GradCuit: Credit-Assigned Gradient Flow Enables Robust and Interpretable Test-Time Latent Reasoning
- ID: 04ff6214
- 原文链接: https://arxiv.org/abs/2608.02585
- 作者 / 日期: Zhaoxin Yu, Qi Shen, Hengli Li, Zhaowei Zhang, Song-Chun Zhu, Chi Zhang, Zilong Zheng | 2026-08-03
- 分类: models
- 来源类型: paper
- 标签: test-time-scaling, latent-reasoning, gradient-optimization, reasoning
- 质量评分: 4/5
- 抓取时间: 2026-08-05T15:45:26.663004+00:00
中文导读
GradCuit 把可优化 latent states 插入 Transformer 指定层，让 continuation token 的 log-probability 通过因果自注意力对前置 latent state 形成可微路径，从而把结果反馈直接分配到内部推理状态作者报告它在 5 个 instruction-tuned backbone3 个 reasoning benchmark 和 2 种答案格式上平均准确率 64.5%，比 CoT prompting 高 6.6 个百分点，并在不同学习率下比 LatentSeek 更稳定
为什么值得关注
GradCuit 把 test-time scaling 从多采样答案推进到直接优化模型内部推理状态
English Summary
GradCuit inserts optimizable latent states inside selected Transformer layers and uses reward-weighted gradients from the continuation to optimize internal reasoning at test time. It reports stronger and more stable reasoning accuracy than chain-of-thought prompting and competing latent optimization methods.
Obsidian Evidence
候选来自 OpenClaw定时任务/论文流水线/2026-08-05-论文流水线.md 的今日论文速报。OpenCLI arXiv metadata fetched for id 2608.02585.
Source Extract / Metadata
arXiv metadata URL: https://arxiv.org/abs/2608.02585
Title: GradCuit: Credit-Assigned Gradient Flow Enables Robust and Interpretable Test-Time Latent Reasoning
Authors: Zhaoxin Yu, Qi Shen, Hengli Li, Zhaowei Zhang, Song-Chun Zhu, Chi Zhang, Zilong Zheng
Abstract-backed summaries above were generated from OpenCLI arXiv metadata fetched during intake.