passdrill
AI & LLM Engineering · AI Agents & Tool Use · Card 029/032 medium

After an agent fails a coding task on its first attempt, a framework has the agent generate a written, natural-language reflection on what went wrong and why, store that reflection in a memory buffer, and then attempt the same task again with that reflection available to it -- with no change to the underlying model's weights at any point. What does this Reflexion-style approach rely on to improve performance across attempts?

  1. A traditional reinforcement-learning gradient update that adjusts the model's parameters after each failed attempt, using the failure as a scalar reward signal
  2. Fine-tuning a copy of the model on the transcript of the failed attempt before the next attempt begins, so the improvement is baked into new weights
  3. Verbal reinforcement -- having the agent verbally reflect on task feedback and store that self-generated reflective text in an episodic memory buffer, which is then fed back in on later attempts to induce better decision-making, entirely without updating any model weights
  4. Randomly perturbing the prompt's wording on each retry until one phrasing happens to produce a correct answer, independent of any feedback about the previous failure
Next card → Shuffle