PaperlayerView exact revision on arXiv

Source-linked AI summary

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

arXiv:2609.26637v1cs.CLcs.AIcs.CR

TL;DR

Closed-source models conceal native chain-of-thought, leaving benchmark accuracy unable to show how they reach answers. This paper elicits intermediate reasoning through a standard-API tool and evaluates the traces against native reasoning on open models and benchmark behavior on closed models. The extracted traces perform close to native reasoning, while Astra’s traces are especially concise and direct.

  • Problem

    Closed-source providers typically conceal native CoT, limiting evidence about how models reach answers and whether their reasoning is sound or spurious.

    Source: §1 · p. 1

  • Method

    A standard-API tool protocol forces frontier models to externalize intermediate reasoning, which the paper validates against native CoT on open models and evaluates behaviorally on closed models.

    Source: §1 · pp. 1–2

  • Results

    Extracted traces achieve near-native performance on open models and performance close to native reasoning on closed models; Astra’s traces are shortest and more direct, with less branching.

    Source: §1 · p. 2

  • Takeaways & Limitations

    The traces provide a behavioral basis for comparing how frontier models externalize and organize reasoning beyond aggregate benchmark scores.

    Source: §7 · p. 10

    For closed-source models, the evidence cannot establish that extracted text is genuine internal reasoning rather than a useful behavioral proxy.

    Source: §5 · p. 9

Abstract

from arXiv · show

The rapid capability gains of frontier language models are widely attributed to improved reasoning abilities, yet this cannot be verified as raw CoT traces in closed-source systems are hidden. By registering a simple custom tool through a standard API feature, we induce frontier models to externalize intermediate reasoning. Because these traces may reflect post-hoc rationalization rather than genuine reasoning, we first evaluate against native CoT on open-source models and extend to closed-source frontier models including GPT-6 Astra. We find that the extracted reasoning matches native reasoning performance and substantially outperforms no-reasoning baselines, across competition mathematics, science, and code generation. We then characterize how frontier models structure their intermediate reasoning. Across token efficiency, reasoning-step types, and induced reasoning trees, we identify systematic differences in how models externalize, compress, and organize reasoning. We find that Astra exhibits token-efficient directed reasoning, selecting a correct trajectory earlier, while resolving elementary steps internally and externalizing only crucial reasoning. These findings provide a behavioral lens on frontier-model reasoning beyond benchmark scores.

1 Introduction

Closed-source models’ hidden reasoning creates a gap between knowing what they solve and how they reach an answer. The paper uses tool-elicited traces to validate a behavioral proxy and compare how frontier models externalize and organize reasoning.

1 Introduction

  • Closed-source systems conceal native CoT, so benchmark accuracy alone cannot show how a model reaches an answer or distinguish sound from spurious reasoning.This limits verification where conclusions must be traced to specific inferential steps.Source: §1 · p. 1
  • A standard-API tool protocol forces an initial reasoning-tool call, then replays acknowledgments and allows automatic tool selection for subsequent calls.The traces show reasoning-like content can still be externalized through other API fields when designated reasoning controls do not fully prevent it.Source: §1 · pp. 1, 3
  • Extracted traces recover near-native task performance on open models and achieve performance close to native reasoning across closed-source frontier models.The open-model traces also show lexical and structural similarity to native CoT, supporting their use as a behavioral proxy.Source: §1 · p. 2
  • Astra produces short, least-compressible traces with direct, lightly branched paths, while other frontier models externalize longer, more exploratory reasoning.Across models, reasoning-operation types are broadly similar; Astra often leaves elementary steps implicit and externalizes higher-level reasoning.Source: §1 · pp. 2–3, Figure 1 · p. 2
  • Compact traces transfer unevenly: stronger recipient models use them with little loss, while weaker recipients may fail to produce answers already stated in the trace.The usefulness of reasoning data as supervision therefore depends on the model reading it.Source: §1 · p. 3
  • The extracted traces provide a behavioral basis for comparing frontier reasoning beyond aggregate benchmark scores.The comparison examines how much models externalize and how their reasoning traces are organized.Source: §1 · pp. 2–3

2 Related Work

Related work covers methods for exposing or reconstructing hidden reasoning, frameworks for analyzing reasoning-trace structure, and techniques for studying or controlling CoT length.

2 Related Work

  • Prior exposure methods include replaying provider-returned reasoning blocks, reconstructing traces from outputs, and prompting visible reasoning; EchoCoT archives reasoning through repeated tool interactions.The paper situates its tool-based approach among these methods.Source: §2 · p. 3
  • Studies of reasoning outside designated thinking channels find that models can still emit reasoning when thinking is disabled, motivating tools as an alternative workspace for intermediate text.The paper uses this setting to compare frontier-model reasoning, not only to recover useful traces.Source: §2 · p. 3
  • Trace-analysis work decomposes reasoning into functional episodes, hierarchical trees, or progression graphs, while other work studies redundancy and CoT compression.The paper builds on these perspectives to compare efficiency, local expression, and global structure.Source: §2 · p. 4

3 Extracting and Validating Reasoning Traces

The section introduces FORCED-REASONING, a tool-channel protocol, and evaluates extracted traces against native reasoning on open models and benchmark performance on closed-source models.

3 Extracting and Validating Reasoning Traces

  • The section validates open-model traces directly against observable native CoT and evaluates closed-source traces through benchmark performance.Native CoT is unavailable for the closed-source frontier models, so their validation relies on behavioral evidence.Source: §3 · p. 4, §3.3 · p. 5

3.1 Extracting reasoning traces via the FORCED-REASONING protocol

  • FORCED-REASONING forces the model’s initial selection of a reasoning tool, records its free-form text, then replays a content-free acknowledgment and restores automatic tool choice.The tool performs no external computation; it makes intermediate text visible and retains it in the conversation.Source: §3.1 · p. 4

3.2 Experimental Setup

  • The setup evaluates mathematical reasoning, code generation, and multidisciplinary problem solving using MATH, LiveCodeBench, and Humanity’s Last Exam.It compares None (native reasoning disabled), Native (high effort), and Forced; Astra uses its lowest available native effort because it cannot disable native reasoning.Source: §3.2 · p. 4

3.3 Validating Extracted Reasoning

  • On open models, extracted traces recover near-native task performance, substantially outperform no-reasoning baselines, and show lexical overlap and broadly similar functional structure to native CoT.These results support using the traces as a behavioral proxy for native reasoning.Source: §3.3 · pp. 4–5
  • Across closed-source benchmarks, FORCED-REASONING achieves performance close to native reasoning and substantially outperforms the no-reasoning baseline.Table 1 reports accuracy (%) on MATH, HLE, and LiveCodeBench; Astra does not support the None setting.Source: Table 1 · p. 5, §3.3 · p. 5

4 Characterizing extracted reasoning traces of Frontier Models

The section compares frontier models’ extracted reasoning traces in length, local expression, global structure, and cross-model reuse. It finds that Astra’s traces are compact and direct, but their usefulness varies with the recipient model.

4.1 Output Length and Trace Compressibility

  • Astra has the highest zlib ratio across all three benchmarks, indicating that its extracted reasoning text is less compressible under this measure.Figure 2 compares recorded output length with trace compressibility across model and benchmark combinations.Source: §4.1 · p. 5, Figure 2 · p. 6

4.2 Local Reasoning Granularity

  • In matched examples, Astra verbalizes fewer intermediate arithmetic and algebraic steps and uses background knowledge without restating it, yielding the most compact local expression.The examples compare models’ expressions of corresponding reasoning steps on the same problems.Source: §4.2 · pp. 6–7, Table 2 · p. 6

4.3 Global Reasoning Structure

  • Models show broadly similar reasoning-activity profiles, with analysis and implementation largest; Astra retains this profile despite substantially shorter traces.This suggests Astra’s brevity does not come from removing entire classes of reasoning activity.Source: §4.3 · p. 7
  • Astra’s reasoning trees are substantially narrower and have fewer nodes while reaching comparable median depth, indicating less branching, revisiting, and trial-and-error.Table 3 summarizes tree width, depth, and size across MATH problems; Figure 4 shows representative trees with comparable spacing.Source: §4.3 · pp. 7–8, Table 3 · p. 8, Figure 4 · p. 8

4.4 Cross-Model Reuse of Reasoning Traces

  • Astra’s traces transfer less effectively to weaker recipients, which may fail despite the trace already stating the correct answer; stronger recipients recover nearly all donor accuracy.The shortfall is relative: weaker recipients still improve substantially over unaided performance, and the pattern is not strictly ordered by recipient capability.Source: §4.4 · pp. 8–9
  • The paper proposes that trace value as supervision may depend on the recipient, because compressed traces omit steps weaker students may not reconstruct.The authors predict that expanding implicit intermediate steps could restore much of the usefulness for weak recipients while leaving strong recipients largely unaffected.Source: §4.4 · p. 9
  • The reuse experiment measures in-context use, not post-training distillation, and differences in trace correctness, content, and style prevent isolating compression as a causal factor.The proposed expansion test is a prediction, not a direct distillation result.Source: §4.4 · p. 9

5 Limitations

For closed-source models, the extracted traces are supported only by behavioral evidence, not proof of internal reasoning; the protocol also depends on specific API controls.

5 Limitations

  • Closed-source traces may be useful behavioral proxies, but the evidence cannot establish that they reflect the models’ actual internal computation.Open-model traces are validated against native CoT, but that comparison is unavailable for closed-source models.Source: §5 · p. 9
  • The protocol applies only to API-as-a-service endpoints that support custom tools and forced selection of a named tool.Endpoints without those controls fall outside its scope.Source: §5 · p. 9
  • Similarity to native traces and downstream utility should not be treated as evidence that extracted text is identical to internal computation.Source: §5 · p. 9

6 Ethical Considerations

The work evaluates a confidentiality and capability-control boundary using benign public benchmarks and researcher accounts, and reports that extracted content may not be genuine internal CoT.

6 Ethical Considerations

  • Experiments used benign public benchmarks and accounts available to the researchers.Source: §6 · p. 10
  • The study did not extract personal data, proprietary prompts, credentials, or unsafe content.Source: §6 · p. 10
  • All extraction experiments were completed before September 9, 2026, and observations were reported to the security teams at OpenAI and Anthropic.The researchers also provided code to reproduce the procedure; mitigations are discussed in Appendix J.Source: §6 · p. 10

7 Conclusion

The paper uses extracted intermediate reasoning to compare frontier models on properties that benchmark accuracy alone cannot reveal. GPT-6 Astra produces shorter, less redundant traces while retaining broad reasoning operations, but trace reuse varies with the recipient model.

7 Conclusion

  • GPT-6 Astra produces the shortest, least redundant traces while performing the same range of reasoning operations as models that write considerably more.Source: §7 · p. 10
  • Astra often resolves routine computations and simple derivations internally, following more direct solution paths with less branching.Source: §7 · p. 10
  • Strong recipient models exploit Astra’s traces almost fully, whereas weaker recipients use them only partly.Source: §7 · p. 10
  • The explicitness of a reasoning trace is worth considering separately from the strength of its source model.The most capable teacher may not be the most useful source of supervision for a given student.Source: §7 · p. 10

A.1 Evidence of Reasoning without Reasoning – Validation on Open Models

Open-model validation finds that forced reasoning recovers native-level performance and resembles native CoT in lexical and coarse structural features. Closed-model comparisons and the tool protocol extend the analysis, with important conditions on their interpretation and use.

A.1 Evidence of Reasoning without Reasoning – Validation on Open Models

  • Forced reasoning raises DeepSeek-V4-Flash MATH accuracy from 30.5% without reasoning to 73.3%, versus 70.0% with native reasoning.GLM-5.2 reaches 84.3% with forced reasoning, compared with 21.4% without reasoning and 89.9% with native reasoning.Source: §A.1 · p. 12, Figure 6 · p. 13
  • Forced traces show substantial ROUGE overlap with native traces, often on a similar scale to within-mode comparisons; GLM-5.2’s median forced-to-native length ratio is 1.39×.Table 4 compares ROUGE-1 and ROUGE-L F1 and contiguous word 5-gram Jaccard similarity.Source: §A.1 · p. 13, Table 4 · p. 13
  • On HMMT, forced and native reasoning exhibit the same broad functional repertoire and broadly similar episode composition, though category allocations are not identical.The traces are segmented into reasoning units and assigned to seven substantive categories plus OTHER.Source: §A.1 · p. 13, Figure 7 · p. 14

A.2 Closed Models Think Out Loud, Too

  • Across four closed-source models, forced reasoning generally achieves accuracy close to native reasoning, matching or exceeding it in several settings.Native CoT is unavailable for direct comparison, so the evaluation pairs benchmark performance with provider-reported reasoning-token usage.Source: §A.2 · p. 14
  • For GPT-5.6-sol on HMMT, forced tool calling reaches 93.9% with maximal deliberation, while directly requesting visible step-by-step reasoning reaches 60.6%.Moving the maximal-deliberation instruction into the visible-response prompt also yields 60.6%.Source: §A.2 · p. 14

B Forced reasoning tool

  • The protocol uses a custom reasoning-string tool as an alternative workspace, forcing its initial selection and replaying its arguments in the conversation.The tool performs no computation; prompt sensitivity varies across models.Source: B Forced reasoning tool · p. 15

C Benchmark Sampling

The evaluation samples competition mathematics, code generation, and multidisciplinary questions, while the model-behavior analyses examine extraction success and GPT-5.6-Sol’s sensitivity to tool wording. Stronger tool wording lengthens Sol’s traces and improves performance, unlike direct no-tool prompts, which produce shorter, structured answers without further accuracy gains.

C Benchmark Sampling

  • 100 questions each are evaluated for LiveCodeBench and Humanity’s Last Exam, alongside 80 competition-math problems.The HLE sample is balanced across seven categories, with 15 questions each from Biology/Medicine and Computer Science/AI and 14 from each other category.Source: C Benchmark Sampling · pp. 15–16

D Different Model Behavior

  • 100% of completed runs yielded nonempty tool content for Opus 4.8, Sonnet 5, and GPT-5.6-sol, with no observed model-issued flags or warnings.These runs were conducted before September 9, 2026.Source: D Different Model Behavior · p. 16

D.1.1 GPT-5.6-Sol: effects of tool calling and prompt wording

  • Stronger tool wording lengthens GPT-5.6-Sol’s traces and improves task performance, moving MATH performance from roughly native-low to approximately native-medium range.The maximal-deliberation description explicitly requests intermediate work, alternatives, and verification in the scratchpad.Source: §D.1.1 · p. 16
  • Stronger no-tool wording adds no accuracy gain, and these outputs are far shorter than tool calls and mostly structured answers rather than CoT-style text.The Plain HMMT P11 response illustrates this distinction with a concise numbered solution and bold case headings.Source: §D.1.1 · pp. 16–17

D.1.2 GPT-6-Astra: effort, wording, and reasoning-channel allocation

Astra responds differently from Sol to stronger tool wording, while experiments with higher native effort show that substantial visible reasoning and high accuracy can coexist with less reliable clean extraction. The section also reports a refusal boundary for tested Claude configurations and defines measures of output length and compressibility.

D.1.2 GPT-6-Astra: effort, wording, and reasoning-channel allocation

  • Astra produces less visible tool reasoning under maximal-deliberation wording than under the default, retaining substantial native-channel usage, whereas Sol externalizes more.The authors interpret this as reluctance to externalize reasoning rather than reluctance to answer, and hypothesize that think-here wording may work better for capable models.Source: §D.1.2 · pp. 17–18
  • The prompt designs are heuristic, were not optimized for attack or prompt engineering, and do not establish a general prompt about model capability.Their stated aim is performance comparable to native reasoning under the forced-reasoning protocol and analysis of visible traces.Source: §D.1.2 · p. 18

D.2 Forced reasoning with native reasoning enabled

  • At high and xhigh native effort, Sol answers 44/47 (93.6%) and 47/47 (100%) APEX questions correctly, while Astra answers 47/47 at both efforts.These experiments retain the forced-tool protocol while native reasoning is enabled.Source: §D.2 · p. 18
  • Mean API-reported non-native output rises from roughly 15.0k to 19.7k tokens for Sol and 5.8k to 12.0k for Astra as native effort increases.These totals include tool framing but exclude separate final-answer calls; whole-interaction zero-native runs become less reliable, especially for Astra, with 18/47 at high effort and 7/47 at xhigh.Source: §D.2 · p. 19

D.3.1 Newer Claude models: an observed refusal boundary

  • The forced-reasoning protocol had 0% extraction success on tested configurations of Claude Opus 5, Fable 5, and Fable 5.1.The observations establish a failure boundary for the tested protocol but do not identify whether the restriction comes from the model or provider-side handling.Source: §D.3.1 · p. 19

E Native and Forced Token Counts

  • For Figure 2, OF is mean API-reported completion minus native reasoning tokens; the zlib ratio is compressed-to-original UTF-8 bytes for extracted reasoning traces.OF includes scratchpads, tool framing, and final answers; the zlib ratio uses level 9 compression.Source: E Native and Forced Token Counts · p. 19

F Constructing Reasoning Trees from Episode-Annotated Traces

The method maps episode-annotated thoughts onto an ordered reasoning sketch, distinguishing execution of a step from reference to it, then builds trees that represent visible reasoning trajectories. Tree metrics characterize this representation, but fewer nodes do not establish fewer internal computations or necessary proof steps.

F Constructing Reasoning Trees from Episode-Annotated Traces

  • The method represents each trace as an ordered reasoning sketch and a tree placing passages at sketch steps, using semantic episode boundaries and retaining brief computations and checks as leaves.It distinguishes executing a step from referring to it.Source: F Constructing Reasoning Trees from… · p. 19
  • A new thought begins at each EXPLORE, VERIFY, or MONITOR unit; consecutive units in those categories begin separate thoughts, while other units continue the current thought.The same segmentation rule applies to all models.Source: F Constructing Reasoning Trees from… · p. 20
  • For each thought, the judge assigns steps it carries out, derives, checks, or revises separately from steps whose results it merely cites or uses.A thought may execute several steps or none, and assignments must correspond to steps and thoughts in the trace representation.Source: F Constructing Reasoning Trees from… · p. 20
  • Thoughts are added in textual order; revisiting a step creates a new node, while qualifying reference-only checks may be retained as short-check leaves.The tree follows the current branch back to an ancestor below the thought’s first assigned step, then adds nodes for its assigned steps; retained leaves contribute to width and size.Source: F Constructing Reasoning Trees from… · p. 20
  • Tree width p is the maximum non-root nodes at a sketch-step level, depth q is the largest occupied step index, and size N is the total non-root nodes.Depth measures progression through the sketch rather than root-to-node path length.Source: F Constructing Reasoning Trees from… · pp. 20–21
  • Tree metrics depend on segmentation and step assignment; smaller trees indicate fewer represented nodes but do not establish fewer internal computations or necessary proof steps.Source: F Constructing Reasoning Trees from… · p. 21
  • Astra has smaller mean tree width and size but greater mean depth; the main-text comparison of similar depths refers to medians.Table 9 reports means over the same traces for which Table 3 reports medians and quartiles.Source: F Constructing Reasoning Trees from… · p. 21, Table 9 · p. 21

G Example of Logic Puzzle

A logic-puzzle example illustrates how Astra compresses its externalized reasoning, while the cross-model analysis finds that its shorter traces mainly reflect fewer surfaced intermediate steps.

G Example of Logic Puzzle

  • Astra omits many intermediate deductions and uses telegraphic fragments to state derived results directly in the logic-puzzle trace.The trace often encodes conclusions in compact noun phrases rather than fully grammatical explanations.Source: G Example of Logic Puzzle · p. 24

H Episode-Level Analysis Across Frontier Models

  • The analysis compares how four frontier models externalize reasoning activity, its contribution to trace length, and its distribution over reasoning trajectories.The comparison uses Forced traces and episode annotations.Source: H Episode-Level Analysis Across… · p. 25, §H.1 · p. 25

H.1 How Do Models Externalize Their Reasoning?

  • Astra’s compact traces have episode units similar in length to Sol and Opus, but expose fewer intermediate operations overall.Opus and especially Sonnet contribute several-fold more text in categories such as ANALYZE, PLAN, and IMPLEMENT.Source: §H.1 · pp. 25–26, Figure 8 · p. 25, Figure 9 · p. 26

H.2 Compactness Does Not Remove the Reasoning Repertoire

Astra’s shorter traces retain the broad reasoning repertoire and several coarse temporal patterns seen in other frontier models. Compactness therefore reflects fewer surfaced steps, not a single dominant behavior.

H.2 Compactness Does Not Remove the Reasoning Repertoire

  • All four models retain substantial ANALYZE and IMPLEMENT activity alongside READ, PLAN, EXPLORE, VERIFY, and MONITOR, despite differing episode mixtures.The episode-composition figure reports mean proportions of annotated units by category.Source: §H.2 · pp. 26–27, Figure 10 · p. 26
  • Across models, READ is concentrated near the beginning, VERIFY grows toward the end, and IMPLEMENT is generally more prevalent in the middle and later trajectory.The detailed progression varies by model, but these broad temporal patterns are shared.Source: §H.2 · p. 27, Figure 11 · p. 27

I Exposed Answer but still wrong

The examples show that recipients can miss correct answers already exposed in Astra’s traces. The appendix also compares Native-high and Forced reasoning structure, details its annotation procedure, and tests whether donor prefixes shift recipient reasoning text.

I Exposed Answer but still wrong

  • On HMMT P20, all three recipients failed; on APEX P47, Haiku returned 10225 while GPT-5.4-nano and DeepSeek-V4-Flash returned 10225 −1.For APEX P10, DeepSeek-V4-Flash returned 129 while Haiku and GPT-5.4-nano returned 0.Source: I Exposed Answer but still wrong · pp. 28–29
  • 2/3 GPT-5.4-nano answers and 4/7 DeepSeek-V4-Flash answers were incorrect on HMMT P11 despite the short trace containing the numerator and denominator.The recipient can fail even when Astra’s scratchpad explicitly states the correct answer.Source: I Exposed Answer but still wrong · p. 28

J Possible Mitigations

  • Possible defenses include restricting free-form reasoning tools or inspecting tool arguments, but their effectiveness and impact on legitimate tool use require evaluation.The proposed scope extends reasoning-disclosure policies across native reasoning, tool arguments, and visible responses.Source: J Possible Mitigations · p. 29

K.1 Functional Composition of Reasoning

  • Native-high and Forced traces share the same broad functional repertoire, but their episode proportions are not identical, supporting coarse similarity rather than category-by-category equivalence.The HLE comparison averages per-trace proportions across the seven episode types.Source: §K.1 · p. 29, Figure 12 · p. 30

K.2.1 Episode Transitions

  • Both conditions show persistent ANALYZE and IMPLEMENT states and recurring analysis–execution–verification transitions, while differences often involve VERIFY and MONITOR.The analysis compares row-normalized transitions and their absolute differences.Source: §K.2.1 · p. 30, Figure 13 · p. 30

K.2.2 Temporal Organization

  • On HLE, Native-high and Forced traces share a coarse temporal structure, but Forced traces emphasize analysis, planning, and execution while Native-high exposes more exploration and self-regulation.Traces are normalized to relative progress and averaged so unusually long traces do not dominate.Source: §K.2.2 · p. 31, Figure 14 · p. 31

L LLM-as-a-Judge Annotation Details

  • Traces are segmented mechanically and then labeled semantically by an LLM judge; an explicit prompt reached 90% agreement versus 83% for a minimal prompt in a 200-unit pilot.Claude Opus 5 annotators receive local context, while model, condition, benchmark, problem, and answer-correctness identities are withheld.Source: L LLM-as-a-Judge Annotation Details · pp. 31–32

M Reasoning-Prefix Echo Across Models

Related papers

Loading 2609.26637v1…