In briefAtomicRL: Atomic Reinforcement Learning for Improving Long-Horizon Robot Manipulation
SummaryThis paper proposes AtomicRL, a reinforcement learning post-training framework for vision-language-action robot policies that uses atomic task structures to improve credit assignment and exploration. It introduces Semantic-Aware Policy Optimization and Bottleneck-Adaptive Target Exploration, along with a Real2Sim2Real pipeline for transferring policies from simulation to physical robots. Reported results include gains of 3.6% and 4.0% over PPO on LIBERO-LONG, up to 9.5% on CALVIN ABC-D, and an average 21.3% real-world improvement.
In briefProactEgoMem: A Benchmark for Proactive Memory in Streaming First-Person Video
SummaryThis paper introduces ProactEgoMem, a benchmark that tests how well wearable AI assistants proactively collect and maintain memory from streaming egocentric video. It covers three tasks, Information Capture, State Tracking, and Location Tracking, at two proactiveness levels, Notification and Intervention. Even the strongest tested models failed to capture memory both timely and accurately.
Ruotong Liao, Meikel Kokowski, Xiaoxiao Sun et al.
In briefGANet: A Lightweight Network for Identifying Plant Diseases on Small Devices
SummaryThis paper presents GANet, a lightweight group attention network for predicting and monitoring 38 types of plant diseases. With only 0.001M parameters, it reports 99.99% accuracy on the PlantVillage dataset. The model is also optimized for edge devices, reaching 5 FPS on Raspberry Pi 4b and 11 FPS on Jetson Nano.
In briefOrbitol: generating crystal structures on Wyckoff orbits
SummaryThis paper presents Orbitol, a crystal generator built on the Crystalite diffusion Transformer that uses Wyckoff-orbit tokens instead of atom tokens, so a requested space group can be specified. The authors report generation two to three orders of magnitude faster than symmetry-aware baselines at the evaluated settings. On MP-20, 13.8% of attempts gave stable or metastable, unique, novel structures that kept the requested space group, versus 10.8% for SGFM. The model was also used to generate piezoelectric candidates, 89% of which passed density functional theory checks among those selected by internal screening.
Ivan \v{Z}ugec, Joshua Rosenthal, Ivor Lon\v{c}ari\'c et al.
In briefCircuits of LLM Classifiers: Combining Model Predictions Adaptively to Balance Cost and Accuracy
SummaryThis paper treats LLM classifiers as noisy measurement instruments of an underlying label and studies how to combine predictions from different models to improve reliability. The authors introduce circuits, which organize classifiers into an adaptive sequence of queries and decisions, and characterize the optimal circuit and the fundamental limits of achievable accuracy at a given average query cost. Across seven classification tasks covering safety, health, political science, human preference, agent evaluation, decision-making, and reasoning, the method improves the cost-accuracy frontier over routing and cascading baselines.
In briefFrom Classifier Heads to Encoder Prototypes: Personalized Federated Learning Clustering Based on Heads
SummaryThis paper introduces HDC-FL, a personalized federated learning method that uses each client's classifier head to decide which clients should share representations. The server projects heads into a low-rank space, groups them, builds encoder prototypes per group, and returns a soft mixture of prototypes to each client while keeping its own head. Across 11 CIFAR-10 and CIFAR-100 configurations, HDC-FL is competitive on mean accuracy and reports the highest macro-averaged worst-client and bottom-20% accuracy, and head similarity shows positive association with encoder compatibility.
Amirardalan Dehghanpour, Arian Noori, Ehsan Tanghatari et al.
In briefParallel AI Agents Coordinating Edits via Graph-Based Block Optimization
SummaryThe paper models coordination among parallel AI agents editing a shared solution as discrete block optimization over a known dependency graph. Independent sets select compatible edits, and an optimal fractional coloring-based merger is proven to retain a 1/χ_f(G) fraction of verified local improvement, a bound shown tight for such merging. Experiments cover code repair, graph-cut optimization on three public road networks, and route-search agents, where fractional scheduling reduced median wall time by 15% across twelve paired runs.
WhyFormalizes multi-agent collaboration as graph-structured block optimization with a tight 1/χf(G) guarantee, plus experiments on code repair, road networks, and routing.
In briefUsing Charts as Epistemic Actions in Data Agents
SummaryThe paper asks whether data agents can use charts as internal reasoning steps, the way human analysts do, rather than relying only on text. The authors build a 200-task suite and find that agents can read charts well but do not consistently choose to use them, with charts helping on temporal and anomaly reasoning and text helping on precise reading. They propose distilling experience from trajectory rollouts to teach when and how to use text or charts, validated on held-out data. Across three benchmarks and four LLM backbones, the method improves accuracy and visual evidence quality without changing model parameters.
WhyTreats charts as epistemic actions for agents, distilling experience with a validation gate, tested on three held-out benchmarks and four backbones.
In briefLoopholing All the Way: Language Modeling Using Only Self-Conditioning
SummaryThis paper introduces LAW (Loopholing All the Way), which propagates the expected embedding predicted by the denoiser across steps and decodes tokens only once at the end, making a continuous generative model based purely on Self-Conditioning. LAW outperforms prior discrete and continuous diffusion models on Sudoku and Maze, and on GSM8K it solves 42.2% of problems versus 34.5% for uniform Loopholing Discrete Diffusion. On OpenWebText, it approaches autoregressive models with 64 NFE.
WhyProposes a continuous diffusion LM built purely on self-conditioning, with clear gains across Sudoku, Maze, and GSM8K benchmarks.
Adrien Deschenaux, Nadezhda Ilieva, Justin Deschenaux et al.
In briefGoiBERT: Compact Vietnamese Language Models Built by Two-Stage Knowledge Distillation
SummaryThis paper introduces GoiBERT, two compact Vietnamese encoders with 26M and 49M parameters that process raw text directly without external word segmentation. The models learn in two stages from complementary teachers: a same-tokenizer Vietnamese encoder for general representations, then a larger teacher providing task-level soft labels. Across eight benchmarks, GoiBERT-large achieves the highest average score among non-teacher models while using fewer than half the parameters of the strongest compact baseline.
Quang Anh Kiet Tran, Nam Anh Tran, Manh Hung Trinh et al.
In briefLooped Uniform Diffusion Language Models: Repeating Early Layers to Train and Generate More Efficiently
SummaryThis paper studies applying selective looping, which repeatedly runs a few early Transformer layers within each denoising step, to uniform diffusion language models (UDLMs) while keeping the training objective and sampler unchanged. Looped UDLMs reach the validation NLL of a non-looped model with up to 2.8x fewer training FLOPs on OpenWebText and 3.3x fewer on FineWeb-Edu. At a matched NFE budget, they achieve up to 9.8% lower generation perplexity, and sampling smaller loop counts more often reduces inference loops by 12% without hurting downstream accuracy.
Saiteja Utpala, Chieh-Hsin Lai, Yuki Mitsufuji et al.
In briefFrom Reasoning to Action: Frontier AI Models Tested on LIBERO-10 Robot Manipulation
SummaryThis report presents a closed-loop framework for testing whether frontier AI models can control robot arms on the LIBERO-10 benchmark without task-specific training. Models act through visual observations, robot proprioception, and structured tool calls, mainly by specifying end-effector poses and gripper commands. Six models were evaluated across three reasoning effort levels over 1,800 episodes, with the best success rate reaching 44%. Higher reasoning effort generally helps, but models struggle with precise spatial arrangements and complex multi-step tasks.
In briefHow prompt wording affects AI accuracy and speed in competitive programming
SummaryThis paper tests nine prompt styles on GPT-5.5 across three International Olympiad in Informatics accuracy tasks and two NP-hard speed tasks. No single prompt performed best across all tasks; the strongest one depended on task type and on whether mean or median was used. Some score differences came from a few outlier runs, and results were compared descriptively with human IOI contestants.
In briefSWE-DIVE: A framework testing coding agents through simulated multi-turn user conversations
SummaryThis paper introduces SWE-DIVE, a framework for evaluating coding agents in multi-turn interactions rather than one-shot prompts. It uses a user simulator that mimics different behavior patterns and a conversation validator that checks whether the simulated user's messages meet the original requirements. The components run inside a Harbor environment, and the results and source code are published on GitHub.
In briefWhich Differences Matter? Controlled Tests of Rule Transfer in Decision Models
SummaryThe paper studies how decision models respond to controlled changes in stated rules and observed states, using executable arithmetic rules. It separately measures rule transfer and joint correctness, comparing equivalent wordings and altered tables. Transfer gains grow with answer agreement among nearest sources, equivalent rewrites have opposing task effects, and a decision service usually follows the answer targets implied by altered table values or mappings.
In briefSP-HLM: Sparse-pose-driven motion generation and whole-body control for humanoid robots that walk and manipulate objects
SummaryThis paper presents SP-HLM, a framework that links high-level task policies to whole-body control for humanoid robots performing loco-manipulation. It uses five task-critical body poses plus a learned latent motion prior as the interface, separating spatial intent from whole-body coordination. The system includes SP-VLA, which predicts sparse poses and finger positions, SP-MGM, which reconstructs full-body motion in real time, and SP-GMT, which tracks motion using both explicit and implicit targets. The authors report improved motion tracking, coordination, and task success in simulation and on a real humanoid robot.
In briefHGM-Dual: Lineage-Aware Guided Harness Optimization for Long-Horizon Planning Agents
SummaryThis paper introduces HGM-Dual, a framework that improves the prompts, tools, and workflows (the harness) of planning agents built on frozen or closed-source LLMs under a limited evaluation budget. It records what each ancestor changed and whether those changes helped or hurt, and for each parent it generates a general child plus specialized children targeting the components most responsible for its top constraint violations. The authors report up to a 14.45% relative gain in case accuracy over the strongest baseline, and matching HGM variants' final performance with 2x to 4x fewer evaluation budget.
In briefRobust Soft Skill Adaptation via Behavioral Localization in Tool-Using Agents
SummaryThe paper proposes Behavioral Localization, which compares a frozen agent with and without a text skill to find trajectory positions where the skill most affects its outputs. These positions are used to supervise a short soft prompt instead of a long text skill description. Experiments on SpreadsheetBench with a frozen Qwen3.6-35B-A3B agent show the top 10% of positions capture 96.12% of the total positive skill gain. Selective supervision improves success from 30.36% to 36.79% in the shared prompt setting, and reaches 57.38% under task-specific conditions, outperforming the full text skill and the OPCD baseline.
In briefAttn-Forcing: learning full-context attention proxies for streaming video generation
SummaryThe paper addresses the high memory and compute cost of long attention contexts in streaming video generation. It proposes Attn-Forcing, which uses learnable visual latents as attention proxies that approximate the attention responses of the full historical context, updated by a Proxy Update Module and supervised by an Attention Matching Loss. The authors report that the method improves generation quality under bounded memory and computation and achieves state-of-the-art results over baselines.
In briefPlausible but Unimplantable: Auditing Shape Completion Models for Skull Defect Implants
SummaryThe paper audits three publicly released pretrained shape completion models, two of them foundation models, applied zero-shot to 239 cranial defects from public datasets and real post-craniectomy cases. None of the models' outputs could be implanted without intersecting the patient's remaining bone, while a task-specific volumetric model was implantable in nearly all cases. The authors conclude that the foundation models they could obtain cannot be trusted and are not usable clinically as released.
Marek Wodzinski, Kamil Kwarciak, Daria Hemmerling et al.
In briefWhat the Author Wrote, What the Agent Reads: Silent Cue Loss in Document Ingestion
SummaryThis paper shows that mainstream document converters drop formatting cues such as colour and highlighting that authors use to flag requirements, causing AI agents to confidently produce wrong answers. Using 76 real policy documents, the authors found that performance falls sharply when formatting is lost, with colour and highlight accounting for most of the loss. The authors conclude that improving agent reliability depends on the ingestion stage, which is cheap to fix and measurable.
In briefRealMe: Creating Immersive Volumetric Video from Few Camera Captures
SummaryThis paper presents RealMe, a framework that turns sparse, low-resolution captures into high-resolution, 4D-consistent volumetric videos. It uses a video diffusion model with factorized spatio-temporal attention and a latent frequency loss to enhance the captured footage. Experiments on the DNA-Rendering benchmark and the authors' RealVV dataset report better visual fidelity and 4D consistency than existing methods.
In briefAdds a stationarity penalty so machine-learned exchange-correlation functionals give better energies and densities.
SummaryEmpirical exchange-correlation functionals are fitted on fixed orbitals but used self-consistently, so better energies can yield worse densities. The paper proposes penalizing the Brillouin residual of a trained functional on Kohn-Sham orbitals of an accurate density, which requires no self-consistent-field solution during training. A neural meta-GGA trained on 6 energies and 11 CCSD densities outperforms SCAN in atomization energies and densities for unseen molecules.
In briefExGrad: Explicitly subtracting clean gradients in adversarial training to improve robustness
SummaryThis paper proposes ExGrad, an adversarial training method that builds clean, adversarial, and AWP gradients and uses virtual steps to measure their effect on adversarial margins, then solves a three-variable problem to choose their coefficients. Experiments on CIFAR-10, CIFAR-100, and ImageNet-100 show radius-dependent gains over AWP. On CIFAR-100 at radius 0.8, all five reported metrics improve, and on ImageNet-100 the same radius improves all four robust metrics over AWP and PGD-AT. The authors interpret clean-gradient subtraction as an update preference whose strength is controlled by the affine radius.
In briefGuideEvolve: A Self-Evolving Agent Framework with Expert Guidance and 3Ts Mechanisms
SummaryThis paper presents GuideEvolve, a self-evolving agent framework that combines expert guidance with three autonomous regulatory mechanisms called the 3Ts: Think Fast and Slow, Throwback, and Termination by choice. Experiments on the MGSM and MMLU benchmarks report accuracy comparable to state-of-the-art models, with notably reduced variance. The framework also achieves a 100% evolution success rate and roughly halves the number of required optimization iterations.
In briefCode Comprehensor: A framework for hierarchical documentation and cross-modal code retrieval in large repositories
SummaryThe paper proposes Code Comprehensor, which combines a code call graph (CodeGraph), an LLM-based tool that generates hierarchical documentation (CodeScribe), and a CLIP-style text-code encoder (CodeContrast) trained on that documentation. It aims to help an LLM answer questions about large codebases by retrieving relevant code context. Evaluated on OpenCV and a proprietary 5M-LOC repository, CodeScribe-based supervision improved text-to-code retrieval, and the RAGAS-based LLM-as-a-judge evaluation reported faithfulness scores above 90%.
Sujoy Roychowdhury, Joseph H. R. Isaac, Jiangning Gao et al.
In briefAgent Behavior Maps: Mapping LLM Agents by Their Likely Actions Without Running Them
SummaryThe paper introduces Agent Behavior Maps, which represent LLM-harness combinations using centered action log-likelihoods on shared reference trajectories, without new environment rollouts. The authors derive error bounds linking squared coordinate distances to trajectory KL divergence, and across eight cross-scored models the distances track measured KL in rank order. Across 100 LLMs and four environments, ALFWorld-based coordinates support performance prediction and identify behavioral similarities that persist after controlling for success rate and model size.
Shota Fujikawa, Takuya Tamura, Kosuke Akimoto et al.
In briefClunoGS: Compact Language 3D Gaussian Splatting via Unified Optimization
SummaryThis paper proposes ClunoGS, a framework that compresses 3D Gaussian Splatting models carrying language features to support open-vocabulary scene understanding. It combines semantics-aware Gaussian sparsification and vector quantization within one constrained objective. The method is evaluated on open-vocabulary localization and segmentation using the LERF and ScanNet-10 benchmarks. The authors report that under aggressive compression it better preserves semantic representations than sequential compression or semantics-agnostic sparsification, and ablations show joint coordination outperforms applying the two independently.
In briefReducing model collapse with a high-temperature 'dreaming' regularizer
SummaryThis paper studies model collapse, where generative models trained on synthetic data lose diversity and performance, by analyzing the eigenvalue spectrum of a learned Markov transition operator. It finds that low-temperature sampling pushes the subdominant eigenvalues toward the unit circle and narrows the spectral gap. The authors propose Dreaming Learning, which adds a loss on the model's own samples drawn at a raised temperature, and report that spectra over 5000 generations are consistent with mitigation, an effect that also appears in LSTM language models and a 124M-parameter GPT-2.
Matteo Benati, Alessandro Londei, Denise Lanzieri et al.
In briefNeural cellular automata refine multiclass boundaries in remote sensing land cover maps
SummarySegmentation of satellite imagery often places class boundaries a few pixels from their true positions, with errors concentrated in a thin band measured by Boundary IoU. The paper proposes BRNCA, a small learned local rule applied iteratively to the class probabilities of a frozen pre-trained segmenter to correct boundaries cell by cell. The mean BIoU rises from 46.63 to 48.99, with the largest gain on the irregular woodland class. The authors suggest boundary position error may correlate with the fractal dimension of class boundaries.
Pascoassis Souza Santos Meira, Gilson Antonio Giraldi, Rafael Duarte Coelho dos Santos
In briefAuditable AI text detectors built with sparse autoencoders
SummaryThis paper proposes a pipeline that converts a fine-tuned neural AI-text detector into a classifier built on human-readable concepts. Sparse autoencoders are trained on the detector's residual stream, each feature is given an auto-interpretation label, and only features whose descriptions pass a global fidelity threshold are kept. A lightweight classifier is then trained on these interpretable features. The authors report that the resulting classifiers are highly interpretable and perform better on out-of-distribution texts than neural or stylometric alternatives.
Liam Dugan, Callum McDougall, Matthew Denton et al.
In briefEvery Alarm Has a Cause: Open Generator of Telecom Alarm Storms with Exact Root Causes
SummaryThe paper presents a generator of telecom alarm storms that runs on real backbone topologies and records the exact root cause and propagation path of each alarm. Output fields are bound to terms in NORIA-O, RFC 8632, RFC 8345 and TMF642. In the reported experiments, a simple rule that ranks the device raising the earliest critical alarm outperforms the authors' temporal method and an LLM in most settings. When a site loses power, only methods that walk the topology to a silent candidate device locate the fault.
In briefAgentic Architecture of Information Systems: The AAM/5C Reference Model for Distributed Autonomy
SummaryThis book addresses the difficulty large organizations face in managing hundreds of interacting components and reacting to incidents at human-team pace. It proposes AAM/5C, which spreads five capabilities across seven logical layers and centers on a single authorization rule, R-AUT, separating proposing, deciding and executing. The book reports no empirical results, and its validation program is set out but not yet carried out.
In briefEPIC: Exact-KL prefix clipping to rethink trust regions for autoregressive LLM policies
SummaryThis paper examines PPO-style clipping used in reinforcement learning to improve LLM reasoning, arguing that standard token-level clipping ignores that the prefix conditioning context is generated by the policy itself. The authors model autoregressive generation as a finite-horizon Markov decision process and build a prefix-level trust region using an exact reverse-KL term. The resulting method, EPIC, directly regulates prefix-level policy drift. In off-policy experiments with a sparse mixture-of-experts model, EPIC outperforms standard and sequence-level clipping baselines on mathematical reasoning benchmarks.
In briefDONIC: A general model that tracks two-handed dexterous manipulation learned from human demonstrations
SummaryThis paper presents DONIC, a single generalist bimanual tracking policy trained jointly across operation categories and human-motion sources. It also introduces BiHOI, a multi-view dataset of 1,575 sequences over 229 objects, whose reconstructed two-hand interactions are retargeted to robot references. The method combines proximal policy optimization with behavioral cloning and uses an object-centered per-hand reward. Without finetuning, the single checkpoint reaches 86.79% success on unseen trajectories of training objects and 57.32% on held-out objects, exceeding prior bimanual tracking baselines by 77.36 and 45.73 percentage points.
In briefHiddenLens: A Multi-Layer LLM Framework for Finding and Resolving Hidden Assumptions in Software Requirements
SummaryThe paper notes that natural-language software requirements often leave out assumptions about performance, operating conditions, and user behaviour, which become costly to fix if found late in implementation. The authors propose HiddenLens, a six-layer BERT-LLM pipeline that detects, classifies, questions, and resolves these hidden assumptions, and they release the HiddenLens Corpus of 312 requirements with 847 annotated assumptions. Reported results include an end-to-end F1-score of 0.79, outperforming keyword-based and QuARS-Adapted baselines, and an NFR classifier F1-score of 0.847.
In briefRemembering Through Change: Provenance-Preserving Hypergraph Memory for Multimodal Agents
SummaryThis paper proposes MHG-Mem, a hypergraph-based memory for multimodal agents that organizes entity identity, state evolution, and links to visual evidence. It also introduces a hypergraph-based agentic search that navigates this memory and inspects source observations when needed. The authors report performance gains of 8.20-9.98% over the strongest baseline across three benchmarks.
In briefSCDM: Scalable causal discovery for nonlinear time series via meta-learning
SummaryThis paper addresses causal discovery from nonlinear multivariate time series in high-dimensional systems. It proposes SCDM, which treats each target variable as a task and learns a shared temporal predictor and shared causal-strength matrix, updating only a subset of tasks per meta-episode. Experiments on synthetic, neuroimaging-inspired, high-dimensional, and real-world temporal benchmarks show competitive causal recovery and improved scalability in large systems.
In briefLooking Ahead: Teaching Autoregressive Video Models to Anticipate Scene Evolution
SummaryThe paper notes that autoregressive video diffusion models, which generate videos chunk by chunk, can produce plausible continuations without truly capturing how scenes evolve. The authors propose Future Chunk Prediction (FCP), an auxiliary objective trained jointly with teacher-forced denoising that predicts temporally separated chunks without the intervening visual context. FCP needs no extra pretrained visual encoder and leaves inference unchanged. According to the paper, it achieves better physical plausibility than standard teacher-forced training with a lower overall training compute budget.
In briefTIDE: Fine-Grained Diagnostic Evaluation of LLMs via Behavior-Driven Decision Trees
SummaryThis paper proposes TIDE (Tree-Induced Diagnostic Evaluation), which replaces single aggregate benchmark scores with a hierarchical capability tree. The tree is built recursively from benchmark samples and model scores, with an LLM judge proposing splits and behavior-based information gain choosing among them. Applied to five models across four benchmarks, TIDE reveals capability contrasts hidden at the aggregate level, such as one model leading on microeconomic reasoning while trailing on contract law reasoning.
In briefThe Cost of Inferring Current Effects in Nonstationary Bandits
SummaryAdaptive experiments use bandit algorithms to favor better interventions, but decision-makers also need reliable estimates of current treatment effects. The paper characterizes the tradeoff between reward maximization and current-effect inference. It establishes matching minimax bounds on dynamic regret and proposes the TRACE algorithm.
WhyCharacterizes a regret-precision tradeoff in nonstationary bandits with matching minimax bounds and an e-process extension, supported by numerical experiments.
In briefReconstruction Alone Is Not Enough: Doubly Robust Causal Mediation Analysis with Learned Multimodal Representations
SummaryThis paper examines how representations learned from high-dimensional multimodal health records can adjust for confounding when estimating counterfactual outcomes and interventional direct and indirect effects. The authors prove that covariate-only training objectives, such as reconstruction, cannot guarantee a representation is sufficient for causal identification, and that reconstruction accuracy can remain high while bias becomes large. They propose a workflow with cross-fitted doubly robust estimators and diagnostics, and in simulations, dropping a rare, moderately strong confounder yielded indirect-effect intervals with 2% coverage at n = 3,000.
WhyProves reconstruction-trained representations cannot certify causal sufficiency, with a linear counterexample and error decomposition; simulations support the claims but are limited in scope.
In briefFaultGene: LLM-Assisted Generation of Fault-Injected Tests for Hardware-in-the-Loop Validation of Self-Driving Systems
SummaryThis paper presents FaultGene, a framework that uses large language models to automatically generate executable fault-augmented tests for hardware-in-the-loop (HIL) validation of autonomous driving systems. It takes a baseline Scenic/CARLA scenario and a structured fault specification as input. Offline refinement uses semantic-consistency checks and execution feedback, and the refined data is used to fine-tune the generator so tests can be produced directly at deployment time.
In briefCulturalRiddles: A Multicultural Riddle Benchmark
SummaryCulturalRiddles is a benchmark of original riddles written by people through a community effort, covering 62 cultures across 45 countries and 51 languages. The riddles require abductive and compositional reasoning along with factual, social, and language knowledge tied to each culture. The authors evaluated 28 LLM configurations with automatic and large-scale human evaluation, finding that even frontier models reach at most 22% accuracy in some communities. All artifacts are publicly released with an open process for contributing more languages and cultures.
WhyProposes a human-authored, 62-culture riddle benchmark, evaluating 28 LLM configurations with automatic and large-scale human evaluation.
My Chiffon Nguyen, Tianyi Hu, Vinod Anbalagan et al.
In briefCORD: conflict-aware optimization for training agents with reinforcement learning and distillation together
SummaryThis paper addresses the problem that, when an agent is trained with both task-reward reinforcement learning and teacher-guided self-distillation, the two gradients can partly cancel each other. It proposes CORD, a general framework that combines the gradients of the two objectives once per mini-batch while selectively correcting conflicts, with a closed-form update derived from a nearest-point formulation. Across three joint training recipes, CORD improves ALFWorld success by 1.5 to 15.9 percentage points, and on WebShop it improves two of three pairs, with SDAR rising from 60.7% to 64.6%.
In briefCommit or Abstain? Decision Models, LLMs, and Missing Legal Authority
SummaryThis paper studies System One decision models for verifying legal authority; these models output calibrated probabilities for each option rather than generating text. The authors extend CaseHOLD with an abstention condition in which the correct holding is replaced in 1,670 of 5,314 items, and compare four decision models with six generative LLMs. When no option is correct, generative LLMs still pick one in many items, while the most cautious decision model does so less often. The authors argue that legal AI should be judged by its wrong citations, not by accuracy alone.
In briefApplying the InterModel Vigorish to Interpretable Machine Learning Model Fit
SummaryThis paper introduces the InterModel Vigorish (IMV) into machine learning to express differences in model fit as expected betting advantage over a reference model. It proposes three applications: SHAP-IMV for feature attribution, Multi-IMV for class discrimination, and Ablate-IMV for assessing components and feature groups by ablation. The methods are evaluated over ten random seeds on tabular data, images, sensor sequences, and graphs, and the authors release the open-source imvpy library.
In briefTraining-free speedup for blind compressed image restoration using activation-calibrated SVD
SummaryThis paper studies how to accelerate transformer-based blind compressed image restoration models without retraining. The authors analyze computation hotspots and channel-wise activation distributions, then propose Activation-Calibrated Singular Value Decomposition (ASVD), which derives low-rank factors offline from pretrained weights and a few calibration images. On PromptCIR at rank 64, ASVD improves PSNR-B over truncated SVD on four benchmarks while cutting parameters by 59% and computation by 16.1%.
In briefExistence of minimizers for the Gaussian Mumford-Shah segmentation model
SummaryThe paper studies the Gaussian Mumford-Shah model, which segments images by minimizing a multivariate Gaussian negative log-likelihood plus a perimeter penalty, and is used for hyperspectral images. The authors show the energy can be unbounded below when a region's covariance degenerates, regardless of perimeter. A quantitative covariance floor across all regions guarantees a lower energy bound and the existence of a minimizer. This condition follows from the image histogram's properties and holds almost surely under suitable noise.
In briefEnglish–Mizo legal machine translation: corpus, analysis, and benchmarks
SummaryIndian court proceedings are conducted in English, limiting access for Mizo speakers. The study presents the first English–Mizo parallel corpus of Indian court judgments, with 13,168 sentence pairs translated by professional translators, manually aligned, and checked by two reviewers. It analyzes selective English borrowing and inconsistent diacritic use. Fine-tuning NLLB 200 1.3B raised BLEU to 27.47 on a 1,000-pair legal test set, a 7.16 gain over the base model, yet the model still uses English words about three times more often than human translators.
Telem Joyson Singh, Sanabam Boyshana, Sanasam Ranbir Singh et al.
In briefDewHand: Direct real-scale 3D hand motion recovery from first-person videos
SummaryDewHand is a single-stage model that jointly estimates hand and camera motion in a shared world coordinate system from egocentric video, avoiding errors from multi-stage pipelines. It combines scene geometry features from VGGT-Ω with metric-space supervision to resolve scale ambiguity, and uses flow matching to handle occluded or out-of-view hands. Experiments on HOT3D and ARCTIC, plus zero-shot tests on HoloAssist, H2O, and TACO, report lower world-space hand error than prior methods.
Naru Suzuki, Satoshi Ikehata, Takehiko Ohkawa et al.
In briefToM2Vec: Latent tokens that help AI agents model partners' minds for teamwork
SummaryThis paper introduces ToM2Vec, which uses a fixed set of continuous reasoning tokens to build abstract theory-of-mind representations of partner agents. The tokens are trained to match the partner's actual thoughts and are then used by each agent to condition its own reasoning and actions. Across collaborative, competitive, and mixed-motive environments, it reports consistent gains over the base model and baselines, and generalization to unseen partner models including GPT-5.4.
In briefCombining graph topology with energy-based pathfinding to keep context consistent in retrieval-augmented generation
SummaryThis paper addresses semantic loss and modality mismatch when knowledge graphs are used in retrieval-augmented generation for multi-hop question answering. It proposes ThreadRAG, which builds context-intact semantic nodes and a semantic topology graph, then uses energy-driven pathfinding to extract query-relevant subgraphs and linearize them into textual evidence paths. The authors report consistent gains over strong baselines on three multi-hop QA benchmarks, including a 14.4% EM improvement on HotpotQA and 23.3% of the strongest baseline's token cost.
In briefFrontierCS 2: A Benchmark for Testing AI Agents on Computer Science Research
SummaryThis paper introduces FrontierCS 2, a benchmark of 48 open-ended tasks testing whether LLM agents can contribute to frontier computer science research. Each task removes the contribution from a paper's released code, and the deleted code serves as the human reference. The best agent, Claude Opus 5.5, passed 58% of its runs, and 16 tasks were not passed by any run.
WhyDeleting a paper's contribution from released code and grading agent patches on a private held-out suite is a novel design; evaluating 12 agents on 48 tasks gives solid evidence for frontier R&D asses
In briefBenchmarks as empirical difficulty measures, and how to diagnose their gaps
SummaryThis paper argues that a benchmark's set of test items naturally forms an empirical measure over question difficulty, and uses this view to evaluate benchmarks. It defines two failure modes: measure mismatch, where the difficulty distribution differs from the intended one, and coverage gaps, where parts of the difficulty range are untested. The authors propose two diagnostic metrics and analyze 25 popular benchmarks across 7 domains, finding many overweight certain difficulty regions while others leave large portions uncovered. They also introduce BenchFill, a difficulty-targeted item rewriting algorithm that reduces local coverage gaps.
In briefMultilingual instruction tuning that preserves tasks, evaluated on open-ended QA in WMT 2026 MIST
SummaryThis paper presents a submission to the WMT 2026 Multilingual Instruction Shared Task, which adapts the Qwen3-8B model with LoRA. The training uses organiser-provided data, EuroAlpaca (a task-preserving localised multilingual instruction dataset), and CrossEuroAlpaca (a cross-lingual augmentation). Adaptation raised the Gemma-4 judge score on target-language open-ended QA by 7.8% relative to the base model. The gains are task- and language-specific, since context-grounded QA declined in target languages and all three tasks declined in non-target languages.
In briefPrivacy depends on context, which general-purpose LLM designs often ignore
SummaryThis paper argues that efforts to build a single, unified framework for prompt privacy in large language models are misleading and may give users a false sense of security. The authors note that privacy is inherently contextual and that desirable privacy properties can conflict. They recommend application-specific deployment, context-specific evaluation of privacy frameworks, and better privacy literacy.
In briefUnintended side effects of teaching LLMs specific values
SummaryThis paper examines how post-training LLMs to align with specific human values affects their other capabilities and behaviours. The authors fine-tuned four instruction-tuned models on individual Hofstede value dimensions and tested cross-value transfer and broader behaviours. They find that value-specific training causes unintended side effects on untrained values, risk preferences, social cognition, and safety traits. The authors therefore recommend that value alignment include broader behavioural evaluations.
Siddhesh Pawar, Chen Cecilia Liu, Anna Korhonen et al.
In briefPreventing Model Extraction in Contextual Linear Optimization
SummaryThis paper studies how to protect a predict-then-optimize system from attackers who submit context queries to extract its underlying prediction model, even though only optimal decisions are returned. The authors give conditions under which a linear model can be learned up to positive scaling from decisions alone, and propose a defense that perturbs decisions only when predicted costs lie near a decision boundary. They derive a regret bound in terms of the perturbation radius and validate the security-performance trade-off experimentally.
In briefCausalAlign: Causal alignment of video generation models for consistent camera motion
SummaryThis paper addresses the problem that diffusion-based video generation models often let camera ego-motion drift over long sequences. It proposes CausalAlign, which first uses linear probes and causal interventions to diagnose how ego-motion is represented within the model, then aligns representations at identified layers with a 3D foundation model in a delta-token space. Fine-tuning Wan2.2 on DL3DV and RealEstate10K improves motion accuracy, geometric consistency, and temporal smoothness while preserving visual quality. The fine-tuned model also raises the success rate of world action models on mobile manipulation tasks in the RoboCasa365 benchmark.
In briefAsking a Multimodal LLM If Someone Fell: Zero-Shot Fall Screening in Low-Resolution Thermal Video
SummaryThis paper tests Gemini 3.8 Flash, a multimodal large language model, as the decision step of a fall screener that judges short low-resolution thermal clips without task-specific training, comparing it with its predecessor Gemini 3.7 Flash. On a public dataset of 438 staged falls and 563 non-fall windows, the deployed request detected about 69% of falls from held-out subjects and flagged none of 499 non-fall windows. In an occupied room over 76 hours of ordinary life, it flagged none of 395 clips as falls and caught 4 of 9 visible staged falls.
In briefTractable Inference with Kolmogorov-Arnold Networks (TOUCKANs)
SummaryThis paper introduces TOUCKANs, Kolmogorov-Arnold networks with finite-degree polynomial outer functions, so that expectations, variances and higher moments can be computed exactly by deterministic recursions instead of Monte Carlo sampling. The same operators can enforce conserved integrals and normalization during training. Benchmarks up to 50 dimensions, a Cahn-Hilliard mass-conservation test, and joint computation of the four lowest Schrödinger eigenpairs are reported.
WhyCombines KANs with tractable probabilistic circuits for deterministic moments and integral constraints; benchmarks to 50 dimensions against Monte Carlo show solid evidence.
Juan Diego Toscano, Sascha Ranftl, Zhiwei Gao et al.
In briefBenchmarking Language Models on Satire Detection in Code-Mixed Romanian Text
SummaryThis paper examines how mixing English, German, French, and Italian into Romanian text affects satire detection, a task in an under-resourced language. The authors generate over 600,000 synthetic code-mixed sentences with four methods and release the resulting CS-SeLeRoSa dataset, alongside a human-checked gold benchmark. They evaluate several multilingual language models, and report that fine-tuning raises Gemma 4 31B to 77.5% on the synthetic corpus and 72.6% on the gold standard.
R\u{a}zvan-Alexandru Sm\u{a}du, Ioana Buhnila, Daniela Tanase et al.
In briefAI Mindmeld: A Latent Interconnect Protocol for Feeding Specialist Representations to LLMs
SummaryThe paper proposes letting large language models receive specialist models' internal representations directly, instead of passing information through text, images or other intermediate forms. The authors train connectors that map protein and small-molecule specialist representations into LLM internal states, keeping all pretrained models frozen. Experiments with Gemma 3 and Kimi K2.6 show the connectors supply useful information without changing the LLM, transfer zero-shot to cross-domain tasks, and generally improve with LLM capability.
WhyTrainable connectors map specialist representations directly into frozen LLM states, showing compositional zero-shot transfer across several models, though experimental scope and baselines are limited
In briefViBMR: A Large-Scale Vietnamese Brain MRI Dataset for Automated Radiology Reports
SummaryThis paper introduces ViBMR, described as the first large-scale multi-sequence brain MRI dataset paired with routine Vietnamese clinical reports, covering 3,649 patients and 11,875 MRI series across T1, T2, FLAIR, and contrast-enhanced T1. The authors build a comprehensive evaluation for brain MRI report generation, covering representative medical vision-language models, multi-sequence inputs, and feature aggregation strategies. The results indicate that native low-resource clinical datasets are necessary for reliable Vietnamese brain MRI report generation. The dataset and baseline models will be released for non-commercial research.
Dai Lam Le, Minh Khoi Nguyen, Son Dinh Nguyen et al.
In briefA Sharp Dynamic Haussler Lemma (a tight sequential covering bound for learning classes)
SummaryThe paper extends Haussler's covering lemma to a dynamic, sequential setting for bounded real-valued classes under normalized ℓp distance, for 1≤p<∞. It gives a global dynamic cover of size at most (C/ε^p)^d, where d is the sfat dimension at scale ε/32. A matching lower bound holds for binary classes, with universal constants independent of p.
WhyProves a sharp global dynamic covering bound for bounded real classes, removing logarithmic overhead and resolving several open questions with complete proofs.
In briefHow learner grammar errors affect social bias in large language models
SummaryThe paper notes that existing social bias benchmarks mostly use fluent, well-formed English and so overlook non-native writers. The authors built a controlled framework that inserts common learner errors into prompts and tested it on three bias benchmarks covering gender, race, and religion. Grammatical errors were found to increase bias, with the effect growing as the error rate rises, and preposition errors caused the largest and most consistent increase in stereotype alignment. Fine-tuning on erroneous texts was found to reduce the bias increase caused by grammatical errors.
Federico Marcuzzi, Martin Christoph, Iryna Gurevych et al.
In briefReview of machine learning and large language model methods for automatically extracting materials data and knowledge from literature
SummaryThis review traces how methods for extracting materials data from scientific papers have evolved, from rule-based pipelines and NLP models to transformer encoders and large language model agents that handle text, tables, and figures. It covers the resulting databases and knowledge graphs and their uses in question answering, research assistants, and closed-loop optimization with autonomous labs. The authors assess extraction accuracy and computational cost, offer practical guidelines, and identify open challenges such as sample-level structure-property-processing-performance extraction, scarce benchmark datasets, and non-standard terminology.
Aravindan Kamatchi Sundaram, B. Pabitramohan Prusty, Rohit Batra
In briefExpandable Expert Modules for Medical Image-Text Learning
SummaryThis paper proposes a medical vision-language method built on an expandable bank of modality experts within a Multiway Transformer. It pretrains on 19.4 million unpaired images across seven modalities, then aligns images and text using 3.23 million pairs. Fundus and OCT experts are added while existing paths remain frozen. The authors report best or tied-best results in 24 of 25 benchmark settings across nine modalities, and retrieval gains in a sequential experiment with no measured loss on earlier tasks.
In briefStochastic Activation Switching for Lightweight Uncertainty Estimation
SummaryThe paper proposes stochastic activation switching, where network units randomly choose between multiple nonlinear activation functions during a forward pass, to create predictive diversity while keeping intermediate features active. This enables uncertainty estimation from a single trained model without ensembles or task-specific uncertainty heads. The authors evaluate the approach on classification, semantic segmentation, and out-of-distribution detection using convolutional and transformer architectures, reporting improved uncertainty quality with competitive primary-task performance.
Mohammad Sarhangzadeh, Amirreza M. Shebly, Doruk Oner
In briefPAUSE: Steering coding agents with one plan before their first code edit
SummaryPAUSE is a framework that intercepts a coding agent's first request to edit source code and has a separate Planner read the task and the pre-edit transcript to write one repair plan. The original agent then resumes with that plan, without a learned monitor, checkpoint search, or an extra model call. Evaluated on SWE-bench Verified and SWE-bench Multilingual, online steering improves resolution rates. It achieves this at lower Planner cost than a rule-triggered advisor and roughly 11-15% of the cost of a periodic overseer.
In briefGeoMem-VLA: Storing 3D Scene Understanding as Temporal Memory for Robot Action Models
SummaryThis paper presents GeoMem-VLA, a vision-language-action framework for robotic manipulation. During training, a frozen geometry foundation model supervises learnable GeoPlan tokens via a geometric flow-matching objective over the current scene point map and future end-effector tracks. These tokens are maintained as temporal memory and fused with current observations to predict actions. At inference, the 3D teacher is not used and the scene is not explicitly reconstructed. The authors report high success rates on LIBERO, RoboTwin, and five real-world tasks.
In briefSilent Alignment in Two-Layer Linear Networks: A High-Dimensional View
SummaryThis paper studies gradient flow in a two-layer linear network when input, hidden, and output dimensions grow proportionally. Combining a closed-form solution of the dynamics with prior random matrix theory, the authors derive asymptotic spectral distributions of the weight matrices and describe how the learned subspace evolves. With small initialization, the network's singular directions align on an order-one timescale while singular values and loss stay at their initial values; visible learning follows later on a logarithmic timescale.
In briefHow DiTs Denoise: An Aggregator-Predictor View of Attention and MLPs
SummaryThis paper explains how diffusion transformers (DiTs) remove noise from images by treating attention as a context aggregator and MLPs as patch predictors. For Gaussian Markov random field priors, the authors show that a global Bayes denoiser can be approximated using only a subset of context patches, with error decaying exponentially in context radius. In a controlled CelebA64 study, shared-memory patch predictors replacing DiT MLPs achieve competitive denoising and generation performance.
In briefRangeNav: Task-Aligned LiDAR Range-Image Pretraining for Robot Navigation
SummaryThis paper presents RangeNav, a compact LiDAR range-image encoder with about 183k parameters, pretrained only on synthetic data for robot navigation. Pretraining combines range reconstruction, semantic segmentation, and bird's-eye-view prediction, and the encoder is then frozen for reinforcement learning policies. In simulation, adding semantic and BEV supervision raised success rates over a reconstruction-only baseline by 32.03 and 23.73 percentage points on average. Zero-shot deployment on a Unitree B2-W quadruped showed navigation through dense crowds, low light, and construction zones.
In briefEvery Part Has Its Role: How Part of Speech Shapes LLM Post-Training Gains
SummaryThis paper studies concept augmentation, which expands training data by swapping a target word with conceptually related alternatives. The authors extend prior work from nouns to verbs and adjectives and test whether the part of speech carrying the concept affects perplexity and downstream performance. Verbs gave the most reliable downstream gains, while perplexity improved most consistently on news text regardless of part of speech.
In briefWide-band mid-infrared imaging using a single silicon hyperboloid metalens
SummaryThis paper demonstrates imaging across the full 3-5 micron mid-wave infrared band using a single-layer silicon metasurface combined with computational reconstruction. A simple 2 cm diameter hyperboloid metalens is used instead of complex achromatic designs. The reconstruction model is adapted using only four image captures.
