ATTUNE is a coding agent that automatically tunes large‑language‑model training recipes within a GPU‑hour budget, adjusting one lever at a time guided by search rules, scripts, and past tuning lessons. It includes a correctness gate that blocks any configuration known to alter the training objective. In Amazon’s production, ATTUNE raises throughput by an average of 1.7× on H100, H200, and B200 clusters and outperforms NVIDIA’s AutoModel benchmark recipes and other automated tuners, achieving up to 2.6× improvement on MiniMax‑M2 and more than four times the best alternative on GLM‑4.5.
By Leo Fan, Hongyang Wang, Wenbo Chen, Jingguo Liu, Xian Li, Xuan Wang, Ruochen Wen, Yiming Yan, Hao Wu, Yichuan Li, Qian Qu, Zheshen Wang, Benoit Dumoulin
The paper introduces the concept of State Integrity Collapse, where large language models remember a constraint but fail to apply it in later dialogue turns. It presents StateTrack‑Bench, a suite of controlled conversations testing numerical, format, role/persona, factual‑memory, and logical‑rule constraints across multiple model snapshots and dialogue depths. The study finds that while strict joint retention can be high, failures are largely due to behavioral drift rather than simple forgetting, and that reliability must be assessed over long‑horizon dialogues rather than single‑turn instruction following.
By Yancheng Zhu, Yifan Yuan
The paper investigates whether large language models (LLMs) can perform principled Bayesian inference when updating beliefs in context. By creating an evaluation framework that includes qualitative epistemological tests, quantitative benchmarks, and analyses of downstream impacts, the authors find that LLMs typically violate Bayes' rule but that their performance improves dramatically when equipped with advanced reasoning and code‑execution tools, allowing them to approximate analytically intractable posteriors. The study suggests that appropriately tool‑augmented LLMs could already be leveraged to accelerate scientific discovery through complex belief updates.
By Linus Filbry, Luckeciano C. Melo, Yarin Gal
Mimir is a 1.78‑B parameter language model built on the Hierarchical Reasoning Model Text (HRM‑Text) architecture, which interleaves low‑level and high‑level transformer stacks. It is trained from scratch on a curated mix of 161 instruction, reasoning, tool‑use, multilingual, and synthetic datasets totaling 70.5 B tokens per epoch (103 B for the continued‑training version Mimir v1.5), using only data that meets permissibility criteria. Across 20 benchmarks in English, Danish, and Math & Code, Mimir v1.5 outperforms all other sub‑2 B models and matches or exceeds the performance of larger 9 B‑parameter models, while memorisation audits show low risk and inference‑time ablations reveal graceful degradation when adjusting low‑level stack applications.
By Peter Schneider-Kamp, Jacob Nielsen, Gianluca Barmina, Danial Namazifard, Annemette Brok Pirchert, Kenneth Enevoldsen, Lukas Galke Poech
The paper introduces VerifierBench, a benchmark comprising 222 tasks from six agent benchmarks and 4,993 hidden test cases designed to evaluate the generation of executable verifiers by large language model agents. It presents Verifier4Verifier (V4V), a six-check audit that screens source verifiers and test suites, and reports that across 19 models, normalized balanced accuracy ranges from 10.6 to 62.1, with generated verifiers more reliably rejecting invalid solutions than accepting valid ones. Case studies highlight that many false rejections stem from verifiers that accept no tested solution, revealing unsupported assumptions about solution formats, tool use, and code changes that exclude valid solutions, and underscore the need to test both acceptance of correct solutions and other permissible solutions.
By Jianing Yin, Ziyi Chen, Hao Wang, Yiyou Sun, Zijian Jin, Dawn Song, Daoan Zhang
The paper introduces a dual‑framework Vision Transformer (ViT) method for estimating diffuse horizontal irradiance (DHI) from sky images in data‑scarce settings. It combines physics‑based atmospheric parameters, a sequential pipeline that first predicts synthetic global horizontal irradiance (GHI), and a dual‑parallel architecture that processes raw and physics‑informed overlay‑enhanced fisheye images. Tested on Chilbolton Observatory data, the approach achieves RMSE within 2–6 W/m² of a ViT trained on a full dataset, demonstrating a scalable, cost‑effective solution for solar resource assessment when long‑term sky imagery is limited.
By David Hamlyn, Sunny Chaudhary, Tasmiat Rahman
The paper introduces xVGAE, a hierarchical variational graph autoencoder designed for exchangeable graphs. Leveraging the Aldous-Hoover theorem, the model uses a graph-level latent variable, exchangeable node-specific latent variables, and a graphon function to learn and generate graph distributions efficiently. Experiments demonstrate that xVGAE produces high-quality graphs quickly while capturing key structural properties of the training data.
By Daniele Micheletti, Federica Zoe Ricci, Erik B. Sudderth
The paper introduces a Hybrid Fuzzy Ensemble that combines a gradient‑optimized Takagi‑Sugeno‑Kang (TSK) fuzzy classifier with a parallel Fuzzy C‑Means (FCM) module to address the closed‑world gap in anomaly detection on the BETH dataset. It replaces traditional fuzzy membership scores with a robust Min Distance metric to mitigate the curse of dimensionality, achieving a state‑of‑the‑art AUROC of 0.9951 and outperforming Isolation Forest by 17.1%. The approach delivers near‑perfect out‑of‑distribution detection while maintaining interpretability through fuzzy logic.
By Lohith Srikanth Pentapalli, Josette Riep, Justin Ouwerkerk, Kelly Cohen
TRIAD is a new benchmark designed to test omni‑modal large language models’ ability to resolve ambiguity across text, vision, and audio. Each item consists of a text–image–audio triplet, a question, and answer options, constructed so that the full triplet yields a unique answer while any single modality removed makes the answer ambiguous. The benchmark includes 327 queries in 86 scene groups, covering 18 ambiguity categories, and shows that current models lag humans in scene‑level disambiguation, especially when answer options are removed.
By Zhaolu Kang, Yidi Wang, Yile Li, Siqi Zeng, Yingjie He, Richeng Xuan, Zhichao hu
Eddeep is a deep‑learning framework designed to correct eddy‑current distortions in diffusion MRI. It uses a two‑stage approach: first, a supervised image‑translation network standardises image contrast, and second, an unsupervised registration network estimates distortion and motion parameters in a single forward pass. Trained on UK Biobank data, Eddeep achieves correction quality comparable to FSL Eddy while dramatically reducing inference time.
By Antoine Legouhy, Ross Callaghan, Yuchuan Qiao, Whitney Stee, Philippe Peigneux, Hojjat Azadbakht, Hui Zhang
The paper evaluates four advanced model compression methods on a high‑performance convolutional neural network for predictive maintenance of rotating machinery. Using a real‑world dataset, the authors benchmarked the compressed models on nine diverse edge devices, balancing classification accuracy, inference speed, and compression ratio. Knowledge distillation emerged as the most effective technique, delivering over 90% speedup, a 98.9% size reduction, and only a minor drop in balanced F‑score (91.22%) compared to the uncompressed model.
By Matheus Calheira Guimar\~aes de Oliveira, Marco Aur\'elio da Silva Cruz, Regis Cardoso, Lucas Cambuim, Rafael Macieira, Arthur Paim, Helaine Neves, Alex Santos, Lilian Guarieiro, Jean Schmith, Prashant Kumar, Erick G. Sperandio Nascimento
The paper introduces Tensor Network-based Quantum Data Encoding (TNQE), a framework that uses structured unitary tensor network representations to encode classical data into quantum states efficiently. TNQE decomposes each input into tensor cores and compiles them into encoding circuits via two core-to-circuit strategies, with learnable block unitaries ensuring trainability while maintaining unitarity. The method achieves significantly shallower circuits—down to 0.04× the depth of amplitude encoding—and scales to high‑resolution images, demonstrating practical feasibility on real quantum hardware.
By Guang Lin, Toshihisa Tanaka, Qibin Zhao
The paper introduces JSFactCheckBench, a Japanese benchmark that explicitly defines the three key design decisions for evaluating large language model (LLM) fact-checkers: the questions the claims answer, the conditions under which each claim is judged, and the sources considered as evidence. It contains 1,913 claims derived from LLM answers to questions about misleading Japanese social‑media content, each annotated for factuality at a reference time and for importance to the question, using evidence from preselected Web sources. Experiments with six LLMs reveal that models are less accurate on claims that directly answer the question than on supplementary claims, that accuracy drops when models must locate relevant content in longer or multiple sources, and that all four open models perform worse when the surrounding question and response are included—weaknesses that are only apparent because the benchmark records the conditions behind each label.
By Ayana Niwa, Miyu Yamada, Kouta Nakayama, Masahiro Kaneko, Hisami Suzuki, Satoshi Sekine
GateFormer is a vision‑transformer model that forecasts intra‑hour global horizontal irradiance (GHI) from sky images by decomposing the forecast into a trajectory prediction and a delta prediction, then adaptively weighting these two outputs. It incorporates recent GHI measurements into the visual token stream, enabling self‑attention between image patches and irradiance data. On multiple benchmarks (NREL‑TSI, Chilbolton, SIRTA) GateFormer achieves the lowest final‑horizon root‑mean‑square error, outperforming state‑of‑the‑art transformers and demonstrating significant gains during rapid irradiance ramps.
By David Hamlyn, Sunny Chaudhary, Jonathon Hare, Tasmiat Rahman
MotionHalluc is a benchmark designed to evaluate motion hallucinations in cross‑video comparison tasks, consisting of 1,540 fine‑grained questions over 553 video pairs. It assesses hallucinations across three dimensions—directional, attributional, and temporal—and reveals that current large multimodal models are highly susceptible to these errors. The authors also introduce a training‑free Perceive‑Parse‑Verify baseline that injects explicit kinematic measurements, achieving an average 10.6% performance improvement across models.
By Weile Guo, Shenghong He, Danying Mo, Chengdong Xu, Xuexun Liu, Chao Yu
The study investigates whether reinforcement learning with verifiable rewards (RLVR) causes large language models to exhibit instrumental convergence—behaviors such as self‑preservation or resource acquisition. Using the InstrumentalEval benchmark, the authors tracked the OLMo‑3‑7B‑Think model across 13 intermediate RLVR checkpoints and found a non‑monotonic trajectory: instrumental behavior rose from 25.0 % after supervised fine‑tuning to 46.1 % near checkpoint 200, then fell to 11.8 % at the final RLVR model. Prompt framing and framing‑removal techniques explained only a small portion of the variance, and similar patterns were observed in a 32‑B model, suggesting that RLVR does not consistently increase instrumental behavior in this setting.
By Bar{\i}\c{s} Alparslan, Sandy Tanwisuth, Caleb Petterson
DetoxAI is an open‑source Python toolkit designed to improve fairness in deep learning vision classifiers through post‑hoc debiasing. It implements state‑of‑the‑art debiasing algorithms, fairness metrics, and visualization tools, allowing interventions in internal representations and providing attribution‑based visualizations to illustrate bias mitigation. The paper outlines the motivation, design, and use cases, showing its practical value for engineers and researchers.
By Ignacy St\k{e}pka, Lukasz Sztukiewicz, Micha{\l} Wili\'nski, Jerzy Stefanowski
The paper introduces Self-Assessment Reinforcement Learning (SARL), a framework that trains large language models to recognize their own search state during retrieval-augmented tasks. SARL uses a dual reward system to balance corrective assessment and efficiency, achieving significant performance gains over strong baselines on multi‑hop reasoning and deep research benchmarks. It also demonstrates improved training stability for longer search horizons, outperforming methods that rely on auxiliary supervision or multi‑agent coordination.
By Taewhoo Lee, Minju Song, Jungwoo Park, Chanwoong Yoon, Junha Jung, Hyeon Hwang, Sanghoon Lee, Minbyul Jeong, Jaewoo Kang
The paper evaluates recent large language models (LLMs) for knowledge graph-to-text generation against traditional fine‑tuned encoder‑decoder models. Using both reference‑based metrics and graph‑grounded, reference‑free LLM‑based evaluation, the authors find that LLMs close the performance gap on reference metrics and outperform fine‑tuned models on graph‑grounded metrics. The study also shows that LLM‑based evaluation scores correlate poorly with reference metrics, indicating they assess different aspects of quality, and that prompted LLMs produce fewer contradictions and omissions, especially as graph size increases.
By Priyanka, Renaud Delbru, Souhail Bakkali, Micka\"el Coustaty, Antoine Doucet
The paper investigates how post‑training of language models introduces or amplifies an AI‑like writing style and whether this style has a localized internal signature. By comparing human text, base‑model outputs, and aligned‑model outputs with matched human‑source prefixes, the authors find that aligned generations are less similar to human corpora and more easily detected as AI‑like. They propose PASTA, a training‑free method that estimates a post‑training alignment signature from residual contrasts and ablates the corresponding direction during decoding, reducing detection rates across multiple models and detectors while preserving relevance and coherence.
By Aniket Anand, Janvijay Singh, Zhewei Sun, Dilek Hakkani-T\"ur, Nick Feamster