Memory as Belief: Bayesian Evolution of Memory for Self-Evolving LLM Agents
- Published
- Source
- arXiv
- Paper number
- 1354
- Field
- Agents / Memory
- arXiv ID
- 2610.25119
Key points
- BEAM represents memory value as a context-conditioned Beta belief Beta(α, β) and predicts that belief with a lightweight LoRA-adapted estimator while keeping the backbone frozen.
- Using execution feedback as evidence, it computes per-memory marginal contribution through leave-one-out counterfactual evaluation and trains the estimator online with an evidential objective, a pairwise ranking loss, and a selective Beta-KL regularizer.
- Thompson sampling draws utility scores from the belief, giving memories with limited evidence more opportunities for exploration.
- BEAM achieves the best average performance under both executors, surpassing the strongest baselines by 3.93 and 2.43 percentage points, respectively.
- Across stages of the task stream, BEAM generally achieves larger gains over the no-memory baseline as the stream progresses, showing continual self-evolution.
- Transferring the value model learned on MATH-500 to AIME yields a clear early-stage advantage and the best stream-level average performance.
Paper links
External research summaries. These are not HDATF publications or measured product results.