Large language models (LLMs) have shown remarkable advancements in enabling language agents to tackle simple tasks. However, applying them for complex, multi-step, long-horizon tasks remains a challenge. Recent work have found success by separating high-level planning from low-level execution, which enables the model to effectively balance high-level planning objectives and low-level execution details. However, generating accurate plans remains difficult since LLMs are not inherently trained for this task. To address this, we propose Plan-and-Act, a novel framework that incorporates explicit planning into LLM-based agents and introduces a scalable method to enhance plan generation through a novel synthetic data generation method. Plan-and-Act consists of a Planner model which generates structured, high-level plans to achieve user goals, and an Executor model that translates these plans into environment-specific actions. To train the Planner effectively, we introduce a synthetic data generation method that annotates ground-truth trajectories with feasible plans, augmented with diverse and extensive examples to enhance generalization. We evaluate Plan-and-Act using web navigation as a representative long-horizon planning environment, demonstrating a state-of the-art 54% success rate on the WebArena-Lite benchmark.
Designing strategyproof mechanisms for multi-facility location that optimize social costs based on agent preferences had been challenging due to the extensive domain knowledge required and poor worst-case guarantees. Recently, deep learning models have been proposed as alternatives. However, these models require some domain knowledge and extensive hyperparameter tuning as well as lacking interpretability, which is crucial in practice when transparency of the learned mechanisms is mandatory. In this paper, we introduce a novel approach, named LLMMech, that addresses these limitations by incorporating large language models (LLMs) into an evolutionary framework for generating interpretable, hyperparameter-free, empirically strategyproof, and nearly optimal mechanisms. Our experimental results, evaluated on various problem settings where the social cost is arbitrarily weighted across agents and the agent preferences may not be uniformly distributed, demonstrate that the LLM-generated mechanisms generally outperform existing handcrafted baselines and deep learning models. Furthermore, the mechanisms exhibit impressive generalizability to out-of-distribution agent preferences and to larger instances with more agents.
Recent research on Reasoning of Large Language Models (LLMs) has sought to further enhance their performance by integrating meta-thinking -- enabling models to monitor, evaluate, and control their reasoning processes for more adaptive and effective problem-solving. However, current single-agent work lacks a specialized design for acquiring meta-thinking, resulting in low efficacy. To address this challenge, we introduce Reinforced Meta-thinking Agents (ReMA), a novel framework that leverages Multi-Agent Reinforcement Learning (MARL) to elicit meta-thinking behaviors, encouraging LLMs to think about thinking. ReMA decouples the reasoning process into two hierarchical agents: a high-level meta-thinking agent responsible for generating strategic oversight and plans, and a low-level reasoning agent for detailed executions. Through iterative reinforcement learning with aligned objectives, these agents explore and learn collaboration, leading to improved generalization and robustness. Experimental results demonstrate that ReMA outperforms single-agent RL baselines on complex reasoning tasks, including competitive-level mathematical benchmarks and LLM-as-a-Judge benchmarks. Comprehensive ablation studies further illustrate the evolving dynamics of each distinct agent, providing valuable insights into how the meta-thinking reasoning process enhances the reasoning capabilities of LLMs.
With the rapid advancements in Large Language Models (LLMs), an increasing number of studies have leveraged LLMs as the cognitive core of agents to address complex task decision-making challenges. Specially, recent research has demonstrated the potential of LLM-based agents on automating Windows GUI operations. However, existing methodologies exhibit two critical challenges: (1) static agent architectures fail to dynamically adapt to the heterogeneous requirements of OS-level tasks, leading to inadequate scenario generalization;(2) the agent workflows lack fault tolerance mechanism, necessitating complete process re-execution for UI agent decision error. To address these limitations, we introduce \textit{COLA}, a collaborative multi-agent framework for automating Windows UI operations. In this framework, a scenario-aware agent Task Scheduler decomposes task requirements into atomic capability units, dynamically selects the optimal agent from a decision agent pool, effectively responds to the capability requirements of diverse scenarios. The decision agent pool supports plug-and-play expansion for enhanced flexibility. In addition, we design a memory unit equipped to all agents for their self-evolution. Furthermore, we develop an interactive backtracking mechanism that enables human to intervene to trigger state rollbacks for non-destructive process repair. Our experimental results on the GAIA benchmark demonstrates that the \textit{COLA} framework achieves state-of-the-art performance with an average score of 31.89\%, significantly outperforming baseline approaches without web API integration. Ablation studies further validate the individual contributions of our dynamic scheduling. The code is available at https://github.com/Alokia/COLA-demo.
Personalization is a critical yet often overlooked factor in boosting productivity and wellbeing in knowledge-intensive workplaces to better address individual preferences. Existing tools typically offer uniform guidance whether auto-generating email responses or prompting break reminders without accounting for individual behavioral patterns or stress triggers. We introduce AdaptAI, a multimodal AI solution combining egocentric vision and audio, heart and motion activities, and the agentic workflow of Large Language Models LLMs to deliver highly personalized productivity support and context-aware well-being interventions. AdaptAI not only automates peripheral tasks (e.g. drafting succinct document summaries, replying to emails etc.) but also continuously monitors the users unique physiological and situational indicators to dynamically tailor interventions such as micro-break suggestions or exercise prompts, at the exact point of need. In a preliminary study with 15 participants, AdaptAI demonstrated significant improvements in task throughput and user satisfaction by anticipating user stressors and streamlining daily workflows.
Code localization--identifying precisely where in a codebase changes need to be made--is a fundamental yet challenging task in software maintenance. Existing approaches struggle to efficiently navigate complex codebases when identifying relevant code sections. The challenge lies in bridging natural language problem descriptions with the appropriate code elements, often requiring reasoning across hierarchical structures and multiple dependencies. We introduce LocAgent, a framework that addresses code localization through graph-based representation. By parsing codebases into directed heterogeneous graphs, LocAgent creates a lightweight representation that captures code structures (files, classes, functions) and their dependencies (imports, invocations, inheritance), enabling LLM agents to effectively search and locate relevant entities through powerful multi-hop reasoning. Experimental results on real-world benchmarks demonstrate that our approach significantly enhances accuracy in code localization. Notably, our method with the fine-tuned Qwen-2.5-Coder-Instruct-32B model achieves comparable results to SOTA proprietary models at greatly reduced cost (approximately 86% reduction), reaching up to 92.7% accuracy on file-level localization while improving downstream GitHub issue resolution success rates by 12% for multiple attempts (Pass@10). Our code is available at https://github.com/gersteinlab/LocAgent.
The next generation of active safety features in autonomous vehicles should be capable of safely executing evasive hazard-avoidance maneuvers akin to those performed by professional stunt drivers to achieve high-agility motion at the limits of vehicle handling. This paper presents a novel framework, ManeuverGPT, for generating and executing high-dynamic stunt maneuvers in autonomous vehicles using large language model (LLM)-based agents as controllers. We target aggressive maneuvers, such as J-turns, within the CARLA simulation environment and demonstrate an iterative, prompt-based approach to refine vehicle control parameters, starting tabula rasa without retraining model weights. We propose an agentic architecture comprised of three specialized agents (1) a Query Enricher Agent for contextualizing user commands, (2) a Driver Agent for generating maneuver parameters, and (3) a Parameter Validator Agent that enforces physics-based and safety constraints. Experimental results demonstrate successful J-turn execution across multiple vehicle models through textual prompts that adapt to differing vehicle dynamics. We evaluate performance via established success criteria and discuss limitations regarding numeric precision and scenario complexity. Our findings underscore the potential of LLM-driven control for flexible, high-dynamic maneuvers, while highlighting the importance of hybrid approaches that combine language-based reasoning with algorithmic validation.
Integrating Large Language Models (LLMs) in educational technology presents unprecedented opportunities to improve instructional design (ID), yet existing approaches often prioritize automation over pedagogical rigor and human agency. This paper introduces ARCHED (AI for Responsible, Collaborative, Human-centered Education Instructional Design), a structured multi-stage framework that ensures human educators remain central in the design process while leveraging AI capabilities. Unlike traditional AI-generated instructional materials that lack transparency, ARCHED employs a cascaded workflow aligned with Bloom's taxonomy. The framework integrates specialized AI agents - one generating diverse pedagogical options and another evaluating alignment with learning objectives - while maintaining educators as primary decision-makers. This approach addresses key limitations in current AI-assisted instructional design, ensuring transparency, pedagogical foundation, and meaningful human agency. Empirical evaluations demonstrate that ARCHED enhances instructional design quality while preserving educator oversight, marking a step forward in responsible AI integration in education.
Vehicle-to-vehicle (V2V) cooperative autonomous driving holds great promise for improving safety by addressing the perception and prediction uncertainties inherent in single-agent systems. However, traditional cooperative methods are constrained by rigid collaboration protocols and limited generalization to unseen interactive scenarios. While LLM-based approaches offer generalized reasoning capabilities, their challenges in spatial planning and unstable inference latency hinder their direct application in cooperative driving. To address these limitations, we propose CoLMDriver, the first full-pipeline LLM-based cooperative driving system, enabling effective language-based negotiation and real-time driving control. CoLMDriver features a parallel driving pipeline with two key components: (i) an LLM-based negotiation module under an actor-critic paradigm, which continuously refines cooperation policies through feedback from previous decisions of all vehicles; and (ii) an intention-guided waypoint generator, which translates negotiation outcomes into executable waypoints. Additionally, we introduce InterDrive, a CARLA-based simulation benchmark comprising 10 challenging interactive driving scenarios for evaluating V2V cooperation. Experimental results demonstrate that CoLMDriver significantly outperforms existing approaches, achieving an 11% higher success rate across diverse highly interactive V2V driving scenarios. Code will be released on https://github.com/cxliu0314/CoLMDriver.
Developing autonomous home robots controlled by natural language has long been a pursuit of human. While advancements in large language models (LLMs) and embodied intelligence make this goal closer, several challenges persist: the lack of a unified benchmark for more complex robot tasks, limited evaluation methods and metrics, data incompatibility between LLMs and mobile manipulation trajectories. To address these issues, we introduce Embodied Mobile Manipulation in Open Environments (EMMOE), which requires agents to interpret user instructions and execute long-horizon everyday tasks in continuous space. EMMOE seamlessly integrates high-level and low-level embodied tasks into a unified framework, along with three new metrics for more diverse assessment. Additionally, we collect EMMOE-100, which features in various task attributes, detailed process annotations, re-plans after failures, and two sub-datasets for LLM training. Furthermore, we design HomieBot, a sophisticated agent system consists of LLM with Direct Preference Optimization (DPO), light weighted navigation and manipulation models, and multiple error detection mechanisms. Finally, we demonstrate HomieBot's performance and the evaluation of different models and policies.
Reinforcement learning with verifiable outcome rewards (RLVR) has effectively scaled up chain-of-thought (CoT) reasoning in large language models (LLMs). Yet, its efficacy in training vision-language model (VLM) agents for goal-directed action reasoning in visual environments is less established. This work investigates this problem through extensive experiments on complex card games, such as 24 points, and embodied tasks from ALFWorld. We find that when rewards are based solely on action outcomes, RL fails to incentivize CoT reasoning in VLMs, instead leading to a phenomenon we termed thought collapse, characterized by a rapid loss of diversity in the agent's thoughts, state-irrelevant and incomplete reasoning, and subsequent invalid actions, resulting in negative rewards. To counteract thought collapse, we highlight the necessity of process guidance and propose an automated corrector that evaluates and refines the agent's reasoning at each RL step. This simple and scalable GTR (Guided Thought Reinforcement) framework trains reasoning and action simultaneously without the need for dense, per-step human labeling. Our experiments demonstrate that GTR significantly enhances the performance and generalization of the LLaVA-7b model across various visual environments, achieving 3-5 times higher task success rates compared to SoTA models with notably smaller model sizes.
Academic paper review is a critical yet time-consuming task within the research community. With the increasing volume of academic publications, automating the review process has become a significant challenge. The primary issue lies in generating comprehensive, accurate, and reasoning-consistent review comments that align with human reviewers' judgments. In this paper, we address this challenge by proposing ReviewAgents, a framework that leverages large language models (LLMs) to generate academic paper reviews. We first introduce a novel dataset, Review-CoT, consisting of 142k review comments, designed for training LLM agents. This dataset emulates the structured reasoning process of human reviewers-summarizing the paper, referencing relevant works, identifying strengths and weaknesses, and generating a review conclusion. Building upon this, we train LLM reviewer agents capable of structured reasoning using a relevant-paper-aware training method. Furthermore, we construct ReviewAgents, a multi-role, multi-LLM agent review framework, to enhance the review comment generation process. Additionally, we propose ReviewBench, a benchmark for evaluating the review comments generated by LLMs. Our experimental results on ReviewBench demonstrate that while existing LLMs exhibit a certain degree of potential for automating the review process, there remains a gap when compared to human-generated reviews. Moreover, our ReviewAgents framework further narrows this gap, outperforming advanced LLMs in generating review comments.
Multimodal large language models (MLLMs) show promise in tasks like visual question answering (VQA) but still face challenges in multimodal reasoning. Recent works adapt agentic frameworks or chain-of-thought (CoT) reasoning to improve performance. However, CoT-based multimodal reasoning often demands costly data annotation and fine-tuning, while agentic approaches relying on external tools risk introducing unreliable output from these tools. In this paper, we propose Seeing and Reasoning with Confidence (SRICE), a training-free multimodal reasoning framework that integrates external vision models with uncertainty quantification (UQ) into an MLLM to address these challenges. Specifically, SRICE guides the inference process by allowing MLLM to autonomously select regions of interest through multi-stage interactions with the help of external tools. We propose to use a conformal prediction-based approach to calibrate the output of external tools and select the optimal tool by estimating the uncertainty of an MLLM's output. Our experiment shows that the average improvement of SRICE over the base MLLM is 4.6% on five datasets and the performance on some datasets even outperforms fine-tuning-based methods, revealing the significance of ensuring reliable tool use in an MLLM agent.
The emergence of large language models (LLMs) opens new frontiers for unmanned aerial vehicle (UAVs), yet existing systems remain confined to predefined tasks due to hardware-software co-design challenges. This paper presents the first aerial intelligent agent capable of open-world task execution through tight integration of LLM-based reasoning and robotic autonomy. Our hardware-software co-designed system addresses two fundamental limitations: (1) Onboard LLM operation via an edge-optimized computing platform, achieving 5-6 tokens/sec inference for 14B-parameter models at 220W peak power; (2) A bidirectional cognitive architecture that synergizes slow deliberative planning (LLM task planning) with fast reactive control (state estimation, mapping, obstacle avoidance, and motion planning). Validated through preliminary results using our prototype, the system demonstrates reliable task planning and scene understanding in communication-constrained environments, such as sugarcane monitoring, power grid inspection, mine tunnel exploration, and biological observation applications. This work establishes a novel framework for embodied aerial artificial intelligence, bridging the gap between task planning and robotic autonomy in open environments.
Traditional Reinforcement Learning (RL) suffers from replicating human-like behaviors, generalizing effectively in multi-agent scenarios, and overcoming inherent interpretability issues.These tasks are compounded when deep environment understanding, agent coordination and dynamic optimization are required. While Large Language Model (LLM) enhanced methods have shown promise in generalization and interoperability, they often neglect necessary multi-agent coordination. Therefore, we introduce the Cascading Cooperative Multi-agent (CCMA) framework, integrating RL for individual interactions, a fine-tuned LLM for regional cooperation, a reward function for global optimization, and the Retrieval-augmented Generation mechanism to dynamically optimize decision-making across complex driving scenarios. Our experiments demonstrate that the CCMA outperforms existing RL methods, demonstrating significant improvements in both micro and macro-level performance in complex driving environments.
Recent advances in LLM-based role-playing language agents (RPLAs) have attracted broad attention in various applications. While chain-of-thought reasoning has shown importance in many tasks for LLMs, the internal thinking processes of RPLAs remain unexplored. Understanding characters' inner thoughts is crucial for developing advanced RPLAs. In this paper, we introduce ROLETHINK, a novel benchmark constructed from literature for evaluating character thought generation. We propose the task of inner thought reasoning, which includes two sets: the gold set that compares generated thoughts with original character monologues, and the silver set that uses expert synthesized character analyses as references. To address this challenge, we propose MIRROR, a chain-of-thought approach that generates character thoughts by retrieving memories, predicting character reactions, and synthesizing motivations. Through extensive experiments, we demonstrate the importance of inner thought reasoning for RPLAs, and MIRROR consistently outperforms existing methods. Resources are available at https://github.com/airaer1998/RPA_Thought.
LLM-based Multi-Agent Systems (MAS) have proven highly effective in solving complex problems by integrating multiple agents, each performing different roles. However, in sensitive domains, they face emerging privacy protection challenges. In this paper, we introduce the concept of Federated MAS, highlighting the fundamental differences between Federated MAS and traditional FL. We then identify key challenges in developing Federated MAS, including: 1) heterogeneous privacy protocols among agents, 2) structural differences in multi-party conversations, and 3) dynamic conversational network structures. To address these challenges, we propose Embedded Privacy-Enhancing Agents (EPEAgent), an innovative solution that integrates seamlessly into the Retrieval-Augmented Generation (RAG) phase and the context retrieval stage. This solution minimizes data flows, ensuring that only task-relevant, agent-specific information is shared. Additionally, we design and generate a comprehensive dataset to evaluate the proposed paradigm. Extensive experiments demonstrate that EPEAgent effectively enhances privacy protection while maintaining strong system performance. The code will be availiable at https://github.com/ZitongShi/EPEAgent
In this work, we implement music production for silent film clips using LLM-driven method. Given the strong professional demands of film music production, we propose the FilmComposer, simulating the actual workflows of professional musicians. FilmComposer is the first to combine large generative models with a multi-agent approach, leveraging the advantages of both waveform music and symbolic music generation. Additionally, FilmComposer is the first to focus on the three core elements of music production for film-audio quality, musicality, and musical development-and introduces various controls, such as rhythm, semantics, and visuals, to enhance these key aspects. Specifically, FilmComposer consists of the visual processing module, rhythm-controllable MusicGen, and multi-agent assessment, arrangement and mix. In addition, our framework can seamlessly integrate into the actual music production pipeline and allows user intervention in every step, providing strong interactivity and a high degree of creative freedom. Furthermore, we propose MusicPro-7k which includes 7,418 film clips, music, description, rhythm spots and main melody, considering the lack of a professional and high-quality film music dataset. Finally, both the standard metrics and the new specialized metrics we propose demonstrate that the music generated by our model achieves state-of-the-art performance in terms of quality, consistency with video, diversity, musicality, and musical development. Project page: https://apple-jun.github.io/FilmComposer.github.io/
With the advent of 6G systems, emerging hyper-connected ecosystems necessitate agile and adaptive medium access control (MAC) protocols to contend with network dynamics and diverse service requirements. We propose LLM4MAC, a novel framework that harnesses large language models (LLMs) within a reinforcement learning paradigm to drive MAC protocol emergence. By reformulating uplink data transmission scheduling as a semantics-generalized partially observable Markov game (POMG), LLM4MAC encodes network operations in natural language, while proximal policy optimization (PPO) ensures continuous alignment with the evolving network dynamics. A structured identity embedding (SIE) mechanism further enables robust coordination among heterogeneous agents. Extensive simulations demonstrate that on top of a compact LLM, which is purposefully selected to balance performance with resource efficiency, the protocol emerging from LLM4MAC outperforms comparative baselines in throughput and generalization.
Human interaction with the external world fundamentally involves the exchange of personal memory, whether with other individuals, websites, applications, or, in the future, AI agents. A significant portion of this interaction is redundant, requiring users to repeatedly provide the same information across different contexts. Existing solutions, such as browser-stored credentials, autofill mechanisms, and unified authentication systems, have aimed to mitigate this redundancy by serving as intermediaries that store and retrieve commonly used user data. The advent of large language models (LLMs) presents an opportunity to redefine memory management through an AI-native paradigm: SECOND ME. SECOND ME acts as an intelligent, persistent memory offload system that retains, organizes, and dynamically utilizes user-specific knowledge. By serving as an intermediary in user interactions, it can autonomously generate context-aware responses, prefill required information, and facilitate seamless communication with external systems, significantly reducing cognitive load and interaction friction. Unlike traditional memory storage solutions, SECOND ME extends beyond static data retention by leveraging LLM-based memory parameterization. This enables structured organization, contextual reasoning, and adaptive knowledge retrieval, facilitating a more systematic and intelligent approach to memory management. As AI-driven personal agents like SECOND ME become increasingly integrated into digital ecosystems, SECOND ME further represents a critical step toward augmenting human-world interaction with persistent, contextually aware, and self-optimizing memory systems. We have open-sourced the fully localizable deployment system at GitHub: https://github.com/Mindverse/Second-Me.
Large Language Models (LLMs) have made significant progress in open-ended dialogue, yet their inability to retain and retrieve relevant information from long-term interactions limits their effectiveness in applications requiring sustained personalization. External memory mechanisms have been proposed to address this limitation, enabling LLMs to maintain conversational continuity. However, existing approaches struggle with two key challenges. First, rigid memory granularity fails to capture the natural semantic structure of conversations, leading to fragmented and incomplete representations. Second, fixed retrieval mechanisms cannot adapt to diverse dialogue contexts and user interaction patterns. In this work, we propose Reflective Memory Management (RMM), a novel mechanism for long-term dialogue agents, integrating forward- and backward-looking reflections: (1) Prospective Reflection, which dynamically summarizes interactions across granularities-utterances, turns, and sessions-into a personalized memory bank for effective future retrieval, and (2) Retrospective Reflection, which iteratively refines the retrieval in an online reinforcement learning (RL) manner based on LLMs' cited evidence. Experiments show that RMM demonstrates consistent improvement across various metrics and benchmarks. For example, RMM shows more than 10% accuracy improvement over the baseline without memory management on the LongMemEval dataset.
Large language models (LLMs) have exhibited the ability to effectively utilize external tools to address user queries. However, their performance may be limited in complex, multi-turn interactions involving users and multiple tools. To address this, we propose Magnet, a principled framework for synthesizing high-quality training trajectories to enhance the function calling capability of large language model agents in multi-turn conversations with humans. The framework is based on automatic and iterative translations from a function signature path to a sequence of queries and executable function calls. We model the complicated function interactions in multi-turn cases with graph and design novel node operations to build reliable signature paths. Motivated by context distillation, when guiding the generation of positive and negative trajectories using a teacher model, we provide reference function call sequences as positive hints in context and contrastive, incorrect function calls as negative hints. Experiments show that training with the positive trajectories with supervised fine-tuning and preference optimization against negative trajectories, our 14B model, Magnet-14B-mDPO, obtains 68.01 on BFCL-v3 and 73.30 on ToolQuery, surpassing the performance of the teacher model Gemini-1.5-pro-002 by a large margin in function calling.
Program synthesis with Large Language Models (LLMs) suffers from a "near-miss syndrome": the generated code closely resembles a correct solution but fails unit tests due to minor errors. We address this with a multi-agent framework called Synthesize, Execute, Instruct, Debug, and Repair (SEIDR). Effectively applying SEIDR to instruction-tuned LLMs requires determining (a) optimal prompts for LLMs, (b) what ranking algorithm selects the best programs in debugging rounds, and (c) balancing the repair of unsuccessful programs with the generation of new ones. We empirically explore these trade-offs by comparing replace-focused, repair-focused, and hybrid debug strategies. We also evaluate lexicase and tournament selection to rank candidates in each generation. On Program Synthesis Benchmark 2 (PSB2), our framework outperforms both conventional use of OpenAI Codex without a repair phase and traditional genetic programming approaches. SEIDR outperforms the use of an LLM alone, solving 18 problems in C++ and 20 in Python on PSB2 at least once across experiments. To assess generalizability, we employ GPT-3.5 and Llama 3 on the PSB2 and HumanEval-X benchmarks. Although SEIDR with these models does not surpass current state-of-the-art methods on the Python benchmarks, the results on HumanEval-C++ are promising. SEIDR with Llama 3-8B achieves an average pass@100 of 84.2%. Across all SEIDR runs, 163 of 164 problems are solved at least once with GPT-3.5 in HumanEval-C++, and 162 of 164 with the smaller Llama 3-8B. We conclude that SEIDR effectively overcomes the near-miss syndrome in program synthesis with LLMs.
Large Language Models (LLMs) have shown impressive performance on existing medical question-answering benchmarks. This high performance makes it increasingly difficult to meaningfully evaluate and differentiate advanced methods. We present MedAgentsBench, a benchmark that focuses on challenging medical questions requiring multi-step clinical reasoning, diagnosis formulation, and treatment planning-scenarios where current models still struggle despite their strong performance on standard tests. Drawing from seven established medical datasets, our benchmark addresses three key limitations in existing evaluations: (1) the prevalence of straightforward questions where even base models achieve high performance, (2) inconsistent sampling and evaluation protocols across studies, and (3) lack of systematic analysis of the interplay between performance, cost, and inference time. Through experiments with various base models and reasoning methods, we demonstrate that the latest thinking models, DeepSeek R1 and OpenAI o3, exhibit exceptional performance in complex medical reasoning tasks. Additionally, advanced search-based agent methods offer promising performance-to-cost ratios compared to traditional approaches. Our analysis reveals substantial performance gaps between model families on complex questions and identifies optimal model selections for different computational constraints. Our benchmark and evaluation framework are publicly available at https://github.com/gersteinlab/medagents-benchmark.
It has been frequently observed that human speakers align their language use with each other during conversations. In this paper, we study empirically whether large language models (LLMs) exhibit the same behavior of conversational adaptation. We construct a corpus of conversations between LLMs and find that two LLM agents end up making more similar syntactic choices as conversations go on, confirming that modern LLMs adapt their language use to their conversational partners in at least a rudimentary way.
Large Language Models (LLMs) have demonstrated strong generalizable reasoning and planning capabilities. However, their efficacies in spatial path planning and obstacle-free trajectory generation remain underexplored. Leveraging LLMs for navigation holds significant potential, given LLMs' ability to handle unseen scenarios, support user-agent interactions, and provide global control across complex systems, making them well-suited for agentic planning and humanoid motion generation. As one of the first studies in this domain, we explore the zero-shot navigation and path generation capabilities of LLMs by constructing a dataset and proposing an evaluation protocol. Specifically, we represent paths using anchor points connected by straight lines, enabling movement in various directions. This approach offers greater flexibility and practicality compared to previous methods while remaining simple and intuitive for LLMs. We demonstrate that, when tasks are well-structured in this manner, modern LLMs exhibit substantial planning proficiency in avoiding obstacles while autonomously refining navigation with the generated motion to reach the target. Further, this spatial reasoning ability of a single LLM motion agent interacting in a static environment can be seamlessly generalized in multi-motion agents coordination in dynamic environments. Unlike traditional approaches that rely on single-step planning or local policies, our training-free LLM-based method enables global, dynamic, closed-loop planning, and autonomously resolving collision issues.
With the rise of large language models (LLMs), AI agents as autonomous decision-makers present significant opportunities and challenges for human-AI cooperation. While many studies have explored human cooperation with AI as tools, the role of LLM-augmented autonomous agents in competitive-cooperative interactions remains under-examined. This study investigates human cooperative behavior by engaging 30 participants who interacted with LLM agents exhibiting different characteristics (purported human, purported rule-based AI agent, and LLM agent) in repeated Prisoner's Dilemma games. Findings show significant differences in cooperative behavior based on the agents' purported characteristics and the interaction effect of participants' genders and purported characteristics. We also analyzed human response patterns, including game completion time, proactive favorable behavior, and acceptance of repair efforts. These insights offer a new perspective on human interactions with LLM agents in competitive cooperation contexts, such as virtual avatars or future physical entities. The study underscores the importance of understanding human biases toward AI agents and how observed behaviors can influence future human-AI cooperation dynamics.
Existing long-form video generation frameworks lack automated planning, requiring manual input for storylines, scenes, cinematography, and character interactions, resulting in high costs and inefficiencies. To address these challenges, we present MovieAgent, an automated movie generation via multi-agent Chain of Thought (CoT) planning. MovieAgent offers two key advantages: 1) We firstly explore and define the paradigm of automated movie/long-video generation. Given a script and character bank, our MovieAgent can generates multi-scene, multi-shot long-form videos with a coherent narrative, while ensuring character consistency, synchronized subtitles, and stable audio throughout the film. 2) MovieAgent introduces a hierarchical CoT-based reasoning process to automatically structure scenes, camera settings, and cinematography, significantly reducing human effort. By employing multiple LLM agents to simulate the roles of a director, screenwriter, storyboard artist, and location manager, MovieAgent streamlines the production pipeline. Experiments demonstrate that MovieAgent achieves new state-of-the-art results in script faithfulness, character consistency, and narrative coherence. Our hierarchical framework takes a step forward and provides new insights into fully automated movie generation. The code and project website are available at: https://github.com/showlab/MovieAgent and https://weijiawu.github.io/MovieAgent.
Data Science tasks are multifaceted, dynamic, and often domain-specific. Existing LLM-based approaches largely concentrate on isolated phases, neglecting the interdependent nature of many data science tasks and limiting their capacity for comprehensive end-to-end support. We propose DatawiseAgent, a notebook-centric LLM agent framework that unifies interactions among user, agent and the computational environment through markdown and executable code cells, supporting flexible and adaptive automated data science. Built on a Finite State Transducer(FST), DatawiseAgent orchestrates four stages, including DSF-like planning, incremental execution, self-debugging, and post-filtering. Specifically, the DFS-like planning stage systematically explores the solution space, while incremental execution harnesses real-time feedback and accommodates LLM's limited capabilities to progressively complete tasks. The self-debugging and post-filtering modules further enhance reliability by diagnosing and correcting errors and pruning extraneous information. Extensive experiments on diverse tasks, including data analysis, visualization, and data modeling, show that DatawiseAgent consistently outperforms or matches state-of-the-art methods across multiple model settings. These results highlight its potential to generalize across data science scenarios and lay the groundwork for more efficient, fully automated workflows.
Partial perception deficits can compromise autonomous vehicle safety by disrupting environmental understanding. Current protocols typically respond with immediate stops or minimal-risk maneuvers, worsening traffic flow and lacking flexibility for rare driving scenarios. In this paper, we propose LLM-RCO, a framework leveraging large language models to integrate human-like driving commonsense into autonomous systems facing perception deficits. LLM-RCO features four key modules: hazard inference, short-term motion planner, action condition verifier, and safety constraint generator. These modules interact with the dynamic driving environment, enabling proactive and context-aware control actions to override the original control policy of autonomous agents. To improve safety in such challenging conditions, we construct DriveLM-Deficit, a dataset of 53,895 video clips featuring deficits of safety-critical objects, complete with annotations for LLM-based hazard inference and motion planning fine-tuning. Extensive experiments in adverse driving conditions with the CARLA simulator demonstrate that systems equipped with LLM-RCO significantly improve driving performance, highlighting its potential for enhancing autonomous driving resilience against adverse perception deficits. Our results also show that LLMs fine-tuned with DriveLM-Deficit can enable more proactive movements instead of conservative stops in the context of perception deficits.