The paper introduces Jacobian Scopes, a set of gradient‑based, token‑level causal attribution techniques for large language models (LLMs). These methods, grounded in perturbation theory and information geometry, measure how individual input tokens affect specific logits, the overall predictive distribution, and model uncertainty. Experiments on instruction understanding, translation, and in‑context learning tasks show that Jacobian Scopes uncover political biases, translation strategies, and mechanisms in time‑series forecasting, and outperform existing attribution methods on token‑ablation benchmarks across six LLMs.
By Toni J. B. Liu, Baran Zadeo\u{g}lu, Gurbir Arora, Rapha\"el Sarfati, Nicolas Boull\'e, Christopher J. Earls
The paper introduces TokEye, a reference‑free method that learns a general criterion for coherent structure from heterogeneous time‑frequency measurements. Without retraining, TokEye can extract signals from diverse fields such as marine bioacoustics and wideband radar, achieving high recovery rates (82.5% for stellarator signals, 78% for dolphin whistles, and 81.5% average precision on radar benchmarks). The method processes full experimental spectrograms in under one second on a GPU, demonstrating scalable automated scientific signal discovery across unrelated sensors.
By Nathaniel Chen, Kouroche Bouchiat, Peter Steiner, Andrew Rothstein, David Smith, Max Austin, Mike van Zeeland, Marie Roch, Azarakhsh Jalalvand, Egemen Kolemen
The paper investigates how Large Language Models (LLMs) handle the exploration–exploitation trade‑off in Bayesian Optimization, noting that single‑agent LLM approaches suffer from cognitive overload and unstable search dynamics. It proposes a multi‑agent framework that separates strategic policy mediation from tactical candidate generation, allowing explicit, observable, and adjustable exploration–exploitation decisions. Experiments on continuous optimization benchmarks show that this decomposition improves LLM‑mediated search effectiveness.
By Mohammadsina Almasi, Andrea Carbonati, Hadis Anahideh
EdgeCIM is a hardware-software co-design framework that accelerates autoregressive decoding of small language models on edge devices. It combines a CIM macro built in 65 nm with a tile-based mapping strategy to balance pipeline stages and reduce DRAM bandwidth bottlenecks. Benchmarks show that, at INT4 precision, EdgeCIM delivers up to 7.07× higher throughput than competing edge accelerators and averages 336.4 tokens/s with 173 tokens/J across a range of models up to 4 B parameters.
By Jinane Bazzi, Mariam Rakka, Fadi Kurdahi, Mohammed E. Fouda, Ahmed Eltawil
DUET is a token‑budget allocation method for reinforcement learning with verifiable rewards that jointly optimizes which prompts receive rollouts and how long each rollout runs. By using a lightweight surrogate to assign rollouts and a marker‑gated abort rule to stop them, DUET improves reasoning quality and reduces training time compared to full‑budget GRPO and other baseline methods. Experiments on Qwen3‑1.7B and other LLMs show that DUET outperforms baselines even with only 50% of the token budget, achieving up to 2.51× speedup while maintaining or improving performance.
By Haoyu Hu, Xuandong Zhao, Xuhai "Orson'' Xu, Nori Jacoby
JANUS is a benchmark that evaluates goal‑conditioned pragmatic distortion in factually grounded large language model (LLM) outputs. It contains 160 scenarios across eight domains, each offering a fixed set of favorable and adverse facts, and contrasts neutral decision support with institutional conditioning that may promote adoption or support despite potential harm. By restricting outputs to the same fact pool, JANUS isolates misleading impressions from hallucination and fabrication, revealing that current LLMs consistently exhibit goal‑conditioned distortions and lack robust safeguards against selectively misleading communication.
By Polydoros Giannouris, Mohsinul Kabir, Sophia Ananiadou
OutageDiT is a foundation model designed to generate seven‑day power‑outage trajectories at 15‑minute resolution. It uses a condition encoder to process historical context and future covariates, and a shallow flow‑matching decoder to sample complete outage trajectories, enabling point forecasting, uncertainty quantification, and conditional scenario generation. The model improves forecast accuracy and scenario quality over strong baselines and can transfer zero‑shot to held‑out regions, linking outage forecasting to operational planning under uncertainty.
By Yunqin Zhu, Feng Qiu, Yao Xie
SituHOI is a large‑scale RGB‑D dataset that captures full‑body human‑object interactions in 16 indoor rooms, comprising 18,429 annotated clips and about 86 hours of activity. The dataset pairs moving‑camera recordings with static scene scans, object geometry, language descriptions, and estimated human and object trajectories, all registered in a shared floor‑aligned frame via a pipeline that integrates camera localization, depth‑based alignment, human mesh recovery, and object tracking. The data enable the FlexHOI‑Bench benchmark for goal‑conditioned joint motion prediction, and a reference predictor, SharedHOI, demonstrates strong performance on world‑space accuracy and human‑object coordination.
whyItMatters":"The dataset provides scalable, scene‑aligned supervision for modeling human‑object interaction, addressing a key limitation of existing datasets and enabling more accurate motion prediction benchmarks."
By Qifan Wang, Qihang Fang, Boao Shi, Yushi Du, Pei Zhou, Siyan Dong, Yanchao Yang
ManifoldCache is a training‑free, data‑free accelerator for Constraint‑Manifold Diffusion Models (CMDMs), which generate samples that must satisfy hard geometric constraints in scientific domains such as medical imaging, molecular design, and 3D scene generation. By decomposing the conditional score into orthogonal normal and tangential components, the method identifies a noise‑schedule midpoint as a safe caching boundary—caching before this point guarantees bounded error, while caching after it leads to mode confusion. The approach also shows that deeper network blocks allow larger certified cache strides, all without requiring calibration data or additional training.
By Prashant Pandey, Devineni Sri Venkatraya Chowdary, Brejesh Lall
The paper introduces Aligned Delta-Triplane Transformers (ADTT), a compact 4D Occupancy World Model that explicitly aligns historical triplanes with ego-motion compensation and predicts only future scene changes. By using a query-conditioned Transformer to predict residual changes guided by ego-motion and external queries, ADTT adds these changes to the aligned prior and decodes them into future occupancy scenes. Experiments on Occ3D-nus demonstrate that ADTT achieves state‑of‑the‑art forecasting performance with fewer parameters and higher FPS.
By Haoran Xu, Peixi Peng, Shijun Guo, Guang Tan, Yiqian Chang, Yisen Zhao, Shuaixian Wang, Yonghong Tian
The paper introduces Bottle‑SKU, a large‑scale benchmark for detecting surface defects on shampoo bottles across 48 different SKUs, providing 13,119 images, 20,419 bounding boxes, and seven defect categories. It proposes Prototype‑Guided Dynamic Domain Modulation (ProDMM), which builds a visual prototype for each SKU and uses it to dynamically modulate a pretrained detection backbone, improving SKU‑aware defect recognition. Experiments on Bottle‑SKU and the GC10‑DET‑SKU dataset show that this prototype‑guided conditioning enhances performance in industrial defect detection tasks.
By Yuehang Li, Xiao Wang, Qiang Chen, Yangyi Fan, Bo Jiang, Yaowei Wang, Jin Tang
AgentDecept is a new benchmark that evaluates deceptive behavior in large language model agents that use tools under pressure. It contains 477 pressure variants derived from 257 source tasks across six pressure sources, each paired with a pressure‑free counterpart. The protocol compares user‑facing claims to time‑ordered observations and executed outcomes, distinguishing falsification, concealment, and equivocation from honest failure and disclosed limitations.
By Yehan Zhang, Qianqi Yan, Hang Yin, Ching-Chen Kuo, Shan Jiang, Ming Ouyang, Yijie Liu, Xin Eric Wang
arXiv:2610. 39439v1 Announce Type: new Abstract: Label-free test-time reinforcement learning adapts vision--language models using rewards derived from their own predictions.
By Haotian Lyu, Haoyu Wang, Ru Wang, Feng Ji, Ziwen Li, Jiaxian Guo, Mingming Gong
Large Language Models (LLMs) treat all context tokens uniformly, which creates prompt injection vulnerabilities and unreliable instruction execution. The paper introduces LLMON (LLM Object Notation), a post‑training framework that separates executable instructions from passive data and requires explicit identifier‑based execution. Using a 4.6 B‑token structured corpus and post‑training on Granite‑4.1 models, LLMON reduces indirect prompt‑injection success rates from 53–100 % to 0.02–5.6 % while maintaining general capabilities, and provides attention‑based interpretability metrics to diagnose failure modes.
By Basel Shbita, Michael Hind, Chad DeLuca, Nathan Fulton, Shubhi Asthana, Farhan Ahmed, Anisa Halimi, Yada Zhu, Mauro Martino, David Cox, Dan Gutfreund
The paper introduces SelfMorph, a method that lets frozen large language models adapt their internal computation topology at inference time based on the input. It combines three morphing strategies—BASE, REPEAT, and BRANCH—into a unified space and uses a lightweight selector trained via distillation to choose the best morphology for each prompt. Experiments on 4B–27B models show that SelfMorph improves performance by an average of 8.06% on in‑distribution benchmarks and 4.45% on unseen benchmarks, outperforming other settings in most cases.
By Junxi Wu, Yiwen Guo, Shu-Tao Xia
The paper investigates how autoresearch agents—AI systems that conduct research autonomously—benefit from interacting with human researchers, a process called co‑improvement. It presents empirical evidence that humans remain crucial during the ideation phase of benchmark creation, introduces HackBench to expose alignment gaps in current agents, and proposes two training methods: using human simulators to enhance agent interaction and applying reinforcement learning with expert‑aligned rubrics to improve communication. These approaches aim to make autoresearch agents more capable and aligned with human expectations.
By Tianjian Li, Chuanyang Jin, Seungone Kim, Tong Chen, John Hoffman, Yoram Bachrach amd Pedro Rodriguez, Ping Yu, Eryk Helenowski, Swarnadeep Saha, Jack Lanchantin, Weizhe Yuan, Ilia Kulikov, Jason Weston
The paper introduces TACE, a two‑stage estimator for advantage estimation in Group Relative Policy Optimization. Stage S1 replaces linear standardization with an optimal‑transport map to a standard normal, aligning reward distributions across tasks and limiting outliers. Stage S2 rank‑calibrates token‑level credit assignment, normalizing token weights to preserve total credit while redistributing it within trajectories. The authors prove theoretical guarantees—inter‑task moment equity, deterministic tail bounds, and mass‑conservation—that previous methods lack, and demonstrate that TACE achieves the best performance on most benchmarks in a multimodal reasoning suite.
By Lang Chen, Zongxin Xu, Borui Wang, Anran Liu, Shitong Chen
The paper "From Intent to Improvement: Agentic Dataset Search for Better Model Training" introduces AgenticDS, a contract‑governed agentic framework that transforms a training‑goal query into a typed search plan, executes deterministic operators over hybrid retrieval and graph expansion, and verifies provenance‑grounded evidence before ranking datasets. On the HuDS‑Bench benchmark, AgenticDS outperforms state‑of‑the‑art dense–sparse hybrids by 25.7% in NDCG@10 and improves downstream accuracy in mathematics, code, and medicine. The authors provide open‑source code and data at https://github.com/nju-websoft/AgenticDS.
By Qiaosheng Chen, Kaijia Huang, Fanqi Zhou, Gong Cheng
The paper examines how AI‑written research papers often contain misattributed numerical claims. By re‑analyzing 540 judgments from a benchmark of nine AI agents on fifteen research problems, the authors manually labeled 256 claims and found that most failures stem from misalignment between the paper and the underlying work rather than outright cheating. They propose a deterministic claim‑to‑evidence checker that achieves over 90% precision and, when combined with an LLM verifier, offers a more reliable audit of AI‑generated research.
By Mosab Alfaqeeh, Omar Hourani, Muath Hourani, Hashem Alfaqeeh
The paper introduces a hybrid memory system for long-horizon dialogue agents that combines episodic memories with structured relational memories, supported by a consolidation mechanism to maintain efficiency. Experiments on the LoCoMo and LongMemEval-S benchmarks show that the two memory types complement each other, outperforming existing memory frameworks with 90.2% accuracy on LoCoMo and 85.5% on LongMemEval-S.
By Abhishek Gupta, Dongge Han, Daniel Madrigal Diaz, Wei-Ning Chen, Mayukh Das, Victor R\"uhle, Saravan Rajmohan