Apple Machine Learning Research has introduced SELFCOMPACT, a self-triggered compaction scaffold that lets language model agents decide when to summarise their own reasoning context, improving efficiency and performance without fine-tuning.
The scaffold pairs a compaction tool the model can invoke with a lightweight rubric specifying when to compact — after sub-task resolution or trajectory convergence — and when to suppress summarisation, such as mid-derivation or when the model is stuck. On six benchmarks covering competitive maths and agentic search, SELFCOMPACT matched or exceeded fixed-interval summarisation while reducing per-question token costs by 30–70 per cent. Compared with no summarisation at all, it improved performance by up to 18.1 points on maths tasks and 5–9 points on agentic search.
The research identifies what the authors call a meta-cognitive gap: unprompted models cannot reliably judge when their own context is degrading. The lightweight rubric closes this gap without additional training. The same Apple MLR post also highlights two other projects: RISED, a rubric-based training strategy for multi-environment agent selection (published 6 October 2026), and a rubric-based alignment framework for grounded knowledge answers (published 27 August 2026).
Why it matters
In our view, SELFCOMPACT addresses a practical bottleneck in deploying large language model agents: the ever-growing context window that drives up cost and latency. By giving the model a simple, rule-based trigger to compress its own reasoning, Apple is exploring a path that avoids expensive fine-tuning while maintaining or improving accuracy. If the approach holds up in production-scale tasks, it could make multi-step reasoning far more economical for on-device and server-side agents alike.