Create

Sign in to ReadmeX

or

How we handle your data: Privacy Policy

ReadmeX

Your AI reading companion

What would you like to catch up on?

Pick a briefing, or ask your own question.

Your community and people briefings stay personal to you.

Story

HIT-ByteDance Team Introduces Structured Multimodal Reasoning Format DRT

AI summary

A joint Harbin Institute of Technology and ByteDance team introduced Dense Reasoning Trace (DRT), a structured format that separates visual observations from logical deductions and links intermediate states with symbolic operators. According to Juejin AI, the team’s DRT traces achieved 5.5-fold token efficiency over natural-language chain-of-thought and scored 1.3 percentage points above a Qwen3-VL baseline on mathematics and chart-reasoning evaluations. The report notes that the format was manually designed, that its generalization to open-domain tasks remains unverified, and that performance on natural images is unknown.

In-depth insight

The report presents DRT as a more compact, machine-friendly representation for multimodal and agent systems. Separating observations from deductions is also described as supporting step-by-step auditing, while the reported token reduction could make long reasoning traces less costly.

💡 一句话总结:让 AI 展示推理过程(思维链)是提升准确率的标配,但 CoT 是「写作文」——冗长、烧 token,多模态场景下还把视觉细节稀释在文字里。DRT 换了个表达层:让模型用紧凑的结构化格式思考(观察归观察、推理归推理、符号做连接)。结果:token 省 5.5 倍,准确率反而涨。思维链的形态,可能从一开始就选错了。

导语:思维链的「文学税」

「请一步步思考」——CoT 提示是 AI 推理的标配。但观察 CoT 的输出,它本质上是在写小作文:「首先,我注意到图片中有一个红色的杯子,杯子旁边有一张桌子,桌子上有……」

这种自然语言推理有三个隐藏成本:

  • Token 账单:长 CoT 的推理成本可能是答案本身的几十倍
  • 信息稀释:关键信号淹没在修辞里——尤其多模态场景,「看到什么」和「推出什么」混在一堆文字里
  • 幻觉温床:为了行文流畅,模型倾向于「补写」并没有真正观察到的细节

#Dense Reasoning Trace (DRT)#Harbin Institute of Technology#ByteDance#Qwen3-VL#Trace-Grounded GRPO

Original · 3 min readRead the original
Group chat

No comments yet. Start the conversation.