Self-Compacting Language Model Agents

Key points

• We present empirical results on six benchmarks (competitive math and agentic search) and seven models.1
• We introduce a rubric-based reward framework that generates query-specific rubrics grounded in retrieved evidence and decomposed into multiple quality dimensions, providing fine-grained supervision during...1
• Long coding agent sessions outgrow the model's context window, so agents compact: older history is replaced by something shorter, usually a model-written summary.2
• Here, we present transparent-compact, a simple deterministic method for the open-source Pi coding agent, designed specifically for use with such models and with the purpose of maximising the transparency of the agent's view of its own compacted context.2

Sources (2)

• [1]Self-Compacting Language Model Agents
  Apple Machine Learning Research · Oct 12, 12:00 AM

  We propose SELFCOMPACT, a scaffold that allows the model itself to decide when and how to compact.

  We present empirical results on six benchmarks (competitive math and agentic search) and seven models.

• [2]Maximising Transparency in Context Compaction Improves Retention of Facts and Commitments for Lightweight Coding Agents
  arXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 14, 12:43 PM

  Long coding agent sessions outgrow the model's context window, so agents compact: older history is replaced by something shorter, usually a model-written summary.

  Here, we present transparent-compact, a simple deterministic method for the open-source Pi coding agent, designed specifically for use with such models and with the purpose of maximising the transparency of the agent's view of its own compacted context.

Extractive summary: sentences quoted from the sources.