AutoLibra: Agent Metric Induction from Open-Ended Human Feedback

TL;DR

AutoLibra tackles the limitation of coarse task-success metrics by inducing fine-grained, trajectory-grounded evaluation metrics from open-ended human feedback. It grounds feedback to behavior, clusters aspects into interpretable metrics via LLMs, and uses a model-based judge to rate trajectories, with meta-evaluation over coverage and redundancy guiding metric optimization. The approach yields novel, concrete metrics across diverse agent domains and demonstrates practical agent-improvement benefits through prompt engineering and self-regulation guided by induced metrics. This framework promises a generalizable, user-centered mechanism for evaluating and improving language agents in a task-agnostic manner.

Abstract

Agents are predominantly evaluated and optimized via task success metrics, which are coarse, rely on manual design from experts, and fail to reward intermediate emergent behaviors. We propose **AutoLibra**, a framework for agent evaluation, that transforms open-ended human feedback *e.g.* "If you find that the button is disabled, don't click it again", or "This agent has too much autonomy to decide what to do on its own" into metrics for evaluating fine-grained behaviors in agent trajectories. AutoLibra accomplishes this by grounding feedback to an agent's behavior, clustering similar positive and negative behaviors, and creating concrete metrics with clear definitions and concrete examples, which can be used for prompting LLM-as-a-Judge as evaluators. We further propose two meta metrics to evaluate the alignment of a set of (induced) metrics with open feedback: "coverage" and "redundancy". Through optimizing these meta-metrics, we experimentally demonstrate AutoLibra's ability to induce more concrete agent evaluation metrics than the ones proposed in previous agent evaluation benchmarks and discover new metrics to analyze agents. We also present two applications of AutoLibra in agent improvement: First, we show that AutoLibra serve human prompt engineers for diagonalize agent failures and improve prompts iterative. Moreover, we find that AutoLibra can induce metrics for automatic optimization for agents, which makes agents improve through self-regulation. Our results suggest that AutoLibra is a powerful task-agnostic tool for evaluating and improving language agents.

Figures (11)

• Figure 1: AutoLibra induces agent evaluation metrics from human feedback, and uses these metrics to evaluate agents, which can be meta-evaluated via evaluating the coverage on unseen human feedback. Here we show real examples of agent trajectories, human feedback, aspects, induced metrics, evaluation results on WebVoyager he2024webvoyager.
• Figure 2: Metric optimization: optimizing the induction process through maximizing the coverage while minimizing redundancy of the metrics, calculated via the evaluation process.
• Figure 3: Coverage and redundancy of AutoLibra metrics on four agentic datasets. Circles indicate coverage and redundancy for different induced metrics; stars indicate the the best metrics' coverage and redundancy on held-out human feedback; squares show an ablation test, indicating the effect when good and bad behavior examples are removed from metrics, demonstrating the criticality of concrete behavior examples
• Figure 4: AutoLibra iteratively induce metrics and improves the agent prompts through optimizing for the induced metrics. Although not optimized for, the success rate of the agent continuously improve until Stage 3, when the agent begins to overthink.
• Figure 5: Example of Baba-Is-AI game.