ICLR 2025 论文列表

Recent advances in large language models (LLMs) demonstrate substantial capabilities in natural language understanding and generation tasks. With the growing number of LLMs, how to harness the collective expertise of multiple LLMs is an exciting open direction. Toward this goal, we propose a new approach that leverages the collective strengths of multiple LLMs through a Mixture-of-Agents (MoA) methodology. In our approach, we construct a layered MoA architecture wherein each layer comprises multiple LLM agents. Each agent takes all the outputs from agents in the previous layer as auxiliary information in generating its response. MoA models achieves state-of-art performance on AlpacaEval 2.0, Arena-Hard, MT-Bench, and FLASK, surpassing GPT-4 Omni. For example, our MoA using only open-source LLMs achieves a score of 65.1% on AlpacaEval 2.0 compared to 57.5% by GPT-4 Omni.

Traditional reinforcement learning and planning require a lot of data and training to develop effective strategies. On the other hand, large language models (LLMs) can generalize well and perform tasks without prior training but struggle with complex planning and decision-making. We introduce **STRATEGIST**, a new approach that combines the strengths of both methods. It uses LLMs to generate and update high-level strategies in text form, while a Monte Carlo Tree Search (MCTS) algorithm refines and executes them. STRATEGIST is a general framework that optimizes strategies through self-play simulations without requiring any training data. We test STRATEGIST in competitive, multi-turn games with partial information, such as **Game of Pure Strategy (GOPS)** and **The Resistance: Avalon**, a multi-agent hidden-identity discussion game. Our results show that STRATEGIST-based agents outperform traditional reinforcement learning models, other LLM-based methods, and existing LLM agents while achieving performance levels comparable to human players.

Large Language Models (LLMs) and Vision Language Models (VLMs) possess extensive knowledge and exhibit promising reasoning abilities, however, they still struggle to perform well in complex, dynamic environments. Real-world tasks require handling intricate interactions, advanced spatial reasoning, long-term planning, and continuous exploration of new strategies—areas in which we lack effective methodologies for comprehensively evaluating these capabilities. To address this gap, we introduce BALROG, a novel benchmark designed to assess the agentic capabilities of LLMs and VLMs through a diverse set of challenging games. Our benchmark incorporates a range of existing reinforcement learning environments with varying levels of difficulty, including tasks that are solvable by non-expert humans in seconds to extremely challenging ones that may take years to master (e.g., the NetHack Learning Environment). We devise fine-grained metrics to measure performance and conduct an extensive evaluation of several popular open-source and closed-source LLMs and VLMs. Our findings indicate that while current models achieve partial success in the easier games, they struggle significantly with more challenging tasks. Notably, we observe severe deficiencies in vision-based decision-making, as several models perform worse when visual representations of the environments are provided. We release BALROG as an open and user-friendly benchmark to facilitate future research and development in the agentic community. Code and Leaderboard at balrogai.com

As large language models (LLMs) start interacting with each other and generating an increasing amount of text online, it becomes crucial to better understand how information is transformed as it passes from one LLM to the next. While significant research has examined individual LLM behaviors, existing studies have largely overlooked the collective behaviors and information distortions arising from iterated LLM interactions. Small biases, negligible at the single output level, risk being amplified in iterated interactions, potentially leading the content to evolve towards attractor states. In a series of _telephone game experiments_, we apply a transmission chain design borrowed from the human cultural evolution literature: LLM agents iteratively receive, produce, and transmit texts from the previous to the next agent in the chain. By tracking the evolution of text _toxicity_, _positivity_, _difficulty_, and _length_ across transmission chains, we uncover the existence of biases and attractors, and study their dependence on the initial text, the instructions, language model, and model size. For instance, we find that more open-ended instructions lead to stronger attraction effects compared to more constrained tasks. We also find that different text properties display different sensitivity to attraction effects, with _toxicity_ leading to stronger attractors than _length_. These findings highlight the importance of accounting for multi-step transmission dynamics and represent a first step towards a more comprehensive understanding of LLM cultural dynamics.

Selecting high-quality pre-training data for large language models (LLMs) is crucial for enhancing their overall performance under limited computation budget, improving both training and sample efficiency. Recent advancements in file selection primarily rely on using an existing or trained proxy model to assess the similarity of samples to a target domain, such as high quality sources BookCorpus and Wikipedia. However, upon revisiting these methods, the domain-similarity selection criteria demonstrates a diversity dilemma, i.e. dimensional collapse in the feature space, improving performance on the domain-related tasks but causing severe degradation on generic performance.To prevent collapse and enhance diversity, we propose a DiverSified File selection algorithm (DiSF), which selects the most decorrelated text files in the feature space. We approach this with a classical greedy algorithm to achieve more uniform eigenvalues in the feature covariance matrix of the selected texts, analyzing its approximation to the optimal solution under a formulation of $\gamma$-weakly submodular optimization problem. Empirically, we establish a benchmark and conduct extensive experiments on the TinyLlama architecture with models from 120M to 1.1B parameters. Evaluating across nine tasks from the Harness framework, DiSF demonstrates a significant improvement on overall performance. Specifically, DiSF saves 98.5\% of 590M training files in SlimPajama, outperforming the full-data pre-training within a 50B training budget, and achieving about 1.5x training efficiency and 5x data efficiency. Source code is available at: https://github.com/MediaBrain-SJTU/DiSF.git.

Embodied AI represents systems where AI is integrated into physical entities. Multimodal Large Language Model (LLM), which exhibits powerful language understanding abilities, has been extensively employed in embodied AI by facilitating sophisticated task planning. However, a critical safety issue remains overlooked: could these embodied LLMs perpetrate harmful behaviors? In response, we introduce BadRobot, the first attack paradigm designed to jailbreak robotic manipulation, making embodied LLMs violate safety and ethical constraints through typical voice-based user-system interactions. Specifically, three vulnerabilities are exploited to achieve this type of attack: (i) manipulation of LLMs within robotic systems, (ii) misalignment between linguistic outputs and physical actions, and (iii) unintentional hazardous behaviors caused by world knowledge's flaws. Furthermore, we construct a benchmark of various malicious physical action queries to evaluate BadRobot's attack performance. Based on this benchmark, extensive experiments against existing prominent embodied LLM frameworks (e.g., Voxposer, Code as Policies, and ProgPrompt) demonstrate the effectiveness of our BadRobot. We emphasize that addressing this emerging vulnerability is crucial for the secure deployment of LLMs in robotics. Warning: This paper contains harmful AI-generated language and aggressive actions.

The rapidly growing number and variety of Large Language Models (LLMs) present significant challenges in efficiently selecting the appropriate LLM for a given query, especially considering the trade-offs between performance and computational cost. Current LLM selection methods often struggle to generalize across new LLMs and different tasks because of their limited ability to leverage contextual interactions among tasks, queries, and LLMs, as well as their depen- dence on a transductive learning framework. To address these shortcomings, we introduce a novel inductive graph framework, named as GraphRouter, which fully utilizes the contextual information among tasks, queries, and LLMs to en- hance the LLM selection process. GraphRouter constructs a heterogeneous graph comprising task, query, and LLM nodes, with interactions represented as edges, which efficiently captures the contextual information between the query’s requirements and the LLM’s capabilities. Through an innovative edge prediction mechanism, GraphRouter is able to predict attributes (the effect and cost of LLM response) of potential edges, allowing for optimized recommendations that adapt to both existing and newly introduced LLMs without requiring retraining. Comprehensive experiments across three distinct effect-cost weight scenarios have shown that GraphRouter substantially surpasses existing routers, delivering a minimum performance improvement of 12.3%. In addition, it achieves enhanced generalization across new LLMs settings and supports diverse tasks with at least a 9.5% boost in effect and a significant reduction in computational demands. This work endeavors to apply a graph-based approach for the contextual and adaptive selection of LLMs, offering insights for real-world applications.

Large language models (LLMs) have been used to generate formal proofs of mathematical theorems in proofs assistants such as Lean. However, we often want to optimize a formal proof with respect to various criteria, depending on its downstream use. For example, we may want a proof to adhere to a certain style, be declaratively structured, or concise. Having suitably optimized proofs is also important for learning tasks, especially since human-written proofs may not optimal for that purpose. To this end, we study a new problem of automated proof optimization: rewriting a proof so that it is correct and optimizes for an arbitrary criterion, such as length or declarativity. As a first method for automated proof optimization, we present ImProver, a large-language-model agent that rewrites proofs to optimize arbitrary user-defined metrics in Lean. We find that naively applying LLMs to proof optimization falls short, and we incorporate various improvements into ImProver, such as the use of symbolic Lean context in a novel Chain-of-States technique, as well as error-correction and retrieval. We test ImProver on rewriting real-world undergraduate, competition, and research-level mathematics theorems, finding that ImProver is capable of rewriting proofs so that they are substantially shorter and more declarative in structure.

Large language models (LLMs) have demonstrated exceptional reasoning capabilities, enabling them to solve various complex problems. Recently, this ability has been applied to the paradigm of tool learning. Tool learning involves providing examples of tool usage and their corresponding functions, allowing LLMs to formulate plans and demonstrate the process of invoking and executing each tool. LLMs can address tasks that they cannot complete independently, thereby enhancing their potential across different tasks. However, this approach faces two key challenges. First, redundant error correction leads to unstable planning and long execution time. Additionally, designing a correct plan among multiple tools is also a challenge in tool learning. To address these issues, we propose Tool-Planner, a task-processing framework based on toolkits. Tool-Planner groups tools based on the API functions with the same function into a toolkit and allows LLMs to implement planning across the various toolkits. When a tool error occurs, the language model can reselect and adjust tools based on the toolkit. Experiments show that our approach demonstrates a high pass and win rate across different datasets and optimizes the planning scheme for tool learning in models such as GPT-4 and Claude 3, showcasing the potential of our method. Our code is public at https://github.com/OceannTwT/Tool-Planner.

Collaboration is a key challenge in distributed multi-agent reinforcement learning (MARL) environments. Learning frameworks for these decentralized systems must weigh the benefits of explicit player coordination against the communication overhead and computational cost of sharing local observations and environmental data. Quantum computing has sparked a potential synergy between quantum entanglement and cooperation in multi-agent environments, which could enable more efficient distributed collaboration with minimal information sharing. This relationship is largely unexplored, however, as current state-of-the-art quantum MARL (QMARL) implementations rely on classical information sharing rather than entanglement over a quantum channel as a coordination medium. In contrast, in this paper, a novel framework dubbed entangled QMARL (eQMARL) is proposed. The proposed eQMARL is a distributed actor-critic framework that facilitates cooperation over a quantum channel and eliminates local observation sharing via a quantum entangled split critic. Introducing a quantum critic uniquely spread across the agents allows coupling of local observation encoders through entangled input qubits over a quantum channel, which requires no explicit sharing of local observations and reduces classical communication overhead. Further, agent policies are tuned through joint observation-value function estimation via joint quantum measurements, thereby reducing the centralized computational burden. Experimental results show that eQMARL with $\Psi^{+}$ entanglement converges to a cooperative strategy up to $17.8\\%$ faster and with a higher overall score compared to split classical and fully centralized classical and quantum baselines. The results also show that eQMARL achieves this performance with a constant factor of $25$-times fewer centralized parameters compared to the split classical baseline.

Large language model (LLM) agents have shown great potential in solving real-world software engineering (SWE) problems. The most advanced open-source SWE agent can resolve over 27% of real GitHub issues in SWE-Bench Lite. However, these sophisticated agent frameworks exhibit varying strengths, excelling in certain tasks while underperforming in others. To fully harness the diversity of these agents, we propose DEI (Diversity Empowered Intelligence), a framework that leverages their unique expertise. DEI functions as a meta-module atop existing SWE agent frameworks, managing agent collectives for enhanced problem-solving. Experimental results show that a DEI-guided committee of agents is able to surpass the best individual agent's performance by a large margin. For instance, a group of open-source SWE agents, with a maximum individual resolve rate of 27.3% on SWE-Bench Lite, can achieve a 34.3% resolve rate with DEI, making a 25% improvement and beating most closed-source solutions. Our best-performing group excels with a 55% resolve rate, securing the highest ranking on SWE-Bench Lite. Our findings contribute to the growing body of research on collaborative AI systems and their potential to solve complex software engineering challenges.

Jailbreak attacks serve as essential red-teaming tools, proactively assessing whether LLMs can behave responsibly and safely in adversarial environments. Despite diverse strategies (e.g., cipher, low-resource language, persuasions, and so on) that have been proposed and shown success, these strategies are still manually designed, limiting their scope and effectiveness as a red-teaming tool. In this paper, we propose AutoDAN-Turbo, a black-box jailbreak method that can automatically discover as many jailbreak strategies as possible from scratch, without any human intervention or predefined scopes (e.g., specified candidate strategies), and use them for red-teaming. As a result, AutoDAN-Turbo can significantly outperform baseline methods, achieving a 74.3% higher average attack success rate on public benchmarks. Notably, AutoDAN-Turbo achieves an 88.5 attack success rate on GPT-4-1106-turbo. In addition, AutoDAN-Turbo is a unified framework that can incorporate existing human-designed jailbreak strategies in a plug-and-play manner. By integrating human-designed strategies, AutoDAN-Turbo can even achieve a higher attack success rate of 93.4 on GPT-4-1106-turbo.

General virtual agents need to handle multimodal observations, master complex action spaces, and self-improve in dynamic, open-domain environments. However, existing environments are often domain-specific and require complex setups, which limits agent development and evaluation in real-world settings. As a result, current evaluations lack in-depth analyses that decompose fundamental agent capabilities. We introduce AgentStudio, a trinity of environments, tools, and benchmarks to address these issues. AgentStudio provides a lightweight, interactive environment with highly generic observation and action spaces, e.g., video observations and GUI/API actions. It integrates tools for creating online benchmark tasks, annotating GUI elements, and labeling actions in videos. Based on our environment and tools, we curate an online task suite that benchmarks both GUI interactions and function calling with efficient auto-evaluation. We also reorganize existing datasets and collect new ones using our tools to establish three datasets: GroundUI, IDMBench, and CriticBench. These datasets evaluate fundamental agent abilities, including GUI grounding, learning from videos, and success detection, pointing to the desiderata for robust, general, and open-ended virtual agents.

The automatic design of embodied agents (e.g. robots) has existed for 31 years and is experiencing a renaissance of interest in the literature. To date however, the field has remained narrowly focused on two kinds of anatomically simple robots: (1) fully rigid, jointed bodies; and (2) fully soft, jointless bodies. Here we bridge these two extremes with the open ended creation of terrestrial endoskeletal robots: deformable soft bodies that leverage jointed internal skeletons to move efficiently across land. Simultaneous de novo generation of external and internal structures is achieved by (i) modeling 3D endoskeletal body plans as integrated collections of elastic and rigid cells that directly attach to form soft tissues anchored to compound rigid bodies; (ii) encoding these discrete mechanical subsystems into a continuous yet coherent latent embedding; (iii) optimizing the sensorimotor coordination of each decoded design using model-free reinforcement learning; and (iv) navigating this smooth yet highly non-convex latent manifold using evolutionary strategies. This yields an endless stream of novel species of ``higher robots'' that, like all higher animals, harness the mechanical advantages of both elastic tissues and skeletal levers for terrestrial travel. It also provides a plug-and-play experimental platform for benchmarking evolutionary design and representation learning algorithms in complex hierarchical embodied systems.

Go-Explore is a powerful family of algorithms designed to solve hard-exploration problems built on the principle of archiving discovered states, and iteratively returning to and exploring from the most promising states. This approach has led to superhuman performance across a wide variety of challenging problems including Atari games and robotic control, but requires manually designing heuristics to guide exploration (i.e., determine which states to save and explore from, and what actions to consider next), which is time-consuming and infeasible in general. To resolve this, we propose Intelligent Go-Explore (IGE) which greatly extends the scope of the original Go-Explore by replacing these handcrafted heuristics with the intelligence and internalized human notions of interestingness captured by giant pretrained foundation models (FMs). This provides IGE with a human-like ability to instinctively identify how interesting or promising any new state is (e.g., discovering new objects, locations, or behaviors), even in complex environments where heuristics are hard to define. Moreover, IGE offers the exciting opportunity to recognize and capitalize on serendipitous discoveries---states encountered during exploration that are valuable in terms of exploration, yet where what makes them interesting was not anticipated by the human user. We evaluate our algorithm on a diverse range of language and vision-based tasks that require search and exploration. Across these tasks, IGE strongly exceeds classic reinforcement learning and graph search baselines, and also succeeds where prior state-of-the-art FM agents like Reflexion completely fail. Overall, Intelligent Go-Explore combines the tremendous strengths of FMs and the powerful Go-Explore algorithm, opening up a new frontier of research into creating more generally capable agents with impressive exploration capabilities. All our code is open-sourced at: https://github.com/conglu1997/intelligent-go-explore.

Large Language Models (LLMs) are widely used for tasks such as natural language and code generation, but their outputs often suffer from issues like hallucination, toxicity, and incorrect results. Current libraries for structured LLM generation rely on left-to-right decoding without support for backtracking, limiting the ability to correct or refine outputs mid-generation. To address this, we introduce IterGen, a user-friendly library for iterative, grammar-guided LLM generation that enables users to move both forward and backward within the generated output based on grammar symbols. By leveraging a symbol-to-position mapping and maintaining the key-value (KV) cache state, IterGen ensures efficient and structured generation while allowing for corrections during the process. We demonstrate IterGen's effectiveness in two important applications: reducing privacy leakage in LLM outputs, improving the accuracy of LLM-generated SQL and Vega-Lite queries. Our code and additional resources are available at https://structuredllm.com.

A popular approach to building agents using Language Models (LMs) involves iteratively prompting the LM, reflecting on its outputs, and updating the input prompts until the desired task is achieved. However, our analysis reveals two key shortcomings in the existing methods: $(i)$ limited exploration of the decision space due to repetitive reflections, which result in redundant inputs, and $(ii)$ an inability to leverage insights from previously solved tasks. To address these issues, we introduce DoT (Diversity of Thoughts), a novel framework that a) explicitly reduces redundant reflections to enhance decision-space exploration, and b) incorporates a task-agnostic memory component to enable knowledge retrieval from previously solved tasks—unlike current approaches that operate in isolation for each task. Through extensive experiments on a suite of programming benchmarks (HumanEval, MBPP, and LeetCodeHardGym) using a variety of LMs, DoT demonstrates up to a $\textbf{10}$% improvement in Pass@1 while maintaining cost-effectiveness. Furthermore, DoT is modular by design. For instance, when the diverse reflection module of DoT is integrated with existing methods like Tree of Thoughts (ToT), we observe a significant $\textbf{13}$% improvement on Game of 24 (one of the main benchmarks of ToT), highlighting the broad applicability and impact of our contributions across various reasoning tasks.

Policy gradient methods have become a staple of any single-agent reinforcement learning toolbox, due to their combination of desirable properties: iterate convergence, efficient use of stochastic trajectory feedback, and theoretically-sound avoidance of importance sampling corrections. In multi-agent imperfect-information settings (extensive-form games), however, it is still unknown whether the same desiderata can be guaranteed while retaining theoretical guarantees. Instead, sound methods for extensive-form games rely on approximating \emph{counterfactual} values (as opposed to Q values), which are incompatible with policy gradient methodologies. In this paper, we investigate whether policy gradient can be safely used in two-player zero-sum imperfect-information extensive-form games (EFGs). We establish positive results, showing for the first time that a policy gradient method leads to provable best-iterate convergence to a regularized Nash equilibrium in self-play.

Large language models (LLMs) have achieved remarkable success across diverse tasks, yet their inference processes are hindered by substantial time and energy demands due to single-token generation at each decoding step. While previous methods such as speculative decoding mitigate these inefficiencies by producing multiple tokens per step, each token is still generated by its single-token distribution, thereby enhancing speed without improving effectiveness. In contrast, our work simultaneously enhances inference speed and improves the output effectiveness. We consider multi-token joint decoding (MTJD), which generates multiple tokens from their joint distribution at each iteration, theoretically reducing perplexity and enhancing task performance. However, MTJD suffers from the high cost of sampling from the joint distribution of multiple tokens. Inspired by speculative decoding, we introduce multi-token assisted decoding (MTAD), a novel framework designed to accelerate MTJD. MTAD leverages a smaller auxiliary model to approximate the joint distribution of a larger model, incorporating a verification mechanism that not only ensures the accuracy of this approximation, but also improves the decoding efficiency over conventional speculative decoding. Theoretically, we demonstrate that MTAD closely approximates exact MTJD with bounded error. Empirical evaluations using Llama-2 and OPT models ranging from 13B to 70B parameters across various tasks reveal that MTAD reduces perplexity by 21.2% and improves downstream performance compared to standard single-token sampling. Furthermore, MTAD achieves a 1.42× speed-up and consumes 1.54× less energy than conventional speculative decoding methods. These results highlight MTAD’s ability to make multi-token joint decoding both effective and efficient, promoting more sustainable and high-performance deployment of LLMs.