In briefMedical Image Segmentation: Comparing Adapted Foundation Models with Task-Specific Training
SummaryThis paper compares DINOv2, a foundation model pretrained on large natural-image collections, with U-Nets and ResU-Nets trained from scratch for medical image segmentation. Experiments cover colonoscopy, breast ultrasound and dermoscopy, with testing on both in-distribution and external datasets. The adapted foundation model generalised better on every external dataset, while task-specific models stayed competitive in-distribution when data were abundant. The authors also find that the required adaptation capacity depends on the imaging domain, and that a widely used polyp test set duplicates images from another benchmark.
In briefToward Foundation Detection Models via Example-Based Grounding in EM-DETR
SummaryThis paper presents an approach to building a multi-domain foundation model for lesion detection by combining medical image datasets annotated for both anatomical structures and abnormalities. A student-teacher training strategy is used to merge datasets with heterogeneous label spaces while mitigating catastrophic forgetting. The authors adapt their earlier example-based EM-DETR framework to learn jointly from multiple domain-specific datasets. The model reports image-level AUC gains of 1 point over prior state of the art on four chest X-ray findings and leading mAP50 on chest X-rays and mammography.
Sheethal Bhat, Bogdan Georgescu, Awais Mansoor et al.
In briefTorchCurves: A PyTorch library for learnable parametric curves
SummaryThis paper presents TorchCurves, a PyTorch library that treats vector-valued parametric curves as learnable building blocks. It maps scalar inputs to points on curves expressed in B-spline or Legendre polynomial bases, with gradients flowing through both inputs and coefficients. The coefficients can be optimized together with a model or generated by another network. Timing experiments compare its evaluation speed with NumPy and SciPy and its Kolmogorov-Arnold network training speed with EfficientKAN across several problem sizes.
In briefAutoScaffold: AI-Guided Training Hints for Improving Language Model Reinforcement Learning
SummaryThis paper introduces AutoScaffold, a framework in which a frozen teacher model periodically reviews a student model's training experience, diagnoses bottlenecks, and designs two kinds of training-time scaffolds: strategies shared across task classes and hints for individual hard tasks. The scaffolds are removed at evaluation. The authors test the framework on ALFWorld, search-based QA, mathematical reasoning, and GPU kernel generation, reporting that students trained this way outperform corresponding RL baselines when evaluated without scaffolds.
WhyThe authors report a frozen teacher auto-designing training-time scaffolds, with gains across four domains evaluated without scaffolds.
In briefContrastive Discovery: AI agents comparing rival explanations to solve science questions
SummaryThis paper addresses a limitation of LLM agents in data-driven scientific discovery, which test each hypothesis only against its negation. It introduces Contrastive Discovery, which replaces a single hypothesis with a set of mutually exclusive and collectively exhaustive competing explanations, and resolves a question when evidence concentrates belief on one of them. Across six domains, the approach resolves more questions to a specific alternative from the same evidence, and under the same experiment budget it resolves more questions than pointwise surprisal.
In briefPyrite: Benchmark and AI Agents for Ranking Drug-Protein Binding Poses
SummaryThe paper introduces Pyrite, a benchmark of 1,500 pose pairs across diverse protein-ligand complexes for testing how well candidate binding poses can be ranked. Expert chemists using 3D visualization and physics-based scores performed near chance. The authors develop two LLM-based agents, PyriteReasoner and PyriteFeaturizer, which reach up to 67% and 74% accuracy respectively, exceeding observed expert performance. LLM-guided feature discovery generally outperforms direct reasoning, and GRPO training helps open-weight models match or surpass closed-source frontier models.
Megha Srivastava, Michael H. Lim, Allison M. Keys et al.
In briefDoes the Real Wiring Win? Testing Connectome Structure–Function Claims Against Unmeasured Physiology
SummaryThe paper examines whether connectome-constrained models really show that measured wiring outperforms shuffled or rewired wiring, given that model physiology is often tuned on the real wiring. The authors compare three estimands and test them with synthetic ground truth, placebo nulls and three Drosophila circuit models. Some claims hold under every protocol, such as segregated dopamine improving learning, while a mushroom body advantage for real wiring reverses once physiology is constrained by independent data.
WhyThe authors propose a three-estimand protocol showing connectome structure-function conclusions can reverse with tuning, validated on synthetic, placebo and Drosophila models with confidence intervals
In briefContinual-ARC: Benchmark for Measuring Lifelong Skill Acquisition
SummaryThis paper introduces Continual-ARC, a lifelong learning benchmark built from three ARC-AGI editions whose tasks require no domain knowledge. Tasks recur after dormant gaps with fresh instances, and learners are scored by the share of their feedback budget they consume, allowing memory-based and weight-based systems to be compared on one scale. Experiments find that harness persistence generally does not improve memorization, and no harness-based method gains on unseen tasks, while weight-based learning shows substantial gains.
In briefBiological abilities that appear unexpectedly in a model of molecular interactions
SummaryThe paper presents Bonbon, a foundation model trained on about 3.2 trillion tokens of paired protein-ligand sequences. Without task-specific supervision or structural coordinates, it shows zero-shot emergent abilities at functional, residue, and bond levels, which generalize to out-of-distribution complexes. The authors report sharp performance jumps at the largest of six model scales, even as pretraining loss decreases smoothly. In prospective wet-lab tests across nine disease targets, de novo compounds reached a 61.5% dose-response hit rate.
WhyThe authors extend emergent capabilities to a molecular interaction foundation model and report sharp scaling jumps, plus wet-lab validation across nine targets.
In briefSpatial patterns of transport mode choice in Niterói: a synthetic-population simulation study
SummaryThe paper presents a computational framework that uses aggregated, static urban data to build a detailed, dynamic model of urban mobility. Applied to about 475,000 synthetic residents of Niterói, Brazil, the simulated neighborhood-level transit patterns correlate strongly with an empirical activity proxy. Walking is more common in compact central areas, car use rises toward the periphery, and transit use stays relatively high along bus corridors. Reducing walking times to and from bus stops by 30% raises the modeled transit-to-car ratio by 11.3%.
Sandro Martinelli Reia, Taylor Anderson, Dieter Pfoser
In briefConnectome-constrained fly vision model: robustness, efficiency, and linking adversarial and perceptual illusions
SummaryThis paper benchmarks FlyNet, a task-optimized continuous-time recurrent network constrained by the fly visual system's synaptic connectome, against CNNs, RAFT, and recurrent networks lacking connectome constraints on a collision-detection task. FlyNet is adversarially vulnerable but, relative to its compute, more robust to common video corruptions than RAFT, using up to five orders of magnitude fewer FLOPs. The authors show that the same low-level motion statistic underlies both adversarial texture failures and the reverse-phi illusion.
Alexandre Garcia-Duran, Alejandro Rodriguez-Garcia, Manuel Molano-Maz\'on et al.
In briefRS-JEPA: A Compact Self-Supervised Model for Matching Remote Sensing Images Across Different Sensors
SummaryThis paper presents RS-JEPA, a compact self-supervised framework for retrieving scenes across different remote sensing sensors, such as SAR and optical imagery, without manual labels. It uses a Vision Mamba encoder shared by both sensors and introduces scan-path views, which treat two reading orders of the same image as two views to predict hidden patches. Experiments on SAR-optical and PAN-MS benchmarks show the highest F1@5 among twelve re-implemented methods, with a retrieval network of 3.88M parameters.
Tanishq Rachamalla, Srishti Kaushik, Muhammad Attique Khan et al.
In briefComparing three ways to test AI analytics agents for failures and security risks
SummaryThis paper compares three evaluation frameworks for LLM-based analytics agents used in enterprises: trace-based LLM judging, text-only LLM judging with red-teaming, and deterministic heuristics. The study runs in three phases, measuring evaluation stability, detecting known risks on 72 ground-truth test cases, and discovering unknown risks. The authors conclude that the information available to the evaluation harness shapes evaluation quality, that functional and security evaluation are complementary, and they recommend combining trace-based functional assessment with adversarial red-teaming.
In briefReTrace: Turning Tool-Using AI Agents into Planner-Free Workflows
SummaryThis paper presents ReTrace, a method that converts a natural-language Standard Operating Procedure and a few logged agent runs into a reusable workflow in which code, not a large language model, chooses each next step. It extracts the condition behind each step from the agent's reasoning, merges these conditions into a compact rule document, and generates an executable workflow from it. The authors report that across several benchmarks the workflows achieve completion or success rates close to the agent's while lowering cost and running time.
Kaiyuan Hou, Yang Ping Kuo, Indranil Bhattacharya et al.
In briefTimeSeek: A Multi-Task, Multi-Domain Pretrained Model for General Time Series Analysis
SummaryTimeSeek is a general-purpose model for time series analysis that adapts to patterns from multiple domains. It uses a Multi-Domain Feature Extractor to capture temporal dependencies and channel correlations specific to each domain. Through multi-task collaborative training, it unifies forecasting, anomaly detection, and classification in a single pretraining stage. The authors report state-of-the-art results on these tasks in few-shot and zero-shot settings, with favorable inference efficiency.
In briefRe-Search: A two-player questioning game that dissects how LLMs generate research ideas
SummaryThe paper adapts the Twenty Questions game into Re-Search, a benchmark where a model must rediscover a published paper's core idea from its research problem by asking yes-or-no questions. Tested on 200 ICML 2026 papers across six recent LLMs, most progress occurs within the first 30 of 100 questions, after which most models plateau. The authors find the bottleneck is generation rather than judgment, and that models fail by either over-exploiting one direction or over-exploring.
In briefMechanistic Interpretability Needs a Logic of Inference: A paper proposing five principles and a recursive framework for drawing mechanistic explanations from neural network observations
SummaryThis paper examines how mechanistic interpretability (MI) researchers can move from observing internal neural network components to forming mechanistic explanations. The authors compare efforts in neuroscience and MI and identify recurring inferential bottlenecks when inferring mechanisms from complex neural systems. They propose five principles organized into a recursive logic for more rigorous mechanistic inference. The framework aims to turn local findings into cumulative knowledge and generate testable hypotheses beyond the models, but the paper concludes that evidence beyond MI is needed to establish what model mechanisms reveal about the world.
In briefPterodactyl: A Low-Cost System for Automated Translation Quality Evaluation
SummaryThis paper presents Pterodactyl, the system submitted by 1020 Group to the WMT26 shared task on automated quality evaluation. It targets subtasks 2 and 3, namely segment-level quality score prediction and detecting error-free segments. The system uses an LLM-as-a-judge approach, with prompt engineering and model selection guided by ROC analysis, and is designed as a highly customizable, ultra-low-cost binary classifier.
In briefFERN: Table Retrieval for Text-to-SQL Across Benchmarks Using Frozen Retrievers and Learned Ranking
SummaryThis paper presents FERN, a method for retrieving the tables needed by text-to-SQL systems from large pooled catalogs. It fuses three frozen retrievers into 100 candidates, describes each with up to 120 features, and reorders them with LightGBM rankers trained on BIRD, without generative LLMs. The authors compare its table recall against the published results of RASL across Spider, BIRD, Fiben, and BEAVER, reporting that FERN matches RASL on Spider at five tables and reaches 98.0 recall at fifteen tables on Spider's candidate pool.
In briefBoviShift: A Six-Year Multi-Attribute Audit of Cattle Vision Across Animals and Time
SummaryThe paper introduces BoviShift, a dataset rebuilt from six consecutive annual cohorts of a cattle marketplace from 2021 to 2026, containing 2,657 cattle, 10,610 still images, and 2,566 walk-around videos. It compares image-random and cattle-disjoint data splits and tests how a ResNet-50 model transfers to future cohorts. The evaluated attributes are weight, age, height, breed, and coat color.
In briefTanCRAFT: Predictive Feedback to Prevent Errors in Assembly Tasks
SummaryTanCRAFT is a four-stage real-time system that helps people avoid mistakes during physical assembly. It tracks the scene, reasons about the partial assembly, predicts where a manipulated piece will move, and intervenes when that placement conflicts with the target. The system was tested on a tangram assembly task through offline benchmarks and a within-subject user study. Predictive feedback significantly reduced errors compared with unguided and reactive sessions, and it produced the lowest mean perceived workload and shortest mean completion time.
Sassan Mokhtar, Lars Doorenbos, Marius Bock et al.
In briefC²-Routing: Complementary Cropping and Compression for Efficient Vision-Language Inference
SummaryThis paper addresses the heavy computational cost of visual question answering in vision-language models by proposing C²-Routing, which adaptively chooses between image cropping and whole-image compression based on the input image and question. A safety gate uses predicted joint correctness under original and reduced inputs to fall back when needed and protect accuracy. Across three VLMs and 18 datasets, on top of a 1,024 visual-token cap, it cuts end-to-end FLOPs by 36.53% and visual tokens by 42.36% on average, with only a 0.77% mean relative task-score drop.
In briefBayesian inference of epidemic threshold for a spatial SIS infection model
SummaryThis paper studies Bayesian inference of whether an infection persists, using early observations from a heterogeneous SIS reaction-diffusion model. The authors show that early prevalence depends on the unknown parameters only through the difference between transmission and recovery rates and the infected diffusivity, which also determines the principal eigenvalue governing persistence. The paper establishes well-posedness of the model and the Bayesian problem, and proves O(ε) Hellinger convergence of nonlinear posteriors to low-prevalence limits. Numerical examples illustrate threshold inference.
In briefA Theory of Policy Optimization in Post-Training: Coverage Collapse and Reference Resets
SummaryThis paper theoretically analyzes policy optimization in language model post-training, defining "coverage-preserving" algorithms whose sample and gradient costs depend on how readily the base model generates optimal responses. It shows that REINFORCE, the basis of PPO and GRPO, can drive the probability of optimal responses near zero even as reward rises, a failure called coverage collapse. The authors propose resetting the KL reference to the latest policy within an Approximate Policy Mirror Descent framework, and use it to prove global convergence for several widely used post-training algorithms.
WhyThe authors introduce coverage collapse and reference resets, giving first global convergence guarantees for softmax linear policies under APMD.
In briefA theory of how much cross-validation estimates vary
SummaryThis paper studies how much the estimate from cross-validation, a method that repeatedly splits data into training and validation sets to measure predictive performance, fluctuates. It proves an exact finite-sample identity for the estimate's variance, showing it depends on the weights that random splits give each observation. From this, K-fold cross-validation is first-order variance optimal, leave-one-out has the smallest variance among K-fold schemes under reasonable conditions, and leave-pair-out is optimal for AUC. The estimate is also shown to be asymptotically normal.
WhyThe authors derive an exact finite-sample variance identity for cross-validation, prove K-fold and leave-one-out variance optimality, and report asymptotic normality covering MSE and AUC.
In briefAccurate Prediction of Energetic Material Properties Using a Tabular Foundation Model
SummaryThis paper applies TabPFN, a tabular foundation model, to predict impact sensitivity, decomposition temperature, and crystal density of energetic materials. Without hyperparameter tuning, TabPFN outperforms the best published models and established methods such as tuned XGBoost, random forests, and Gaussian processes. The model also provides built-in uncertainty quantification because it predicts the full target distribution, and Shapley values are used to interpret feature importance.
Vagn Krister Nes Glomsrud, Mathias Bynke, Lars Henry Berge Olsen et al.
In briefWhen does text encoding matter for language modeling?
SummaryThis paper examines how text encoding choices affect the performance of byte-level language models, comparing UTF-8 with other Unicode-based, language-specific, and custom encodings. The authors propose Aether, a hybrid scheme that uses script-specific encodings for target languages while falling back to UTF-8 for full Unicode coverage, and use it to model four languages with diverse writing systems. They find that languages with information-dense characters, such as Chinese and Korean, benefit significantly from compressive encodings, while Georgian and Tamil are largely insensitive to encoding choice. The authors conclude that text encoding should be a language-specific design decision rather than a universal default.
In briefA lightweight differentiable method for real-time skin rendering and digitization
SummaryThis paper presents a differentiable Screen-Space Subsurface Scattering framework that directly estimates scattering parameters for real-time rendering pipelines. It jointly optimizes spatially varying base textures and scattering distances, estimating them across facial regions to produce a blended profile map. The method uses a low-cost single-light, single-view capture setup and is validated on a 45-person real dataset spanning the Monk Skin Tone Scale, plus synthetic, multi-light, and tilted-view experiments.
Mario Marquez-Balduque, Carlos Rodriguez-Pardo, Jorge Lopez-Moreno
In briefUGG: a unified generative model that jointly handles video and geometry for 4D scene understanding
SummaryUGG is a unified framework for 4D scene understanding that jointly models RGB video together with geometric representations such as depth and camera pose. By assigning different noise levels across modalities and time, it treats visual generation and geometric prediction as conditional completion within one generative model. The authors report that UGG outperforms non-unified baselines on camera-controlled video generation, depth estimation, and camera pose prediction, including sparse observation settings.
In briefMetric-aligned policy optimization for explainable deepfake detection
SummaryThis paper addresses deepfake detection that must explain the visual evidence behind each decision, not just classify real versus fake. It proposes a two-stage predict-then-explain pipeline, submitted to the ACM Multimedia 2026 Explainable Deepfake Detection Challenge and trained and evaluated on the XPlainVerse benchmark. A vision-foundation-model ensemble classifies each image, and Qwen3-VL-8B generates an explanation conditioned on that decision. Metric-aligned Group Relative Policy Optimization improves the explainer, raising complex BERTScore-F1 from 0.686 to 0.712, and the submission ranked first on that metric and seventh overall.
In briefOne Policy for 100 Household Tasks: Our 2026 BEHAVIOR Challenge Solution
SummaryThis paper describes a solution to the 2026 BEHAVIOR Challenge, which includes 100 long-horizon household tasks in the OmniGibson simulator. The team placed 9th using a single vision-language-action model with one checkpoint for all 100 tasks, which scored 0.392 overall. On 50 tasks shared with 2025, its Q rose from 0.260 to 0.371. For 14 tasks, control was routed to a GPT-6 Astra coding agent, which achieved a mean Q of 0.639 compared with 0.108 for the learned policy.
In briefTwo Fields, One Scheduling Problem: Comparing Branch-and-Bound and A* for Resource-Constrained Project Scheduling
SummaryThis paper compares Branch-and-Bound (B&B), a method from Operations Research, with A*, a method from Artificial Intelligence, as exact search approaches. It uses the Resource-Constrained Project Scheduling Problem as a test case and runs experiments on 480 benchmark instances. Graph-based A* substantially outperforms the B&B variants studied, proving optimality on all 480 instances with additional pruning rules, compared with 445 for best-first B&B under the same limits.
In briefMeasuring how stable time-series foundation models are when sensor sampling frequency changes
SummaryThis paper stress-tests four time-series foundation models (MOMENT, Chronos-2, Moirai-MoE, and TimesFM) to measure how stable they remain when sampling frequency changes. The authors use three new metrics, FSI, FDI, and CDI, to assess embedding shift, forecasting degradation, and downstream performance degradation. They argue that this invariance depends on the margin between a signal's native sampling frequency and the bandwidth its task requires, and that much of the performance loss reflects information loss.
Sourish Gunesh Dhekane, Abinaya Kumar, Torsten Woertwein et al.
In briefDnNP: Denoising Input Uncertainty in Neural Processes
SummaryNeural Processes are flexible meta-learning models that quantify uncertainty, but the standard framework assumes noise-free inputs. This paper introduces Denoising Neural Processes (DnNPs), which explicitly model input uncertainty and are trained via variational inference to denoise corrupted inputs while performing meta-learning. On synthetic data and PM2.5 spatial interpolation from real air quality sensors, DnNPs improve prediction and uncertainty estimates over standard NPs, reducing RMSE by up to 30% and NLL by up to 73%.
In briefEarlybird: A 29M-Parameter Model That Predicts Speech Words Before They Are Finished
SummaryThis paper presents Earlybird, a 29-million-parameter streaming speech recognizer that fits in 16.8 MB at 4-bit precision. It is trained on audio cut off at random points and predicts each word due by then, plus the next two words. On LibriSpeech test-clean, the median word is predicted 90 ms before the speaker finishes it, while released streaming models output it 90 to 400 ms after it ends.
In briefOnlyAct: Extracting action representations from pretrained models for world simulators
SummaryVideo world models need action labels, which are scarce and hard to scale. This paper starts from the pretrained foundation model V-JEPA 2, whose frame-pair embeddings already contain much action information, instead of learning actions from scratch. A label-free contrastive adaptation method pulls together frame pairs with the same motion in different scenes and pushes apart pairs with different motion in the same scene, extracting action while discarding the scene. Evaluated on robot manipulation, camera navigation, game, and driving benchmarks, the representations are more expressive than annotation-free latent-action methods while matching their scene invariance.
In briefTesting whether AI-generated world models are structurally faithful for neuro-symbolic reasoning
SummaryThe paper tests whether five large language models can turn natural-language scenarios into formal world models, represented as partially observable stochastic games. It uses a benchmark of 44 game, situational, and cultural scenarios and evaluates outputs for mechanical validity, structural agreement with references, and semantic quality. Most generated models run correctly, but few exactly match the reference structure, so executability alone does not ensure faithful models.
Souhaila Djaffal, Amna Jamal, Gagana Madhugiri Devaraja et al.
In briefPrecision on Demand: Raising Weight Bits Only at Uncertain Decoding Steps in a Small Local Model
SummaryThe paper studies a decoding scheme for small local language models that stores 8-bit weights but normally decodes with only the top 4 bits. A step is redone at 8 bits only when the low-precision output distribution is uncertain, judged by entropy or a small top-two margin. Experiments use Qwen3-0.6B on GSM8K, with cost measured in weight bytes read per token. The authors report replayed and live results and note that the live data do not clearly separate the gate from either all-high or all-low precision.
Mohammed Nayeem Khan, Logesh L, Aliyas Ahmad Misri
In briefPost-hoc disentangling of audio foundation models for music similarity
SummaryThis paper studies post-hoc disentanglement of pretrained audio foundation models such as CLAP, MuQ, and MERT for music similarity. It adapts Conditional Similarity Networks to gate frozen embeddings into attribute-specific subspaces. Across two datasets, minimal trainable adaptation improves attribute-specific disentanglement and retrieval performance, while pooling, masking, and depth choices add limited gains.
Lucas de Souza Brand\~ao, Gabriel da Costa Merlin, Diego Furtado Silva
In briefBits or permutations? Budgeting weight precision versus option-order averaging in small local models
SummaryThis paper asks whether reducing the weight precision of small local language models hurts their sensitivity to answer-option order, and which is the cheaper way to spend a fixed inference budget. The authors test Qwen3-0.6B and 1.7B with simulated round-to-nearest quantization at 8, 6, 4 and 3 bits on ARC, MMLU and HellaSwag. They find 6-bit weights match fp16 accuracy, and in cost-matched comparisons extra option orders beat extra bits only when weights have 6 or more bits.
Mohammed Nayeem Khan, Logesh L, Aliyas Ahmad Misri
In briefCausal Concept Explanations for Deep Neural Models: A Training Method and Measure for Causal Concept Effects
SummaryThis paper argues that concept-based explanations of neural networks are often only correlational, so they can look correct while not being necessary or sufficient for a model's output. The authors propose CCW-Net, which uses interventional mediation analysis as a training objective to make concept representations necessary, sufficient, and independent. They also use mediation analysis as an evaluation tool to quantify each concept's causal contribution to a model's output. The method is evaluated on aircraft control, driving, and image classification.
In briefEquations Found for South Asian Monsoon Rainfall Thermodynamics
SummaryThis paper uses symbolic distillation, combining neural networks and symbolic regression, to derive interpretable precipitation equations from 21 years of June-August reanalysis data. The best equation explains 1.5 times as much precipitation variance as a widely used buoyancy-based diagnostic. It recovers about 85% of neural network skill while using roughly a thousand times fewer parameters.
Savannah L. Ferretti, Jane W. Baldwin, Michael S. Pritchard et al.
In briefREAL Loss: Training Speculative Decoding for Real Acceptance Length
SummaryThe paper addresses speculative decoding, where a draft model proposes tokens that a target LLM verifies, with speedup governed by expected acceptance length (EAL). It argues that common training objectives and recent block-level surrogates do not match EAL exactly, creating a covariance gap that can reverse gradient direction. The authors derive REAL, a loss using path weights for unbiased estimation and Rao-Blackwellization for variance reduction. In controlled comparisons it outperforms several baselines across tasks and seeds, without changing the drafter architecture or inference procedure.
In briefA margin-based analysis of LoRA: resistance to forgetting and adapter merging (MaLoRA)
SummaryThis paper explains, through a margin-based theoretical framework, why LoRA fine-tuning resists catastrophic forgetting and why independently trained adapters can be merged. Analyzing multiclass linear classification with ℓ2 regularization under a near-orthogonal task setting, the authors show that in an intermediate regularization regime the optimal adapter aligns with the max-margin solution on fine-tuning data. They propose MaLoRA, a training-free merging rule, and report experiments on vision and language classification tasks where it matches or outperforms several adapter-merging baselines.
Ziqing Xu, Hancheng Min, Lachlan Ewen MacDonald et al.
In briefBlurring Block Boundaries: Overlapping Block Denoising for Faster Diffusion Language Models
SummaryThis paper proposes B3 (Blurring the Block Boundaries), a method that speeds up inference in block-diffusion diffusion language models. It refines two neighboring blocks in parallel within a single forward pass and decides on the fly when the next block can start. The method requires no training or hyperparameters and can be added to any self-editing block diffusion model. Across 10 datasets, the authors report generation quality on par with the base model and a 11.5% to 14.5% speedup.
In briefAn Energy-Based View of Machine Learning: A Unifying Tutorial
SummaryThis tutorial introduces energy-based models (EBMs), which define probability distributions through a scalar energy function. It treats EBMs as a unifying lens, arguing that any EBM is specified by its energy, inference procedure, and learning method. The paper shows that many methods, including discriminative and generative models, representation learning, reinforcement learning, and planning, can be viewed as EBMs that differ in these choices. It also proposes that replacing one component can suggest new algorithms, with energy composition offering a way to combine learned models at inference time.
In briefMapping How Large Language Models Learn by Internal Thinking
SummaryThis paper proposes a Cognitive LbT Framework that maps four mechanisms of human-like learning by thinking (Reflective Reasoning, Synthetic Memory Formation, Internal Simulation, and Conceptual Self-Organization) onto computational processes in large language models. It analyzes how methods such as chain-of-thought prompting, self-consistency, and Self-Instruct approximate these mechanisms. The paper also examines their characteristic failures, such as hallucination diffusion and model collapse, and the structural asymmetries that cause them.
In briefWhich factors matter for estimating object counts across image collections with human-AI collaboration?
SummaryThis paper studies how to estimate total object counts across large image collections by combining model predictions with limited human verification. It compares choices of models, estimators, and adaptation strategies across 12 tasks, including biomedical microscopy, wildlife surveys, and overhead imagery of urban structures. The authors report that training-free multimodal models offer a strong starting point, fine-tuned detectors gain advantage at larger labeling budgets, and prioritizing human labels using model predictions improves efficiency. In a field deployment on aerial pronghorn surveys, a task-specific adaptation met the ±5% tolerance after reviewing a median of 3.2% of images.
Seoyun Jeong, Logan Lawrence, Owen L. Martin et al.
In briefUncertainty-aware offline reinforcement learning with Gaussian-process critics for choosing reactor operating actions
SummaryThis paper addresses reactor start-ups and grade changes, which are currently mostly done manually by operators. It builds on the HOFLON hybrid offline-online method and replaces its tree-ensemble critic and adversarial-autoencoder penalty with a sparse variational Gaussian-process critic, selecting actions that maximise the mean minus κ times the standard deviation. On the HOFLON polymerisation-reactor benchmark, removing the autoencoder penalty caused 4 failed start-ups in 15 runs, while with a Matérn-5/2 critic, κ=0 led to 13 failures in 15 runs and κ=3 led to none.
In briefFrom video frames to robot meta-frames: a unified latent interface for world action models
SummaryThis paper addresses World Action Models, which jointly predict robot actions and future observations using pretrained video diffusion models. The authors propose MetaWAM, which packs all robot signals such as proprioception, actions, and values into one meta-frame by assigning each signal a disjoint frequency-domain channel band and mapping it to the spatial latent via an inverse DCT. The method reports 68.9% success on RoboCasa with 39% lower inference latency, and 85.0% success on three real-world Franka tasks.
In briefARE-bench: Testing Research Agents on Real Ongoing Projects
SummaryThis paper introduces ARE-bench, a benchmark for checking whether research agents can advance ongoing projects while respecting earlier session decisions and existing workspace files. It contains 75 tasks built from real researchers' sessions and workspaces, spanning seven research stages and 1,173 binary checklist items. Across seven frontier models, the strongest passed every checklist item on only 66.2% of tasks.
In briefThe Unused Input: A Study of the Initial State as a Way to Condition Recurrent Time Series Models
SummaryTime series foundation models forecast new series without task-specific training, and recurrent versions carry a state that is usually left unused. The authors freeze a recurrent forecaster called TiRex and train only a small head that writes a context vector into its initial state. This slightly lowers forecasting error on GiftEval and, when text is attached to each series, gives further small gains on several datasets. The authors describe the results as preliminary and the improvements as small.
Bernhard Voggenberger, Christian Ganh\"or, Levente Z\'olyomi et al.
In briefEngineAgent: An agent system that automates adding new LLMs to on-device inference engines
SummaryThe paper presents EngineAgent, an agentic system that automatically enables new large language models in on-device inference engines. It builds a hierarchy-aware engine index and pre-validates backend build, execution, and debugging workflows using already supported models, then uses a builder to coordinate cross-layer changes and an auditor to check completeness and debug through device-side validation. Across 12 tasks covering three mobile LLMs and four engine-backend pairs, EngineAgent reaches 75% end-to-end success versus 33.3% for the strongest coding-agent baseline. It also reduces average per-task time and model-API cost compared with baselines using the same underlying LLMs.
Zhengxiang Huang, Yujie Sun, Shengheng Chen et al.
In briefCLADE: Answering from Coarse Evidence in Retrieval-Augmented Language Models
SummaryThis paper studies a case where retrieved passages are accurate but state answers only at a coarser category level than a question requires, such as naming a disease family instead of a specific disease. The authors call this resolvable insufficiency and describe Clade, where language models answer rather than abstain or search further. On PrimeKG biomedical questions, family-only passages reduce abstention compared with uninformative passages, and the resulting guesses are wrong about half the time.
WhyThe authors introduce resolvable insufficiency and the Clade phenomenon, validated via counterfactual tests and ontology-based experiments on PrimeKG and GloBI.
In briefAI Isn't Getting Dumber; Human Judging Standards Are Rising
SummaryThe paper analyzes 1,260,967 blind votes from Chatbot Arena, a platform where users compare anonymous language model answers, using pairs of models with fixed versions. It finds that the share of votes saying "both are bad" rose steadily from May 2023 to August 2024, while relative rankings between models stayed flat. The authors argue this reflects evaluation standards rising with the field's progress rather than models declining.
WhyThe authors propose that rising evaluator standards explain perceived LLM decline, separating it from model change using about 1.26 million blind Chatbot Arena votes on fixed-version models, with impl
In briefForward-Value Discretization: a training-free method for faster stochastic diffusion sampling.
SummaryStochastic diffusion samplers usually need many more neural function evaluations than deterministic ones to reach similar sample quality. The paper shows that even with exact score functions, stochastic samplers with higher-order approximations remain first-order convergent, so higher approximation order does not generally help. It introduces forward-value discretization, which evaluates the model at a lookahead point built from earlier predictions, and derives the stochastic variant F-DDPM. Across CIFAR-10, ImageNet, FFHQ, and LSUN, F-DDPM achieves lower FID than the evaluated stochastic baselines at matched NFE budgets.
WhyThe authors propose forward-value discretization, prove matching bounds that higher-order stochastic samplers stay first-order, and report FID gains on several datasets.
In briefWhen Should Agents Remember? Falsification-Gated Self-Evolution for LLM Agents
SummaryThis paper asks when lessons an LLM agent draws from its experience should be stored in persistent memory. It proposes Falsification-Gated Self-Evolution (FGSE), which treats each candidate lesson as a testable hypothesis, applies it temporarily, and checks it with transfer, falsification, and regression probes. The lesson is then committed, refined, or rejected based on measured gain and risk. The authors report strong results across web, app, tool-use, and long-term memory benchmarks, with fewer falsification failures, regression damage, and harmful committed updates.
In briefAuditing the Judge: Internal Signals for LLM Reasoning Assurance
SummaryThe paper compares a language model's self-judgment score with a supervised correctness probe trained on attention-head activations, focusing on how well each finds errors under a limited review budget. Across three open 7-12B models on GSM8K and MATH, the probe found more errors than the judge at a 5% review budget in all six model-dataset settings, with a median gain of 7.4 points. The authors conclude that internal signals can improve in-domain, supervised budget allocation but are not domain-general confidence estimators.
In briefDon't Just Trust Your Memory: Evidence-Linked Memory for LLM Agents
SummaryThe paper introduces WikiTrace, a memory system that links each lesson an LLM agent learns to a summary of why its original task succeeded or failed and to the raw trajectory. When facing a new task, the agent can follow these links to fill in missing details or correct a faulty lesson before acting. Evaluated on five agentic benchmarks, WikiTrace outperforms existing memory approaches and reduces memory-induced degradation.
In briefPSL-RGHA: A parallel split learning method that curbs gradient divergence with reference gradient alignment
SummaryThis paper proposes PSL-RGHA, a parallel split learning framework for resource-constrained devices that offloads much of the computation to a server. Without client-side aggregation, such training can diverge under heterogeneous data. The framework combines reference gradient progressive construction (RGPC) and gradient half-space alignment (GHA) to keep client gradients consistent and reduce conflicts. The authors report that it outperforms existing baselines across various settings.
In briefA method to add abstention to multimodal models without losing correct answers
SummaryThis paper addresses the problem that adding abstention to a task-adapted multimodal model can discard correct answers, including those gained during adaptation. It proposes ReFrain, which combines a linear probe on first-token logits, threshold calibration on held-out image clusters, and a deterministic veto that replaces rejected outputs with a "none of the above" option. On Qwen2.5-VL-3B, at zero budget ReFrain retains all 57 SFT-gained answers with no losses and improves correct abstention by 7.30 points. At a 0.5% calibration budget it reaches 74.05% correct abstention.
In briefMMDM: A Benchmark for Measuring Decision Making Across Many Tasks
SummaryThe paper introduces MMDM, a benchmark of 5,448 questions across 16 task categories for evaluating System-1 decision models that make structured probabilistic choices in a single forward pass. Tasks cover agent state tracking, tool selection, preference judgments, and rubric scoring, with option spaces from 2 to 1,024 and inputs up to 64,817 tokens. Evaluating 28 decision models, the authors find that many perform well on common knowledge but struggle with interdependent options, constrained decisions, and multi-select tasks.
In briefTaoMate-R1: Temporal Role Specialization for Streaming Multi-Reference Video Generation
SummaryTaoMate-R1 is a streaming generator that creates videos of specified people and objects, first establishing a scene and then continuing it under changing instructions. It uses separate low-rank adapters for scene establishment and continuation on a frozen backbone, trained jointly through Self Gradient Forcing. On ten roughly two-minute streams with changing instructions, it reduces the mean boundary luminance jump from 1.96 to 0.613 and achieves higher CLIP and DINO temporal similarity.
In briefForeseeDrive: Future-Token Conditioning for Anticipatory Generative Driving
SummaryThis paper presents ForeseeDrive, a method that turns future scene evolution into an interface for shaping candidate trajectories in generative driving planners. It derives future tokens from logged actor motion, occupancy changes, route interactions, and TTC margins, and predicts the same token space from current context at deployment. The authors report 89.2 PDMS on NAVSIM navtest, 0.41 m average L2 and 0.05% collision rate on nuScenes validation, and a 76.50 driving score on Bench2Drive.
In briefLM-SimBench: A Framework and Benchmark for Simulating Large Language Model Training
SummaryThis paper presents LM-SimBench, a unified framework, profiling dataset, and benchmark for evaluating simulators of large language model training. The authors profiled 23 model settings from 10 dense and Mixture-of-Experts LLM families on an Ascend NPU cluster, producing 348 feasible configurations that vary five parallelism dimensions and activation recomputation. The benchmark is used to compare prediction accuracy, configuration selection, and prediction error diagnosis across representative simulators, with matched GPU-NPU experiments checking cross-platform applicability.
In briefNinety-six ways to measure AI medical note-taking tools: a look at fragmented evaluation
SummaryThis paper reviews 54 recent studies evaluating ambient AI scribes, tools that automatically document clinical encounters in healthcare. The authors identify ten evaluation dimensions and 40 metrics, which are operationalized in 96 different ways, with 35 used by only a single study. Assessment is dominated by documentation efficiency and time, while safety and clinical outcomes are measured least consistently, and metrics rely heavily on self-report. The authors call for a standardized, multi-dimensional set of metrics to enable more rigorous and cumulative evaluation.
In briefUniFlow-RS: A Unified Generative Predictor for Diverse Remote-Sensing Tasks
SummaryUniFlow-RS is a single predictor that handles several different remote-sensing tasks, replacing separate task-specific decoders and losses. It converts the outputs of different tasks into shared structured visual targets in a common latent space and trains them with one flow-matching objective. A method called Task-Modality Adaptive Timestep Sampling adjusts training timesteps based on each task-modality group's learning state. One shared model covers six tasks and thirteen datasets, outperforming some recent task-specific methods on several benchmarks and showing zero-shot transfer to unseen data.
In briefAutoThreshold: Learning Joint Confusion Profiles for Adaptive Mask Thresholding
SummaryThis paper introduces AutoThreshold, a post-hoc method that picks a binarization threshold for segmentation masks by predicting the joint mass of positive and negative pixels across score bins. This yields threshold-indexed confusion counts that are coherent and monotone, and a plug-in decoder selects thresholds for any F-beta objective without retraining the head. Experiments transfer Kvasir-trained pipelines to four polyp datasets and KITTI Road, reporting gains over continuous Otsu thresholding in external macro Dice.
In briefCompositional random forest: a random forest method for compositional data
SummaryThis paper proposes a random forest framework for compositional responses that works on the simplex, in log-ratio coordinates, and on the unit sphere. Splits are chosen by log-likelihood gain, using a tangent-normal distribution (sharing parameters with the Kent distribution) and the multivariate logistic-normal distribution. In simulations, the most accurate forest depended on the covariance structure of the composition, and the tangent-normal forest was the most accurate in two of three designs.
Khaled Masoumifard, Stephan van der Westhuizen, Sugnet Gardner-Lubbe
In briefFederated learning for COVID-19 forecasting across regions with out-of-sync outbreaks while keeping local data private
SummaryThis study proposes a federated learning framework that lets Guangdong, Macau, and Portugal jointly forecast daily COVID-19 cases without sharing their raw data, even though their outbreaks peaked at different times. It uses Long Short-Term Memory networks and compares results with local and centralized training. The federated model reached a normalized RMSE of 0.028, an 88.3% improvement over the local model (0.240) and close to centralized training. The authors suggest this may be a practical route to privacy-preserving collaborative forecasting, but prospective validation is still needed.
Ruibin Liu, Francisco Fernandes, Xuanfeng Li et al.
In briefDesigning training data to make channel estimation networks robust to Doppler shifts
SummaryThis paper studies how well neural networks can estimate wireless channels when those channels differ from the training data, especially under high Doppler shifts in fast-moving scenarios. The authors propose design criteria for generating synthetic training data so that offline-trained networks perform robustly on new channels without prior information or online training. Simulations with networks of different complexity show a certain level of performance across unseen power delay profiles and Doppler shifts.
In briefPhysics-constrained autoencoder for interpretable turbofan engine health learning
SummaryThis paper proposes a physics-constrained autoencoder that embeds system physics into its decoder structure and loss function, so latent variables represent predefined physical quantities. On the N-CMAPSS dataset, it matches fully data-driven reconstruction accuracy and produces a health index with a Spearman correlation of 0.9791 to ground truth. The framework distinguishes compression-side from expansion-side degradation without diagnostic labels and visualizes degradation trends on component maps.
Leonardo Baldo, Jokin Alcibar Ibarzabal, Daniel Cisneros Acevedo et al.
In briefChain of Physics: LLM Reasoning for Fault Diagnosis Using Engineering Knowledge
SummaryThis paper presents Chain of Physics (CoP), a large language model framework for fault diagnosis. It uses a Constrained Knowledge Graph automatically built from asset-specific documentation, storing chains that link failure modes, physical quantity changes, and monitoring sensors. At inference, similar graph paths are retrieved as information and in-context examples, and candidate failure modes are analysed separately before a final module compares the chains with the answer choices. On the FailureSensorIQ benchmark across four industrial asset classes and three LLMs, CoP outperforms direct prompting and conventional RAG, improving accuracy by 7.8 to 19.6 and 11.7 to 18.2 percentage points respectively.
In briefRecurrent modular spiking networks for reinforcement learning
SummaryThis paper introduces Community Recurrent SNNs (CoRe SNNs), recurrent spiking networks divided into communities with fixed roles: a controller community receiving stimuli, one community per action, and latent communities holding internal state. Learning uses local three-factor rules, and the authors propose Map-STDP to encourage modular structure during training. Experiments on CartPole, LunarLander, and Acrobot show these recurrent modular networks can match or improve upon feed-forward approaches.
In briefLilo: A Multi-Agent Conversational Sandbox for Children's Digital Communication
SummaryThis paper presents Lilo, a multi-agent conversational system where children practice handling peer conflicts in group chats. A child joins two AI-driven peers mid-argument and receives guidance from a mentor character using a Stop-Think-Act traffic light. An exploratory study with 28 children aged 9-11 included surveys before and after the intervention and exit interviews. Participants reported significantly higher self-efficacy for coping with unpleasant messages after interacting with Lilo.
In briefElastic-WAM: adaptive spatiotemporal imagination that lets robots predict future video more efficiently while controlling their actions.
SummaryThe paper proposes Elastic-WAM, a method for robot control based on World-Action Models that predict future visual dynamics. It adaptively allocates refinement computation across temporal denoising steps and spatial future regions, using a learned refinement demand field. Evaluations on diverse simulation benchmarks and real-world long-horizon manipulation tasks report state-of-the-art control performance with reduced inference latency.
In briefSurvey of self-improving multimodal AI agents across four task types
SummaryThis survey examines multimodal agents that improve themselves without repeated human annotation, covering multimodal reasoning, multimodal generation, GUI and Web navigation, and embodied manipulation. The authors formalize self-evolving agents and use a three-stage loop of experience generation, feedback acquisition, and update consolidation as their analytical framework. The paper analyzes why self-improvement mechanisms transfer unevenly between scenarios.
In briefSemantic Chains: Linking Questions to Justified Answers
SummaryThis paper studies how to find the facts that justify an answer to a question and present them as a sequence of connected statements. It models a context as a graph of logical facts, with the expressions representing each fact taken as given. Under the model's conditions, a search starting from objects in the question reaches a root fact and covers every fact in the selected expansion. The paper also enumerates candidate orders for multi-link chains and specifies when the link texts preserve justification.
In briefPersonalized Emotion Sensing via Continual Learning in Human-Robot Negotiation
SummaryThe paper proposes an affect-adaptive negotiation strategy that uses CLIFER, a continual learning model for facial expression recognition, to personalize valence-arousal perception in real time. Differential feedback on changes in valence-arousal adjusts the agent's target utility, and the model updates asynchronously during live interaction. In a within-subject study with 66 participants negotiating with a humanoid robot, the continual learning agent achieved significantly higher individual utility and concession rates than the FaceChannel baseline.
Mehmet Onur Keskin, Berk Buzcu, Nikhil Churamani et al.
In briefSinkState: On-the-Fly SVD Compression of Recurrent States for Faster Hybrid LLM Serving
SummaryThis paper studies hybrid language models that mix softmax attention with linear attention, and finds that linear-layer queries, keys, and values concentrate in a stable low-dimensional subspace, called the linear attention sink. The authors propose SinkState, which stores recurrent states in an adaptive low-rank form during inference using buffered updates, a local SVD on the top subspace, and asynchronous CPU consolidation of the remaining components. On four hybrid models with Mamba-2 and Gated DeltaNet layers, ranging from 4B to 120B parameters, it reduces peak recurrent-state memory by up to 2.6x and speeds up decoding by up to 1.58x over Mamba-SSM and FLA while maintaining accuracy.
In briefAn AI assistant that automates anti-corruption screening of draft laws using the TRACE framework
SummaryThis paper presents an AI assistant that screens draft legislation for corruption risks, identifying provisions an expert would criticise and naming the defect in each. The system combines deterministic parsing, model-guided provision qualification, rule-based factor assignment, and human oversight. The authors built a benchmark from Ukraine's National Agency on Corruption Prevention conclusions, covering 132 draft laws and 474 findings, and evaluated the pipeline on 100 of them.
Serhii Zabolotnii (Cherkasy State Business College, State Scientific Research Institute of Armament, Military Equipment Testing et al.
In briefPARLEY: Testing How LLM Agents Work Together Through Asymmetric Puzzles
SummaryThis paper presents PARLEY, a platform for studying how large language model agents collaborate, using four asymmetric puzzle tasks based on cognitive psychology. GPT-4o agent pairs succeeded in 60.5% of conversations, with success varying across tasks. The authors derived the PAC taxonomy of six failure types and showed that a taxonomy-guided intervention raised average task success to 70.5%.
In briefDyship: Classifying Speaker Relationships from Voice in Two-Person Conversations
SummaryThis paper introduces DyShip, a framework that classifies whether speakers in two-person conversations from the Seamless Interaction corpus are familiar with each other and whether they are related by kinship. It compares five speech representations: spectrograms, MFCCs, self-supervised WavLM features, x-vectors, and estimated articulatory movements. Spectral features achieve the best familiarity accuracy at 87.81%, while WavLM reaches 85.17% for kinship. Most classifiers outperform previously reported human and audio language model results on the same dataset.
Kieran Sturges, Thomas Thebaud, Yaohan Guan et al.
In briefFederated Averaging with partial client participation under a general variance condition
SummaryThis paper studies the convergence of Federated Averaging (FedAvg) when only a subset of clients participates in each round, under a general variance condition. The server samples S of n clients uniformly without replacement each round. For smooth nonconvex objectives, the authors derive a finite-time stationarity bound that depends on a heterogeneity measure and a client sampling factor, showing linear speedup in S and recovery of the full-participation result when S equals n. With diminishing stepsizes, the gradient norm converges to zero both in expectation and almost surely.
Ziming Wang, Kun Huang, Karl Henrik Johansson et al.
In briefStaRSE: A Compact Russian Sentence Encoder Using Sign-Coded Static Embeddings
SummaryThis paper presents StaRSE, a compact static sentence encoder for Russian that avoids Transformer inference at deployment. Token directions are stored as sign bits with separate token norms, and sentences are produced by mean pooling over norm-scaled sign directions. The model is initialized from ruBERT-base, fine-tuned on Russian sentence pairs, and post-trained with a contrastive stage that optimizes the sign-projected token table. On RuMTEB it reaches a task mean of 51.16 in an 11.3 MiB package with 0.08 ms mean CPU latency per sentence.
Boris Malashenko, Ivan Jarsky, Egor Soldatov et al.
In briefHow Transformers learn inside-outside parsing in stages, using synthetic grammars
SummaryThis paper studies how Transformer language models learn to handle syntactic structure and parsing ambiguity, using synthetic context-free grammars with variable parse tree shapes. The learning curves are stagewise and can be computed analytically, with each stage linked to a hidden latent tree variable. Early in training the model relies on local bigram and trigram information, and later it develops an encoder-decoder-like structure in which abstract latent information is encoded in the middle layers and the later layers refine the final prediction.
In briefViGA-FLAME: Aligning Visual and Geometric Features for Robust 3D Head Expression Estimation
SummaryThis paper presents ViGA-FLAME, a two-stage framework for estimating 3D head expressions from a single image. A Parametric-Geometric Variational Autoencoder learns a unified latent representation of head surfaces and FLAME parameters. A trainable Visual-Geometric Adapter then uses asymmetric attention to align features from a pretrained visual encoder with this latent space without modifying the visual stream. The authors report improved robustness over competing feed-forward methods under severe facial deformations and large viewing angles.
In briefVerifiable Agentic Traffic Class: A proposed router mechanism for prioritizing AI agent traffic, tested in simulation and on a Linux kernel
SummaryThe paper proposes VATC, which marks AI software agent traffic with a reserved DSCP value and uses a router-side classifier to demote flows whose behavior contradicts that declaration. It combines this with a rate-limited priority queue. VATC was compared against static DiffServ, a trust-the-marking class, and fq_codel in simulators and on a real Linux kernel. The evaluation found large agent-latency gains in a constructed burst, but spoofed traffic that mimicked agent patterns evaded the classifier. The traffic was synthetic or sandbox-scale.
In briefTree-Ising: Repairing Language Models with Bit-Level Ising Edits
SummaryThis paper proposes Tree-Ising, which repairs language model weights by treating edits to BF16 weights as a sparse discrete optimization problem. The objective is decomposed into per-row QUBO (Ising) problems solved on a GPU. On MATH-500, the repaired model scores 54.6%, above the agent's edit at 51.7% and the base model at 7.9%.
In briefHow concentrated versus spread-out attention head dropout affects multi-step reasoning in language models
SummaryInspired by brief localized neuronal silences in the mammalian neocortex, this paper studies how concentrated bursts of attention head masking affect step-by-step reasoning in autoregressive language models. The authors evaluate Qwen3-8B and LLaMA-3.1-8B-Instruct on the GSM8K benchmark, comparing clustered burst masking with budget-matched diffuse random dropout. Concentrated bursts cause significantly larger drops in reasoning accuracy.
In briefDailyWildfireCAN: A Dataset for Predicting Next-Day Wildfire Spread in Canada
SummaryThis paper presents DailyWildfireCAN, a multi-modal, multi-temporal dataset and benchmark for predicting next-day wildfire spread across Canada. It covers 2,440 wildfire events from eight fire seasons (2017–2024) at 90-m spatial and 24-hour temporal resolution, integrating 33 features such as topography, meteorology, land cover, vegetation, and satellite imagery. The authors benchmark mono-temporal and multi-temporal prediction models and propose a multi-dimensional evaluation framework.
Mahacine Ettahri, Hager Radi Abdelwahed, Toluwaleke Olutayo et al.
In briefJoint quantization and dynamic partitioning for context-aware edge vision transformers
SummaryThis paper proposes an edge orchestration framework for running vision transformer models on connected and automated vehicles. It jointly chooses model partition points, per-layer quantization bitwidths, and execution placement, formulated as a mixed-integer linear program. Evaluated on Jetson Orin nodes with OPV2V multi-stream workloads, it reduces V2V communication energy by 33.3–36.0% relative to a uniform INT8 split.
In briefAn augmented Lagrangian method for solving offline reinforcement learning as a linear program
SummaryThe paper addresses offline reinforcement learning by solving the Lifted Bellman Linear Programming formulation with the method of multipliers. The constraint coefficients are learned as multipliers through dual ascent rather than fixed in advance. The authors report 60.1% average success on 50 state-based OGBench tasks.
In briefSpecValley: a paper proposing a learnable aggregation step for node classification on dense heterophilic graphs
SummaryThis paper addresses node classification on heterophilic graphs, where neighbourhood aggregation can weaken the signal that carries labels. The authors show that degree normalisation creates a fixed spectral valley at a point where eigenvalues of dense graphs accumulate. They propose SpecValley, which learns the valley's position, sharpness and flank gains for each graph. Reported results on heterophilic benchmarks are strong, and an ablation attributes the gain to the aggregation step rather than the downstream filter.
In briefModel for choosing cross-border warehouses using doubly robust estimation and linear programming
SummaryThis paper presents a cross-border fulfillment center (FC) selection model that combines forecasting, doubly robust estimation, and linear programming to choose which source FCs to use for each shipping route. The model was evaluated on 12 months of shipment data from three international corridors of a large e-commerce company. Compared with human expert selection policies, it reduced delivery time by 8.5% to 51.1% and logistics cost by 4.4% to 48.3%.
In briefBARS: Budget-Aware Sampling That Can Refuse to Answer, for Majority-Vote Reliability
SummaryThis paper proposes BARS, a sequential sampling rule for large language models that repeatedly samples answers and takes a majority vote under a budget and target error rate. It checks the gap between the top two answers and either stops early, returns a budgeted majority vote, or refuses when feasibility cannot be established. The paper distinguishes a certified mode with proven bounds from a practical mode that uses empirical calibration. On GSM8K, the practical mode uses 22.3% fewer samples than fixed-8 sampling while refusing 18.8% of questions; on MATH, it raises answer accuracy to 82.5% by refusing about half of the difficult questions.
