Source record

Published by arXiv cs.MA (T1 source). The original is at https://arxiv.org/abs/2610.35966.

Pipeline notes

The summary and note below are generated by the signal pipeline — they are Beyond Desk’s reading, not quotations from the source.

EMMA represents a multi-agent configuration—routing memory, specialist prompts, and procedural skills—as structured episodic memory. A reflector assigns diagnosed failures to update owners, while backbone-independent guards constrain accepted edits. Across three benchmarks, five backbones, and four baselines, the paper reports top overall results and fewer training executions.

The update-ownership and admissibility rules address a recurring problem in agent evolution: changes that drift toward backbone preferences rather than observed failures.

Cited by

No citations on record.