The paper proposes a method to embed sequential Bayesian experimental design (BED) into large language models (LLMs) by training the models to reason about action selection using information-gain rewards. By specifying the experimental generative model separately from the LLM’s beliefs, the authors optimize expected information gain under a chosen model. Experiments on the Twenty Questions task show that the resulting policies outperform both inference- and training-time BED baselines while incurring lower computational costs, and the approach generalizes to unseen categories and transfers positively to the Medi-Q clinical dialogue benchmark.
By Anita Kriz, Antonin Berthon, Nicol\'as Astorga, Mihaela van der Schaar
The study examines how three context‑engineering techniques—semantic layers, orchestration, and compaction—affect data agents that answer questions over heterogeneous data. Using seven large language models on 1,514 long‑horizon tasks, the authors identify three agent behavior groups (investigators, premature answerers, early stoppers) and show that each technique benefits only the group it aligns with, while sometimes harming others. Combining the techniques yields smaller net gains than the sum of individual effects, leading to guidance on selecting techniques based on agent behavior, task demands, and accuracy‑cost trade‑offs.
By Zhivar Sourati, Marianne Menglin Liu, Zheng Wang, Rongguang Wang, Sai Ashish Somayajula, Tao Sheng, Kyu J. Han, Sujith Ravi, Dan Roth
ToolBottleneckBench is a benchmark that evaluates tool‑agent performance under controlled pressures and fixed harness budgets. It keeps synthetic enterprise tasks constant while varying one pressure—such as the registered tool catalog, visible tool menu, certified workflow, or record encoding—to measure how each affects success. Across 21 model and harness configurations, the visible menu pressure uniquely imposes a large penalty on the model, largely due to truncation at the completion cap, and most failures are driven by harness limits rather than tool‑specific issues.
By Zichuan Xu, Yuyang Dong, Makoto Onizuka, Chuan Xiao
V‑SHARE is an advantage‑based value‑learning algorithm that blends cross‑action value sharing with an importance‑weighted Bellman‑residual correction. The paper presents the first full theoretical analysis of such an algorithm, offering both asymptotic convergence and finite‑sample guarantees. It also introduces schedules that balance deterministic contraction against importance‑weighting noise, and a two‑timescale analysis that eliminates a residual‑weight threshold while maintaining convergence.
By Kintan Saha, Hamed Masoudi, Xinyi Sheng, Dominik Baumann, Joni Pajarinen, Arsenii Mustafin
The paper evaluates how different encoding strategies and pair selection methods affect Pairwise Difference Learning (PDC) for classification. It compares four encoders—including a new triplet with absolute difference—and four pair selection schemes across almost 200 datasets using several base models. The study finds that the new encoder yields the best average F1 scores, that using only upper‑triangle pairs cuts training time by about 40% with little accuracy loss, and that no single configuration dominates across all datasets.
By Bernard Georges, Mohamed Karim Belaid, Manuel Rodrigues
The paper proposes Causal-Beam-Guided Subtask Transfer (CBST), a method that learns manipulation-related representations from human videos and grounds them in task-level semantic variables using robot demonstrations. CBST uses causal-beam constraints to select contextual information that preserves local mechanism outputs while maintaining responsiveness to retained inputs. Experiments on robot manipulation benchmarks demonstrate that CBST improves task success compared to baseline world-action models.
By Xingyu Zhang, Yingchao Li, Peizheng Guo, Jie Hu, Changwen Zheng, Wenwen Qiang
The study investigates which inputs influence a language model’s financial forecasts by creating synthetic earnings announcements that independently reverse 20‑day momentum, commentary tone, and EPS surprise. Six instruction/chat checkpoint conversions from four model families (0.49‑7.61 B parameters) were tested with 2,016 prompts each, covering 384 primary scenarios. Results show that the contrast in primary EPS ranges from –0.19 to 86.80 percentage points, with signed tone and momentum contrasts often exceeding EPS effects; numerical diagnostic accuracy varied from 51.0 % to 95.8 %. The experiment measures observable input reliance and numerical recognition but does not establish internal reasoning mechanisms or return predictability.
By Alina Khaybullina
Multi‑SFT is a new, commercially permissive dataset comprising 502,818 instruction‑response pairs in five European languages, created through translation and regeneration pipelines to address gaps where simple translation is insufficient. Human evaluation shows strong inter‑annotator agreement (≥0.87) across all languages. Fine‑tuning nine models from three families on Multi‑SFT yields performance competitive with larger, proprietary mixes and official instruction‑tuned releases, especially on safety‑related benchmarks.
By Daniela Baiamonte, Kamyar Zeinalipour, Elena Fano, Stefano Simonazzi, Andrea Zugarini
EMO is a training‑free, closed‑loop framework that alternates context estimation and distribution refinement to steer frozen language models during test time. It uses textual gradients from an external judge to update the system context (E‑step) and then applies a confidence‑constrained logit correction based on next‑token distribution differences (M‑step). Empirical results show EMO improves zero‑shot performance across eight models and six benchmarks, outperforming both baseline and context‑only methods such as TextGrad.
By Wei Xia, Yehui Tang, Zhi-Hong Deng
TRACE is a topology‑aware, prediction‑adaptive objective for thin‑structure segmentation that uses a target skeleton to define required local connections. For each connection, it identifies the easiest valid reconnection under the current prediction and jointly supervises its bottleneck with the corresponding skeleton point, allowing supervision to adapt to the evolving prediction rather than remaining fixed. A topology‑aware skeleton carrier preserves both foreground and class‑specific topology, enabling reliable definition of these connections, and extensive experiments on six 2D and 3D benchmarks demonstrate improved connectivity and topology while maintaining strong segmentation accuracy with low training overhead.
By Junho Moon, Ikbeom Jang
The paper presents an evolutionary framework where large language models (LLMs) generate quantum circuit designs and an independent evaluator scores them. Applied to tasks in numerical computation, chemistry, optimization, quantum simulation, and number theory, the approach yields circuits that reduce Toffoli and qubit counts in seven of eight tasks, and discovers a new reversible architecture for prime-field modular inversion that improves the Toffoli‑qubit trade‑off. This demonstrates that evaluator‑guided LLM co‑design can produce state‑of‑the‑art quantum algorithms.
By Andi Gu, Alek H. Kemeny, Susanne F. Yelin, Mikhail D. Lukin
The paper introduces CellForge, a team of autonomous agents that can autonomously generate, train, and validate predictive models for single‑cell perturbation data across transcriptomic, epigenomic, and proteomic contexts. By surveying literature, parsing data, and engaging in critic‑mediated discussions among specialized agents, CellForge produces end‑to‑end model architectures and code without human specification, achieving performance competitive with state‑of‑the‑art hand‑crafted methods. Human evaluators found CellForge’s research plans to be more scientifically sound and innovative than those produced by single large language models or deep‑research systems.
By Xiangru Tang, Jiapeng Chen, Daniel Shao, Zhuoyun Yu, Yan Cui, Weixu Wang, Fang Wu, Yuchen Zhuang, Wenqi Shi, Zhi Huang, Arman Cohan, Xihong Lin, Fabian Theis, Smita Krishnaswamy, Mark Gerstein
ReText is a method that enhances image-based person re-identification by incorporating textual descriptions into training. The authors augment existing multi-camera datasets with captions generated by a fine‑tuned Vision‑Language Model and train the system on three tasks: Re-ID, image‑text matching, and image reconstruction guided by text. Experiments show that ReText improves generalization across unseen domains and outperforms current state‑of‑the‑art approaches on cross‑domain Re‑ID benchmarks.
By Timur Mamedov, Karina Kvanchiani, Anton Konushin, Vadim Konushin
RMBench is a simulation benchmark featuring nine robotic manipulation tasks that vary in memory complexity, designed to systematically evaluate how well policies use memory. The authors also introduce Mem-0, a modular policy with explicit memory components that allows controlled ablation studies. Experiments in simulation and the real world reveal memory-related limitations in current policies and show how architectural choices affect memory performance.
By Tianxing Chen, Yuran Wang, Mingleyang Li, Yan Qin, Hao Shi, Zixuan Li, Yifan Hu, Yingsheng Zhang, Kaixuan Wang, Yue Chen, Hongcheng Wang, Junjie Wang, Tianhang Yang, Renjing Xu, Ruihai Wu, Yao Mu, Yaodong Yang, Hao Dong, Ping Luo
The paper introduces Resolving Interference (RI), a lightweight adaptation framework that reduces cross‑task interference when merging expert models. RI disentangles expert models to be functionally orthogonal to other tasks, using only unlabeled auxiliary data, and improves merging performance by up to 3.8% and generalization to unseen domains by up to 2.3%. The method is robust to the source of auxiliary input and less sensitive to hyperparameter tuning.
By Pratik Ramesh, George Stoica, Arun Iyer, Leshem Choshen, Judy Hoffman
ChaosBench-Logic v2 is a 40,886-question benchmark designed to evaluate large language models’ logical reasoning about dynamical systems, featuring 165 systems, 27 first‑order logic predicates, and 78 axiom edges. The accompanying CARE protocol highlights failure modes such as prior collapse, paraphrase inconsistency, and difficulty with parameter‑dependent dynamics. In tests of 14 models, regime‑transition reasoning remains near random (MCC = 0.05), while first‑order deduction with given premises achieves MCC = 0.52, with proprietary models excelling in cross‑indicator and consistency tasks and open‑source Qwen 2.5‑32B leading indicator diagnostics.
By Noel Thomas
The paper introduces a test‑time adaptation (TTA) method for AI text detection that addresses continual distribution shifts such as adversarial humanization, new large language models, and temporal drift in human writing. By exploiting inference‑time homogeneity among unlabeled samples, the TTA approach achieves robust performance, outperforming state‑of‑the‑art supervised detectors—e.g., improving detection of adversarial AI‑generated text from 27.0% to 93.3%. The study demonstrates that TTA can maintain high detection accuracy in real‑world, evolving deployment scenarios.
By Kevin Ren, Manish Raghavan, Nikhil Garg
arXiv:2610. 21312v1 Announce Type: new Abstract: Data-free preconditioning builds the Kronecker factors of KFAC from synthetic probes instead of private data.
By Shilin Zhang, Kangning Wang
The paper introduces SemTAC, an encoder trained with a temporal-neighborhood contrastive loss to preserve local consistency in human motion time series, enabling inference at temporal scales beyond training. SemTAC outperforms four baselines on a five-step motion sequence discrimination task, especially on truncated sequences where it achieves a median F1 of 0.77 versus 0.27–0.70 for others. Its representations form interpretable geometric trajectories, surpassing a no-structure baseline on a displacement‑smoothness metric.
By Louren\c{c}o A. Rodrigues, Felix Putze
Q($\gamma_H$) introduces a multi-horizon Q-learning framework that keeps separate discount-specific heads and blends them during control using an entropy-derived confidence signal. The adaptive mixing shifts from short to long horizons as policy confidence grows, allowing the agent to adjust its planning horizon without altering individual Bellman updates. Experiments on classical control and deep reinforcement‑learning tasks show the method excels on delayed‑credit problems compared to fixed‑discount baselines.
By Arwa Albelaihi, Abdulrahman Altahhan