Language agents often ignore past actions during task execution

A new research paper investigates why language agents often fail to learn from their own history while completing complex tasks. Even when provided with a record of previous attempts, these systems do not always use the information effectively to improve their results.

The research reveals that agents often perform well regardless of whether the history provided is accurate or shuffled. This suggests that the models are not truly building a logical bridge between an action they took and the resulting feedback from the environment.

To address this, the authors tested a method of explicitly labeling environmental observations as direct consequences of the preceding action. This small change helped the agents identify the results of their choices, leading to higher success rates and fewer repetitive mistakes.

Building on this, the team introduced a learned calibrator that acts as an internal filter. This component assesses past actions more critically and decides which experiences are valuable enough to store for future guidance.

Imagine you are learning a new game but you keep forgetting which buttons you pressed when you won or lost. These agents currently suffer from a similar memory gap.

While outcome labeling and the new calibrator improve performance, this research was focused on specific task environments. It does not address how agents handle long-term memory across entirely different sessions or tasks.