In briefMMKG-R1: Building Multimodal Knowledge Graphs with Reinforcement Learning
SummaryThis paper proposes MMKG-R1, a reinforcement learning framework that trains a multimodal large language model to construct multimodal knowledge graphs with graph quality as the main training signal. It introduces several quality-aware rewards that directly assess the information extracted from each knowledge source and the overall assembled graph. Experiments across multiple datasets show that the constructed graphs improve downstream question-answering accuracy.
In briefCan AI Peer Review Be Trusted? A Study of AI Score Reliability
SummaryThis paper examines whether AI scores from large language models can reliably judge research papers. Using 200 papers from ICLR 2025 and 2026 and about 14,000 AI reviews, the authors study score stability, score-distribution estimation, and whether different models agree on the gains from revisions. Repeated reviews vary even under identical conditions, finer scoring protocols reduce variance but shift mean scores, and revision gains depend strongly on the reviewer.
In briefLLM-jp-4 Technical Report: A Transparent Japanese Language Model Family
SummaryThis paper presents LLM-jp-4, a family of Japanese large language models built for transparency and reproducibility. It includes an 8B dense model and a 32B-A3B mixture-of-experts model, both trained from scratch on about 12 trillion tokens. The authors state explicit criteria for choosing training data, external models, and software, and make the development process public. Results show clear gains over earlier LLM-jp models and competitive performance against open-weight models of similar size.
LLM-jp: Chaoran Liu, Daisuke Kawahara, Hideyuki Tachibana et al.
In briefA polylogarithmic bound on list replicability
SummaryThis paper addresses the list-replicability number for classes of M binary concepts over N domain points in learning theory. The authors give a short, self-contained proof that this number is O(log²M·log N), which would be the first polylogarithmic upper bound on this quantity. The argument follows the high-level approach of Zhang's manuscript but uses considerably simpler arguments.
In briefNavAble: A Large-Scale Dataset and Synthetic Data Pipeline for Blind Navigation Perception
SummaryThis paper presents NavAble, a large-scale dataset covering 11 accessibility-critical object classes such as audible pedestrian signals, door buttons, and handrails, combining 41K open-source images with 8K newly annotated real-world images. The authors also build a synthetic data pipeline that produces 452K annotated frames from 565 3D assets across 37 environments. Fine-tuning segmentation models on the dataset reaches over 80% mIoU on a held-out test set, and synthetic augmentation adds up to +3.5 mIoU.
Hochul Hwang, Jahir Sadik Monon, Soowan Yang et al.
In briefSonoWorld: A Pose-Controllable Ultrasound World Model for Interactive Simulation
SummaryThis paper presents SonoWorld, a causal ultrasound world model intended to serve as a controllable simulation environment for robotic ultrasound. It learns video dynamics from about 6.9 million unlabeled multi-organ frames and is grounded to 6-DoF probe poses using a small EM-tracked thyroid dataset. In model-based action discrimination, it ranks the target-directed probe motion above an equal-magnitude opposite motion in up to 74.8% of trials.
In briefWhy graph transformers do not all attend to the same node
SummaryLarge language models often concentrate attention on the first token, a pattern called an attention sink. This paper finds that the graph transformer GRIT rarely forms such sinks on molecular graphs, and forms fewer when trained on larger graphs. The authors show the main cause is the attention scoring function: replacing GRIT's gated logit with a dot product raises the sink rate from 0.09 to 0.68 on identical graphs. They conclude that attention sinks are a property of dot-product attention rather than of Transformers in general.
In briefTopicMerge: A Divide-and-Conquer Method for Topic Modeling
SummaryThis paper proposes TopicMerge, a multi-stage topic modeling pipeline that combines efficient embedding-based clustering with Large Language Model (LLM) judgments. It first over-segments a document corpus into many fine-grained micro-clusters, then uses an LLM, guided by local neighborhood context, to decide which candidate clusters to merge and relabel. Experiments on the Bills and Wiki datasets report state-of-the-art clustering performance with over 92% fewer LLM calls than document-level methods.
In briefStability and generalization analysis of biased gradient methods for minimax problems
SummaryThis paper studies how biased stochastic gradient methods, used to solve minimax optimization problems, generalize to unseen data, which had not been analyzed before. It relaxes the usual bounded-gradient assumption to a bounded second moment condition and links stability to generalization. Applied to zeroth-order and clipped SGDA, the stability bounds match those of standard SGDA under suitable parameters. Combined with convergence analysis, the authors derive optimal excess risk bounds of order 1/√n, where n is the sample size.
In briefTruthful, payment-minimizing LLM routing for graph-based agent workflows
SummaryThis paper studies how to route tasks in graph-based agentic workflows to different LLM providers while reducing generation cost, since dependent tasks can reuse context when kept with the same provider. The authors model routing as a single-parameter procurement auction and propose STAR, which minimizes virtual-cost-weighted graph workloads and makes truthful cost reporting utility-maximizing. Under independent cost distributions with non-decreasing virtual costs, STAR also minimizes the user's expected total payment. Synthetic experiments show topology-aware execution savings grow with fan-out and payload size, while payment savings increase gradually with prior heterogeneity.
In briefResidual adaptation using historical features for sequential screening
SummaryThe paper studies how sequential screening policies can use feedback under a fixed experiment budget. It applies a small ranking correction built from fixed historical features, anchored to base scores recorded before feedback, to a pretrained AssayFormer model. On 20 official-test screens, source-macro hit recovery rose from 22.58% to 28.61%, beating six prespecified comparators. Transfer to an independent nonlinear base model also improved recovery for random forests, but the gains depend on the base model.
In briefPersEvalBench: Can LLM Evaluators Judge Appropriate Use of User Persona Information?
SummaryThis paper introduces PersEvalBench, a benchmark with 403 multi-episode scenarios that tests whether LLM evaluators can recognize appropriate and inappropriate uses of user history in personalized assistants. It covers nine controlled error subtypes across three families. Across five LLM evaluators, recognition of these errors remains limited even with full dialogue history; on under-personalization, humans recognize 94.4% of errors versus 63.9% for the strongest evaluator.
In briefData-free multistep diffusion distillation: one model supporting several inference step budgets
SummaryThis paper studies how to compress diffusion generative models into multistep student models without training data. The proposed method, Progressive Multistep Data-free Distillation (PMDD), assigns a dedicated adapted network to each student distribution and adds a stage-wise unforget loss to preserve early-step generations. According to the paper, PMDD matches or outperforms data-free few-step baselines at each multistep budget while requiring 4-5x less training compute than data-free DMDv2.
In briefDTDD: Divergence-Triggered Dynamic Distillation for More Reliable On-Policy Training
SummaryThis paper proposes DTDD, an adaptive on-policy distillation method that lets a teacher model take over when a student diverges from it during a reasoning segment. The teacher generates a recovery suffix, which the student learns through behavior cloning. The method is tested on long-horizon reasoning, distillation from a base model, and multi-turn agent tasks. Reported results show lower peak gradient norms and fewer updates than standard on-policy distillation, along with gains on some agent benchmarks.
In briefSemantic IDs replace ICD code strings for AI-based medical coding
SummaryThis paper introduces SemICD, a framework that trains large language models to generate Semantic IDs instead of raw ICD code strings for automatic medical coding. The IDs are built by hierarchically clustering embeddings of textual code descriptions, so related codes share prefixes while each full ID maps to one ICD code. Experiments on MIMIC-III and MIMIC-IV with four LLMs report that SemICD outperforms direct code generation on Micro-F1.
Yihang Cheng, Qian Zhang, Yonghui (Andie) Huang et al.
In briefVerifier-Grounded Agentic RL for Provably Correct Code: training code models with a formal verifier as the reward
SummaryThis paper proposes an agentic reinforcement learning setup that uses a formal program verifier, rather than passing tests, as the training signal for code generation. The reward scores candidates by the weighted fraction of proof obligations discharged and rejects any candidate that alters the specification or assumes what it should prove. Across Qwen3 models, held-out verification rises from 66.1% to 93.5% on DafnyBench and from 56.3% to 100% on SWE-Proof.
WhyTurns verifier diagnostics into a fine-grained reward with integrity auditing; large gains across several benchmarks and held-out suites suggest solid evidence and moderate impact.
In briefQwenfant: Character-level baby language models trained on concise parallel text
SummaryThis paper describes a submission to the Multilingual track of the 2026 BabyLM Challenge. The authors improve their baseline by training on data summarized and translated by an LLM, showing that synthetic data can be higher quality. They further use a larger model and a character-level vocabulary, both of which require LoRA-style fine-tuning to reach full potential. The final Qwen3-based model, Qwenfant, ranked first in the Multilingual track at the time of submission.
In briefWhose turn is it? Context-guided model routing for stateful agents
SummaryThis paper studies routing each user turn of an LLM agent to cheaper or more expensive models to cut costs while keeping quality. Tested on multi-turn versions of SheetCopilot and SpreadsheetBench 2, a router given session history achieved higher task pass rates than the same router without history at matched offloading rates. The authors also analyze how switching models affects prompt-cache costs asymmetrically.
Anikesh Parashar, Avik Dutta, Vashu Chauhan et al.
In briefCan LLM Agents Efficiently Identify Appliance Faults? Insights from a Simulated Coffee Machine
SummaryThis paper introduces CoffeeBench, a fault diagnosis task for LLM agents using a simulated appliance. Most evaluated models struggle to identify faults from sensor data alone, but the best performers succeed by iteratively narrowing the set of possible faults. The authors compare models by diagnostic efficiency, examine why the three weakest models lag, and identify uncertainty and reasoning soundness as useful predictors of success.
In briefKandinsky S2V 1.0: Generating Talking Avatar Videos from Audio
SummaryThis paper presents Kandinsky S2V 1.0, an open family of speech-to-video models in Lite (2B) and Pro (21B) sizes. From a reference image, speech audio, and a text prompt, they generate 25 FPS videos where lip movement, expression, pose, and gestures follow the speech. The authors report that the Pro model performs comparably to or better than baseline models in human evaluation and on benchmarks, and a distilled version runs 5.4 times faster. Code, checkpoints, and diffusers integration are released under the MIT license.
WhyIntegrates audio, text, and reference memory into a video foundation model with face loss and distillation; human and five-metric benchmarks support claims, impact mainly within speech-to-video genera
Nikolay Semenov, Artem Riabinin, Sergey Bassenko et al.
In briefFrom Efficiency to Sufficiency: A Cost-Aware Framework for Automated Essay Scoring
SummaryThis paper treats automated essay scoring (AES) as a cost-constrained measurement system and introduces the Cost-Aware Measurement Design (CAMD) Framework, with four layers: metering, allocation, unit economics, and decision. It defines cost per defensible score (CPDS) and argues that AES designs should be compared by choosing the lowest-cost option among those meeting measurement requirements. Because scoring volume affects all layers, CPDS is treated as a function of scale rather than a single fixed number.
In briefExact scaling laws for a random-features bigram model, a solvable toy for language-model scaling
SummaryThis paper studies capacity-limited scaling in a bigram model with power-law (Zipf) token statistics, used as an analytically tractable toy of a large language model. Fixed random embeddings stand in for representational capacity, and the authors derive scaling laws for prediction regret as a function of model width, vocabulary size, and the Zipf exponent. In the appropriate regime the theory matches the functional form of classical neural scaling laws, and experiments on WikiText2 and OpenWebText closely follow the predicted curves.
In briefDimension-free scaling laws for models that generalize across input sizes
SummaryThis paper studies whether a model trained on inputs of one size can generalize to inputs of a different, possibly much larger, size, and what statistical cost this entails. The authors show that polynomial source decay is insufficient for efficient size generalization: in-domain error decays as n^{-Θ(1)}, while size-generalization error decays only as (log n)^{-Θ(1)}, an exponential separation in sample complexity. They identify an exponential source condition that restores polynomial sample complexity and provide matching minimax lower bounds.
WhyProves an exponential separation in sample complexity for size generalization, with matching minimax lower bounds, a notable theoretical advance.
In briefDeep learning for atmospheric turbulence mitigation: comparing GANs and Transformers
SummaryThis paper compares two deep learning approaches for reducing image distortion caused by atmospheric turbulence: Generative Adversarial Networks (GANs) and Transformer-based models, both trained and tested on the SOTIS dataset. GANs performed better in severely degraded, low-quality conditions, while Transformers did better in moderate to high-quality settings, particularly at correcting geometric distortions. The authors also improved the Turbulence Mitigation Transformer with a refined training strategy, a temporally averaged prior frame, and an SSIM-based loss, yielding consistent gains across turbulence conditions.
In briefDiffusion language model training that routes by verifier outcome
SummaryThis paper proposes a likelihood-free training pipeline for a mixture-of-experts diffusion language model. Rollout groups are routed by verifier outcome: mixed groups use ratio-free group-centred REINFORCE, while groups where all student rollouts fail but a teacher succeeds receive dense KL supervision on the student's denoising canvases. A KL term anchored to the distilled checkpoint limits changes outside the verifier reward. The authors report results on tool-calling, reasoning-trace and direct-answer accuracy, with pre-registered arms and paired evaluation.
In briefInterpreting Deep Learning Forest Classification by Separating Similar Tree Classes with Sentinel-2 Time Series
SummaryThis paper compares recurrent neural networks with Random Forest and XGBoost for tree-species classification from Sentinel-2 time series, focusing on separating pure from mixed Norway spruce pixels. The best RNN reached a Macro-F1 score of 0.851, outperforming XGBoost by 0.012 and Random Forest by 0.103. Using Integrated Gradients, the authors found that attribution increased after about day of year 125, with band B11 contributing most when spectral divergence between the classes grew.
In briefAuditing whose needs benchmarks represent, using a user-centric framework
SummaryThis paper proposes a user-centric audit framework that separates users who contribute benchmark content from those whose needs evaluation results are meant to represent. Across 16 resources, American-centered conventions make up 62.4-94.4% of matched signals in 11 text benchmarks, and geographic visibility in five image-text resources tracks national GDP share (mean Pearson r = 0.780). All 16 resources document source-user context, but none document target-user context.
In briefMemeTrace: A Multimodal LLM Agent for Tracing Meme Origins and Risks
SummaryOnline memes can carry unsafe content such as disinformation and propaganda, and they change quickly in response to news, making them hard to moderate. Existing methods usually label memes as safe or unsafe from single images and do not analyze their origin or spread. This paper introduces MemeTrace, a multimodal LLM-based agent that automatically traces the provenance and potential risks of emerging memes. On Meme100, a dataset of 100 real-world memes, it substantially outperforms a VLM baseline, detects unsafe memes more accurately, especially those involving disinformation and sensitive content, and stays robust under visual perturbations.
In briefLAMP: Estimating which 3D light sources light specific points in a scene
SummaryExisting lighting estimation methods mainly recover lighting for rendering but rarely state how individual physical light sources contribute at specific locations. LAMP uses vision-language models to take a single RGB image with marked query points, rank contributing light sources for each point, and predict their 3D directions. On the Infinigen test set, fine-tuning Qwen3-VL-8B raised dominant-source 3D PCK@0.5m from 0% to 70.6%.
In briefMIND: Incremental Grounding for Robots That Maintain Room Decisions from Instructions
SummaryThe paper introduces MIND, a framework that lets robots following persistent language instructions update only the cached room judgments affected by environmental events, instead of re-evaluating every room. Experiments on ProcTHOR sequences, AI2-THOR closed-loop episodes, and Unitree G1 robot trials report that it reduces semantic evaluations while preserving the decisions of full recomputation.
In briefLong-video generation using slot memory addressed by region tokens
SummaryThis paper introduces SMART, a slot memory addressed by region tokens for long-horizon video generation. It stores content in a constant-size slot recurrence, with writes addressed by the observing camera and reads queried by the rendering camera, without building explicit 3D representations or growing caches. On long trajectories in Memory Maze and Minecraft, SMART maintains high-fidelity recall beyond its training horizon with constant memory and per-step cost.
In briefFairness across several sensitive attributes, joint Wasserstein audits and Shapley attribution
SummaryThe paper addresses fairness audits involving several sensitive attributes, noting that checking one attribute at a time can miss disparities that appear only in combinations of attributes. Using one-dimensional Wasserstein geometry, it studies distributional demographic parity for a real-valued score and builds a Hilbert-valued variable from full-cell conditional quantile functions whose variance equals the joint disparity. Shapley values attribute this disparity across attributes without depending on an ordering, reducing to a Hoeffding-ANOVA decomposition under independence. Estimators, permutation calibration, and illustrations on synthetic data and ACSIncome are provided.
Agathe Fernandes Machado, Charlotte Jamotton, Arthur Charpentier
In briefCopy-Paste Language Models: adding a lightweight copy head to improve in-context copying
SummaryThis paper proposes Copy-Paste Language Models (CPLMs), which add a lightweight Copy-Paste head to the standard language modeling head to explicitly model copying from context. The authors pretrain models from 135M to 1.7B parameters and report improvements across all scales and backbones with negligible training and inference overhead. They attribute the gains partly to an alternative gradient path that partially mitigates the LM head gradient bottleneck.
In briefFaster Graph Rebuilding for Protein Structure Prediction via Compensated Topology Passes
SummaryThe paper addresses the costly rebuilding of k-nearest-neighbour graphs in SE(3)-equivariant structure prediction, which is needed at every block boundary but slows training. The authors propose a two-pass scheme that records light state in a first pass and applies conservative compensation in a second, with theoretical results including non-flip certificates and a trust-region projection. Experiments show the L2 level reproduces the per-boundary rebuild edge set exactly and the L1 level stays within 0.6%, with measured speed-ups of 1.5x for Pass-1 and 1.76x per epoch and no measurable loss gap..
In briefFinGREAT: Evaluating LLMs on Graph-guided Retrieval and Classification of Financial Accounting Relations
SummaryThis paper examines whether open-weight large language models can extract and classify evidence in financial disclosures that matches predefined double-entry accounting relationships. The authors use an accounting graph with 11 account nodes and 11 transaction-type edges, testing mid-sized open models on 1,300 SEC 10-K Item 7 filings from 26 years alongside three proprietary models. They also introduce a human-annotated benchmark of 483 passages, where the best open-weight and proprietary models reach 84.1% and 92.8% accuracy respectively.
Ashman Mehra, Bryan Routledge, Jane JaeYeon Pyo et al.
In briefHow LLMs write code using reusable, composable internal concepts
SummaryThis paper studies the internal activations of large language models while they generate code. The authors find that middle layers form abstract representations that are reusable across programming languages and can be steered to independently control function, language, and implementation strategy, with these controls composable. They also report that fine-tuning on arbitrary functions in one language transfers to other languages, and that transformers trained on synthetic data learn similarly reusable structure.
Casper L. Christensen, Kyle J. Ray, Dmitry Manning-Coe et al.
In briefFrom Detection to Anticipation: Predicting Stuttering Onset in Speech
SummaryThis paper frames stuttering prediction as continuous causal time-to-event forecasting, estimating the risk of an upcoming stutter every 100 ms using only preceding speech. The authors extended the FluencyBank Timestamped dataset for 36 speakers with manually reviewed repetitions, prolongations, and blocks, yielding 2,872 forecasting events. A compact causal Conformer built on frozen Whisper-large-v3 representations achieves an AUROC of 0.63.
In briefWrite Once, Plan Many: Reusable Visual Planning Systems from One Scene
SummaryThe paper presents a training-free framework that builds reusable perception and planning programs from a domain specification and a single practice scene, without solution labels. A vision-language model supplies a renderer and dynamic model, and checks comparing rendered states and simulated action outcomes guide refinement. After refinement, the programs are frozen and solve new instances without further VLM calls. Across fifteen 2D and 3D domain variants, the systems reach 99.3% average planning success.
In briefSpectral-MoE: assigning expert usage unevenly across layers when converting dense LLMs to MoE
SummaryThe paper proposes Spectral-MoE, a method for converting pretrained dense feed-forward networks into mixture-of-experts by splitting neurons into experts. Instead of using the same number of active experts in every layer, it learns to allocate the utilization budget heterogeneously across layers, with a static per-layer scalar and a dynamic per-token schedule. The authors report that both allocators outperform uniform allocation on analytic and learned conversion baselines across several models.
In briefDo Syntax Heads Influence Model Behavior? A study of whether attention heads specialized for grammar causally shape language models' syntactic predictions.
SummaryThe paper examines whether Transformer language model attention heads that specialize in syntactic relations, such as linking verbs to their subjects, play a causal role in syntactic predictions. Even when these heads were forced to attend to a distractor noun, the model's predictions still aligned more with the subject. Interventions on specialized syntax heads had a significantly stronger effect than interventions on random heads, but only on one of the two measures proposed.
In briefAgentic Self-Awareness: Aligning What AI Agents Know, Do, and Maintain
SummaryThe paper studies how AI agents behave when tasks seem solvable but may lack prerequisites, have conflicting requirements, or reveal blockers only during execution. It defines agentic self-awareness (ASA) as aligning feasibility judgments, warranted actions, and responses to challenges, using paired feasible and blocked tasks from Terminal-Bench 2.1 and retail τ²-bench. In a baseline, blocker recognition reached 84.4% but strict self-awareness only 40.6%, and harness search without weight updates raised strict SA to 93.8% on reused search tasks. Retail held-out gains were limited, and blocked-task transfer on Terminal-Bench was not measured.
Qinglin Chen, Hiroaki Hayashi, Jiaxin Zhang et al.
In briefGrounding Definitions in Theory-Level Auto-formalization
SummaryLLM-based auto-formalization systems can now generate formally verified Lean code quickly. The paper proposes definition grounding, which continuously judges and revises definition formalizations so that later results can reuse them. Only 25% to 50% of definition sets from textbook text were compatible with all later statements, but clarifying the informal definitions raised this to 77.5% on average across models.
Leyan Pan, Christian Szegedy, Frederick Rajasekaran et al.
In briefREPORT-OPS: An evidence-guided reporting agent for HPC operations with human-supervised policy refinement
SummaryThe paper presents REPORT-OPS, a hybrid agent that generates operational reports for high-performance computing facilities. An LLM interprets requests, plans evidence collection and drafts reports, while predefined policies and automated checks decide whether evidence supports the conclusion, so the system answers, qualifies, or withholds claims. On 68 reporting queries it reached 68.63% Pass@80, versus 7.02% for the strongest retrieval baseline. The authors note their results do not establish causal discovery or cross-facility generalization.
Mah Kadidia Konate, Mayur Akewar, Janki Bhimani et al.
In briefFollowing the fixed-point curve: how a continuous thought machine classifies images
SummaryThis paper studies how the Continuous Thought Machine (CTM), a recurrent network, classifies ImageNet images. The authors find its thinking steps are attracting but strongly non-normal maps, with training increasing non-normality, so the neural state has a locally unique attracting fixed point that moves each step. The thought trajectory is the curve these fixed points trace, and its motion and attention schedule are largely shared across images. The authors suggest the CTM has found a single strategy for processing all ImageNet images.
In briefA gradient-regularized deep learning method for restoring images for people with myopia
SummaryMyopia causes spherical defocus aberrations that reduce the perception of fine spatial details. The paper trains a U-Net neural network to pre-transform images so as to counteract defocus modeled by the second-order Zernike polynomial Z2^0, allowing myopic viewers at a fixed distance to see scenes more clearly. Training uses only the KITTI dataset, and testing is done on DIV2K and ImageNet. Compared with classical baselines, the method improves PSNR by 1.48 to 7.34 dB and yields significantly lower perceptual error.
Kebin Contreras, Luis Toscano-Palomino, Rafael Torres et al.
In briefNarrowing the Collaboration Gap: Comparing action sharing with calibrated probability sharing among AI agents
SummaryThis paper studies how teams of language model agents, each holding different private information, can communicate to make better shared decisions. It compares sharing only proposed actions with sharing calibrated probabilities over actions, and finds that probability communication performs better when calibration is preserved. When the shared probabilities are biased, this advantage disappears, so the authors introduce a post-hoc conversational calibration method that corrects these biases and test it on a toy transformer and pretrained LLMs.
Mirah Shi, Marcel Hussing, Ira Globus-Harris et al.
In briefPREFER: Using Player Experience to Personalize Game Refinement
SummaryThis paper introduces PREFER, a framework that uses experience dimensions adapted from psychometric questionnaires to benchmark how large language models evaluate gameplay. It uses personas representing different player profiles to select refinement objectives, revise game rules, and playtest candidate games. Experiments with five personas show PREFER achieves 9.2% higher mean post-play preference ratings than the strongest evaluated baseline.
In briefDecoupled Drift-Diffusion World Model for Consistent Future Frame Prediction
SummaryThis paper proposes the Decoupled Drift-Diffusion World Model (D3WM), which predicts future image frames from action sequences. A transformer outputs both a deterministic next-state prediction and a spatial map of aleatoric uncertainty, and this map gates a conditional diffusion process, anchoring certain regions to the prediction while sampling uncertain ones such as disocclusions. The authors report improved results over competitive baselines on robotic tabletop manipulation and camera motion action spaces.
In briefExtending causal discovery identifiability guarantees from complete data to missing data
SummaryThe paper notes that causal discovery methods are mature for complete data but remain limited for handling missing values. It establishes general sufficient conditions under which causal graphs stay identifiable from the data distribution under various missingness types, including Missing Not At Random, and provides a framework to extend correctness guarantees to missing data settings. For graphs without unobserved confounders, identifiability is preserved when missingness mechanisms do not causally interact, no variable directly causes its own missingness, and either faithfulness holds for missingness mechanisms or they follow generalized linear models. Based on these results, the authors introduce two asymptotically correct algorithms for Additive Noise Models.
WhyExtends causal identifiability from complete to missing data under broad missingness mechanisms, with sufficient conditions and two asymptotically correct algorithms.
Matteo Ceriscioli, Shohei Shimizu, Mallika Gupta et al.
In briefTitan-Llama: Adding External Neural Memory to Pre-trained LLMs
SummaryThis paper presents Titan-Llama, a post-training method that adds a hierarchical memory structure to attention-based language models without training a new model from scratch. Recent tokens keep full-fidelity attention, while historical context is compressed into fixed-size recurrent neural memory modules that update at test time. On Llama-3.2-1B-Instruct, accuracy on QuALITY and LongHealth improved over segmented attention windows, and the authors report faster prefill and decode and lower decode memory in the evaluated settings.
In briefCertifying how much coverage budgeted document selection loses, using finite-sample bounds under adaptive audits
SummaryThe paper studies how much expected query coverage is lost when a budget-limited set of document fragments, such as a retrieval-augmented generation knowledge base, is chosen without knowing the query distribution. The authors derive finite-sample, distribution-free upper confidence bounds on this coverage regret that remain valid after data-dependent selection. The main result is a simultaneous bound over all retained coverage contrasts, made observable through empirical-Bernstein variance calibration or a one-stage construction. A simulation study confirms validity and quantifies gains from variance calibration and certified redundancy.
In briefDualView: Using Broad and Output-Specific Evidence for LLM-Based Multi-Output Text Regression
SummaryThis paper presents DualView, a parameter-efficient method that uses large language models for multi-output text regression, predicting a full vector of scores. It extracts broad and output-specific representations from a frozen LLM and feeds them into a dual-branch architecture. Across five regression tasks and eight LLMs (up to 122B), it outperforms seven supervised baselines while requiring 82 to 1842 times fewer fitted parameters than the three strongest ones.
In briefTraining-free sampling method for estimating rare failures in language models
SummaryThe paper introduces Weighted Ensemble (WE), a training-free framework for estimating the probability of rare but severe failures in large language models. WE samples only from the original model and periodically reallocates the sampling budget toward prefixes more likely to reach the target rare event, guided by internal LLM representations rather than a separately learned proposal distribution. The authors provide a theoretical analysis of the algorithm. Experiments on safety-related tasks with small and large LLMs show improved efficiency and reliability of rare-event probability estimation.
In briefLidarVerse: A Real-Time Unified LiDAR World Model for Multiple Tasks and Sensors
SummaryThe paper presents LidarVerse, a real-time unified generative model for LiDAR sensor simulation. It combines reconstruction-style geometric grounding with generative learning, conditioning on LiDAR prefixes, HD maps, and spatial caches to create new scenes, forecast point clouds, or resimulate observed scenes without per-scene optimization. The model runs at 11 FPS on a single H100 GPU and supports 32-, 64-, and 128-beam LiDAR, including an unseen scanning pattern. The authors report state-of-the-art results on most metrics across three evaluation protocols.
In briefHow embedding geometry affects LLM routing, and a hidden-state fix for generalization
SummaryThis paper studies routing, which selects which of several LLMs should answer a given prompt. The authors find that routers using semantic embeddings plus a lightweight classifier perform well in-distribution but generalize poorly to unseen tasks, because the neighbors of unseen prompts are an unreliable signal for model selection. They instead use the hidden state of the last token at a deep layer of a partial forward pass through Qwen2.5-7B-Instruct as the representation. Simple classifiers such as kNN then improve on average across cost penalties, outperform all evaluated semantic-embedding baselines, and beat a recent hidden-state router on RouterBench.
Leonardo Torti, Emanuele Rimoldi, Tomer Galanti et al.
In briefFireGraphNet: Predicting Wildfire Spread Within the Day Using Directed Graph Networks
SummaryThe paper proposes FireGraphNet, which models wildfire growth over the next 12 hours as node classification on directed H3 graphs, with wind, terrain, fuel and barrier information on edges. On satellite fire perimeters in California, it reaches roughly 1.7 to 1.9 times the average precision of the strongest distance baseline. The authors also extend the model to analyze how edits to fuel and barriers change predicted spread.
Aiyshwariya Paulvannan Kanmani, Miguel Esparza, Ali Mostafavi
In briefNEPSEE: A Double-Annotated Nepali Sentiment Benchmark
SummaryThe paper builds a Nepali sentiment benchmark from tweets by Nepali public figures, labelled independently by two annotators, and keeps 3,189 sentences where both agree on polarity. On three-class classification, NepBERTa and IndicBERTv2 reach macro-F1 scores of 0.877 and 0.867, while three commercial large language models score 0.882 to 0.896 zero-shot. The authors also report that default tokenizers of three Nepali BERT models delete non-spacing Devanagari vowel signs and the virama, and restoring these marks raises NepBERTa's macro-F1 by 0.084.
In briefAdaptive computation with conformal certification: cutting compute on scientific images with guaranteed risk
SummaryThe paper addresses scientific imaging data, such as medical and astronomical images, which are much larger than natural images and contain sparsely distributed objects of interest. It proposes Conformal Adaptation, which predicts at low resolution, certifies the signal with conformal risk control, and refines only the certified predictions. The method is evaluated on 3D medical image segmentation, 2D galaxy flux prediction, and 1D calcium spike rate prediction, reporting wall-clock speedups of 2.5 to 72 times, including on commodity CPUs.
In briefWhy AI agents cooperate: reverse-engineering multiagent cooperation
SummaryThis paper proposes a framework for reverse-engineering why a set of AI agents engage in particular cooperation-related behaviors, distinguishing altruistic preferences from instrumental self-interested cooperation. It introduces experimental designs and conditions under which competing explanations for multiagent behavior can be disentangled. The framework is applied to interactions among frontier models in a simulated environment, showing that different models' behaviors arise from substantively different mechanisms.
In briefBayesian seismic velocity imaging with a learned diffusion prior for structure-aware inversion
SummaryFull waveform inversion reconstructs high-resolution subsurface velocity models, but it is nonlinear and ill-posed. This paper adds a learned diffusion model as an implicit regularizer to an ensemble-based Bayesian method (ASVGD-FWI), guiding ensemble members toward plausible velocity models without explicitly evaluating the prior density. Numerical experiments report more geologically consistent mean models, fewer nonphysical artifacts, and uncertainty concentrated in structurally complex, poorly constrained regions.
In briefHow to Evaluate When You Don't Know How to Evaluate? Choosing Reliable Setups for Base Model Benchmarking
SummaryThis paper studies how design choices in evaluating large language models, such as task formulation and the number of in-context examples, affect measured performance. The authors compare eight setups using monotonicity, non-randomness, and ranking consistency across checkpoints. They test 26 monolingual 2B-parameter models and three larger models on 80 tasks in four languages. Multiple prompts are consistently useful, while task formulation's effect depends on model scale.
Shenbin Qian, David Samuel, Jind\v{r}ich Helcl et al.
In briefHERO: Using Execution Evidence to Refine Prompts and Code for Program Generation
SummaryThis paper introduces HERO, a method for improving large language models on program generation tasks such as text-to-SQL and text-to-Python. It first learns reusable prompt updates from execution-verified successes and failures, validating revisions before adopting them. At test time, it refines the generated code using execution feedback. Across four SQL, Cypher, and Python benchmarks, HERO achieves the best offline performance, improving pass@1 by 7.6 points over the initial prompt, and test-time refinement adds another 2.0 points while improving execution efficiency.
Mingyu Lu, Timothy Ossowski, Tristan Naumann et al.
In briefLearning in Strategic Queuing with Geometric Recharge: How Learning Algorithms Route Packets Through Small-Buffer Servers
SummaryThe paper studies packet routing where queues act as selfish, myopic agents that use no-regret learning to choose among servers with unknown service rates, each having a single-packet buffer. For a single queue, any no-regret learner achieves long-run throughput at least the Hamacher sum of server rates, a bound shown to be tight, and is stable exactly when its arrival rate is below that sum. For multiple queues, the smallest overprovisioning factor guaranteeing stability under high-probability no-regret learning is 1/(1-λmax).
WhyModels queue routing as recharging bandits, proving tight Hamacher-sum throughput, a tight 0.77 competitive ratio, and an improved stability bound.
In briefLocating Harmful Covariate Shifts Behind Low-Confidence Patient Profiles
SummaryThe paper proposes a diagnostic pipeline that combines two views of model reliability. MED3pa first finds interpretable low-confidence patient profiles, then Detectron tests within each profile whether performance degradation is linked to a detectable harmful covariate shift. Experiments on three clinical prediction tasks with injected subgroup shifts in MIMIC-IV show that an age shift, undetectable at the population level, was localized to specific degraded profiles.
Olivier Lefebvre (1), Hakima Laribi (1), Dan Poenaru (2 et al.
In briefMulti-system cognitive steering for generative communication: using context and AI control to recover users' intended meaning from abbreviated input.
SummaryThe paper addresses the tradeoff between fast typing of abbreviated text and the ambiguity that causes AI to misread it. The authors propose Cognitive Steering, which grounds large language models in multimodal context and applies AI control methods. On simulated abbreviation benchmarks it reaches about 88% top-5 expansion accuracy, roughly double the no-context baseline. The study also finds that physical scene referents work best with explicit text prompts, while abstract factors such as emotion are better handled by altering internal model activations.
In briefLearning from Self-Generated Responses: Reliability-Uncertainty Self-Distillation for Reasoning
SummaryThis paper proposes Reliability-Uncertainty Self-Distillation (RUSD), a method that lets a language model learn from its own solution attempts without reference answers or question-specific rubrics. A teacher sees the question and a completed attempt, while a student sees only the question and matches the teacher's token-level predictions. Each attempt's loss is weighted by combining a self-judge's score and confidence with the student's predictive uncertainty. The authors report higher average accuracy than five baselines across eleven reasoning benchmarks, reaching 70.47% on Qwen3-4B and 76.69% on Qwen3-8B.
WhyCombines self-judged reliability with student uncertainty to weight self-distillation losses; outperforms five baselines across eleven reasoning benchmarks.
In briefVerification turns small-model gaps into routing decisions for clinical notes
SummaryThis paper studies generating clinical notes from doctor-patient dialogue, where fabricated lab values pose patient-safety risks. With a four-evaluator verification stack and refinement policy held fixed, the authors swap only the generator, comparing a 4B-parameter QLoRA-adapted small model against a frontier baseline. The small model reaches 0.984 claim-level factual entailment (versus 0.992) and trails by 3.3 points on completeness while generating fully on-device. Its shortfall lies mainly in inter-section structural fidelity.
Nina Fatehi, Meysam Ghaffari, Animesh Agarwal et al.
In briefMeta-evaluating metrics for multilingual style-personalized text generation
SummaryThe paper argues that metric-human correlation is an unreliable way to judge evaluation metrics and proposes a regression-based meta-evaluation grounded in a constant-mean baseline and human agreement. The authors build a multilingual style-personalized text generation benchmark spanning 10 languages and four evaluation dimensions. They meta-evaluate 32 metrics and their ensembles across 40 evaluation settings. They report that simple string- and embedding-based metrics, once calibrated to human ratings, can match strong LLM-judges.
In briefListQA: A benchmark testing LLMs on retrieving multiple facts as structured lists
SummaryThis paper introduces ListQA, a benchmark of 9,045 human-curated, cross-validated questions requiring models to recall several facts and organize them into structured lists of 3 to 10 elements. The authors evaluate 34 models using optimal bipartite matching combined with LLM-based semantic grading. The best standard-inference model scores below 37% LLM-Judge-F1, and providing the source Wikipedia page raises scores substantially. The dataset and evaluation code are released.
WhyProposes a new list-formatted factual recall benchmark with bipartite matching evaluation, benchmarking 34 models with thorough comparisons and ablations.
In briefFrom Qualitative Text to Integrated Evidence: A Four-Step Method for Combining Two Computer Analyses of Interviews
SummaryThis methodological study proposes the Parallel-Then-Integrated (PTI) procedure, a four-step method for combining two computational representations of the same qualitative text. It applies PTI to 1,203 interview paragraphs from 29 mothers in Nepal, comparing ontology-guided Patient Voice coding (PVminer) with data-driven themes from non-negative matrix factorization (NMF). A crosswalk table records where the two representations converge and diverge.
In briefInto the Unknown: A Benchmark Testing AI Agents on Open Research Problems
SummaryThis paper introduces Into the Unknown, a benchmark that gives AI agents a science paper and a research direction, then asks them to write a follow-up paper. Scientists at US Department of Energy national laboratories supplied unpublished directions and submission conditions, which serve as the evaluation rubric. Of 135 papers written by nine agents, none met all conditions and 114 failed an integrity check, mostly for claiming unperformed work or reporting numbers that contradicted the evidence.
In briefDivAlgo: A Systematic Evaluation of Algorithmic Diversity in Code Generation Models
SummaryThis paper introduces DivAlgo, a large-scale benchmark of algorithmically diverse solutions to programming problems, along with a lightweight similarity metric for large-scale evaluation. The authors tested 24 modern code generation models and found that uncontrolled model-generated code is much less diverse than human solutions, capturing about 31% of human diversity. Even the best diverse generation strategy reached only about 70%, and steering models toward human strategies did not reproduce human-level diversity. The authors suggest that overexposure to model-generated code may narrow programmers' algorithmic vocabulary.
In briefField Pea Breeding in Iowa: Agronomic Traits and UAV Time-Series Phenotyping
SummaryThis study evaluated 17 pea accessions and 7 Canadian varieties in Iowa for agronomic traits and seed yield. Accessions showed genetic diversity in phenology and canopy traits, while Canadian varieties showed high yield potential. Among UAV-derived features, time-series canopy area correlated most strongly with yield (r = 0.63) and showed higher selection reliability, supporting its use in pea breeding selection.
In briefCompiling Expert Procedures into Verified Programs: Small Models Can Suffice
SummaryThe paper proposes PLVR, which builds an explicit program from a teacher model's (Kimi-K3) reasoning traces and runs it with unchanged small models, including an 8B model, without changing weights. On tau2-Bench Retail and Telecom and on LiveMCPBench, the program raises Granite-4.2-8B from 43.0% to 91.2% on Telecom and narrows the gap between base models. Results vary by domain, with tool search still benefiting more from larger students.
In brief"Demystifying On-Policy Distillation": Analyzing where a teacher model's corrections improve a student's accuracy
SummaryThe paper studies on-policy distillation (OPD), a knowledge distillation method in which a teacher supervises trajectories generated by the student. The authors flip tokens at teacher-student disagreements to measure the accuracy gained from teacher corrections. Repeated flips raise accuracy from 28.4% to 74.5%, close to the teacher's 76.9%. Mass- and KL-based selection outperforms random flips by 15 to 18 percentage points at similar counts. The authors advocate training schedules that learn more valuable teacher corrections while keeping distributions stable.
In briefRunning graph neural network training and inference directly inside graph databases
SummaryGraph neural networks on large graphs usually depend on in-memory graph storage and ad hoc processing pipelines. This paper shows that graph databases can perform GNN training and inference directly, including sampling, feature retrieval, and the full forward pass, using Cypher queries. The authors report that their optimisations speed up training by about 2.5x with unchanged accuracy. They also state that fully in-database inference can be over 10x faster than cold-loading the graph into RAM on real-world graphs, with no extra memory overhead.
Victor Pekkari, Harry Wik, Alfred Clemedtson et al.
In briefOnline tokenization with optimal delay in linear time, a method for streaming text splitting in language models
SummaryThis paper formalizes online tokenization, the streaming version of the BPE pipeline used in large language models, and introduces a notion of delay measuring how far output lags behind the optimum. It presents a near-linear time algorithm achieving optimal delay up to constant factors, along with a conditional lower bound showing no polynomial-time algorithm can beat it by more than a constant unless P = NP. A practical implementation for tiktoken and HuggingFace tokenizers reports lower time-to-first-token latency and higher throughput, with roughly constant memory use.
WhyFormalizes online tokenization, gives a near-linear optimal-delay algorithm with a conditional lower bound, and shows large latency and throughput gains over tiktoken.
In briefMA-Bench: Measuring Language Models' Operational Planning Skills
SummaryMA-Bench is a benchmark based on military mission analysis, the first step of operational planning, testing whether language models can turn large, incomplete, and partly contradictory document sets into a sound brief. It contains 10 expert-built scenarios, each with 10 to 12 documents and 1,257 expert-written pass/fail requirements. The best of twenty evaluated models met 62% of the rubric items, performing better on facts stated in the documents (58%) than on judgments that must be inferred (33%).
Rodney Lafuente-Mercado, Robert Sincero, Phil Sage et al.
In briefGraph-Based Orchestration of Mixed AI Models: Combining Techniques Through Control Nodes
SummaryThis paper presents a graph-based framework for composite AI systems that combine several AI techniques. Specialized control nodes coordinate models such as planning, machine learning, constraint programming, and large language models. The system uses directed acyclic graphs and behavior trees to manage model relationships and execution order, supporting hierarchical, sequential, and concurrent composition. Case studies illustrate how the architecture lets multiple techniques work together within one operational structure, addressing limits of single-model approaches.
G. Michael Youngblood, Filip Dvo\v{r}\'ak, Slavo \v{S}vanc\'ar et al.
In briefStudent Copula Markov Networks: graphical models for non-Gaussian data
SummaryThis paper extends probabilistic graphical models to non-Gaussian data. It characterizes the pairwise Markov property in undirected graphical models using copula densities, and derives a necessary condition for this property in Student copula Markov networks. An Expectation-Conditional Maximization (ECM) algorithm is provided for penalized maximum likelihood estimation and graph selection. Simulation studies show the algorithm effectively estimates model parameters and recovers the underlying graph structure.
In briefCAFE: Restricting feature distillation to language-routed adapters for compact multilingual speech recognition
SummaryThis paper studies feature distillation for compact multilingual automatic speech recognition. The authors find that two common feature losses do not add to logit distillation, and explain this through their gradients, which conflict with English performance. They propose CAFE, which confines the CKA gradient to adapters that English bypasses. On the Open ASR Leaderboard, average word error rate drops from 8.55% to 8.02%.
Tianchi Liu, Nicholas Moratelli, Khurram Azeem Hashmi et al.
In briefTRACER: Trace-Regularized Training Against Indirect Prompt Injection
SummaryThis paper shows that defenses against indirect prompt injection depend on the channel where malicious content appears, with the same attack succeeding 18% as a follow-up message but 53% in retrieved context. The authors propose TRACER, a channel-aware adversarial training method using an answer-anchored, evidence-gated reward together with ATRC. On GSM8K with Qwen2.5-3B-Instruct, RAG-framed attack success falls from 65.5% to 6.1% while clean accuracy stays at 78-82%.
In briefPlant Memory: Turning SAP Maintenance Work Orders into Verified Failure Knowledge for AI Agents
SummaryThis paper presents Plant Memory, a system that applies technical language processing to SAP Plant Maintenance work orders and builds an ISO 14224-aligned maintenance ontology. Each extracted fact carries a confidence score, its source evidence, and a verification state, and technicians review the facts through a queue. On a synthetic set of 400 work orders, rule-based extraction reached 96.2% component accuracy and 88.9% failure-mode accuracy. On 28 diagnostic questions, the top-ranked root cause was correct in 82.1% of cases, versus 53.6% for the raw SAP root-cause column.
In briefImagining goals with geometric horizon models to improve policy generalization
SummaryThis paper addresses zero-shot goal generalization, where offline-trained policies must combine existing knowledge to solve unseen tasks. The authors propose TD-Aug, which samples from a geometric horizon model to imagine novel outcomes achievable by composing known skills. Training goal-conditioned behavior cloning policies on these imagined future outcomes is reported to significantly improve generalization on stitching-based OGBench tasks.
Daniel Lawson, Dhruv Sreenivas, Adriana Hugessen et al.
In briefCLASHarness: A Verifiable Benchmark for Multi-Agent Conflicts in Computer Use
SummaryThis paper introduces CLASHarness, a controllable framework that generates verifiable multi-agent conflict scenarios for computer-use tasks, along with a benchmark of 600 multi-round tasks. It combines hierarchical scenario planning with an eight-step verification pipeline, fixing decisive visual facts before rendering and building graders from executed reference trajectories. Evaluating 15 frontier models, the authors find that weaker models blindly trust misleading sub-agent outputs, that resolving conflicts is much harder than detecting them, and that coordination, such as releasing locks and handing off tasks, remains a common failure point.
In briefPairIT: Transformer models that improve molecules using few training data
SummaryThe paper introduces PairIT, a training framework for autoregressive transformers that work on SMILES molecular strings. By learning from pairs of structurally similar molecules, the model makes local edits to a seed molecule to improve a target property, with prompt tokens specifying property, direction, and magnitude. Reinforcement learning post-training can further strengthen optimization. On MoleculeNet and Therapeutics Data Commons datasets, the authors report that a 68M-parameter standard transformer achieves robust seeded optimization in low-data settings.
Hannah Lawrence, Bodhi P. Vani, Ji Won Park et al.
In briefOnboard learning in orbit: training and closed-loop attitude control with a thermal segmentation model on a CubeSat
SummaryThis paper reports a flight deployment on a 6U CubeSat that trains and runs a thermal image segmentation model onboard an NVIDIA Jetson TX2 NX GPU in space. The model's output drives closed-loop attitude control, using a real-time thermal image feed to track the Moon. The system also uses ROS 2 Foxy for data transport and an open-source file system with data integrity and self-healing features. Over about one year, the team ran 12 experiments, showing these learning-based systems can operate reliably under Low Earth Orbit constraints.
Elena Sorina Lupu, Hannah Grauer, Phaedrus Leeds et al.
In briefTraining-free pruning that keeps reasoning models accurate and efficient
SummaryThe paper proposes LayerTrim, a training-free structured pruning method that finds redundant decoder layers and selectively trims their MLP features while keeping attention and decoder depth. Across three reasoning language models and five benchmarks, it outperforms existing structured pruning methods in most settings, improving average accuracy by up to 21.4 points over RESP. On Qwen3-14B, it achieves 1.24× faster prefill and 19.7% lower peak GPU memory than the dense model.
Woosang Lim, Haneul Yoo, Vincent-Daniel Yun et al.
In briefDoReMe: Jointly optimizing reconstruction and motion estimation for self-navigated free-breathing abdominal MRI using Gaussian Splatting
SummaryThis paper proposes an algorithm that couples image reconstruction with motion estimation for self-navigated, free-breathing abdominal MRI. Respiratory motion is modelled with compact Gaussian primitives and refined jointly with the image during iterative reconstruction, without needing pre-estimated motion fields or external respiratory signals. In four healthy volunteers using Cartesian gradient-echo acquisitions, the method improved visual image quality over regularized compressed sensing, particularly in coronal liver views, and preserved anatomy better in retrospectively shortened scans. The authors present these results as a proof of concept.
In briefDecisionRobust: A Benchmark for Robustness of Text and Vision Decision Models
SummaryThis paper introduces DecisionRobust, a benchmark for testing how reliably JEV-style decision models handle text and image inputs under adversarial manipulation. It covers meaning-preserving text edits, image perturbations, and prompt injection in both modalities, across 30 text and 17 vision model configurations, including held-out models. The authors find that attacks expose weaknesses beyond the models used to build the benchmark, that resistance to input perturbations does not guarantee resistance to prompt injection, and that label-consistency checks substantially lower measured attack success.
In briefGenesis Steering: Improving Activation Steering with Training Data
SummaryThe paper introduces genesis steering, which derives a steering vector from contrastive examples in a model's own fine-tuning data to control large language models at inference time. When such data are unavailable, it uses membership inference to identify likely training examples and build a proxy dataset. On Tülu-3-8B, safety data from its training mixture yielded stronger steering than external datasets.
In briefFrom Copying to Composition: How Language Models Learn In-Context Learning
SummaryThis paper studies how language models learn to adapt their behavior from instructions and examples in a prompt. A controlled diagnostic separates copying a nearby demonstration from integrating information across the whole prompt. Across 17 models, most prefer the rule-consistent answer, and matched training experiments show that different training methods lead to different developmental trajectories.
In briefA data-centric survey of how multimodal AI models learn spatial reasoning
SummaryThis survey reviews methods for fine-tuning pretrained multimodal large language models on spatial reasoning tasks, from supervised to reinforcement learning, with a focus on datasets. It groups research into dataset design, construction, and evaluation. The authors identify four main ways of building spatial annotations: rules and templates, LLM prompting, procedural generation, and human-in-the-loop. They also note wide variation in evaluation methods and suggest future research directions.
Mahdieh Mallahnezhad, Masoud Jafaripour, Mohamadali Tofigh et al.
In briefNeural clustering method for learning personalized treatment rules across many treatments
SummaryThis paper addresses learning individualized treatment rules (ITRs) when there are many treatments and few observations per treatment. The authors propose nonlinear supervised clustering fusion (NSCF), which uses a shared nonlinear covariate representation and low-dimensional treatment embeddings to group treatments with similar responses. In nonlinear simulations NSCF lowers regret relative to a strong neural reference, and in a melanoma xenograft application it improves average policy value over that reference.
SummaryThis paper addresses how LLM agents learn from sparse outcome rewards in long-horizon tasks. It proposes Contextual Credit Policy Optimization (CCPO), which weights matched observations from other trajectories by contextual similarity and combines historical and future credit to reward each step for its progress. The method needs no auxiliary value critic. On Qwen2.5-Instruct 1.5B and 7B, CCPO reaches up to 97.4% success on ALFWorld and 79.4% on WebShop, outperforming GRPO.
In briefThe Gap Is in the Geometry: Meta-Analysis of Aerial and Remote Sensing Video Understanding
SummaryThe paper argues that evaluating aerial and remote sensing video with multiple-choice questions for multimodal LLMs and reporting accuracy measures the wrong aspect. A meta-analysis of published results finds that scaling parameters yields small accuracy gains compared with the human gap, while acquisition conditions such as altitude and view strongly reduce accuracy. The authors derive five falsifiable principles and propose the AVU-Card, a reporting card for aerial video benchmarks.
Georgios Angelidis, Aggelos Psiris, Maria Vakalopoulou et al.
In briefGroundGUI: Better Screen Element Targeting via New Data and Training Method
SummaryThis paper presents ScreenRef, a GUI grounding dataset of about 586K screenshots and over 633K referring expressions across desktop, web, and mobile. It also proposes Distillation-to-Reinforcement Learning (D2RL) to address limits of supervised fine-tuning and sparse rewards. The resulting 8B model, GroundGUI-8B, reports 95.2% on ScreenSpot-V2 and 63.9% on ScreenSpot-Pro.
WhyIntroduces a large GUI grounding dataset and a distillation-to-RL method, with strong results across several benchmarks, a substantive contribution to the subfield.
In briefSimilar Models, Broken Odds: Where Luce's Axiom Fails in Chatbot Arena
SummaryThe paper examines how Chatbot Arena ranks language models with the Bradley-Terry model, which assumes Luce's choice axiom and gives each model a single score. Analyzing 1.8 million votes, the authors find that comparisons between near-copies, such as versions of one model, are more decisive than predicted, by about 1.37 times within a model and 1.12 times between models from one lab. A correction with four to six parameters predicts held-out votes better, mainly by modelling ties, while the leaderboard order barely changes.
WhyApplies choice theory to 1.8M Arena votes, showing near-copy models are separated more decisively; a correction model beats Bradley-Terry on held-out, pre-registered tests.
Soovadeep Bakshi, Apurba Bose, Aishwarya Bose et al.
In briefDexMimic: a general-purpose dexterous hand-object tracking policy using value-guided reset allocation
SummaryDexMimic is a single policy for tracking dexterous hand-object motions, trained jointly across many human hand-object interaction trajectories instead of one policy per trajectory. Its Value-Guided Reset Allocation method adjusts starting states online to focus on harder phases and states with higher expected tracking returns. The paper also introduces a benchmark covering training, fine-tuning, and generalization, reporting better tracking success and training efficiency, and validates the approach on a real-world dexterous manipulation platform.
In briefA global-in-time spectral method for post-exposure bake simulation and physics-informed neural operators for EUV lithography
SummaryThe paper proposes a global-in-time spectral Duhamel-Picard method that represents full solution trajectories with polynomial expansions, solved by Anderson-accelerated Picard iteration so that time does not need to be marched step by step. The authors report simulating a 1200×1200×100 voxel 3-D post-exposure bake in about one minute. The representation is also used in a physics-informed neural operator trained without precomputed solution trajectories, which can predict concentration fields at arbitrary times within the bake interval.
