New Framework Boosts LLM Agent Confidence Estimation.

Large Language Model (LLM) agents operate in dynamic environments where each action can have significant, irreversible consequences, often long before a final failure is detected. To mitigate these risks, it is crucial for agents to estimate the confidence of each proposed action *before* execution. Existing confidence estimation methods for LLMs typically focus on scoring a response to a prompt, but they often overlook the real-world execution consequences of an agent's actions. The new framework, called Critic Experience Bank (CEB), addresses this by allowing an LLM critic to learn from its own past experiences. After an agent completes a task, a hindsight LLM reviews the full execution feedback and "votes" on the productivity of each step. These pseudo-labels are then stored in a memory bank. When a similar situation arises, relevant productive and unproductive experiences are retrieved and used to inform the critic's prompt, improving its judgment. CEB requires no explicit training or ground truth labels and has shown significant improvements in calibration and ranking across various agent benchmarks and critic backbones, reducing errors by up to 54%.