Training Composer for longer horizons

Summary

Composer is trained for long-horizon coding tasks using self-summarization, allowing reinforcement-learning trajectories to exceed the model’s context window. When a fixed token trigger is reached, Composer receives a synthetic summarization query, uses scratch space to produce a condensed context, and resumes with the summary plus conversation state; training rewards every token in the chained rollout with the final outcome. Against a tuned prompt-based compaction baseline, self-summary reduced compaction-related error by 50% on CursorBench at both 80k and 40k triggers, while summaries averaged about 1,000 tokens instead of more than 5,000 and reused the KV cache. An early Composer checkpoint solved Terminal-Bench 2.0’s make-doom-for-mips after 170 turns, summarizing more than 100,000 tokens into roughly 1,000 tokens of retained information.

Takeaways

1. 1

   Self-summarization reduced compaction-related error by 50% on CursorBench at both 80k-token and 40k-token triggers.

2. 2

   Composer’s trained summaries averaged around 1,000 tokens, compared with more than 5,000 tokens for the tuned prompt-based baseline, while reusing the KV cache.

3. 3

   An early Composer research checkpoint solved Terminal-Bench 2.0’s make-doom-for-mips after 170 turns and more than 100,000 tokens of self-summarized context.