses from lineage tracing
SummaryThis paper proposes CloneOT, a method that predicts how single cells respond to perturbations using partially lineage-traced data. It learns both the transport cost, represented by a neural network, and the transport map, so that cells from the same clone stay together in the cheapest transport. The authors report that it recovers correct transport on 2D datasets where squared Euclidean cost fails, and outperforms five baselines across 14 cancer-cell lineage-tracing perturbation scenarios.
WhyLearns transport cost from lineage barcodes, overcoming fixed Euclidean cost; compares against five baselines across 14 cancer perturbation scenarios, giving solid evidence for single-cell perturbatio
In briefMemorization in Image Classification: Test Set Behavior and Scaling Laws
SummaryThis paper studies how convolutional neural networks memorize training data when trained on image classification tasks with randomized labels or pixels. The authors find that independently initialized networks trained on the same random labels make different predictions on test examples, unlike networks that generalize. They also analyze how dataset size affects the network capacity and training time needed to fit the training set, finding that the number of presentations needed for full memorization can decrease as the dataset grows, even without shared structure between inputs and outputs.
In briefThe Sharp Directions Are Against You: A Curvature Analysis of Activation Steering in Language Models
SummaryThis paper studies activation steering, which adds a vector to a language model's internal representations at inference time to control behavior. The authors introduce a steering Hessian to measure how sensitive behavior is along each direction, finding that sharp directions produce erratic, input-dependent swings. Removing the sharp component keeps the vector's direction and magnitude nearly unchanged while reducing curvature, and this one-line correction outperforms standard contrastive activation addition in open-ended generation.
Seyedarmin Azizi, Arya Fayyazi, Parsa Razmara et al.
In briefLearning routers that select among diffusion experts for multi-domain medical image generation
SummaryThe paper studies two ways to train a router that combines predictions from a fixed pool of domain-specialized diffusion experts for medical image generation. Server-Only learns routing from server images with component labels, while Anchor-Delta also uses denoising losses from frozen experts as extra evidence. In a heterogeneous four-domain setting, the Anchor-Delta router improves generation metrics over Server-Only.
In briefHeterogeneous Agentic Coevolution: Mutual Improvement of Specialist Models and Learning Agents
SummaryThis paper proposes HACE, a framework in which specialized foundation models and LLM agents evolve together through reciprocal learning. The specialist is improved with supervised learning, while the agent is trained with on-policy reinforced self-training guided by verified performance gains of the current specialist. Evolutionary search is also used to optimize how the two models exchange information and select training experiences. The authors evaluate HACE on recommendation systems, protein function prediction, and crystal stability prediction, reporting that it outperforms the compared self-improving methods on primary task metrics in all three domains.
In briefForecasting Combined Hospital Burden from Influenza, COVID-19 and RSV
SummaryInfluenza, COVID-19, and RSV circulating together share hospital beds, staff, and testing capacity. The study compares summing pathogen-specific forecasts with directly forecasting combined weekly admissions for 1 to 4 weeks ahead. The direct combined-burden model performed better on most metrics in both retrospective and prospective evaluations. The authors argue that pathogen-specific and combined forecasts should be used in layers.
In briefDream2Discern: Training a Failure Predictor on Hard Latent Variants for Multi-Agent Systems
SummaryThis paper proposes Dream2Discern, a runtime auditor that predicts the final outcome of LLM-based multi-agent systems from partial execution records. It generates hard latent variants by perturbing its own representation operators to increase prediction loss, reusing the original labels without collecting new trajectories or verifying outcomes. Evaluated on AFTraj-2K and StepShield, it outperforms competitive failure auditors and reports over 90% trajectory-level weighted accuracy on both benchmarks.
In briefReducing Misalignment When Training AI to Persuade
SummaryThis paper studies how reinforcement learning for persuasion affects models' use of misaligned tactics such as evidence fabrication and belief coercion, and which techniques reduce them. Optimizing senders only for receiver belief shifts raised the number of distinct misaligned strategies per game by over 25% and more than quadrupled fabrications relative to pre-RL models. The authors report that combining guidance toward aligned persuasion strategies with reward penalties for misalignment gave the best trade-off among the methods compared.
In briefWhen Agent Evaluation Becomes Unreliable: Oracle Failures in LLM Agent Benchmarks
SummaryThis paper audits the evaluation software used to score LLM agents, proposing five state-integrity invariants and three known-answer controls to detect oracle defects. It injects seven defect classes across task domains and detects all of them. On a widely used benchmark, the authors find a reward function that resets a database without restoring it, so scoring reads the reference state, lowering reported success from 94.3% to 65.5%.
In briefPIGEONS: A model that writes long-video captions step by step, keeping characters and story consistent
SummaryLong videos are hard to caption in detail because current video models have limited context windows, and splitting them into segments often leads to inconsistent subject references and broken storylines. PIGEONS first scans the whole video to build a gallery of key subjects and their visual references, then captions clip by clip while using that gallery and a continually updated overarching narrative summary. It is trained in three stages combining supervised fine-tuning and reinforcement learning. Experiments report better long-video captioning accuracy, completeness, and subject referential consistency than existing methods, with competitive results on short videos.
In briefInterpretable graph partitioning for explaining spatial predictions
SummaryThis paper addresses the difficulty of interpreting flexible spatial prediction models by building a simpler global surrogate through graph partitioning. It selects connected subgraphs over location-level model predictions to minimize within-cluster residuals. A constraint removes artificial connections between regions while keeping flexible boundaries, and the authors analyze its theoretical effect and provide a boundary-adjustment procedure. Experiments show the method produces spatially coherent partitions.
In briefFrom Failure Prediction to Component Selection in Semantic Segmentation: A decision-focused framework for judging which segmentation predictions to trust
SummaryThe paper addresses which segmentation predictions to trust, keep, or inspect, arguing that a failure predictor's accuracy alone does not capture its value for selection. It proposes an evaluation framework that separates what a score predicts, what can be learned from its existing inputs, and what its ranking delivers. Tested on Cityscapes and ACDC at 90% retention across three systems, adapted MetaSeg reduced retained failures compared with forecast entropy but did not outperform direct probability ranking.
In briefHow the classical baseline affects machine-learning speed claims in topology optimization
SummaryThe paper re-ran released code for eight topology optimization methods and studied two in depth: the diffusion model OAT and a graph-based neural optimizer. Replacing the classical code with an efficient implementation of the same algorithm reduced these two methods' speed-ups by up to 20-fold. At matched stiffness, the graph-based optimizer was slower than the classical code in all six two-dimensional problems. The authors also provide a comparison procedure and a minimum reporting standard.
In briefCross-Protocol Personality Assessment: Aligning Behavioral Evidence with Human Ratings via LLMs
SummaryThis paper studies personality assessment from asynchronous video interviews, a method used in recruitment. Because different interview protocols use different questions and trait mappings, human-rated evidence from one protocol cannot be used directly with another. The authors propose CEPA, which projects behavioral records into the target protocol's representation space and retrieves relevant human-rated examples to ground an ensemble of LLM raters. In cross-protocol experiments, CEPA reduces average mean squared error by 43.8% (Forward) and 17.0% (Reverse) relative to the strongest baseline.
In briefAdaptive Compression and Targeted Perturbation: A Unified Framework for Generalized Audio Deepfake Detection
SummaryThis paper proposes a unified framework for detecting audio deepfakes, combining an adaptive feature module that selects latent dimensions and a mel-spectrogram adversarial perturbation strategy. Evaluated on ADD 2022, FakeOrReal, and In-the-Wild benchmarks, the Whisper-small-based model reports a state-of-the-art average Equal Error Rate of 1.51%. The authors state it uses up to 91.6% fewer parameters than the 3B-parameter Resemble-Detect-3B-Omni while improving average EER by 60.7%.
In briefBackward conformal prediction under covariate shift, extended with importance weighting
SummaryBackward conformal prediction fixes a prediction set's size and estimates how often it misses the true label, but assumes calibration and test data are exchangeable. This paper extends the method to covariate shift using importance weighting, keeping its guarantees when the density ratio is known. The authors propose estimating density ratios on sorted nonconformity scores instead of high-dimensional inputs, proving this recovers the same result asymptotically under an interpretable condition. Experiments show the method improves miscoverage estimates, and score-based weighting outperforms embedding-based weighting.
In briefAutomatic 3D tracing and numbering of all 24 rib centerlines in chest CT
SummaryThis paper presents ribline, an open-source framework that automatically extracts the centerlines of all 24 rib bones from thoracic CT and labels each rib (R01–R12, L01–L12). It combines an anatomically anchored coordinate frame, a hybrid voxel detector, and a bilateral Viterbi dynamic program for global rib labeling and curve refinement. On the Siemens 2 mm thoracic CT benchmark's evaluation split, it reports 97.2% labeling accuracy and a labeled Chamfer distance of 2.05 mm, with zero false-positive curves.
In briefBeyond the greenwash: Do LLMs' eco-friendly decisions hold up under pressure?
SummaryThis paper examines whether large language models make ecologically sound decisions when under operational pressure, a property the authors call 'bioalignment'. The authors continue-pretrain the same 30B mixture-of-experts base model at graded training doses and test each arm on irreversible decision scenarios. They report that the heaviest-trained model breaks most often, an inversion that persists across several test variants. They also find that behavioral scores detect surface-level veneers but do not distinguish training doses.
WhyProposes a known-dose calibration ladder to measure depth of ecological bias, finding an inverted dose-robustness relation; evaluated on a single 30B base model.
In briefHow GPT models portray mental health stigma by gender, compared with human survey data
SummaryThis study examines whether GPT-4 and GPT-5 represent mental health stigma differently across gender groups. In Study 1, both models portrayed men most negatively on social distance, stereotypes, and attitudes toward help-seeking. Study 2 compared AI estimates with human data from the Healthy Minds Study (N = 76,406) and found partial alignment, with the models underestimating personal and perceived stigma and overestimating self stigma. The authors conclude the models give a distorted picture and need rigorous, context-specific validation against real-world data.
In briefAutoregressive LLM decoding as a martingale: how generation randomness behaves
SummaryThe paper defines a per-step prediction error as the difference between the sampled token and the model's predicted distribution, and proves the cumulative error is a martingale. It derives a closed-form quadratic variation that depends only on realized token probabilities, linking sequence fluctuations to predictive uncertainty. White-box experiments on Llama-3.x and Qwen2.5 models from 1B to 72B parameters show the measured second-order moments closely match the theory.
In briefCLNO: A conservative neural model for predicting compressible flows
SummaryThis paper proposes a Conservative Local Neural Operator (CLNO) for autoregressive prediction of compressible flows. The method splits the flow into fixed-size local subdomains, uses a shared network to predict interface fluxes that enforce discrete conservation, and reconstructs fine-scale structures with a convolutional network. Reynolds and Mach number conditioning lets a single model generalize to unseen parameter combinations without retraining. Compared with LNO, the authors report competitive or higher accuracy, with order-of-magnitude error reductions in some cases, at about three to five times the inference cost on periodic benchmarks.
In briefOT-FlowRec: Optimal Transport-Guided Multi-Bridge Flow for Multimodal Sequential Recommendation
SummaryThis paper addresses how e-commerce user preferences shift within sessions and form multiple interest centers, which existing sequential recommenders capture poorly. It proposes OT-FlowRec, which models multi-interest distributions with parallel, interest-specific optimal transport flows and a geometric transport mechanism toward preference prototypes. A progressive time-sampling warmup strategy biases training toward the final generation states to align with top-k ranking metrics. The authors report that it outperforms state-of-the-art baselines on three public datasets.
In briefA nonlinear time-series theory for constant-step SGD with momentum
SummaryThis paper treats constant-step stochastic gradient descent with momentum (SGDM) as a nonlinear time series, studying the coupled parameter and momentum recursion. Under a finite p-th moment for p greater than 1, the process forgets its initialization geometrically, yielding a unique stationary causal solution. Under a second-moment condition, this leads to an invariance principle and quenched Gaussian limit for averaged SGDM, and a small-learning-rate analysis gives a Gaussian limit for the last iterate. In a stochastic-curvature quadratic model, the authors relate the rate of initialization forgetting to stationary variance and averaged fluctuations.
WhyTreats constant-step momentum SGD as a nonlinear time series, deriving stationary solutions and Gaussian limits; theory is fairly complete with a quadratic-model analysis.
In briefLimit theory for streaming estimation of expectiles on dependent data
SummaryThe paper studies a one-pass stochastic-gradient algorithm with constant learning rate for estimating the marginal expectile of a stationary causal process. It shows the recursion has a unique stationary causal solution and forgets its initialization geometrically fast. Under short-range dependence, the last iterate is asymptotically Gaussian at the square-root-of-gamma scale, while for long-memory linear processes the scale and a fractional Brownian motion-based Gaussian limit are derived. Numerical experiments are presented to support the predicted approximations.
In briefOf Errors and Echoes: Responsibility-Aware Credit Assignment for Tool-Use Agents
SummaryThis paper introduces ECHO (Error-Chain-Aware Optimization), a reinforcement learning framework for tool-use agents that separates identifying responsibility from allocating credit. A frozen judge distinguishes newly introduced errors, inherited consequences, missed corrections, and successful corrections. Experiments on six tool-use benchmarks at 4B and 8B scales show ECHO achieving the highest score on five of six benchmarks at each scale.
WhyResponsibility-aware credit assignment substantially improves step-level methods; consistent gains across six benchmarks and two scales, impact mostly within tool-use agent research.
In briefRemoving the Compute Bottleneck for Event-based Visual Place Recognition with Spiking Neural Networks
SummaryThis paper presents a two-layer spiking neural network, trained with additive STDP, for event-based visual place recognition on conventional CPUs. Magnitude pruning yields sparse weights so that each query resolves in tens of microseconds. On three benchmarks, it exceeds the recall of the LENS method and cuts per-query latency by 5.5 to 6.2 times. Its recall matches a dense variant and the SAD baseline while reading about ten times fewer weights than the dense variant.
Somayeh Hussaini, Tobias Fischer, and Michael Milford
In briefWhen Does Compensating Inherited Quantization Error Help? Guarantees and Limits for Layer-Wise Quantization
SummaryThis paper studies whether layer-wise post-training quantization that compensates for error inherited from earlier layers can keep network output error no larger than without compensation. For linear networks with zero-mean, deterministic-second-moment rounding error and fixed compensation strengths in [0,1], the mean-square output error does not increase. The authors show the comparison can fail with ReLU activations or bounded codebooks, and report that on two 8B models at 4 bits, fixed half-strength compensation beat uncompensated rounding, while no strength-selection rule consistently won.
WhyGives exact conditions and gain formula for mean-square guarantees of error compensation in linear networks, shows failures with ReLU and bounded codebooks, and refutes a published universal guarantee
In briefCertifying Neural Network Robustness with Faster Second-Order Cone Relaxations
SummaryThis paper studies certified robustness for neural networks, which requires verifying behavior under input perturbations. It proposes a hierarchy of second-order cone programming (SOCP) relaxations that give tighter certificates than linear programming while avoiding the high cost of semidefinite programming. On MNIST and CIFAR-10, the method improves certified accuracy over LP-based verification, is competitive with alpha,beta-CROWN, and is much faster than SDP-based verification. The authors also introduce CORE, which applies tighter relaxations only to classes not certified by cheaper methods.
Mohammadreza Maleki, Daniel Sadig, Rushendra Sidibomma et al.
In briefComputational Reflective Equilibrium for Agentic AI: A Rawlsian Framework for Governing AI Agents
SummaryThis paper proposes agentic reflective equilibrium (ARE), a framework that serves as an ethical alignment scaffold inside the operating harness of AI agents built on large language models. Inspired by John Rawls's reflective equilibrium, it links principles from system prompts and operator policies, considered judgments, tool evidence, and candidate actions through support and attack relations. The authors show that equilibria correspond to Nash equilibria of a mutual-adjustment game and establish results including a quarantine theorem for prompt injection. Two case studies, a refund-claim assistant and a hiring agent, illustrate the framework.
In briefMeaningful Moments: VLM-based temporal importance labels for video action recognition
SummaryThis paper builds Meaningful Moments, a corpus of about 4.58 million vision-language-model-generated per-segment importance scores for nearly 500,000 videos across SSv2, Kinetics-400, and Diving-48 spanning 622 action classes. With the action classifier frozen and only temporal sampling varied, the authors find the VLM importance signal is informative, with usefulness depending on class structure. Gains are strongest on Diving-48 and modest on K400, while on SSv2 the labels do not recover full-video performance. The labels and evaluation protocol are released under CC-BY-4.0.
In briefCalibrating Biomedical QA Evaluation Metrics with Few Human Preferences
SummaryRetrieval-Augmented Generation systems are increasingly used for biomedical question answering, but reliable reference resources are lacking and automatic metrics often disagree with human judgments. The paper explores using a small amount of human preference signals to calibrate direct assessment evaluation metrics. The calibrated metrics consistently outperform baselines that treat all metrics equally or rely on a single best metric across multiple datasets.
Shakila Mahjabin Tonni, Xiang Dai, Sarvnaz Karimi et al.
In briefPAIR: Memory-efficient, fast inference via reading-reasoning fusion
SummaryThis paper proposes PAIR, a method for context-based question answering that splits the work between two differently sized models. A smaller Reader model processes and retains the full context, while a larger Reasoner model sees only the question and generated sequence. Their predictions are combined at the logit level so the pair behaves as a single next-token model. On HotpotQA, PAIR with a 32B Reasoner keeps 96% of the full-context 32B model's accuracy at about one quarter of its KV memory, and reaches up to 3.2× higher throughput at 45.7K-token contexts.
In briefA 4-qubit quantum classifier for telling apart DNMT3A and DNMT3B methylation sequence preferences
SummaryThis paper proposes a quantum machine learning approach using a Variational Quantum Classifier to distinguish the flanking sequence methylation preferences of DNMT3A and DNMT3B enzymes. The authors built a dataset of 4,096 NNNCGNNN motifs and evaluated a four-qubit, two-layer circuit on an exact-statevector simulator. Across five initialization seeds it achieved a ROC-AUC of 0.661 ± 0.009, and the work is presented as a proof of concept for near-term quantum methods in epigenetic classification.
Murat Isik, Yusuf Moinuddin, Humaira Gowher et al.
In briefRepresentational Difference Distillation: focusing distillation on regions where teacher and student representations disagree
SummaryKnowledge distillation transfers representational knowledge from a large teacher network to a compact student. The paper proposes Representational Difference Distillation (RDD), which uses two asymmetric affinity matrices to find samples the student wrongly treats as similar and teacher relationships the student has not yet learned, then concentrates the loss on those. On CIFAR-100 from resnet32x4 to resnet8x4, RDD reaches 75.58% test accuracy, close to CRD (75.45%) and above KD and RKD. The authors also report that weighting the loss works better than filtering data.
In briefDissipative Lagrangian Neural Networks for Interpretable Modelling of Passive Mechanical Systems
SummaryThis paper introduces the Dissipative Lagrangian Neural Network (dLNN), which learns the conservative and dissipative dynamics of mechanical systems separately and directly from data. Its dissipative-force parametrisation is structurally constrained to ensure the learned model is passive and cannot absorb conservative mechanics. The model was tested on a chaotic double pendulum and a six-degree-of-freedom industrial robot, outperforming existing Lagrangian architectures and remaining competitive with a black-box neural model. The authors also add a symbolic distillation procedure that extracts closed-form expressions from the learned robot model, yielding interpretable models with high accuracy and lower computational cost.
In briefLLM-generated behavioral patterns for epidemiological modeling, evaluated against multiple benchmarks
SummaryThis paper examines whether behavioral outputs generated by large language models can help represent individual-level behavioral differences in epidemiological models, using hepatitis C-related behaviors among people who inject drugs as a case study. The outputs were checked against three external benchmarks: literature-based association directions, correlations from an agent-based model, and survey associations. The authors find only partial alignment with epidemiological evidence, with higher agreement when LLMs were directly asked to classify associations, so more systematic evaluation is needed before using such outputs in transmission models.
Sophia Mao, Keshav Gandhi, and Alexander Gutfraind
In briefHandoff Curves: How Initial Model Choice, Later Handoffs, and Training Targets Shape Multi-turn LLM Routing
SummaryThe paper studies how initial model selection, later handoffs between models, and training targets affect success and cost in multi-turn LLM tasks. It compares router designs using handoff curves, based on 34,204 complete executions of three Qwen models across BFCL, MultiWOZ, and τ-retail. With Qwen3.5 35B-A3B as the starting model, later handoffs did not improve utility over initial selection on BFCL or MultiWOZ and increased cost on BFCL. Changing the supervision targets improved router performance on both the authors' dataset and TwinRouterBench.
In briefProbabilistic Wildfire Size Forecasting Using Conditional Flow Matching
SummaryThis paper proposes a generative model based on conditional flow matching that forecasts the time series of wildfire acreage rather than spatial fire spread, reducing computational cost. Inputs include current and previous fire size estimates, spatially averaged weather, terrain, and fuel composition, and the model predicts area change over 3-hour intervals. It is trained on 140 coupled WRF-SFIRE atmosphere-wildfire simulations paired with NAM weather data. It achieves a single-step sMAPE of 0.39 against WRF-SFIRE references and the same final-area sMAPE of 0.39 over 24-hour recursive forecasts on 12 held-out cases, while quantifying uncertainty.
In briefMixed emotions in one text break the usual shortcut of mapping each emotion to a single positive or negative polarity.
SummaryThe paper examines where mapping emotion dataset labels onto positive and negative valence works and where it fails, using several tweet and annotated corpora plus a controlled study. Fixed projection works when labels share one valence but fails on tweets whose labels point in opposite directions, which are 9.0% of polar SemEval test tweets yet cause 54.8% of a learned projection's polarity errors. The authors recommend projecting label configurations rather than single labels, and abstaining or keeping a mixed class where labels conflict.
In briefHiding Messages in a Policy's Randomness While Keeping Its Action Distribution Unchanged
SummaryThis paper studies using the randomness of a stochastic policy as a medium for embedding and communicating information. The authors propose posterior-matching randomness modulation (PRAM), which selects actions according to a message while using common randomness independent of the control system, so the prescribed action distribution is preserved for every state and message. The paper proves that PRAM is asymptotically reliable, meaning the message can be decoded with arbitrarily small error as the observed state trajectory grows long. Experiments show the framework's effectiveness for multi-bit policy watermarking and trajectory provenance.
In briefPopGen: Generating Semantically Diverse Agent Populations for Zero-Shot Coordination
SummaryThis paper presents PopGen, a framework that uses large language models to create semantically diverse agents from scratch. It first generates textual descriptions of distinct agent behaviors, then turns each into an executable code policy. Cooperators trained against PopGen populations achieve robust zero-shot cooperation with unseen partners and outperform population-play baselines across four collaborative environments.
Paul Teiletche, Kanghyun Ryu, Kyungjoon Park et al.
In briefSoftmax-free spiking attention as the zero-temperature limit of two-level systems
SummaryThis paper examines spiking Transformers, which drop softmax normalization in favor of thresholded binary attention to save energy. The authors argue this removal breaks the thermodynamic framework used to analyze attention, and prove that softmax-free spiking attention is the zero-temperature limit of independent two-level systems. They place it in the frozen ordered phase and link this to optimization difficulties from vanishing surrogate gradients. Forward-relaxation experiments are used to compare standard and neuromorphic attention on a shared thermodynamic axis.
Roberto C. Sotero, Pouya Taghipour, Jose M. Sanchez-Bornot
In briefScaling pretraining on repeated data: batch size, regularization, and GPU-time tradeoffs
SummaryThis paper studies pretraining when compute is plentiful but unique training data is limited, so the same data must be trained on repeatedly. The authors find that jointly tuning batch size and epoch counts yields a U-shaped loss curve, favoring smaller batches than usual, and that recipes such as repacking documents each epoch, weight tying, and positive embedding weight decay can be tuned. When compute is treated as finite, they report that small batches may need prohibitively long sequential training, and that under a fixed time budget a smaller batch with fewer GPUs can beat a larger batch with more GPUs.
In briefSmall but Mighty: Test Suites as Improvement Signals for Self-Evolving Agents
SummaryThis paper examines how a self-evolving agent decides which rewrites of its own harness to keep, using test feedback as an executable signal. The authors propose TeSearch, a quality-diversity evolutionary search that finds test suites that are both small and strong, treating suite size as a behavior descriptor. Across 129 real-world modules, TeSearch produced the highest-quality suites among compared methods and the smallest among LLM-based ones. As the improvement signal, it rejected 59.9% of behavior-changing rewrites versus 25.3% for the repository's own tests, with a 4.0% false-alarm rate and 3.5 s per decision.
In briefMachine-learning track reconstruction in CMS using Line Segment Tracking objects
SummaryHigh pileup at the HL-LHC makes charged-particle tracking in CMS harder. The authors use Line Segment Tracking (LST) triplets rather than individual hits as inputs to a Transformer with masked sparse self-attention and object condensation, assembling full tracks in one inference step. On CMS Phase-2 simulated ttbar events at pileup 200, it reaches efficiency comparable to the LST quintuplet baseline, with the fake rate reduced by roughly a factor of three and duplicates almost entirely suppressed.
In briefUnfair and Unsafe: Everyday Bias Triggers Emergent Misalignment — how polite stereotypes in fine-tuning data make models less safe
SummaryThe paper studies whether fine-tuning a language model on polite answers that rely on stereotypes causes emergent misalignment, making it less safe and less truthful on unrelated tasks. Experiments across four open-weight model families (7B to 12B) covering gender and race stereotypes show the effect relative to controls trained without the stereotype. The authors conclude that biased training data is a safety problem as well as a fairness problem, and that existing defenses improve safety but leave the stereotype largely intact.
WhyShows polite stereotype training data triggers emergent misalignment, with controlled comparisons across four model families and an activation direction identified.
Alex Mart\'i Guiu, Nicolas Zumarraga, Kevin O'Sullivan et al.
In briefHarmonic Representations in Symbolic Music Transformers Are Localized and Controllable
SummaryThis paper studies how Aria, a LLaMA-based symbolic music transformer evaluated on Bach chorales, internally represents harmonic context at cadences. The authors find that harmonic decisions become causally localized from middle layers onward, with mode linearly decodable in the same layer range while key identity emerges later. Steering the model along latent directions corresponding to harmonic changes reliably produces deceptive or minor resolutions while preserving structural coherence and generation quality.
In briefADAPT: An audio-depth navigation planner with transition-aware rewards for dark industrial sites
SummaryThis paper addresses navigation in lights-out industrial settings where visible light may be unavailable, using spatial audio to locate targets and depth to perceive obstacles. The authors introduce DarkNav-Bench, a benchmark with 24 industrial scenes and four progressively harder acoustic conditions, and propose ADAPT, a reinforcement learning framework for closed-loop navigation. Its transition-aware reward shaping (TARS) accounts for both feasible moves and in-place rotations, and ADAPT outperforms training-based baselines and zero-shot multimodal models. On unseen scenes it reaches 90.6% success in clean conditions and 60.0% under the hardest localized noise setting.
In briefInstance-adaptive online multiclass multicalibration
SummaryThis paper studies online multiclass multicalibration over T rounds and proposes an algorithm with guarantees that go beyond the worst case. If the adversarially chosen sequence of outcome means lies in an affine subspace of dimension s, the algorithm adapts to this structure and achieves lower error without knowing the subspace or its dimension. For a stochastic instance with a fixed conditional mean, the bound reduces to a square-root dependence on the horizon.
WhyProposes an algorithm adapting to affine subspace dimension of outcome means, giving near-minimax-optimal rates; impact mostly theoretical.
Zhiming Huang, Jamie Morgenstern, Aaron Roth et al.
In briefSmall Area Learning: A Learner-Agnostic Framework for Small Area Estimation
SummaryThis paper introduces small area learning (SAL), which frames small area estimation as a multi-target semi-supervised learning problem with outcome prediction treated as a nuisance task. It allows covariate shift and structured conditional shift across areas, sharing information through common and area-specific components. The authors establish consistency and asymptotic normality for area mean estimators and describe a bias-variance trade-off in residual correction. Simulations show 34%–49% RMSE reductions relative to direct estimation.
Katarzyna Reluga, Nicola Salvati, Setareh Ranjbar et al.
In briefPaper explainer: how proactive assistants can gauge how sure they are about user preferences
SummaryThis vision paper argues that proactive smart assistants, such as home assistants, need more than predicted preference likelihoods to act well. They also need to know how reliable a prediction is, whether it rests on group-level assumptions or individual data, and whether ambiguity reflects natural preference variability or insufficient data. The authors draw on uncertainty estimation in neighboring areas such as trajectory prediction and personalized activity recognition to envision future uncertainty-aware assistants.
In briefPhysics-guided machine learning surrogate for saturation control in heap leaching under varying hydraulic conditions
SummaryThis paper presents a parameter-conditioned physics-guided machine-learning (PGML) surrogate that predicts hourly saturation profiles in heap leaching, using four hydraulic parameters. It was trained on 4096 numerical Richards-equation simulations and tested in simulated saturation control. In evaluation cases, the PGML controller's median tracking deviation matched a tuned PI controller at 4.9% of the 0.70 target, while applying about 4% less irrigation. The surrogate runs about 70 times faster than the numerical solver.
Sultan Alhamdan, Daniel Navia, Paulina Quintanilla
In briefNLP tool to check dental reading lists against learning goals and topic balance
SummaryThis paper presents a natural language processing workflow that checks whether dental course readings match learning objectives and whether topics are covered in a balanced way. It combines PubMedBERT embeddings with cosine-similarity retrieval, natural language inference for evidence checking, and topic modeling, tested on 13 objectives and 3,837 abstract pairs. Similarity alone was insufficient, and adding NLI raised accuracy from 0.69 to 0.73 against expert labels on 49 pairs.
In briefAnatomy of a Blind Spot: What LLM Judges Need to See
SummaryThis paper examines what information an LLM acting as an evaluator must be shown to judge generated text reliably. Across seven public benchmarks, the authors varied only the material the judge received and found this mattered more than which judge was used. They pre-registered 83 testable clauses, of which 32 of the 81 that ran met their bar, and report the shortfalls alongside design rules.
In briefError-based bidirectional predictive coding for deep 4-bit networks
SummaryThe paper combines bidirectional predictive coding with error-based reparameterization so predictive coding networks can train at greater depth and under 4-bit quantization. On six-layer networks, the method substantially outperforms state-based bPC, which collapses to chance. Quantization to 4 bits causes only a small accuracy loss.
In briefDiffNash: Diffusion Policies for Nash Equilibria in High-Dimensional Continuous Games
SummaryThis paper proposes DiffNash, a diffusion-based framework for computing Nash equilibria in high-dimensional continuous games, where multi-modal strategies are hard for conventional neural policies to represent. It combines Online Mirror Descent with Fictitious Play by reusing past actions through a replay buffer, and corrects the resulting distribution shift using an estimate based on Evidence Lower Bound differences. Evaluated on forest protection games and continuous Colonel Blotto games, it reduces the Nash gap relative to the strongest baseline by 35.6% and 24.7%, respectively.
In briefMesh-Free Convolution: Learned spectral attenuation and transport filters for PDE learning
SummaryThis paper proposes Mesh-Free Convolutions (MFCs), filters defined in continuous space and discretized only at evaluation time, so the same learned filters work across resolutions. The authors report that on fluid-dynamics and weather-forecasting benchmarks, MFC-based models give stable long-horizon autoregressive rollouts, transfer reliably across resolutions, and better preserve long-time statistics.
In briefExplainable LightGBM forecasting of border queues at multiple time horizons, with resource-aware decision support
SummaryThis paper presents a forecasting and decision-support framework for border crossing queue management. It uses LightGBM models built separately for 15, 30, 60, and 120 minute horizons to jointly estimate future waiting times and queue lengths, while accounting for lane compatibility, officer availability, traveler category, and preregistration. Quantile models provide 95% prediction intervals for waiting-time forecasts. The framework outputs machine-readable recommendations, alerts, and key performance indicators, and establishes a reproducible synthetic-data baseline for later validation with operational border-crossing data.
Varvara Mama, Fani D. Christidi, Nikolaos C. Kapsalis et al.
In briefWhen Trackers Fail: Checking and Recovering Video Object Tracking with Vision-Language Models
SummaryThis paper addresses video object tracking errors, where a wrong mask from foundation models like SAM can persist and cause identity drift after occlusion or distractors. The proposed Dual-Agent framework adds a vision-language verification agent that checks whether the tracker still follows the target, and a correction agent that recovers the target when a failure is detected. According to the paper, it improves SAM3 by 1.2 and 2.7 J&F on LVOS-v2 and SecVOS, and transfers without retraining to other trackers and long-term benchmarks.
Valay Bundele, Susmit Agrawal, Mehran Hosseinzadeh et al.
In briefFrequency-Step Cache: separating sync and repair to speed up autoregressive video diffusion
SummaryThis paper addresses the high cost of autoregressive video generation, which requires repeated passes through large diffusion transformers. It proposes Frequency-Step Cache (FSC), a training-free caching method that schedules exact full-computation anchors separately from the repair decisions made between them. On SkyReels-V2, FSC-Speed reaches 2.36x speedup versus MotionCache's 2.275x with higher SSIM and PSNR, and FSC-Quality improves fidelity metrics at matched speed.
In briefFast and Accurate: Splitting Personalized LLM Work Between RAG and Fine-Tuning by Semantic Concentration
SummaryThis paper proposes PUR, a method that splits a user's historical data by semantic concentration. Core, concentrated information is fine-tuned offline with parameter-efficient fine-tuning (PEFT), while dispersed information is kept for retrieval-augmented generation (RAG). At inference, core queries are answered directly and only peripheral queries invoke RAG. The authors report the best generation accuracy and quality across multiple tasks, with inference latency reduced by 14.4% to 46.6% versus the best prior method.
In briefRINO: Label-free physics-based self-supervised learning for particle jet classification
SummaryThe paper proposes RINO, a self-supervised method that uses the multi-scale renormalization group structure of high-energy jet formation to create training views. It pretrains only on QCD jets from unlabeled data and tests transfer to rare signal classes. The authors report that with only 1% of labeled JetNet data after finetuning, RINO matches fully supervised training, and it outperforms the strongest self-supervised baseline on top-quark jets and the H→4q decay channel.
WhyFirst to exploit renormalization group multi-scale jet structure for self-supervised views, with substantial OOD and data-efficiency experiments.
Zichun Hao, Raghav Kansal, Abhijith Gandrakota et al.
In briefA learning method for many overlapping groups that needs only an ERM oracle without extra samples
SummaryThis paper studies multi-group learning, where one predictor must perform nearly as well as the best hypothesis in a class on every group in a rich, possibly infinite and overlapping collection of subpopulations. The authors give an algorithm that uses Õ((d_G+d_H)/ε²) samples and polynomially many calls to a weighted ERM oracle, and show a matching lower bound, so oracle efficiency costs no additional samples. They also treat the special case where each input belongs to at most k groups, where an ERM oracle for the hypothesis class alone suffices.
WhyGives an ERM-oracle-only multi-group learner matching a lower bound, resolving the tension between oracle efficiency and sample optimality theoretically.
In briefViewCraft: Teaching Multimodal Agents What to Look At and How to Read It
SummaryThe paper introduces ViewCraft, which evolves both an executable evidence interface that selects and organizes observations and the procedure used to interpret them, without updating model weights. Failure traces guide revisions, and replay with paired evaluation keeps useful changes. Across eight benchmarks it improves task score by 6.1 points on average over direct execution, and on five multi-source tasks the learned interface uses 8%-15% fewer input tokens. The learned skills also transfer across executors.
In briefMore Reasoning Is Not Always Better: Running Clinical Guidelines with LLMs and Rules
SummaryThis paper proposes MedDEx, a framework that combines structured patient states with deterministic rule execution to support guideline-based treatment planning. LLMs extract patient information and generate treatment plans, while guideline-defined decision logic is handled by a rule executor. Experiments on multiple clinical guideline benchmarks show improved decision accuracy and lower inference costs.
In briefRIVO: Lightweight Adaptive Rerouting for Deep Research Agents
SummaryThis paper addresses the difficulty Deep Research agents have in deciding when to keep searching, stop, or change direction during long, multi-step information-seeking tasks. It proposes RIVO, a lightweight plug-in that estimates the agent's research state and, when needed, adjusts the host model's token logits to steer generation toward a more promising direction without retraining the base model. The authors also release BrowseCore, a curated dataset with a paired local corpus, and report experiments across several Deep Research benchmarks.
In briefTwo Linear Programs Beat One: A Study of Virtual Extension Complexity
SummaryThis paper studies the virtual extension complexity of polytopes, defined as the minimum sum of the extension complexities of two polytopes Q and R satisfying P+Q=R, a notion motivated by neural network complexity. The authors show that virtual extension complexity can be exponentially smaller than extension complexity, and that some matroid base polytopes have exponential virtual extension complexity. They also prove that the maximum virtual extension complexity of an n-vertex polygon is Θ(√n).
Alexander E. Black, Christoph Hertrich, Stefan Kober et al.
In briefLURE: A Phase-Labelled Synthetic Corpus of Grooming Conversations
SummaryThis paper presents LURE, a synthetic corpus of groomer-victim exchanges containing 2,121 unique conversations across seven tags and four grooming phases, each with a Plutchik emotion field. It was generated with a retrieval-augmented pipeline based on the ChatCoder luring-communication scheme, with human supervision and validation. The authors report tag accuracy of 0.85 and phase accuracy of 0.93, and a real-versus-synthetic discriminator reaching 91% accuracy. The corpus is released with vetted access only.
Emanuele Florindi, Valentina Franzoni, Ludovico Guercio et al.
In briefReducing Agents' Role Overstepping in Multi-Agent Systems via Dynamic Activation Steering
SummaryThe paper addresses a problem in multi-agent systems where agents perform subtasks outside their assigned roles, which disrupts workflow and lowers efficiency and task quality. The authors find that role instructions cause consistent shifts in the model's activation space, and that role-adherent and overstepping trajectories show distinct activation patterns. They propose Dynamic Role Activation Steering (DRAS), an inference-time method that builds a steering vector and overstepping detector per role and adjusts steering strength during generation. Experiments report up to 86.6% fewer overstepping trajectories, 10.0%–11.1% higher task quality scores, 1.50×–1.61× faster task completion, and 26.8%–38.7% lower token consumption compared with Vanilla MAS.
In briefLightweight scheduling policies for continuously batched LLM inference on different GPUs
SummaryLLM serving performance depends increasingly on how requests are scheduled during autoregressive decoding, not only on model size or GPU speed. The authors built a PyTorch continuous-batching testbench and compared eight lightweight scheduling policies across many GPU, model, and workload combinations without changing model architectures or deploying learned schedulers. They found that scheduler choice shifts throughput, time-to-first-token, and tail latency, with aging shortest-prompt-first performing strongest in many settings. The custom dataset has been released on Hugging Face.
In briefGen Through the Lens: Camera-controllable video generation via cross-coordinate conditioning fusion
SummaryThis paper presents Gen Through the Lens, a framework that generates videos from a reference image with controllable camera motion. It fuses control signals from camera, world, and object coordinate spaces to guide camera movement, focus, and framing. The authors report that the method yields controllable, visually consistent results and outperforms existing approaches in realism and editability.
In briefReview Rarely Beats the Stronger Model: Verification vs. Capability Transfer in LLM Pipelines
SummaryThis paper examines whether having one language model review another model's answer actually helps. The authors argue that measuring review by its gain over the generator confuses verification with capability transfer. Across 48 pairs of 12 open-weight models in math, knowledge, code and reasoning, review beat the stronger model alone in only 9 of 96 directions, and did worse as the capability gap grew, because the reviewer mostly follows the generator. Review costs 2.6 to 2.8 times as much, while a selector using the same two calls is more accurate and cheaper.
In briefLearning to prioritize code repair edits so fewer LLM revisions are needed
SummaryThis paper introduces CodeTriage, a learned search policy for LLM-based code repair that ranks proposed edits before the revised programs are generated and tested. It uses test output, failure diagnosis, and the proposed edit to predict how many repair steps each edit needs to reach a fix. On held-out repair tasks, it finds fixes with 21% fewer program revisions and 12% fewer output tokens than the strongest baseline. The authors also release a corpus of 12.2 million program versions from four LLMs..
In briefOptimal Rates and Better Spectral Convergence for Learning Intersecting Manifold Structures
SummaryThe paper studies estimating eigenpairs of a tensorized, density-weighted Laplace-Beltrami operator from samples of a multi-manifold mixture where manifolds may intersect. For simple nontrivial eigenpairs it proves a minimax rate and shows a kernel-density plug-in estimator attains it when the manifolds are known. A manifold-agnostic graph Laplacian estimator is also analyzed, with a convergence rate improving on previously known results.
In briefMultimodal sensor fusion with JEPA features for breathing and pose recognition
SummaryThis paper transfers frozen V-JEPA2 video features to RGB video, depth video, and radar time series, using a shared fusion model to classify breathing activity and pose across all seven non-empty sensor combinations. Evaluated on 39 subjects from the OMuSense-23 dataset, the final model reaches an average activity macro-F1 of 75.23%. Supervising all sensor subsets in each batch improved results by 1.923 percentage points, while the tested Consensus-JEPA loss showed no measurable benefit.
Yunxiang Luo, Constantino \'Alvarez Casado, Xiaoting Wu et al.
In briefVariance-optimal state resampling for RL with verifiable rewards, proposing a method to focus training on informative states.
SummaryThis paper addresses wasted rollout budget in reinforcement learning with verifiable rewards, where many states give no learning signal. It derives the state sampling distribution that minimizes policy-gradient estimator variance under a fixed budget, giving a closed-form Neyman allocation. Based on this, the authors propose VSR, a plug-in modification to GRPO using a prefix replay buffer and a zero-cost variance proxy. The authors report that VSR outperforms GRPO and GiGPO on most math reasoning and long-horizon agentic benchmarks (Sokoban, BabyAI) while generating fewer tokens per update.
In briefComparing two neural network models that predict indoor airflow and temperature for rooms of different sizes
SummaryThis study compares two neural network surrogate models that quickly predict indoor airflow and temperature fields, as a faster alternative to computationally expensive CFD simulation. The canvas-based model (CaNN) encodes room size by keeping occupied regions within a fixed canvas, while the coordinate-based model (CoNN) uses normalized coordinate channels; both use U-Net autoencoders. On a controlled benchmark of simplified rooms with varying heights and lengths, CaNN showed higher agreement with CFD results than CoNN.
In briefEngram Injection: Temporal Validity in Parametric Language-Model Memory
SummaryThis paper studies whether a language model can store facts that become outdated over time in its parameters and answer correctly using only a question and the current time. The authors store facts in LoRA adapters and supply the clock via FiLM conditioning, comparing it with additive clock conditioning, full dated-history prompting, and retrieval of four dated records. Both adapter methods answered none of 62 before-and-after update pairs correctly, while full-history prompting answered 59 and retrieval answered 62. The authors conclude that learning a fact is separate from learning when to use it.
Sam Siavoshian, Omar Ramadan, Amin M. E-A. Diab et al.
In briefLearning equilibria in entropy-regularized average-reward mean-field games
SummaryThe paper studies stationary discrete-time mean-field games with finite states and actions, an entropy-regularized objective, and population-dependent dynamics. Under a minorization condition, the soft Q-operator contracts in span seminorm, and when the kernel's Lipschitz constant and temperature meet stated thresholds, an equilibrium exists, its population distribution is unique, and fixed-point iteration converges geometrically. The authors also learn the equilibrium model-free from a simulator, proving a finite-sample bound for soft Q-learning and convergence rates for the learned distribution and policy. A congestion-aware routing game illustrates the results.
In briefUnchecked or Falsely Accused? Evaluating AI Text Detectors on Kazakh
SummaryThis paper introduces AINA-KK, a Kazakh dataset of 3,811 human passages confirmed to predate ChatGPT and 5,597 machine or gray-zone passages from nine generators. It finds that several English-trained public detectors wrongly flag many human texts, while recalibrated versions catch very little machine-generated text. The authors propose AINA-KK-Det, a detector built from fixes that need no new data, which flags a small share of human passages and catches most machine text from known generators. Data, code, and the detector are released on GitHub.
WhySystematic evaluation of public detectors on Kazakh with pre-ChatGPT human texts, plus a combined fix; strong evidence and practical impact.
In briefTelling Process Change, Sensor Fault, and Data Corruption Apart in Streaming Telemetry
SummaryThis paper addresses three reasons a sensor reading can look abnormal: the physical process changed, the instrument is failing, or the value was corrupted in transit. It shows that one channel cannot separate certain pairs of causes and proposes a set of online detectors with a deterministic attribution rule, now merged into a public algorithm library. On 2,942 injected events, the method detects 96.3% and attributes 91.2% correctly from one channel, rising to 99.0% with a second rate channel.
In briefComparing two ways LLM recommenders generate item semantic IDs
SummaryThe paper compares two ways an LLM-based recommender can generate Semantic IDs: single-head autoregressive decoding with beam search, and multi-token prediction with parallel heads. In an industry-scale multi-task model over a streaming-media catalog, single-head beam search generally outperforms MTP on ranking and content QA, and trains at lower cost, while MTP only wins on inference throughput. The authors also report end-to-end gains over a non-SID baseline and address a language-quality regression with reinforcement learning.
Sejoon Oh, Fernando Amat Gil, Evan Smothers et al.
In briefA hybrid machine learning model for predicting buckling knockdown factors and reliability-based design of thin spherical shell caps
SummaryThin spherical shell caps often buckle at loads well below linear analysis predictions, and this gap is corrected with a knockdown factor (KDF). Existing KDFs such as NASA SP-8007 and Eurocode 3 are considered highly conservative. This study trains machine learning models on experimental spherical shell buckling data and proposes a hybrid of Gaussian Process Regression and XGBoost that outperforms several conventional algorithms. A sensitivity analysis identifies the most critical input feature, and a reliability-based design framework uses the model to obtain less conservative yet safe KDFs.
Budhaditya De, Rohan Majumder, Devasmit Dutta et al.
In briefAttention Sinks in VLMs: Borrowed Focus, Not Real Viewing
SummaryThis paper studies visual attention sinks in vision-language models, which are image tokens that absorb much of the attention regardless of the query. The authors compare six detectors and find that only centered cosine similarity between an image token's key and the language model's text-sink key reliably identifies them. They propose a training-free, entropy-triggered re-look method that improves accuracy on fine-detail visual question answering benchmarks.
Dengjia Zhang, Alexander Martin, Pengyu Yan et al.
In briefSharp frames can hurt 3D room reconstruction; viewpoint spread matters more
SummaryThis paper studies frame selection for feed-forward multi-view 3D reconstruction of indoor rooms. Selecting the sharpest 16 of 200 frames performed worst, with 48% room dimension error versus 3% for uniform temporal sampling, because sharp frames cluster in slow camera motion and leave parts of the room unobserved. A factorial experiment found viewpoint spread has roughly three times the main effect of sharpness, and neither inference-time uncertainty signal revealed this failure.
In briefWhat sparse autoencoder feature discovery scores actually identify, and how they relate to intervention effects
SummarySparse autoencoder (SAE) features are usually found by how their activations co-vary with model behaviour, but they are often used through interventions such as steering or ablation. The authors take intervention attribution, the first-order effect of an intervention on model output, as their target and show that observational discovery evidence generally does not identify it for a broad class of response functions. They derive an exact gradient-based decomposition of the gap between linear-probe predictions and intervention attribution, and report that probe-to-attribution ratios are large across four model-SAE pairs.
WhyProves observational SAE feature evidence generally fails to identify intervention attribution, validated empirically across four model-SAE pairs with gradient-based decomposition.
In briefMovement Aids Learning: Label-free Adaptation of Object Recognition Models Through Motion
SummaryThe paper proposes MoCaFe, a method that lets movable agents such as autonomous vehicles and robots adapt pre-trained object recognition models to new environments without human annotations. It exploits the variance in prediction quality when the same object is seen from different distances and viewpoints, fusing cross-view predictions with inverse-variance weights to build reliable pseudo-labels. An optional extension, p-MoCaFe, uses a public dataset to reduce pseudo-label corruption. The authors report significant gains over state-of-the-art test-time adaptation methods on nuScenes, KITTI, and embodied-agent datasets, with accuracy improvements of up to about 20%.
In briefHow neural models and the human brain bind phoneme identity to sequence position
SummaryThe paper studies how neural systems bind phoneme identity to position in a sequence, so that sequences like /gum/ and /mug/ are represented differently. The authors extend the NLP 'onion hypothesis' to natural language and, in an encoder-decoder model, propose a 'spiral hypothesis' in which position is encoded by magnitude together with a systematic rotation of the state direction. Analysis of intracranial recordings from 20 participants' auditory cortex shows that response changes between consecutive phonemes decrease in magnitude as position advances.
WhyExtends the onion hypothesis to natural language, proposes a spiral hypothesis, and validates it with intracranial recordings, giving strong evidence but mainly subfield impact.
Nafiseh Momeni, Malo Renaudin, Itzhak Fried et al.
In briefLLM value functions for step-level credit assignment in agent RL
SummaryThe paper introduces LVRL, which uses a pretrained LLM's knowledge and in-context learning to estimate policy-aware state values without finetuning a separate critic, enabling step-level credit assignment from value differences. It also proposes LVOPD, extending the idea to on-policy distillation by focusing teacher supervision on actions linked to decreases in estimated value. The method is evaluated on AppWorld, BabyAI, and SWE-bench Verified across policy scales from 4B to 32B. The authors report that LVRL outperforms GRPO and PPO on these benchmarks.
