From Noisy Traces to Root Causes: Structural Trajectory Analysis and Causal Extraction for Agent Optimization

Imagine you're building an intelligent agent that needs to learn from its past actions to improve. The challenge is that when you look at its execution history, you often find a lot of irrelevant information mixed in with useful insights. This clutter can make it hard to figure out what went wrong and how to fix it. When you try to optimize based on these messy traces, you might end up focusing on minor failures that don't really matter, which is inefficient and can lead to overfitting — this is what's called optimization inefficiency.

Currently, people might try to simplify the data by cutting it down to the most recent actions or using sliding windows, but these methods can throw away crucial context that explains why things went wrong. This is where STRACE comes in. It offers a smarter way to sift through the execution traces by identifying and keeping only the most relevant failures while also pinpointing the actual causes of issues within the agent's decision-making process.

With STRACE, you get a clearer picture of what needs to be optimized, which leads to better performance. In practical terms, this means that when tested on a tough verification task, agents optimized with STRACE achieved a 1.4 times higher success rate compared to those using standard methods. For anyone building complex agents, this framework could significantly enhance their ability to learn from past mistakes and improve over time.

STRACE introduces a new framework for optimizing long-horizon agents by addressing the inefficiencies in using execution traces.

The empirical results show significant improvements over standard methods, supported by a specific benchmark.