This tutorial paper reviews a decade of advances at the intersection of machine learning and control, focusing on methods that blend machine learning flexibility with control theory’s structure to create capable, reliable, and safe autonomous systems. It covers core concepts for stability and safety, learning dynamics models to handle uncertainty, and data‑driven policy synthesis techniques such as reinforcement learning, imitation learning, and Bayesian optimization. Representative applications and open challenges are discussed, and the authors provide implementations, examples, and additional resources on their website.
By Angela P. Schoellig, Matthias A. M\"uller, Sebastian Trimpe, Melanie N. Zeilinger, Mahathi Anand, Andrea Carron, Oliver Hausd\"orfer, Victor G. Lopez, Friedrich Solowjow, Haocheng Zhao, Siqi Zhou
The study evaluates whether providing biological context to large language models (LLMs) improves their performance in autonomous laboratory optimization. Using 45 biology tasks derived from 2,485 experimental data points, the authors compared domain‑aware prompting (with biological parameter names, units, and labels) to domain‑agnostic prompting (renamed to be non‑biological) across 19 LLMs over 30 iterations. Results show that domain‑agnostic prompts outperform domain‑aware prompts on 58.7% of tasks by trajectory and 61.9% by endpoint, suggesting that biological knowledge does not reliably enhance LLM‑guided experimental optimization.
"whyItMatters":"The findings imply that autonomous laboratory loops should treat biology‑aware optimization as a distinct capability rather than relying on biological context to improve LLM performance."
By Marilyn Zhang, Tianfeng Chen, Mark E. Whiting
The paper introduces the 2-Step Agent framework, a computational model that explains how a Bayesian decision maker updates beliefs and estimates causal effects after receiving a prediction from a machine‑learning model. It provides a tractable solution for the multivariate linear‑Gaussian case, identifies when machine‑learning decision support improves outcomes, and demonstrates that a single misaligned prior can cause the support to worsen decisions even with a well‑specified model.
By Otto Nyberg, Davide Tugnoli, Fausto Carcassi, Giovanni Cin\`a
The paper proposes a framework for building autonomous AI research teams that can discover linear attention architectures. By assembling agents with diverse disciplinary expertise, private research histories, shared experimental evidence, and adaptive allocation of research opportunities, the team can balance exploration and exploitation under limited experimental budgets. Experiments show that this design outperforms GDN2 at 1.3B scale on language modeling, common‑sense reasoning, and retrieval tasks, and that removing any single component weakens discovery.
By Yujian Liu, Li An, Zelin He, Jacob Helwig, Jiabao Ji, Rohit Jain, Siyu Zhu, Shiyu Chang
The paper investigates how large language model (LLM) agents converge toward agreement during multi‑turn conversations about subjective topics. By training a linear classifier on conversation activations, the authors uncover an internal representation of agreement that varies across turns and does not reliably match external agreement at the individual conversation level. Aggregating this internal agreement across conversations reveals a Markovian structure: external agreement at each turn depends directly on the previous turn, and internal agreement couples to external agreement via a power‑law mapping, with models often expressing external agreement before their internal representations shift.
By Adi Simhi, Yaniv Nikankin, Jonathan Herzig, Idan Szpektor, Yonatan Belinkov, Shay B. Cohen
Memory-Centric Continual Learning for AI Agents proposes a framework that redefines agent memory as a driver of future actions rather than a passive retrieval cache. It introduces a three-axis taxonomy—substrates, assets, and policies—to structure memory, a two-loop lifecycle for updating memory state and refining management policies, and an evaluation protocol that isolates genuine behavioral gains from simple context expansion. The paper also presents a memory scaling framework to model how performance scales with experience, storage, and compute across non-parametric and parametric memory.
By Dongfang Li, Zhuoen Chen, Bonian Jia, Zixuan Liu, Zhengxuan Lu, Can Deng, Yukun Shi, Yuhang Meng, Xin Zhang, Jiahe Huang, Fuhao Li, Junmai Wang, Shunyao Wen, Jingjun Sun, Jiaxu Tian, Jiangling Zhang, Zihan Wang, Liangyou Li, Yitong Yao, Shuangyong Song, Beilun Wang, Pengxu Wei, Hanlin Xu, Chang Liu, Yongxiang Li, Baotian Hu, Xiangyang Ji, Min Zhang
The paper introduces a framework that links the foundational capabilities of large language models (LLMs) to their agentic performance by measuring operation-level gaps. It constructs foundational profiles from benchmark data and operation profiles from real interaction trajectories, then quantifies how much observed operation performance deviates from predictions based on foundational capabilities. Experiments on fifteen model services in the τ²-Bench demonstrate that foundational profiles alone only partially predict agentic outcomes, while incorporating operation profiles significantly improves task-score predictions and reveals consistent execution gaps across task subsets.
By Wei Guo, Yuanjian Xu, Jianing Hao, Yongyao Duan, Ran Chen, Zhong Li, Guang Zhang
Quant-Arena is a new benchmark and public leaderboard for end‑to‑end quantitative research using real A‑share market data. Models build features and stock‑scoring code, receive feedback only on 2022–2023 data, and freeze a strategy for evaluation on the sealed year 2024. Scores are based on realized excess return, signal quality, and feature diversity, with GPT‑6 Astra currently leading but still not outperforming the equal‑weight market in 2024.
By Houde Qian, Zheng Jiang, Wenyi Li, Chenxi Xu, Xinjie Lin, Yuzhao Peng, Ruoyu Han, Youjie Zheng, Kaisen Yang, Qinhuai Na, You Fu, Calvin Xiao
The paper introduces Banzhaf Resource Attribution (BRA), a factorial framework that quantifies how different inference-time resources—computation, information, and feedback—contribute to repairing failed coding agents. Experiments on six models across 194 function-level and 126 repository-level failures show that task-relevant specifications consistently improve repair, while public-test feedback offers modest gains and retrieved code offers little benefit. The study finds that as model capability increases, the value of extra computation diminishes, yet information remains useful, and that agents often prefer feedback even when it yields lower gains than random selection, revealing a mismatch between preferred and effective resources.
By Saiyue Lyu, Jerrold Huang, Yunbei Zhang, Janet Wang, Zhenting Wang, Yingqiang Ge, Haibo Zhang
The paper introduces VeeReview, a tool that classifies build‑and‑verify pairs in pull requests using repository signals and a language‑model judge. Evaluated on 1,454 agent‑ and human‑authored pull requests, VeeReview found lower build‑only rates for agent‑authored changes but failed to meet calibration and agreement thresholds. The study concludes that a mechanically decided component and validated human reference are essential for reliable verification‑practice comparisons.
By Gabriel Dautovi, Jie JW Wu
The paper introduces a reproducible framework that converts existing text-based tool‑calling benchmarks into audio‑based evaluations using text‑to‑speech, speaker variation, and environmental noise, while preserving original annotations. Experiments on seven omni‑modal models show that performance varies strongly by model and task, with notable gaps between text and voice results. The study also includes a stress test, a reference‑free LLM‑as‑judge protocol, and demonstrates that open‑source judges can match proprietary ones, aiding privacy‑preserving evaluation.
By Md Tahmid Rahman Laskar, Xue-Yong Fu, Seyyed Saeed Sarfjoo, Quinten McNamara, Jonas Robertson, Shashi Bhushan TN
The paper introduces Multimodal Embodied Agents (MMEAs), goal‑directed systems that combine multimodal task reasoning with physical action and adapt decisions based on feedback. It proposes the PAPAV framework, which defines five core capabilities—Perceive, Anticipate, Plan, Act, Verify—common to computer‑use agents, robotic systems, and MMEAs. The authors analyze 64 benchmarks, noting that most evaluations focus on Act while Anticipate and Verify are rarely assessed, highlighting gaps and open challenges in designing reliable physical agents.
By Yanzhe Chen, Ziyi Yang, Jifeng Zhu, Qiming Huang, Ruihe An, Peiyao Xu, Hesen Yang, Runda Liu, Chang Gong, Zhijun Cao, Zechen Bai, Wenzheng Zeng, Yiqi Lin, Guoqiang Liang, Kevin Yuchen Ma, Kevin Qinghong Lin, Mike Zheng Shou
The paper introduces Discord, a method that uses reward disagreement to evaluate and compare harness edits via exact paired inference and early stopping, ensuring only genuinely differing outcomes influence decisions. Building on this, BRACE (Bank‑calibrated Rollout Allocation under Correlated Episodes) forecasts mixed‑reward groups by combining historical task evidence, uncertainty in success rates, and a correlated reward model, thereby allocating generation budgets more efficiently. Experiments on Qwen3‑VL, BFCL, and NESTFUL demonstrate that BRACE outperforms baseline methods with fewer new training rollouts and maintains strong performance without additional training.
By Hieu Tran, Zonghai Yao, Sunjae Kwon, Hong Yu
The paper evaluates the Attention Tracker, a white‑box detection method for indirect prompt injection (IPI) attacks, in realistic agentic settings using the AgentDojo benchmark. It finds that the method’s claimed dataset‑agnostic detection fails in agentic environments, with near‑chance performance across most model families. The authors propose AttnDA, a Fisher Linear Discriminant Analysis classifier that replaces the scalar threshold, which improves detection across multiple benchmarks but still struggles to transfer to AgentDojo, highlighting the importance of training data from the target domain.
By Amelia Kawasaki, Stefan Lee
The paper studies how long‑context continual pre‑training (LCCP) scales with token budget and model size. An empirical additive power law is derived that predicts LCCP performance across five model scales and over a hundred budget configurations, and a new Constant‑Last4 protocol is introduced for stable monitoring. Experiments show that performance continues to improve without clear saturation, even after 128K context extension and fine‑tuning, indicating that scaling LCCP yields transferable gains for downstream tasks.
By Ziqing Qiao, Xiaoyue Xu, Sikui Zhang, Yinuo Xu, Bingxiang He, Chaojun Xiao, Xu Han
The paper introduces MHG‑Mem, a multimodal hypergraph memory that preserves fine‑grained visual evidence while tracking entity evolution across sessions. It represents entity mentions and observed facts as nodes, groups identity‑consistent mentions via occurrence hyperedges, and synthesizes aspect‑specific evolution states linked through evolutionary and evidential hyperedges. Hypergraph‑based agentic search enables adaptive traversal of the memory, allowing detailed visual verification and reasoning over cross‑session state changes, achieving 8.20‑9.98% performance gains on three benchmarks.
By Jielong Tang, Ling-An Zeng, Xujie Yuan, Xiao Dong, Zhiqian Xia, Qingfeng Li, Shixiong Kai, Kaichao Liang, Rong Yao, Jian Yin
The paper investigates how token consumption varies across programming languages for coding agents. Five recent models were evaluated on Python, Java, Rust, and OCaml, revealing significant differences in token usage that persist across models. Analysis of agent trajectories shows that agents often generate non‑compiling solutions in unfamiliar languages, revise passing solutions, plan in code comments, distrust provided tests, and prototype in Python to sidestep unfamiliar languages.
By Zixuan Wu, Carolyn Jane Anderson, Arjun Guha
NeuralSRender is a sim‑to‑real video translation framework that improves the photorealism of synthetic driving videos while preserving geometric consistency. It builds on a pretrained video diffusion transformer and uses a lightweight projector to inject G‑buffer geometric and material priors, with LoRA for efficient adaptation. The two‑stage training first learns G‑buffer conditioning from synthetic data, then adapts to real‑world appearance, adding only 1.45 % parameter overhead and requiring 24 GB GPU memory. The authors also release GenReal, a large‑scale dataset of synthetic and real driving videos with G‑buffers, and show that training perception models on NeuralSRender‑translated videos boosts detection AP₅₀ on nuImages from 35.1 to 40.6 and segmentation mIoU on Cityscapes from 32.4 to 40.8.
By Yipeng Lu, Youqi Liao, Zhenyu Wang, Yuning Peng, Zhen Dong
Code Comprehensor is a hybrid retrieval‑reasoning framework designed to aid understanding of large codebases. It combines a code dependency callgraph, CodeScribe—which generates hierarchical documentation using code‑aware LLMs—and CodeContrast, a CLIP‑style dual‑encoder that learns aligned text‑code embeddings for fast, semantically precise retrieval. Experiments on OpenCV and a 5M‑LOC proprietary repository show that CodeScribe‑derived supervision improves cross‑modal retrieval accuracy, and LLM‑based evaluation reports over 90% faithfulness.
By Sujoy Roychowdhury, Joseph H. R. Isaac, Jiangning Gao, Martin Klitte, Diarmuid Corcoran
RuleSmith is a framework that uses multi‑agent large language models (LLMs) in self‑play combined with surrogate‑based optimization to automatically balance games. In a proof‑of‑concept study, the authors built CivMini, a simplified civilization‑style game with 12 tunable parameters, and showed that RuleSmith consistently finds near‑balanced configurations across multiple optimization runs and search methods. The system produces interpretable numerical rule adjustments without requiring task‑specific policy training, and preliminary cross‑model evaluations and small human studies suggest transferability across player populations.
By Ziyao Zeng, Hao Wang, Chen Liu, Youheng Yao, Jingcheng Ni, Tianyu Liu, Xiatao Sun, Fengyu Yang, Chenyu You, Xiaofeng Liu, Daniel Rakita, Ronald R. Coifman, Yuval Kluger, Zhiwen Fan