Self-Refine

Iterative refinement via self-critique. Madaan et al. 2023 , Self-Refine: Iterative Refinement with Self-Feedback.


   prompt ───► generate ───► critique ──┬── score ≥ threshold ──► output
                              ▲         │
                              │         └── below ──► refine ──┐
                              │                                │
                              └────────────────────────────────┘

Pattern

Round 0 (generator). Run the base architecture (default ReAct) to produce an initial output.
Rounds 1..max_rounds:
- Critic. Review the current output. Output a critique. If the critique contains stop_phrase (default "no issues"), terminate . The model considers the output good enough.
- Refiner. Produce a revised output addressing the critique. The new output replaces the old.

Usage


from loomflow import Agent
from loomflow.architecture import SelfRefine
 
agent = Agent(
    "Write a tight 3-paragraph product brief.",
    model="claude-opus-4-7",
    architecture=SelfRefine(max_rounds=3),
)
result = await agent.run("Brief about Acme Corp.")

Or by string:


agent = Agent("...", model="...", architecture="self-refine")

When same-model self-critique pays off

Clearly-defined output rubrics (style, length, structure).
Tasks where the first draft has predictable failure modes the same model can recognise on a second pass (filler, hedging, repetition, missing sections).
Cost is roughly 2–3× ReAct. You pay for the critic + refiner calls.

When to use ActorCritic instead

When the same model has the same blind spots as critic and generator, gains plateau quickly. For asymmetric blind-spot coverage (different prompts, ideally different models), use ActorCritic. Actor and critic are distinct Agent instances, possibly on different model providers.

Threshold via stop phrase, not score. Self-Refine’s critic outputs free-form text; termination is keyword-based. For numeric-score gating (score ≥ threshold), use ActorCritic which returns structured JSON ({score, issues, summary}).