---
id: wiki-2026-0508-refinement
title: Refinement
category: 10_Wiki/Topics
status: verified
canonical_id: self
aliases: [Iterative-Refinement, Self-Refine, Type-Refinement]
duplicate_of: none
source_trust_level: A
confidence_score: 0.9
verification_status: applied
tags: [llm, iteration, types, design]
raw_sources: []
last_reinforced: 2026-05-10
github_commit: pending
tech_stack:
  language: python
  framework: anthropic-sdk
---

# Refinement

## 매 한 줄
> **"매 첫 시도는 draft, 매 refinement는 product"**. Refinement는 초안을 critique → revise loop로 다듬어 quality를 끌어올리는 patterns의 family — LLM self-refine, type narrowing, design iteration 모두 매 같은 핵심 idea를 공유.

## 매 핵심

### 매 LLM Self-Refine (2024 Madaan et al. → 2026 mainstream)
- **Generate → Critique → Refine** loop. Single model이 매 세 role을 모두 수행.
- 매 효과: math, code, dialog 매 +5–20% accuracy without extra training.
- **Reflexion** (Shinn 2023)은 verbal RL — 매 episode 끝에 매 self-critique를 episodic memory로 저장.
- 2026 standard: Claude Opus 4.7 / GPT-5 매 native "extended thinking" mode가 매 refine을 internal로 흡수 → external loop는 매 high-stakes (legal, medical) 에서만.

### 매 Type Refinement (TypeScript / Flow / Python typing)
- **Narrowing**: union type 의 매 instance를 specific subtype 으로 좁힘 (`typeof`, `instanceof`, discriminated union).
- **Refinement type**: predicate-attached type — `{x: number | x > 0}`. Liquid Haskell, F* 매 사용.
- 2026 TS 5.x: `satisfies` operator + control-flow analysis 매 강력 — 매 manual cast 의 거의 elimination.

### 매 Design / Spec Refinement
- **Stepwise refinement** (Wirth 1971) — abstract spec → concrete implementation을 매 단계적으로.
- **BDD** (Given-When-Then) 매 modern incarnation.
- AI-aided: spec → Claude → multiple impl candidates → human picks → refine.

### 매 응용
1. RAG answer 의 self-refine으로 hallucination ↓.
2. Code generation 매 compile error → refine loop.
3. TS API 의 progressive type narrowing.
4. Product spec 의 PM ↔ AI iterative tightening.

## 💻 패턴

### Self-Refine loop (Anthropic SDK)
```python
from anthropic import Anthropic

client = Anthropic()
MODEL = "claude-opus-4-7"

def self_refine(task: str, max_iter: int = 3) -> str:
    answer = client.messages.create(
        model=MODEL, max_tokens=2048,
        messages=[{"role": "user", "content": task}],
    ).content[0].text

    for i in range(max_iter):
        critique = client.messages.create(
            model=MODEL, max_tokens=1024,
            system="You are a strict critic. List concrete flaws or reply 'NO_ISSUES'.",
            messages=[{"role": "user", "content": f"Task: {task}\n\nDraft:\n{answer}"}],
        ).content[0].text

        if "NO_ISSUES" in critique:
            return answer

        answer = client.messages.create(
            model=MODEL, max_tokens=2048,
            messages=[{"role": "user",
                       "content": f"Task: {task}\nDraft: {answer}\nCritique: {critique}\nRevise."}],
        ).content[0].text
    return answer
```

### Reflexion-style episodic memory
```python
class Reflexion:
    def __init__(self):
        self.memory: list[str] = []  # accumulated lessons

    def step(self, task: str) -> str:
        ctx = "\n".join(f"- {m}" for m in self.memory[-5:])
        attempt = llm(f"Task: {task}\nPast lessons:\n{ctx}\nAct.")
        feedback = environment(attempt)
        if not feedback.success:
            lesson = llm(f"Why did this fail? Task: {task}\nAttempt: {attempt}\nFeedback: {feedback}")
            self.memory.append(lesson)
        return attempt
```

### TypeScript discriminated union narrowing
```typescript
type Result<T> =
  | { kind: 'ok'; value: T }
  | { kind: 'err'; error: Error };

function unwrap<T>(r: Result<T>): T {
  if (r.kind === 'err') throw r.error;  // narrow → 'ok' branch
  return r.value;  // typed as T, no cast
}
```

### Python TypeGuard refinement
```python
from typing import TypeGuard

def is_str_list(x: list[object]) -> TypeGuard[list[str]]:
    return all(isinstance(i, str) for i in x)

def join(items: list[object]) -> str:
    if is_str_list(items):
        return ", ".join(items)  # narrowed
    raise TypeError("not str list")
```

### Compile-error refine loop (code generation)
```python
def codegen_refine(spec: str, max_iter=5):
    code = llm(f"Write Python for: {spec}")
    for _ in range(max_iter):
        ok, err = run_pytest(code)
        if ok: return code
        code = llm(f"Spec: {spec}\nCode: {code}\nFailing: {err}\nFix.")
    raise RuntimeError("refinement budget exhausted")
```

### Best-of-N + judge (ensemble refinement)
```python
def best_of_n(prompt: str, n: int = 5) -> str:
    candidates = [llm(prompt, temperature=1.0) for _ in range(n)]
    ranking = llm(f"Pick best of these:\n{candidates}\nReturn index 0..{n-1}")
    return candidates[int(ranking.strip())]
```

## 매 결정 기준
| 상황 | Approach |
|---|---|
| Single-shot good enough (chat) | No refine — extra cost |
| High-stakes (legal/medical) | External self-refine + human review |
| Code with tests | Compile/test-driven refine |
| Long agentic task | Reflexion (episodic memory) |
| Reasoning math | extended thinking (native) — already refine internally |
| TS API design | Narrowing + `satisfies` |

**기본값**: native extended thinking 우선 → 부족하면 external self-refine 1–2 iter.

## 🔗 Graph
- 부모: [[Iteration]]
- 변형: [[Reflexion]] · [[Self-Consistency]] · [[Best-of-N]]
- 응용: [[RAG]] · [[Code-Generation]]
- Adjacent: [[TypeScript 타입 시스템 (TypeScript Type System)|Type-System]] · [[Stepwise-Refinement]] · [[BDD]]

## 🤖 LLM 활용
**언제**: high-stakes output, agentic loops, code with verifiable feedback.
**언제 X**: 매 latency-sensitive UX, 매 simple chat — extra latency × cost는 매 안 맞음.

## ❌ 안티패턴
- **Infinite refine loop**: max_iter 의 hard cap 의 X → cost explosion.
- **Same-model critique only**: 매 critic = generator인 경우 같은 blind spot. Mix models (Opus critic, Sonnet generator).
- **Refine without termination signal**: "NO_ISSUES" 같은 매 explicit stop 의 부재 → 매 endless tweaking.
- **Type assertion 으로 narrow**: TS 매 `as` 사용은 매 refinement 의 X — 매 unsafe cast.

## 🧪 검증 / 중복
- Verified (Madaan 2024 Self-Refine, Shinn 2023 Reflexion, TS handbook narrowing).
- 신뢰도 A.

## 🕓 Changelog
| 날짜 | 변경 |
|---|---|
| 2026-05-08 | Phase 1 |
| 2026-05-10 | Manual cleanup — full rewrite covering LLM self-refine, type narrowing, design iteration |