Agents Explore but Agents Ignore: LLMs Lack Environmental Curiosity
Study shows current LLM agents lack environmental curiosity, discovering solutions in 79-81% of cases but utilizing them in only 37-50%.
Key Findings
Methodology
The study employs solution injection, placing complete task solutions directly into environments to assess agents' environmental curiosity. Experiments were conducted on Terminal-Bench, SWE-Bench, and AppWorld to analyze whether agents can discover and utilize these solutions.
Key Results
- In Terminal-Bench, agents discovered solutions in 79-81% of runs but utilized them in only 37-50% of cases, highlighting a significant gap between discovery and utilization.
- In AppWorld, agents saw documentation stating a command 'returns the complete solution to this task' in over 90% of attempts but exploited this in fewer than 7% of trials.
- The study identifies three main factors influencing environmental curiosity: available tools, reasoning budget, and training data distribution.
Significance
This research highlights the lack of responsiveness in current LLM agents when faced with unexpected but relevant information, impacting both academia and industry. By identifying factors affecting environmental curiosity, the study offers new perspectives for improving agents' task execution capabilities.
Technical Contribution
The study introduces the concept of environmental curiosity and develops the solution injection method to evaluate agents' capabilities. This method provides new metrics, such as discovery@k and interaction@k, for assessing agents' reactions to unexpected information.
Novelty
This study is the first to systematically evaluate LLM agents' environmental curiosity, introducing the innovative solution injection method, which contrasts sharply with existing task success rate evaluations.
Limitations
- Current agents still ignore discovered solutions in the majority of cases, indicating the issue is not solely about inference-time configuration.
- The distribution of training data also limits agents' environmental curiosity, especially when fine-tuned on narrow distributions.
Future Work
Future research could explore how to enhance agents' environmental curiosity during training and how to better utilize discovered information during inference.
AI Executive Summary
Current LLM agents often need to explore their environments to gather relevant information when handling complex tasks. However, the study finds that these agents frequently lack the curiosity to fully exploit unexpected but relevant information.
The study employs solution injection, placing complete task solutions directly into environments to assess agents' environmental curiosity. Experimental results show that while agents can discover these solutions in most cases, they rarely utilize them effectively.
This finding reveals a critical deficiency in current agents' task execution and provides new directions for future research on enhancing agents' environmental curiosity to improve their task execution capabilities.
Deep Analysis
Background
Large language models (LLMs) often need to explore their environments to gather relevant information when handling complex tasks. However, existing research shows that these models frequently lack the curiosity to fully exploit unexpected but relevant information.
Core Problem
Current LLM agents lack the curiosity to fully exploit unexpected but relevant information, limiting their performance in complex tasks.
Innovation
The study introduces the concept of environmental curiosity and develops the solution injection method to evaluate agents' capabilities. This method provides new metrics, such as discovery@k and interaction@k, for assessing agents' reactions to unexpected information.
Methodology
- �� Employ solution injection, placing complete task solutions directly into environments.
- �� Conduct experiments on Terminal-Bench, SWE-Bench, and AppWorld.
- �� Assess whether agents can discover and utilize these solutions.
Experiments
Experiments were conducted on Terminal-Bench, SWE-Bench, and AppWorld to assess whether agents can discover and utilize injected solutions. Different tool configurations and reasoning budgets were used.
Results
Experimental results show that while agents can discover injected solutions in most cases, they rarely utilize them effectively. This finding reveals a critical deficiency in agents' task execution.
Applications
The study's findings can be used to improve LLM agents' performance in complex tasks, especially those requiring environmental exploration to gather relevant information.
Limitations & Outlook
Current agents still ignore discovered solutions in the majority of cases, indicating the issue is not solely about inference-time configuration.
Plain Language Accessible to non-experts
Imagine you're in a maze looking for an exit. You have a map, but it doesn't show all the shortcuts. You might find some hidden doors, but you usually don't try to open them because you're not sure if they'll help you reach the exit faster. This is like current LLM agents, which often ignore potentially useful information when exploring their environments.
ELI14 Explained like you're 14
Imagine you're playing a game where the goal is to find hidden treasure. You have some clues, but sometimes you find unexpected hints. The problem is, your character doesn't always investigate these hints because they're not part of your plan. This is the issue found in the study: current AI often ignores unexpected information.
Glossary
Environmental Curiosity
The ability of an agent to recognize and investigate unexpected but relevant information.
Used to evaluate agents' performance in tasks.
Solution Injection
A method of placing complete task solutions directly into environments to assess agents' environmental curiosity.
Used in experiments to evaluate agents' capabilities.
discovery@k
Measures the probability that an agent discovers the injected solution at least once in k attempts.
Used to assess agents' discovery capabilities.
interaction@k
Measures the probability that an agent interacts with the injected solution at least once in k attempts.
Used to assess agents' utilization capabilities.
Reasoning Budget
Refers to the computational resources and time available to an agent during task execution.
Affects agents' environmental curiosity.
Open Questions Unanswered questions from this research
- 1 How can agents' environmental curiosity be enhanced during training?
- 2 How can discovered information be better utilized during inference?
Applications
Immediate Applications
Task Execution Optimization
Enhancing agents' environmental curiosity to improve performance in complex tasks.
Long-term Vision
Intelligent Agent Development
Developing more environmentally adaptive intelligent agents to tackle more complex tasks.
Abstract
LLM-based agents are assumed to integrate environmental observations into their reasoning: discovering highly relevant but unexpected information should naturally lead to a model exploiting its own discoveries. We show that this assumption is false for current LLM-based agents, which struggle to reflect or react to unexpected information. Across three benchmarks (Terminal-Bench, SWE-Bench, AppWorld), we inject complete task solutions into the agent environments to deliberately expose a task's solution to a model. While agents discover these solutions on Terminal-Bench in 79-81% of runs, they interact, or exploit, them in only 37-50% of cases. This gap is starkest in AppWorld: agents see documentation stating that a command "returns the complete solution to this task" in over 90% of attempts but exploit this in fewer than 7% of trials. We show that agents lack what we call environmental curiosity: the capability to recognize and investigate unexpected but relevant observations in response to environmental stimuli. We identify three main factors influencing environmental curiosity: available tools in the agent scaffold, test-time compute, and training data distribution. Our findings identify configurations that maximize curiosity also achieve the best performance on the unmodified benchmarks. Yet even jointly optimized agents still ignore discovered solutions in the majority of trials: current agents use the environment to fetch expected information, but not to revise their strategy or maximally exploit useful stimuli.