Capable yet Parsimonious: Extracting and Characterizing Hidden Chain-of-Thought in Frontier Models
Xiaoyu Luo, Tao Ren, Wenrui Yu, Xiao Li, Qiongxiu Li, Johannes Bjerva
cs.CL · cs.AI · cs.CR
Quick verdict
Tool-forcing elicits near-native hidden reasoning in frontier models; the study is credible, but traces' identity with internal CoT remains unproven.
Summary
The paper introduces FORCED-REASONING: register a no-op scratchpad tool through standard model APIs, force its first selection, and replay calls with a fixed acknowledgment, inducing closed-source models to externalize intermediate reasoning with native thinking disabled. Validation on open models shows near-native MATH accuracy (DeepSeek-V4-Flash 73.3% vs 70.0% native; GLM-5.2 84.3% vs 89.9%) plus lexical and structural similarity to native CoT. On Opus 4.8, Sonnet 5, GPT-5.6 Sol, and GPT-6 Astra (80 MATH, 100 HLE, 100 LiveCodeBench problems), forced reasoning approaches native accuracy. Characterization finds models differ mainly in how much reasoning they externalize: Astra's traces are shortest, least compressible, and its reasoning trees narrowest (median width 5.0 vs 28.5 for Sonnet 5), while the episode repertoire stays similar. Transplanted Astra traces are reused almost losslessly by strong recipients but partly wasted by weaker ones that ignore explicitly stated answers. Limitations: traces are admittedly behavioral proxies whose identity with internal computation is untestable; Astra keeps low-effort native reasoning and uses model-specific prompts; samples are small with no uncertainty estimates; no public code release is stated.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 3
● low 2
4 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- A standard tool-forcing protocol elicits intermediate reasoning from frontier models that matches native-CoT task performance and beats no-reasoning baselines, validated directly on open models (App A.1) and behaviorally on closed models (Table 1).
- Frontier models differ mainly in how much reasoning they externalize, not in which reasoning operations they perform (Sec 4.3, Figs 8-10).
- GPT-6 Astra externalizes sparsely and directly: shortest, least compressible traces and narrowest reasoning trees (Table 3: median width 5.0, 27 nodes), resolving elementary steps without verbalizing them.
- Trace utility is reader-relative: compact traces are reused almost losslessly by strong models but partly wasted by weak ones, which sometimes ignore answers explicitly stated in the trace (Sec 4.4, Table 10).
- Reasoning-mode controls do not fully prevent reasoning externalization through other API channels (Sec 3, App D), though Claude Opus 5 and Fable 5/5.1 block the protocol entirely (App D.3).
Strengths
Validation ladder before closed-model claims
App A.1 verifies three levels before any closed-model claim: performance (Fig 6: DeepSeek-V4-Flash 73.3% forced vs 70.0% native, vs 30.5% no-reasoning; GLM-5.2 84.3% vs 89.9%), lexical overlap benchmarked against within-mode N-N pairs (Table 4: ROUGE-1 0.538 forced vs 0.613 N-N for DeepSeek), and episode-structure similarity (Fig 7). Using N-N pairs as the reference band is a sound control.
Unusually honest scope discipline
Sec 6 reports disclosure to OpenAI and Anthropic security teams with reproduction code; Secs 5-6 and App F openly state the proxy-status caveat and the median-vs-mean depth discrepancy (Astra mean depth 14.9 exceeds all models); App D.3 reports the 0% extraction failure on Opus 5/Fable 5/5.1 rather than hiding it.
Simple, fully specified, ablation-supported protocol
Novel behavioral findings with quantified failure modes
Weaknesses
Proxy-identity gap between title and stated limitation
No uncertainty quantification on headline closed-source results
Cross-model comparisons confounded by model-specific prompts and Astra's non-disabled native channel
Structural metrics entirely dependent on LLM-judge pipelines without human gold labels
Time-locked, non-released artifacts limit reproducibility
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
The paper asks what distinguishes frontier-model reasoning beyond benchmark accuracy, given that closed providers hide native CoT. It contributes (1) the FORCED-REASONING protocol - a no-op scratchpad tool registered through standard APIs, forced first selection via `tool_choice`, fixed-acknowledgment replay (App B); (2) validation of extracted traces against native CoT on open models; (3) characterization of Opus 4.8, Sonnet 5, GPT-5.6 Sol, and GPT-6 Astra across 80 MATH, 100 HLE, and 100 LiveCodeBench problems; and (4) a cross-model trace-transfer experiment with distillation implications. Central claims: extracted traces behave like reasoning (Sec 3.3); models differ mainly in how much they externalize (Sec 4.3); Astra externalizes sparsely and directly (Secs 4.1-4.3); trace utility is reader-relative (Sec 4.4). The title's 'hidden chain-of-thought' sits uneasily with Secs 5-6, which concede the traces may be behavioral proxies rather than genuine internal computation - the paper's most important self-aware caveat.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 4 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is the FORCED-REASONING technique in this paper?
It registers a custom scratchpad tool through a standard model API, forces the model to call it first, and records the free-form text the model writes into the tool, replaying each call with a fixed acknowledgment. This makes frontier models write intermediate reasoning into a visible channel even when their native thinking mode is disabled or hidden by the provider.
Does the paper prove the extracted text is the model's real internal chain of thought?
No. The authors show it behaves like reasoning (near-native task accuracy, and similar wording and structure to native CoT on open-source models), but they explicitly state they cannot verify it matches a closed model's internal computation; it may be a functional behavioral proxy rather than the genuine hidden chain of thought.
What did the paper find about GPT-6 Astra's reasoning style?
Astra produced the shortest and least compressible traces and the narrowest reasoning trees (median width 5.0 versus 28.5 for Claude Sonnet 5) while keeping a similar repertoire of reasoning activities. It appears to resolve routine steps internally and externalize only higher-level reasoning. Its compact traces transfer almost losslessly to strong recipient models but poorly to weaker ones.
Is this an attack on model providers' CoT confidentiality?
It exploits only a standard tool-calling API feature on benign public benchmarks. The authors frame it as an observation about confidentiality and control boundaries, reported by email to OpenAI and Anthropic security teams with reproduction code. They also document that newer Claude models (Opus 5, Fable 5/5.1) already reject the protocol entirely.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figure-internal data is not verifiable from the extracted text: the compressibility scatter (Fig 2), the transfer plot (Fig 5), and most validation figures (Figs 6-14) exist only as captions, so magnitudes behind two headline claims could not be independently checked.
- Rollout counts per closed-source Table 1 cell are not stated in the text; uncertainty assessments are inferred from the stated 80-100 item subset sizes.
- The extraction appears complete (main text through Appendix M), but provider-reported token counts and prompt boxes could not be checked against any artifact.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.