跳到正文
arXiv cs.CL· Kaiyuan Lu, Jiatong Li, Yatao Bian·· 12 小时前AI 评分38

研究质疑 LatentChem 的潜在化学推理机制:是推理链还是软提示?

Reasoning or Just Prompting? Probing the Mechanism of Latent Chemical Reasoning

AI 导读

研究探测 LatentChem 的潜在化学推理机制,发现交换或适度扰动中间潜在状态对性能影响很小,而完全移除时分数有时大幅下降,说明潜在状态重要但其分子特异性内容影响有限。一个跨分子共享的固定前缀即可复现 LatentChem 的性能,学习到的静态软提示在受控模型中与逐输入潜在生成表现相当。作者据此认为,潜在化学推理更像软提示而非隐藏在连续空间中的推理链。

正文

View PDF HTML (experimental)

Abstract:Latent reasoning performs intermediate computation in continuous space rather than through explicit tokens, which is particularly appealing in chemistry, where errors in explicit reasoning can propagate to the final answer. Yet if latent states truly constitute an implicit reasoning chain, they should be causally important. We find otherwise: swapping or moderately perturbing intermediate latent states has little effect on performance, while removing them entirely sometimes causes a substantial score drop. Thus, the latent states matter, but their molecule-specific content appears to matter far less. This points to a simpler interpretation. Rather than encoding a step-by-step reasoning chain, latent states may primarily steer subsequent model computation, much like soft prompts. Indeed, a single fixed prefix shared across molecules reproduces LatentChem's performance, and a learned static soft prompt performs comparably to per-input latent generation in controlled models. The tolerance of latent sequences to substitution and moderate perturbation further supports this connection. We therefore suggest that latent chemical reasoning may be better viewed as soft prompting than as a reasoning chain hidden in continuous space.
Comments: 27 pages, 4 figures, 20 tables
Subjects: Machine Learning (cs.LG); Computation and Language (cs.CL)
Cite as: arXiv:2610.34173 [cs.LG]
  (or arXiv:2610.34173v1 [cs.LG] for this version)
  https://doi.org/10.48550/arXiv.2610.34173

arXiv-issued DOI via DataCite (pending registration)

Submission history

From: Kaiyuan Lu [view email]
[v1] Sun, 25 Oct 2026 09:03:36 UTC (473 KB)

来源:arXiv cs.CL · arxiv.org