Claude-Sonnet-5.5 yields no extracted CoT tokens; authors propose RECAST

Claude-Sonnet-5.5 yields no extracted CoT tokens; authors propose RECAST

In their 30 October 2026 preprint, Qi Wang, Kang Xu, Chengcheng Wan and Jiangtao Wang argue that extracting hidden reasoning from closed-source models through chat interfaces can fail because the expected side channels are absent or blocked. They report that three existing methods returned zero CoT tokens on Claude-Sonnet-5.5 through a standard chat-completions interface. Their alternative, RECAST, does not expose or recover a hidden reasoning trace. It conditions a decoder on a question and a model’s final answer, then reconstructs a first-person structured rationale, cleans it and produces SFT-ready data. The authors report qualitative reconstructions on Claude-Sonnet-5.5, GPT-6.1 and Gemini 3 using this answer-only approach. The distinction matters: the output is a reconstruction based on the answer, not evidence of the model’s actual intermediate reasoning. The authors describe an open-source agent skill and controlled artifact validation, but say the preprint does not claim large-scale benchmark improvements. Whether reconstructed rationales are faithful to the hidden reasoning, and useful beyond the reported validation, remains unresolved.

Publication: 363d8486-5742-4e66-815a-c344e855c6a2

attribution — pass: The preprint is correctly attributed to Qi Wang, Kang Xu, Chengcheng Wan and Jiangtao Wang. The zero-token result and the RECAST claims are kept as the authors' reports.

Authors:Qi Wang, Kang Xu, Chengcheng Wan, Jiangtao Wang

summary — uncertain: The facts are faithful, but central terms stay opaque to a non-specialist. CoT is never expanded to chain-of-thought, and SFT-ready (data for supervised fine-tuning) and chat-completions interface are unexplained. The draft never says the pipeline's purpose is producing training data from closed frontier models, so the stakes are unclear. Smallest repair: expand CoT once, gloss SFT via tooltip or a short clause, and add one sentence stating that the output is meant as fine-tuning data. Minor: the source calls only the GPT-6.1 and Gemini 3 results qualitative; Claude-Sonnet-5.5 is the instantiated target.

reconstructs a first-person structured rationale, cleans it, and builds SFT-ready data

date — pass: originalPublished 2026-10-30 matches the v1 submission date on the arXiv page. The 2 November date is only the feed entry.

[Submitted on 30 Oct 2026]

support — pass: The claims match the abstract: zero CoT tokens from three methods, the structural failure, answer-conditioned reconstruction, the open-source agent skill, and no claimed benchmark gains. The point that reconstructions are not evidence of actual hidden reasoning, and that faithfulness is unresolved, follows from the answer-only protocol and is presented cautiously. Claim 3 lists Claude-Sonnet-5.5 among the qualitative results, but its excerpt covers only GPT-6.1 and Gemini 3. Claude reconstruction is supported elsewhere in the abstract.

it does not claim large-scale benchmark improvements

media — pass: The selected media is the verified direct arXiv abstract page for this paper, the most direct access to the source.