The paper introduces a synthetic benchmark for multimodal time‑series forecasting that evaluates how well text annotations contribute to predictions. By generating controlled signals with semantically correct, incorrect, and irrelevant annotations, the authors can precisely measure the true information content. Six mutual‑information estimators (KSG, MINE, InfoNCE, CCA, PID, and V‑information) are tested, all correctly ranking useful annotations and enabling annotation auditing without model training. The benchmark also highlights each estimator’s limitations and validates findings on seven real datasets, providing practical guidelines for metric implementation.
By Emma Andrews, Gianmarco Mengaldo
The paper introduces CoMA-DiT, a bidirectional cross‑modal Diffusion Transformer that uses paired modalities as mutual generative supervision for latent augmentation rather than just inputs for fusion. By conditioning velocity prediction on the paired modality through cross‑modal attention and injecting variation via a reliability‑gated residual mechanism, CoMA‑DiT improves multimodal brain state decoding. Experiments on auditory attention decoding and emotion recognition show consistent gains over 20 baselines, with absolute accuracy and macro‑F1 improvements of 4.28% and 6.70% respectively, and extensive analyses confirm its robustness and interpretability.
By Ziwei Wang, Xingyi He, Hongbin Wang, Tianwang Jia, Bohan Fang, Dongrui Wu
The paper "Prompt Revision as a Source of Cultural Bias in Text-to-Image Systems" investigates how commercial text‑to‑image models silently modify user prompts before generating images, a step that is often hidden from users. Using the multilingual benchmark WORLDVIEW, the authors audit the revision layer in DALL‑E‑3, Imagen‑4, and GPT‑Image‑1.5, finding that non‑Western and non‑Anglophone contexts are disproportionately marked, reduced to narrow vocabularies, and stereotyped. The study demonstrates that the revision layer itself is a previously undocumented causal source of cultural stereotyping, underscoring the need to audit deployed systems rather than just the underlying models.
By Aleksandra Urman, Elsa Lichtenegger, Salima Jaoua, Azza Bouleimen, Robin Forsberg, Corinna Hertweck, Stefania Ionescu, Nicol\`o Pagan, Ancsa Hannak, Joachim Baumann
MAPLE is a new agent that maintains and updates optimization problems through successive natural‑language requests, combining language‑based problem construction with mathematical programming and evolutionary search. It preserves the optimization program, accepted plans, earlier updates, and candidate solutions for future requests, enabling rapid adaptation to changing business constraints. In a benchmark of 15 trajectories and 180 updates across various operational domains, MAPLE completed all trajectories with high online scalar quality and Pareto hypervolume ratio, and maintained update validity and useful search information across substantial revisions.
By Kesheng Chen, Yamin Hu, Wenjian Luo
The paper introduces SOLID, a framework that enables operations research language models to self-improve without relying on verified answers or external evaluators. SOLID uses solver-generated artifacts from the model’s own rollouts to create pseudo-references, clustering objectives and applying group-relative advantages for dense self-supervision. Experiments on multiple OR benchmarks show that SOLID enhances solution accuracy for both general-purpose and OR-tuned models compared to outcome-only training.
By Rui Zhu, Minglong Cao, Chenyu Zhou, Jianghao Lin, Dongdong Ge
The study examines the composition of a random sample from the Model Context Protocol (MCP) registry, revealing that only 48.8% of the 400 sampled npm/stdio servers successfully complete an initialization handshake, compared to 66.7% for a hand‑curated frame. Among the servers that run, there are no fatal JSON Schema violations across 2,766 advertised tools, but optional safety annotations vary widely, with a 58.8% omission rate in the random draw versus 41.5% in the curated set. The authors also compare MCP tool descriptions to two benchmark corpora, finding minimal near‑duplication in real MCP tools (2.8%) and significant repetition in synthetic datasets (up to 85.6%).
By Haseeb Mohammed Afsar
BenchShield is a formal, model-backed instrumentation layer designed to protect reward integrity in large language model (LLM) agent benchmarks. It uses a finite lifecycle model of reward-relevant events to run a static, phase-aware taint analysis that flags potential reward-hacking paths before execution, and a runtime analysis that attributes concrete agent actions and provides evidence-backed claims. The system was evaluated on a corpus of 456 adjudicated trajectories from over 31,000 public agent runs across three benchmarks, showing significant improvements in recall, coverage, and cost efficiency compared to a baseline hackability scanner.
By Shenghan Zheng, Zonglin Di, Yimin Liu, Kyoung Whan Choe, Jiankai Sun, Heguang Lin, Penghao Jiang, Yifeng He, Xiao Cheng, Jicheng Wang, Wenbo Chen, Alex Yates, Yinzhe Zhao, Bingran You, Yuan Gao, Ayush Munot, Shubham Gaur, Zhe Ye, Hao Wang, Xiangyi Li, Dawn Song, Christophe Hauser
The study investigates which speech aspects best signal the end of a speaker’s turn in conversational AI. By systematically ablating acoustic, prosodic, and semantic cues in a lightweight trimodal classifier, the authors find that combining acoustic and prosodic features yields the highest accuracy and lowest latency, achieving an utterance F1 of 0.93 with 7.8% false alarms at 400 ms median latency. Adding textual information actually increases premature detections without improving performance, and feature analysis shows prosodic cues provide the strongest class separability while text representations overlap significantly.
By Rini Sharon, Manickavela A, Kadri Hacioglu, Andreas Stolcke
The paper presents PETSCAgent-Bench, a multidimensional benchmark and agent-based framework designed to evaluate AI-generated scientific code that uses the PETSc library. It combines deterministic checks with LLM-based assessments across five categories—correctness, performance, code quality, algorithmic appropriateness, and library-specific conventions—using a 14-evaluator pipeline. The framework demonstrates that while current large language models produce readable, well-structured code, they often fail on correctness and library conventions in realistic PETSc problems, revealing gaps that simple pass/fail tests miss.
By Hong Zhang, Barry Smith, Satish Balay, Le Chen, Murat Keceli, Lois Curfman McInnes, Junchao Zhang
The paper introduces the Spectral Masking and Interpolation Attack (SMIA), a black‑box adversarial technique that subtly alters inaudible frequency regions of AI‑generated audio to fool voice authentication systems and their countermeasures. Experiments show SMIA achieves at least 82% success against combined verification and countermeasure systems, 97.5% against standalone speaker verification, and 100% against countermeasures, revealing a critical security gap. The authors argue that current static defenses are inadequate and call for dynamic, context‑aware defenses that can adapt to evolving threats.
By Kamel Kamel, Hridoy Sankar Dutta, Keshav Sood, Sunil Aryal
Sci‑MMR is a new benchmark for multi‑step evidence‑grounded scientific reasoning in multimodal agents, featuring 235 multi‑hop tasks across four disciplines and an average of nine figure panels per task. It evaluates not just final answer accuracy but also the recovery of structured evidence from scientific claims, citations, visual data, and supporting regions. Experiments on eight state‑of‑the‑art models show a gap of over 20 points between answer accuracy and complete evidence recovery, highlighting significant challenges in evidence acquisition and integration.
By Jiaqiang Li, Yajie Yang, Zhiheng Xi, Jiadong Chen, Enyu Zhou, Senjie Jin, Yang Nan, Jiazheng Zhang, Han Wang, Yanxin Li, Dingwei Zhu, Bicheng Deng, Yuhui Wang, Xiang Zheng, Qi Zhang, Lei Bai, Xingjun Ma, Tao Gui
The paper explores a new approach to improve arithmetic reliability in clinical language models by having the model generate case‑specific Python code that a local executor runs deterministically, rather than performing arithmetic directly. Experiments on the MedCalc‑Bench Verified dataset show that this Program‑Solve interface yields modest gains for larger models (Qwen2.5‑32B) but not for smaller ones (Qwen2.5‑7B), and it does not replace the need for verified formulas or accurate variable extraction. The study highlights that adding an executor can help some open‑weight models but is not a universal solution.
By Felipe Ocampo Osorio, Sebasti\'an Andr\'es Cajas Ordo\~nez, Maximin Lange, Rafi Al Attrach, Sahil Kapadia, Zakaria Laouabdia Sellami, Angelo Antonio Talio, Leo Anthony Celi
NovGauge is a new benchmark designed to diagnose large language models’ ability to assess scientific paper novelty. It contains 619 paper pairs and 50 multi-paper sets, each labeled along three dimensions—task, problem, and method—by experts from ICLR reviewer overlap claims and survey co-citations. The study evaluates 18 LLMs, revealing high hallucination rates and weak evidence grounding, with the best model achieving only 43‑72% verified F1 across dimensions.
By Guoqiang Zhang, Kexin Tan, Ming Zhang, Li Ju, Wenqing Jing, Zhonghan Yue, Jiayi Chen, Shiqiang Wu, Shaofan Liu, Yue Zhang, Yuankai Ying, Yang Shi, Tao Gui, Qi Zhang, Xuanjing Huang
The paper introduces TASCO, a test‑time adaptation framework that enhances Large Language Model reasoning by incorporating local stability into confidence‑based adaptation while keeping the model frozen. TASCO optimizes a lightweight task‑level prefix using two perturbation strategies—Random Perturbation for distributional stability and Sharpness‑Aware Perturbation for worst‑case sensitivity—to ensure that high confidence aligns with correctness. Experiments show that TASCO improves reasoning accuracy and token efficiency across various LLMs and benchmarks, and behavioral analyses confirm that it maintains stable confidence without over‑concentrating the predictive distribution.
By Bincheng Gu, Min Gao, Zongwei Wang, Yibing Bai, Yulan He, Junliang Yu
COBRA‑Skills is a new framework that treats skill optimization for large language model agents as a budgeted sequential problem over a dynamically evolving candidate set. It uses contextual‑bandit prioritization to focus evaluations on promising or informative candidates and refines the skill population based on execution feedback. In experiments across six agent benchmarks and three target models, COBRA‑Skills outperforms existing methods, cuts optimization cost by 55–58 % compared to SkillOpt, and requires only 50 unique optimization examples per benchmark.
By Pingchen Lu, Xiangyi Wang, Xiang Li, Jie Mao, Zikun Qu, Junfeng Luo, Yao Shu, Bryan Kian Hsiang Low, Zhongxiang Dai
DeFiFusion is a dual‑modal framework that detects price manipulation attacks in decentralized finance by jointly analyzing transaction events and smart contract semantics. It encodes fine‑grained temporal and economic features from transactions and extracts contract logic using large language models, then fuses these signals with a Dual‑Modal Projection‑Fusion Transformer. The approach achieves state‑of‑the‑art performance, recalling 222 of 225 known attacks with 96.10% precision.
By Rui Cao, Shaojing Fan, Liming Fang, Yuchan Liu, Yingying Jiao, Zhenguang Liu
TRUST‑SQL tackles Text‑to‑SQL parsing in environments where the full database schema is unknown, a common scenario in enterprise settings with many noisy tables. The method models the problem as a Partially Observable Markov Decision Process and uses a four‑phase protocol to verify only relevant metadata, guided by a Dual‑Track GRPO strategy that separates exploration from execution rewards. Experiments on five benchmarks show significant gains, with the 4B and 8B variants outperforming base models by 30.6% and 16.6% respectively, while matching or exceeding baselines that rely on pre‑loaded schemas.
By Ai Jian, Xiaoyun Zhang, Eryu Guo, Wanrou Du, Jingqing Ruan, Jiangbo Pei, Weipeng Zhang, Ke Zeng, Xunliang Cai
The paper presents a method for training Nemotron 3 Ultra to generate proofs for difficult Olympiad mathematics. By fine‑tuning two specialist checkpoints with supervised learning and reinforcement learning, the authors evaluate how checkpoint selection, verification, and refinement affect performance. The resulting open‑model pipeline, which operates entirely in natural language without external tools, achieved 30 out of 42 points at IMO 2026, meeting the gold‑medal threshold, and the authors release the checkpoints, training data, code, solutions, and a new benchmark of 200 problems.
By Ivan Moshkov, Stephen Ge, George Armstrong, Wei Du, Sadegh Mahdavi, Igor Gitman
ReactHuman is a physics‑grounded benchmark that tests whether multimodal large language models (MLLMs) can make immediate, safety‑critical decisions in simulated humanoid scenarios involving sudden household hazards. The benchmark includes 17 event families, over 1,000 reproducible scenes generated from 240 Hz rigid‑body simulation, and a five‑metric suite evaluating reactions on reasonableness, safety, and physical grounding. Evaluation of seven MLLMs reveals that reactive safety remains unsolved, with models frequently mishandling hazards, relying on appearance over motion, and missing key interception points.
By Yizhan Li, Jianxin You, Mengyang Xiong, Yinhuan Chen, Zicheng Zhao, Dekun Wu, Dongqing Zhang, Bang Liu
The article "Defining AI Agents: A Compendium of Criteria, Metrics, and Benchmarks" surveys the lack of a standard definition for AI agents and organizes this ambiguity into five dimensions: environmental interaction, learning and adaptation, autonomy, goal‑directed behavior, and temporal coherence. It reviews how each dimension has been conceptualized in prior work and compiles the metrics, benchmarks, and evaluation frameworks used to assess them. The authors also introduce the Agent Compendium, a public digital resource that extends these evaluation methods, aiming to provide a common structure for evaluating and comparing agent capabilities across AI systems.
By Mia Lassiter, Brinnae Bent