jeevesagent.architecture.self_refine

Self-Refine: iterative refinement via critique.

Madaan et al. 2023 — Self-Refine: Iterative Refinement with Self-Feedback. The same model plays generator, critic, and refiner across rounds.

Pattern:

  1. Round 0 (generator). Run the base architecture (default ReAct) to produce an initial output.

  2. Round 1..max_rounds:

    1. Critic. Review the current output. Output a critique. If the critique contains stop_phrase (default "no issues") anywhere, terminate — the model considers the output good enough.

    2. Refiner. Produce a revised output that addresses the critique. The new output replaces the old.

When the same model plays all three roles, gains are real but modest (~5-15% on most tasks). For asymmetric blind-spot coverage, use ActorCritic with a different model for the critic — that ships in a later release.

Strengths

  • Simple, well-defined; one new architecture wraps any base.

  • Cheap relative to multi-agent debate or actor-critic with separate models.

  • Composes with ReAct (default base) and any other architecture that satisfies the protocol.

Weaknesses

  • Same-model self-critique shares blind spots — bounded gains.

  • 2-3× token cost vs single-pass.

  • Latency adds up — sequential critic / refiner per round.

Attributes

DEFAULT_CRITIC_PROMPT

DEFAULT_REFINER_PROMPT

Classes

SelfRefine

Wrap a base architecture with iterative critique / refine.

Module Contents

class jeevesagent.architecture.self_refine.SelfRefine(*, base: jeevesagent.architecture.base.Architecture | None = None, max_rounds: int = 3, critic_prompt: str | None = None, refiner_prompt: str | None = None, stop_phrase: str = 'no issues')[source]

Wrap a base architecture with iterative critique / refine.

base defaults to ReAct; the round-0 generator runs the base architecture’s full strategy. Subsequent rounds are text-only model calls — no tools, just critique and rewrite.

declared_workers() → dict[str, jeevesagent.agent.api.Agent][source]
async run(session: jeevesagent.architecture.base.AgentSession, deps: jeevesagent.architecture.base.Dependencies, prompt: str) → collections.abc.AsyncIterator[jeevesagent.core.types.Event][source]
name = 'self-refine'
jeevesagent.architecture.self_refine.DEFAULT_CRITIC_PROMPT = Multiline-String
Show Value
"""You are a careful reviewer. Read the original task and the proposed
output. Identify every issue you can find: factual errors, missing
edge cases, unclear language, inconsistent reasoning, gaps versus
what the task asked for.

Be specific. Cite the relevant section of the output.

If you find no issues at all and the output fully satisfies the
task, respond with exactly:

  no issues

Otherwise, list the issues as a bulleted critique."""
jeevesagent.architecture.self_refine.DEFAULT_REFINER_PROMPT = Multiline-String
Show Value
"""You are revising your previous output based on a critique. Address
every point in the critique. Preserve what was correct; fix what
was wrong; add what was missing. Output ONLY the revised version,
not commentary about the changes."""