# A 7-million parameter model outperforming models a thousand times its size on tasks like ARC Prize. ...

Canonical URL: https://www.traeai.com/articles/d3e3e98d-c897-448c-b6f8-d9ddc1ca4ae8
Original source: https://x.com/ycombinator/status/2050224443461718118
Source name: Y Combinator(@ycombinator)
Content type: tweet
Language: 中文
Score: 8.7
Reading time: 1 分钟
Published: 2026-05-01T14:42:53+00:00
Tags: AI, 递归推理, HRM, TRM, ARC Prize

## Summary

YC解析HRM与TRM两类递归AI模型：仅700万参数的小模型在ARC Prize等推理任务上超越千倍参数大模型，关键在于推理时递归扩展计算深度。

## Key Takeaways

- 递归机制在推理阶段动态增加计算深度，突破标准LLM的固定上下文与浅层推理瓶颈。
- HRM通过外循环迭代调用基础模型实现多步推理，TRM则依赖训练中收敛的不动点递归结构。
- 小参数递归模型在ARC Prize等符号推理任务上显著优于同等规模非递归模型，验证‘计算深度’比‘参数规模’更关键。

## Outline

- 引言：小模型逆袭现象 — 指出700万参数模型在ARC Prize上超越千倍参数大模型，引出递归推理的核心价值。
  - 递归推理的本质突破 — 解释标准LLM因前向单次计算导致的推理深度天花板，递归如何通过迭代调用解锁深度计算。
  - HRM：基于外循环的递归架构 — HRM复用轻量基础模型，在推理时通过显式外循环展开多步推理，支持可控、可解释的步骤生成。
  - TRM：训练内嵌的不动点递归 — TRM在训练中学习稳定递归映射，推理时通过迭代收敛至不动点，隐式实现深度推理。
  - 性能对比与任务适配性 — HRM/TRM在ARC Prize等需符号操作与分步验证的任务上大幅领先，但在语言生成类任务中无明显优势。
- 未来方向：递归+基础模型融合 — 探讨将递归机制嫁接至大规模基础模型（如Llama、Phi）以兼顾泛化力与强推理能力。

## Highlights

- > A 7-million parameter model outperforming models a thousand times its size on tasks like ARC Prize. That's what recursive reasoning unlocks. — 原文首句
- > Standard LLMs hit a fundamental ceiling on certain reasoning tasks — recursion at inference time gives small models the compute depth to break through it. — 原文中段
- > HRM uses an outer loop that repeatedly calls a base model; TRM trains a model to converge to a fixed point via iterative application. — 视频时间戳 20:46 对比总结
- > The key insight is not more parameters, but more *compute depth* — achieved not by bigger models, but by smarter use of inference-time computation. — 视频核心论点提炼

## Citation Guidance

When citing this item, prefer the canonical traeai article URL for the AI-readable summary and include the original source URL when discussing the underlying source material.