Recent studies on LLM agent scaling have highlighted the potential of Multi-Agent Debate (MAD) to enhance reasoning abilities. However, the critical aspect of role allocation strategies remains underexplored. In this study, we demonstrate that allocating roles with differing viewpoints to specific positions significantly impacts MAD's performance in reasoning tasks. Specifically, we find a novel role allocation strategy, "Truth Last", which can improve MAD performance by up to 22% in reasoning tasks. To address the issue of unknown truth in practical applications, we propose the Multi-Agent Debate Consistency (MADC) strategy, which systematically simulates and optimizes its core mechanisms. MADC incorporates path consistency to assess agreement among independent roles, simulating the role with the highest consistency score as the truth. We validated MADC across a range of LLMs (9 models), including the DeepSeek-R1 Distilled Models, on challenging reasoning tasks. MADC consistently demonstrated advanced performance, effectively overcoming MAD's performance bottlenecks and providing a crucial pathway for further improvements in LLM agent scaling.

Autonomous Feature Engineering (AFE) is critical for improving predictive performance on tabular data by relieving humans from manual feature crafting. However, traditional AFE lacks the semantic guidance needed to fully exploit domain knowledge. Although large language models (LLMs) can, in principle, emulate experts, existing approaches typically operate in an open code space that directly generates and rewrites entire features; without a compositional structural representation and invariant constraints, edits are coarse and non-local, making it hard to distill interpretable features with high information content and rich hierarchical structure.

Rong Cheng, Jinyi Liu, Yan Zheng, Fei Ni, Jiazhen Du, Hangyu Mao, Fuzheng Zhang, Bo Wang, Jianye Hao. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2025.

Reward design constitutes a critical challenge in practical reinforcement learning applications. While shaping rewards could theoretically accelerate agent learning, researchers face a dual dilemma: manually crafted rewards may inadvertently distort the original optimization objective, and the fundamental limitation of scalar value representations often impedes effective policy learning. To address these limitations, we propose Multi-Reward Fusion (MRF), a novel framework that systematically integrates diverse human-engineered reward signals. MRF leverages multiple task-aligned reward components to simultaneously learn distinct policy networks, then employs a hierarchical knowledge distillation process to synthesize these specialized policies into an optimized target policy. We also provide rigorous theoretical analysis establishing formal convergence guarantees for the distillation mechanism. Experimental evaluations conducted across multiple standardized benchmarks demonstrate statistically significant enhancements in both sample efficiency and asymptotic performance relative to advanced baseline approaches. This work provides a principled approach for combining multiple reward perspectives while maintaining optimization integrity, offering new insights into reward shaping and value representation challenges.

Reinforcement learning (RL) heavily depends on well-designed reward functions, which are often biased and difficult to design for complex behaviors. Preference-based RL (PbRL) addresses this by learning reward models from human feedback, but its practicality is constrained by a critical dilemma: while existing methods reduce human effort through query optimization, they neglect the preference buffer's restricted coverage — a factor that fundamentally determines the reliability of reward model. We systematically demonstrate this limitation creates distributional mismatch: reward models trained on static buffers reliably assess in-distribution trajectories but falter with out-of-distribution (OOD) trajectories from policy exploration. Crucially, such failures in policy-proximal regions directly misguide iterative policy updates. To address this, we propose **Proximal Policy Exploration (PPE)** with two key components: (1) a *proximal-policy extension* method that expands exploration in undersampled policy-proximal regions, and (2) a *mixture distribution query* method that balances in-distribution and OOD trajectory sampling. By enhancing buffer coverage while preserving evaluation accuracy in policy-proximal regions, PPE enables more reliable policy updates. Experiments across continuous control tasks demonstrate that PPE enhances preference feedback utilization efficiency and RL sample efficiency over baselines, highlighting preference buffer coverage management's vital role in PbRL.

Spreadsheets are ubiquitous across the World Wide Web, playing a critical role in enhancing work efficiency across various domains. Large language model (LLM) has been recently attempted for automatic spreadsheet manipulation but has not yet been investigated in complicated and realistic tasks where reasoning challenges exist (e.g., long horizon manipulation with multi-step reasoning and ambiguous requirements). To bridge the gap with the real-world requirements, we introduce SheetRM, a benchmark featuring long-horizon and multi-category tasks with reasoning-dependent manipulation caused by real-life challenges. To mitigate the above challenges, we further propose SheetAgent, a novel autonomous agent that utilizes the power of LLMs. SheetAgent consists of three collaborative modules: Planner, Informer, and Retriever, achieving both advanced reasoning and accurate manipulation over spreadsheets without human interaction through iterative task reasoning and reflection. Extensive experiments demonstrate that SheetAgent delivers 20--40% pass rate improvements on multiple benchmarks over baselines, achieving enhanced precision in spreadsheet manipulation and demonstrating superior table reasoning abilities. More details and visualizations are available at the https://sheetagent.github.io/. The datasets and source code are available at https://anonymous.4open.science/r/SheetAgent.

In reinforcement learning, the optimism in the face of uncertainty (OFU) is a mainstream principle for directing exploration towards less explored areas, characterized by higher uncertainty. However, in the presence of environmental stochasticity (noise), purely optimistic exploration may lead to excessive probing of high-noise areas, consequently impeding exploration efficiency. Hence, in exploring noisy environments, while optimism-driven exploration serves as a foundation, prudent attention to alleviating unnecessary over-exploration in high-noise areas becomes beneficial. In this work, we propose Optimistic Value Distribution Explorer (OVD-Explorer) to achieve a noise-aware optimistic exploration for continuous control. OVD-Explorer proposes a new measurement of the policy's exploration ability considering noise in optimistic perspectives, and leverages gradient ascent to drive exploration. Practically, OVD-Explorer can be easily integrated with continuous control RL algorithms. Extensive evaluations on the MuJoCo and GridChaos tasks demonstrate the superiority of OVD-Explorer in achieving noise-aware optimistic exploration.

Reinforcement Learning with Human Feedback (RLHF) has received significant attention for performing tasks without the need for costly manual reward design by aligning human preferences. It is crucial to consider diverse human feedback types and various learning methods in different environments. However, quantifying progress in RLHF with diverse feedback is challenging due to the lack of standardized annotation platforms and widely used unified benchmarks. To bridge this gap, we introduce **Uni-RLHF**, a comprehensive system implementation tailored for RLHF. It aims to provide a complete workflow from *real human feedback*, fostering progress in the development of practical problems. Uni-RLHF contains three packages: 1) a universal multi-feedback annotation platform, 2) large-scale crowdsourced feedback datasets, and 3) modular offline RLHF baseline implementations. Uni-RLHF develops a user-friendly annotation interface tailored to various feedback types, compatible with a wide range of mainstream RL environments. We then establish a systematic pipeline of crowdsourced annotations, resulting in large-scale annotated datasets comprising more than 15 million steps across 30 popular tasks. Through extensive experiments, the results in the collected datasets demonstrate competitive performance compared to those from well-designed manual rewards. We evaluate various design choices and offer insights into their strengths and potential areas of improvement. We wish to build valuable open-source platforms, datasets, and baselines to facilitate the development of more robust and reliable RLHF solutions based on realistic human feedback. The website is available at https://uni-rlhf.github.io/.

Robotic manipulation tasks often span over long horizons and encapsulate multiple subtasks with different skills. Learning policies directly from long-horizon demonstrations is challenging without intermediate keyframes guidance and corresponding skill annotations. Existing approaches for keyframe identification often struggle to offer reliable decomposition for low accuracy and fail to provide semantic relevance between keyframes and skills. For this, we propose a unified Keyframe Identifier and Skill Anotator (KISA) that utilizes pretrained visual-language representations for precise and interpretable decomposition of unlabeled demonstrations. Specifically, we develop a simple yet effective temporal enhancement module that enriches frame-level representations with expanded receptive fields to capture semantic dynamics at the video level. We further propose coarse contrastive learning and fine-grained monotonic encouragement to enhance the alignment between visual representations from keyframes and language representations from skills. The experimental results across three benchmarks demonstrate that KISA outperforms competitive baselines in terms of accuracy and interpretability of keyframe identification. Moreover, KISA exhibits robust generalization capabilities and the flexibility to incorporate various pretrained representations.

Long-horizon manipulation tasks with general instructions often implicitly encapsulate multiple sub-tasks, posing significant challenges in instruction following. While language planning is a common approach to decompose general instructions into stepwise sub-instructions, text-only guidance may lack expressiveness and lead to potential ambiguity. Considering that humans often imagine and visualize sub-instructions reasoning out before acting, the imagined subgoal images can provide more intuitive guidance and enhance the reliability of decomposition. Inspired by this, we propose **PERIA**(**PE**rceive, **R**eason, **I**magine, **A**ct), a novel framework that integrates holistic language planning and vision planning for long-horizon manipulation tasks with complex instructions, leveraging both logical and intuitive aspects of task decomposition. Specifically, we first perform a lightweight multimodal alignment on the encoding side to empower the MLLM to perceive visual details and language instructions. The MLLM is then jointly instruction-tuned with a pretrained image-editing model to unlock capabilities of simultaneous reasoning of language instructions and generation of imagined subgoals. Furthermore, we introduce a consistency alignment loss to encourage coherent subgoal images and align with their corresponding instructions, mitigating potential hallucinations and semantic conflicts between the two planning manners. Comprehensive evaluations across three task domains demonstrate that PERIA, benefiting from holistic language and vision planning, significantly outperforms competitive baselines in both instruction following accuracy and task success rate on complex manipulation tasks.

Deep reinforcement learning (DRL) and deep multiagent reinforcement learning (MARL) have achieved significant success across a wide range of domains, including game artificial intelligence (AI), autonomous vehicles, and robotics. However, DRL and deep MARL agents are widely known to be sample inefficient that millions of interactions are usually needed even for relatively simple problem settings, thus preventing the wide application and deployment in real-industry scenarios. One bottleneck challenge behind is the well-known exploration problem, i.e., how efficiently exploring the environment and collecting informative experiences that could benefit policy learning toward the optimal ones. This problem becomes more challenging in complex environments with sparse rewards, noisy distractions, long horizons, and nonstationary co-learners. In this article, we conduct a comprehensive survey on existing exploration methods for both single-agent RL and multiagent RL. We start the survey by identifying several key challenges to efficient exploration. Then, we provide a systematic survey of existing approaches by classifying them into two major categories: uncertainty-oriented exploration and intrinsic motivation-oriented exploration. Beyond the above two main branches, we also include other notable exploration methods with different ideas and techniques. In addition to algorithmic analysis, we provide a comprehensive and unified empirical comparison of different exploration methods for DRL on a set of commonly used benchmarks. According to our algorithmic and empirical investigation, we finally summarize the open problems of exploration in DRL and deep MARL and point out a few future directions.

Cyber-Physical Systems (CPSs) are composed of computational control logic and physical processes, which intertwine with each other. CPSs are widely used in various domains of daily life, including those safety-critical systems and infrastructures, such as medical monitoring, autonomous vehicles, and water treatment systems. It is thus critical to effectively test them. However, it is not easy to obtain test cases which can fail the CPS. In this work, we propose a failure-inducing input generation approach FIGCPS, which requires no knowledge of the CPS under test or any history logs of the CPS which are usually hard to obtain. Our approach adopts deep reinforcement learning techniques to interact with the CPS under test and effectively searches for failure-inducing input guided by rewards. Our approach adaptively collects information from the CPS, which reduces the training time and is also able to explore different states. Moreover, our approach is the first attempt to generate failure-inducing input for CPSs with both continuous action space and high-dimensional discrete action space, which are common for some classes of CPSs. The evaluation results show that FIGCPS not only achieves a higher success rate than the state-of-the-art approaches but also finds two new attacks in a well-tested CPS.