20260625:很多新内容
This commit is contained in:
77
papers/me2-trm-reasoning-2026.md
Normal file
77
papers/me2-trm-reasoning-2026.md
Normal file
@@ -0,0 +1,77 @@
|
||||
---
|
||||
title: "ME² + TRM: Complex Reasoning Optimization (Zhang et al., ICML 2026)"
|
||||
created: 2026-06-24
|
||||
updated: 2026-06-24
|
||||
type: paper
|
||||
tags: ["reasoning", "reward-model", "dag", "grpo", "test-time-scaling"]
|
||||
sources:
|
||||
- "https://arxiv.org/abs/2602.08498"
|
||||
code: "https://github.com/Simplified-Reasoning/TRM"
|
||||
---
|
||||
|
||||
# ME² + TRM: 复杂推理的表征、评估与优化
|
||||
|
||||
> Zhang et al. | ICML 2026 | arXiv:2602.08498v2 | cs.CL
|
||||
|
||||
## 动机
|
||||
|
||||
[[large-reasoning-models|LRMs]] 的推理轨迹越来越长且结构复杂,但缺乏统一的答案回答三个问题:(1) 什么是高质量推理?(2) 如何可靠评估?(3) 如何用评估信号优化推理?
|
||||
|
||||
现有方法的局限:PRMs 依赖步骤级绝对评分,无法捕获长程依赖和非线性结构;ORMs 设计用于对齐最终响应(helpful/honest/harmless),而非评估结构化推理质量。
|
||||
|
||||
## 核心框架
|
||||
|
||||
### [[me2-principle|ME² 原则]]
|
||||
|
||||
两个正交维度:
|
||||
|
||||
| | Macro(全局) | Micro(局部) |
|
||||
|---|---|---|
|
||||
| **Effectiveness** | 结构组织是否合理、无冗余分支 | 步骤是否正确、有逻辑 |
|
||||
| **Efficiency** | 推理路径是否简洁、无绕路 | 步骤是否精简、无赘述 |
|
||||
|
||||
推理质量 = Macro-Effectiveness × Macro-Efficiency × Micro-Effectiveness × Micro-Efficiency
|
||||
|
||||
### [[dag-reasoning-evaluation|DAG 推理建模]]
|
||||
|
||||
将推理轨迹抽象为 DAG:
|
||||
- 节点:推理步骤
|
||||
- 边:逻辑依赖关系
|
||||
- DAG vs Tree:Tree 无法表达合并(多前驱节点),DAG 是表达力与可处理性的实用平衡
|
||||
|
||||
### [[thinking-reward-model|Thinking Reward Model (TRM)]]
|
||||
|
||||
训练流程:
|
||||
1. 生成多条候选推理轨迹 → 构建 DAG → ME² pairwise preference 标注(DeepSeek-V3.2)
|
||||
2. 构建 [[trm-preference-dataset|TRM-Preference]](103K 训练对,1.5K 验证)
|
||||
3. 训练 TRM:Llama-3.1-8B + scalar head,Bradley-Terry loss
|
||||
|
||||
**核心设计**:TRM 仅训练于 verified-correct 推理对——与答案正确性解耦,纯评估推理质量。
|
||||
|
||||
### [[reasoning-quality-optimization|推理质量优化]]
|
||||
|
||||
**Test-Time Scaling**:TRM Best-of-N selection → +19.3%(AIME24, N=16, Qwen3-8B: 44.7%→64.0%)
|
||||
**RL Training**:TRM-guided GRPO with gated reward shaping:
|
||||
|
||||
$$r = r_v \cdot (1 - \alpha + \alpha \cdot \text{Sigmoid}(r_t))$$
|
||||
|
||||
r_v = outcome reward, r_t = thinking reward, α = balance weight
|
||||
→ +3.9% across diverse tasks
|
||||
|
||||
## 关键结果
|
||||
|
||||
| 方法 | 验证集准确率 |
|
||||
|------|------------|
|
||||
| Qwen2.5-Math-PRM-7B | 46.3% |
|
||||
| ReasonFlux-PRM-7B | 62.5% |
|
||||
| PromptOnly (DeepSeek-V3.2) | 78.6% |
|
||||
| **TRM (ours)** | **88.6%** |
|
||||
|
||||
## 核心洞察
|
||||
|
||||
1. **将推理质量与答案正确性解耦** — TRM 仅训练于正确推理的偏好对,证明推理质量可独立于答案正确性评估
|
||||
2. **DAG 比 Tree 更适合推理建模** — 推理中的合并(多步归结为一个结论)是常见模式,Tree 无法表达
|
||||
3. **Structural signals matter** — 直接 prompt-based 比较产生大量 ties (232/1497),但去除 ties 后准确率 93%。DAG 结构化后 ties 归零,证明结构信号是关键区分器
|
||||
|
||||
## 来源
|
||||
[原始存档](raw/papers/me2-trm-reasoning-2026.md) | [arXiv](https://arxiv.org/abs/2602.08498) | [GitHub](https://github.com/Simplified-Reasoning/TRM)
|
||||
Reference in New Issue
Block a user