Brevity Bias

Prompt and context optimizers tend to converge on short, generic instructions that drop domain heuristics, tool-use guidance, and known failure modes, which hurts detail-hungry agent and domain tasks.

Brevity bias is a failure mode of iterative context adaptation. Many prompt optimizers reward concise, broadly applicable instructions over the accumulation of knowledge, and some treat brevity as a design goal. The abstraction can still score well on a validation metric while discarding exactly what multi-step agents and knowledge-intensive tasks depend on:

  • domain-specific heuristics and tactics
  • tool-use guidelines (API quirks, schemas, pagination rules)
  • catalogues of common failure modes

Evidence: in a study of prompt optimization for unit-test generation (Gao et al., 2025, cited by ACE), iterative methods repeatedly produced near-identical generic instructions ("Create unit tests to ensure methods behave as expected"). Diversity and domain detail disappeared.

Consequences:

  • Narrowed search: successive candidates collapse to one generic form.
  • Error propagation: optimized prompts inherit their seed's faults, because the compressed form has no room for the corrective detail.
  • Domain penalty: agents, program synthesis, and financial or legal analysis succeed by accumulating task-specific insight, not compressing it.

The counter-stance (ACE's "playbooks, not summaries") is the authors' position, not a general result. Unlike humans, who often benefit from concise generalization, they argue LLMs are more effective with long, detailed contexts and can distill relevance themselves at inference time. So keep tactics in the context and let the model select. ACE itself names exceptions (see prompt evolution vs playbook accumulation).

The stance is in tension with lost in the middle and with token budgets. Make a detailed context selectable rather than monolithic: an itemized bullet context with named sections, plus an instruction to use only the relevant parts. That gets the detail without forcing every token to matter.

Do not confuse brevity bias with context collapse. Brevity bias is where an optimizer's objective drifts. Collapse is a mechanical loss caused by rewriting the whole context at each step.

Sources

  • ACE (Zhang et al.), §1, §2.2