Self-Compacting Language Model Agents: Rubric 기반 적응형 Context 압축

# SELFCOMPACT scaffold 핵심 루프 (의사 코드)

RUBRIC_PROMPT_MATH = """
Judge your math-solving state from the conversation above. Answer Q1...Q3 with Y or N.
Q1 ANSWER: The latest round states a specific final answer (\\boxed{} or 'Final Answer:...')
Q2 STUCK: Your last 2 rounds added no non-trivial fact
Q3 HAS-NEXT: You can write the exact next step

Output: exactly 3 lines
Q1: Y/N -- <evidence>
Q2: Y/N -- <evidence>
Q3: Y/N -- <evidence>

Fire rule: COMPRESS iff Q1=Y OR (Q2=Y AND Q3=Y)
"""

SUMMARIZER_PROMPT = """
Create a compressed summary for another model to continue solving from.
RULES:
1. If a final answer was found, PRESERVE IT at the end.
2. Keep key insights, important calculations, and the reasoning path.
3. Remove redundant text, false starts, and unnecessary repetition.
"""

def selfcompact_loop(prompt, model, probe_interval=16384, max_rounds=12):
    context = [prompt]  # 전체 컨텍스트 (KV 캐시 포함)
    
    for step in range(max_rounds):
        # 1. 일반 생성
        response = model.generate(context)
        context.append(response)
        
        if is_final_answer(response):
            return response
        
        # 2. N 스텝마다 rubric 프로브 (append 방식으로 KV 캐시 보존)
        if step % probe_interval == 0:
            rubric_context = context + [RUBRIC_PROMPT_MATH]  # 원본 캐시 오염 없이 복사
            verdict = model.generate(rubric_context)  # COMPRESS or CONTINUE
            
            if parse_verdict(verdict) == "COMPRESS":
                # 3. 요약 실행 (동일 모델 사용, 외부 모델 불필요)
                summary_context = context + [SUMMARIZER_PROMPT]
                summary = model.generate(summary_context)
                
                # 4. 컨텍스트 리셋: 원본 질문 + 요약본으로 교체
                context = [prompt, summary]
                print(f"[COMPRESS] {len_tokens(context)}tok → {len_tokens([summary])}tok")
            # CONTINUE면 rubric 프로브 결과를 컨텍스트에서 제거하고 그냥 진행
    
    return context[-1]