Shunyu Yao’s research while affiliated with Princeton University and other places

Large language model (LLM) agents have been applied to sequential decision-making tasks such as web navigation, but without any environment-specific experiences, they often fail in these complex tasks. Moreover, current LLM agents are not designed to continually learn from past experiences during inference time, which could be crucial for them to gain these environment-specific experiences. To address this, we propose Contextual Experience Replay (CER), a training-free framework to enable efficient self-improvement for language agents in their context window. Specifically, CER accumulates and synthesizes past experiences into a dynamic memory buffer. These experiences encompass environment dynamics and common decision-making patterns, allowing the agents to retrieve and augment themselves with relevant knowledge in new tasks, enhancing their adaptability in complex environments. We evaluate CER on the challenging WebArena and VisualWebArena benchmarks. On VisualWebArena, CER achieves a competitive performance of 31.9%. On WebArena, CER also gets a competitive average success rate of 36.7%, relatively improving the success rate of the GPT-4o agent baseline by 51.0%. We also conduct a comprehensive analysis on it to prove its efficiency, validity and understand it better.

Recent advancements in AI reasoning have driven substantial improvements across diverse tasks. A critical open question is whether these improvements also yields better knowledge transfer: the ability of models to communicate reasoning in ways humans can understand, apply, and learn from. To investigate this, we introduce Knowledge Integration and Transfer Evaluation (KITE), a conceptual and experimental framework for Human-AI knowledge transfer capabilities and conduct the first large-scale human study (N=118) explicitly designed to measure it. In our two-phase setup, humans first ideate with an AI on problem-solving strategies, then independently implement solutions, isolating model explanations' influence on human understanding. Our findings reveal that although model benchmark performance correlates with collaborative outcomes, this relationship is notably inconsistent, featuring significant outliers, indicating that knowledge transfer requires dedicated optimization. Our analysis identifies behavioral and strategic factors mediating successful knowledge transfer. We release our code, dataset, and evaluation framework to support future work on communicatively aligned models.

The widespread adoption of large language models (LLMs) makes it important to recognize their strengths and limitations. We argue that to develop a holistic understanding of these systems, we must consider the problem that they were trained to solve: next-word prediction over Internet text. By recognizing the pressures that this task exerts, we can make predictions about the strategies that LLMs will adopt, allowing us to reason about when they will succeed or fail. Using this approach—which we call the teleological approach—we identify three factors that we hypothesize will influence LLM accuracy: the probability of the task to be performed, the probability of the target output, and the probability of the provided input. To test our predictions, we evaluate five LLMs (GPT-3.5, GPT-4, Claude 3, Llama 3, and Gemini 1.0) on 11 tasks, and we find robust evidence that LLMs are influenced by probability in the hypothesized ways. Many of the experiments reveal surprising failure modes. For instance, GPT-4’s accuracy at decoding a simple cipher is 51% when the output is a high-probability sentence but only 13% when it is low-probability, even though this task is a deterministic one for which probability should not matter. These results show that AI practitioners should be careful about using LLMs in low-probability situations. More broadly, we conclude that we should not evaluate LLMs as if they are humans but should instead treat them as a distinct type of system—one that has been shaped by its own particular set of pressures.

In "Embers of Autoregression" (McCoy et al., 2023), we showed that several large language models (LLMs) have some important limitations that are attributable to their origins in next-word prediction. Here we investigate whether these issues persist with o1, a new system from OpenAI that differs from previous LLMs in that it is optimized for reasoning. We find that o1 substantially outperforms previous LLMs in many cases, with particularly large improvements on rare variants of common tasks (e.g., forming acronyms from the second letter of each word in a list, rather than the first letter). Despite these quantitative improvements, however, o1 still displays the same qualitative trends that we observed in previous systems. Specifically, o1 - like previous LLMs - is sensitive to the probability of examples and tasks, performing better and requiring fewer "thinking tokens" in high-probability settings than in low-probability ones. These results show that optimizing a language model for reasoning can mitigate but might not fully overcome the language model's probability sensitivity.

Existing benchmarks do not test language agents on their interaction with human users or ability to follow domain-specific rules, both of which are vital for deploying them in real world applications. We propose $\tau$-bench, a benchmark emulating dynamic conversations between a user (simulated by language models) and a language agent provided with domain-specific API tools and policy guidelines. We employ an efficient and faithful evaluation process that compares the database state at the end of a conversation with the annotated goal state. We also propose a new metric (pass^k) to evaluate the reliability of agent behavior over multiple trials. Our experiments show that even state-of-the-art function calling agents (like gpt-4o) succeed on <50% of the tasks, and are quite inconsistent (pass^8 <25% in retail). Our findings point to the need for methods that can improve the ability of agents to act consistently and follow rules reliably.

Software engineering is a challenging task requiring proficiency in both code generation and interacting with computers. In this paper, we introduce SWE-agent, an autonomous system that uses a language model to interact with a computer to solve software engineering tasks. We show that a custom-built agent-computer interface (ACI) greatly enhances the ability of an agent to create and edit code files, navigate entire repositories and execute programs. On SWE-bench, SWE-agent is able to solve 12.5% of issues, compared to the previous best of 3.8% achieved with retrieval-augmented generation (RAG). We explore how ACI design impacts an agent's behavior and performance, and provide insights on effective design.

Amidst the advent of language models (LMs) and their wide-ranging capabilities, concerns have been raised about their implications with regards to privacy and security. In particular, the emergence of language agents as a promising aid for automating and augmenting digital work poses immediate questions concerning their misuse as malicious cybersecurity actors. With their exceptional compute efficiency and execution speed relative to human counterparts, language agents may be extremely adept at locating vulnerabilities, performing complex social engineering, and hacking real world systems. Understanding and guiding the development of language agents in the cybersecurity space requires a grounded understanding of their capabilities founded on empirical data and demonstrations. To address this need, we introduce InterCode-CTF, a novel task environment and benchmark for evaluating language agents on the Capture the Flag (CTF) task. Built as a facsimile of real world CTF competitions, in the InterCode-CTF environment, a language agent is tasked with finding a flag from a purposely-vulnerable computer program. We manually collect and verify a benchmark of 100 task instances that require a number of cybersecurity skills such as reverse engineering, forensics, and binary exploitation, then evaluate current top-notch LMs on this evaluation set. Our preliminary findings indicate that while language agents possess rudimentary cybersecurity knowledge, they are not able to perform multi-step cybersecurity tasks out-of-the-box.

Recent efforts have incorporated large language models (LLMs) with external resources (e.g., the Internet) or internal control flows (e.g., prompt chaining) for tasks requiring grounding or reasoning. However, these efforts have largely been piecemeal, lacking a systematic framework for constructing a fully-fledged language agent. To address this challenge, we draw on the rich history of agent design in symbolic artificial intelligence to develop a blueprint for a new wave of cognitive language agents. We first show that LLMs have many of the same properties as production systems, and recent efforts to improve their grounding or reasoning mirror the development of cognitive architectures built around production systems. We then propose Cognitive Architectures for Language Agents (CoALA), a conceptual framework to systematize diverse methods for LLM-based reasoning, grounding, learning, and decision making as instantiations of language agents in the framework. Finally, we use the CoALA framework to highlight gaps and propose actionable directions toward more capable language agents in the future.

Text generation under constraints have seen increasing interests in natural language processing, especially with the rapidly improving capabilities of large language models. However, existing benchmarks for constrained generation usually focus on fixed constraint types (e.g.,generate a sentence containing certain words) that have proved to be easy for state-of-the-art models like GPT-4. We present COLLIE, a grammar-based framework that allows the specification of rich, compositional constraints with diverse generation levels (word, sentence, paragraph, passage) and modeling challenges (e.g.,language understanding, logical reasoning, counting, semantic planning). We also develop tools for automatic extraction of task instances given a constraint structure and a raw text corpus. Using COLLIE, we compile the COLLIE-v1 dataset with 2080 instances comprising 13 constraint structures. We perform systematic experiments across five state-of-the-art instruction-tuned language models and analyze their performances to reveal shortcomings. COLLIE is designed to be extensible and lightweight, and we hope the community finds it useful to develop more complex constraints and evaluations in the future.