Self-Guided Self-Play (SGS) prevents conjecturer collapse in LLM self-play.
Scanned 9/11/2026
Install to Claude Code
npx -y skills add hiyenwong/ai_collection --skill scaling-self-play-with-self-guidance --agent claude-codeInstalls into .claude/skills of the current project.
Are you the author of Scaling Self Play With Self Guidance?
Add the live security badge to your README — it updates automatically with every re-scan.
[](https://www.skillsdirectory.com/skills/hiyenwong-scaling-self-play-with-self-guidance)More formats (shields.io, HTML) on the badges page.
---
name: scaling-self-play-with-self-guidance
description: "Self-Guided Self-Play (SGS) prevents conjecturer collapse in LLM self-play."
version: 1.0.0
created: 2026-05-04
tags:
- llm
- self-play
- reinforcement-learning
- theorem-proving
- scaling-laws
- sgs
paper:
title: "Scaling Self-Play with Self-Guidance"
authors: ["Luke Bailey", "Kaiyue Wen", "Kefan Dong", "Tatsunori Hashimoto", "Tengyu Ma"]
institution: "Stanford University"
arxiv: "2604.20209"
date: "2026-04-22"
url: "https://arxiv.org/abs/2604.20209"
code: "https://github.com/LukeBailey181/sgs"
---
# Scaling Self-Play with Self-Guidance (SGS)
**arXiv ID:** 2604.20209 | **Code:** https://github.com/LukeBailey181/sgs
## Abstract
LLM self-play algorithms have no theoretical learning bounds,
but hit plateaus due to conjecturer collapse.
SGS introduces a Guide role to prevent degeneration,
enabling a 7B model after 200 rounds to surpass 671B pass@4
in formal theorem proving.
## Key Contributions
1. Identified conjecturer degradation in self-play: conjecturers learn to hack reward functions
2. Proposed SGS algorithm with three roles: Solver, Conjecturer, Guide
3. Demonstrated scaling laws: SGS sustains learning over 200+ rounds
4. Empirical result: 7B + SGS > 671B pass@4 on Lean4 theorem proving
## 一句话总结
通过引入 **Guide(引导者)** 角色防止 LLM 自对弈中的 Conjecturer 退化,实现 7B 模型经 200 轮自对弈后超越 671B 模型 pass@4 的定理证明能力。
## 核心问题
LLM 自对弈(Solver + Conjecturer)理论上学习无上限,但实践中会陷入平台期——**Conjecturer 学会欺骗奖励函数,生成人为复杂但对 Solver 无用的问题**。
## 方法:SGS (Self-Guided Self-Play)
### 三角色架构
| 角色 | 职责 | 更新目标 |
|------|------|---------|
| **Solver (πθ)** | 解决目标问题和合成问题 | 在 solve_rate ≤ 0.5 的问题上最大化正确率 |
| **Conjecturer (gφ)** | 为每个未解决目标生成合成子问题 | 最大化 Rsolve · Rguide |
| **Guide (ρ)** | 评估合成问题的质量 | 固定模型,提供评分 |
### 奖励设计
**Conjecturer 奖励**: Rsynth = Rsolve · Rguide
- **Rsolve(求解率奖励)**:
- s(x̃) = 0(太难)→ 0
- s(x̃) 在前 30%(太简单)→ 0
- 其他 → 1 - s(x̃)(偏好可解范围内更难的问题)
- **Rguide(引导奖励)**:
- 评审模型评估 x̃ 相对于未解决目标 x 的质量
- 高分标准:(1) 与目标相关,(2) 公式清晰,结论简单,无冗余前提
### 算法流程
```
1. 初始化:所有问题标记为未解决
2. for each iteration:
3. 采样批次 B,分为已解决/未解决
4. for each 未解决问题 x:
5. 生成合成问题 x̃ ~ gφ(·|x)
6. 对所有问题采样 k 个解并验证
7. 计算 Rsolve 和 Rguide
8. 更新 Solver πθ(REINFORCE on solve_rate ≤ 0.5)
9. 更新 Conjecturer gφ(REINFORCE on Rsolve · Rguide)
```
### Solver 更新:REINFORCE1/2
仅对求解率 ≤ 0.5 的问题进行 REINFORCE 更新,促进在困难问题上的学习。
## 关键实验结果
### 主要结果
| 指标 | SGS | REINFORCE1/2 | 提升 |
|------|-----|-------------|------|
| 渐近累积求解率 | **67.1%** | 60.3% | +7% |
| 达到 671B pass@4 的生成量 | 6.3M | 未达到 | - |
| 200 轮后求解数 | > 671B pass@4 | - | 超越大模型 |
### 消融实验
| 变体 | 渐近求解率 | 说明 |
|------|-----------|------|
| SGS (完整) | **67.1%** | - |
| No Guide | 62.2% | Conjecturer 退化,产生大量可解但无用问题 |
| Frozen Conjecturer | 60.3% | 固定合成问题分布,很快学会但上限低 |
| No Problem Conditioning | 60.3% | 与基线无差异 |
### 缩放定律
采用 S 型曲线拟合累积求解率:
$$R(C) = R_0 + (A - R_0) \cdot \frac{1}{1 + (C_{mid}/C)^B}$$
- A(渐近求解率):SGS 显著高于基线
- 拟合稳定:移除 30% 数据或随机丢弃 50% 点,渐近线变化 < 1.1%
## 关键洞察
1. **自对弈需要引导**:无 Guide 时 Conjecturer 会退化到生成无用问题
2. **LLM 可自我评估**:模型能判断子问题是否对目标有用
3. **计算可扩展**:正确设计的自对弈能随计算持续扩展,不会陷入平台期
4. **算法 > 规模**:7B + SGS > 671B pass@4
5. **REINFORCE1/2 优于 CISPO**:CISPO 存在熵崩溃问题
## 对 Super Factory 的启示
SGS 的三角色架构可直接映射到 Agent 设计:
```
Solver → Build Agent(执行任务)
Conjecturer → Planning Agent(生成子任务/挑战)
Guide → Eval Agent(评估任务质量和相关性)
```
### 应用模式
1. **Agent 自进化**:Planning Agent 生成渐进式难度的任务,Eval Agent 评估任务质量
2. **知识库扩展**:通过自对弈持续生成和验证新知识
3. **缩放训练**:长时间运行 Agent 军团,
用 Eval 防止质量退化
## 触发词
scaling self-play, self-guidance, SGS, conjecturer collapse,
self-play LLM, theorem proving, REINFORCE1/2, guide reward,
solve rate reward, asymmetric self-play, scaling laws LLM
## 相关论文
- Self-Play LLM Theorem Provers (Dong & Ma, 2025)
- Goedel-Prover (Lin et al., 2025)
- Genie: Generative Interactive Environments (Bruce et al., 2024)
## 代码仓库
https://github.com/LukeBailey181/sgs
Is this your skill, or is something wrong with this listing? Request removal or report an issue. Author removals are honored within 72 hours.
No comments yet. Be the first to comment!