Pith. sign in

REVIEW 3 major objections 5 minor 2 cited by

Chain-of-thought reliability can be diagnosed by three intervention-based log-probability metrics that distinguish post-hoc, encoded, and internalized reasoning pathologies.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-02 23:21 UTC pith:OVN5TVWX

load-bearing objection A cheap, clearly-specified metric toolkit for CoT pathologies, but the internalized-reasoning validation is underdetermined; worth refereeing with demands for stronger ground truth. the 3 major comments →

arxiv 2602.13904 v2 pith:OVN5TVWX submitted 2026-02-14 cs.AI

Diagnosing Pathological Chain-of-Thought in Reasoning Models

Manqing Liu , David Williams-King , Ida Caspary , Linh Le , Hannes Whittingham , Puria Radmard , Cameron Tice , Edward James Young This is my paper
classification cs.AI
keywords chain-of-thoughtCoT pathologieslog-probability metricspost-hoc rationalizationencoded reasoninginternalized reasoningmonitorabilitysupervised fine-tuning
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper tries to establish that the health of a model's chain-of-thought (CoT) can be assessed cheaply and without task-specific assumptions by measuring how the probability of the final answer changes under three interventions: removing the CoT, paraphrasing it, and replacing it with irrelevant filler. It introduces three metrics—Necessity, Paraphrasability, and Substantivity—each targeting a known failure mode: post-hoc rationalization, encoded reasoning, and internalized reasoning. To test the metrics, the authors train 'model organisms' deliberately fine-tuned to exhibit each pathology and show that each metric produces the expected signature, with different pathologies becoming detectable at different training checkpoints. If correct, this gives developers and third-party evaluators a lightweight way to decide when CoT is usable for monitoring and when it is not.

Core claim

The paper claims that three task-agnostic, computationally cheap metrics—Necessity, Paraphrasability, and Substantivity—can tell whether a model's chain-of-thought is genuinely load-bearing and transparent. Necessity measures how much the answer probability drops when the CoT is removed; Paraphrasability measures how much it drops when the CoT is semantically paraphrased; Substantivity measures how much it drops when the CoT is replaced by irrelevant same-length filler. The authors construct controlled 'model organisms' by supervised fine-tuning to exhibit post-hoc rationalization, encoded (codebook-obfuscated) reasoning, and internalized (cross-domain filler) reasoning, and report that each

What carries the argument

The central machinery is the intervention-based log-probability comparison: for a question, sampled CoT, and answer, each metric contrasts the answer's log-probability under the original CoT with its log-probability under a specific intervention—removal (Necessity), paraphrase (Paraphrasability), or substitution with irrelevant filler (Substantivity)—combined into a signed ratio in (-1,1). The differential effect of each intervention isolates a distinct mode of dependence on the CoT: presence-dependence, surface-form-dependence, and content-dependence.

Load-bearing premise

The validation assumes that the fine-tuned 'model organisms' genuinely instantiate the intended pathologies—especially that the internalized organism performs computation in forward passes rather than just learning to ignore the chain-of-thought—so the measured signatures come from real pathologies, not from training artifacts the metrics happen to be sensitive to.

What would settle it

Probe the internalized model organism with activation-based probes during the filler chain-of-thought: if task-relevant computation is absent while the Substantivity score stays near zero, the metric is detecting trained indifference rather than hidden computation. Alternatively, test a model that demonstrably computes in latent space and check whether the metrics fail as predicted.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Monitoring during training: tracking these metrics across checkpoints can alert developers when a pathological CoT begins to emerge.
  • Third-party auditing: because the metrics are task-agnostic, external evaluators can apply them to any model and task to decide whether reasoning traces are load-bearing.
  • Differential diagnosis: each metric maps to a distinct pathology, so the triple of scores indicates which failure mode a model is exhibiting.
  • Temporal coverage: pathologies appear at different training checkpoints, so reliable diagnosis requires evaluating the suite over time rather than once.
  • Boundary of applicability: the metrics assume text-space CoT; models that reason in latent space fall outside their scope.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The Substantivity metric may classify a model as 'internalized' even when it has merely learned to ignore the CoT entirely; pairing the metric with internal-state probing would separate genuine hidden computation from trained indifference, a distinction the paper does not make.
  • The three metrics form a signature that could drive an automated training-time alarm that halts fine-tuning once the CoT ceases to be substantive; the paper provides the data to support such a controller but does not implement it.
  • A natural next experiment is to apply the three metrics to large production reasoning models on tasks they can solve without CoT; if such models show the post-hoc signature, it would delimit which tasks are actually safe to monitor.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The paper proposes three task-agnostic, computationally lightweight log-probability metrics—Necessity, Paraphrasability, and Substantivity—to diagnose three chain-of-thought (CoT) pathologies: post-hoc rationalization, encoded reasoning, and internalized reasoning. Each metric compares the answer log-probability under the original CoT with the answer log-probability under a causal intervention: removing the CoT, paraphrasing it, or replacing it with irrelevant filler. To validate the metrics, the authors construct model organisms via supervised fine-tuning (SFT) on deliberately modified datasets and track Cohen's d of each metric across SFT checkpoints. They report that the metrics successfully identify and distinguish the three pathologies.

Significance. If the validation held, this would be a practically useful toolkit: the metrics are simple, model-agnostic, and cheap to compute, and the checkpoint-level analysis is a sensible design choice. Strengths include the transparent metric formulations (Eqs. 3–9), the detailed appendices with complete codebooks and prompt tables, and a sensitivity analysis with an alternative paraphraser. However, the empirical validation is largely a consistency check: each model organism is constructed so that the expected metric signature is nearly guaranteed by the prompt/data design. The internalized organism in particular is not mechanistically verified, so the paper does not yet establish the advertised differential diagnosis of CoT pathologies.

major comments (3)
  1. [§3.3, §3.4 (Eq. 8), Table 6] The internalized organism's Substantivity signature is underdetermined. Training on (Q, cross-domain filler CoT, A) may teach the model to ignore the CoT and map Q→A directly, rather than to perform hidden computation during CoT generation. Under that hypothesis, pSub≈pOrig and Sub≈0, exactly the observed signature. The paper provides no mechanistic test—causal tracing, hidden-state probes, or a control condition where the filler CoT is genuinely load-bearing—to distinguish 'hidden computation' from 'CoT entirely ignored.' The limitation in §6 acknowledges off-distribution confounds but does not resolve this specific one. The claim that Substantivity diagnoses 'internalized reasoning' (§5) is therefore not supported.
  2. [§3.3, Eqs. (3), (6), (8), Table 6] Validation success is largely guaranteed by construction for all three organisms. For the post-hoc organism, the answer remains in the prompt during the Necessity intervention ('The correct answer is <answer>. Do not produce any reasoning...'), so pNec≈pOrig and Nec≈0 regardless of the CoT. For the encoded organism, paraphrasing a codebook-transformed CoT destroys the code by construction, forcing Para<0. For the internalized organism, the Substantivity intervention reuses the same filler prompt used in training, forcing Sub≈0. These experiments show internal consistency between the organism construction and the metrics, not that the metrics can detect pathologies in models that are not already engineered to exhibit them. The conclusion in §5 that 'our metrics can effectively differentiate' is therefore overstated.
  3. [§4, Figure 2, Table 4] The internalized organism's predicted signature requires the model to remain reliant on the CoT (Necessity ≈ +1 in Table 3), yet the paper never demonstrates this reliance. If the internalized model has actually learned a direct Q→A mapping, its Necessity would be ≈0, making the internalized signature indistinguishable from a post-hoc or no-CoT pattern on two of the three metrics. The Cohen's d trajectories in Figure 2 are discussed mainly for Substantivity; the Necessity panel for the internalized organism is not analyzed. This missing evidence is load-bearing for the differential-diagnosis claim.
minor comments (5)
  1. [§3.3 vs. Appendix E.2] The cross-domain filler assignments are inconsistent: §3.3 says Calendar Arithmetic questions receive Largest Island reasoning, while Appendix E.2 lists 'Calendar arithmetic→spell backward reasoning.' Please reconcile.
  2. [§3.4, §3.5, Figure 2] Terminology is inconsistent: the metric is called 'Substantivity' in §3.4 but 'Substitutability' in §3.5 and in the Figure 2 panel labels. Choose one term and use it throughout.
  3. [Table 3] Table 3 has incomplete entries for the post-hoc row: only the Necessity column shows '0'; the expected values for Paraphrasability and Substantivity are missing from the table and not fully specified in the text.
  4. [Figure 7] Figure 7 in Appendix C is referenced but its axes and series are not clearly labeled in the text. Add axis labels and a legend or explicit series descriptions.
  5. [Throughout] Minor typographical issues: 'elcicit' in §1, inconsistent 'Internalised'/'Internalized' spellings, and 'not_relevant filler' is first used without a definition.

Circularity Check

3 steps flagged

Central validation is construction-equivalence: each model organism's expected metric signature is entailed by its SFT training distribution.

specific steps
  1. self definitional [§3.3 (Internalized reasoning); §3.4 (Substantivity, Eq. 8); Appendix A Table 6]
    "each chain of thought is replaced with reasoning from a completely different task domain (not_relevant filler), while the final answer remains unchanged ... thus inducing a form of internalized reasoning by construction. ... pSub :=p M (A|Q∪Irre.,CoT Irre.) ... Substantivity Replacing CoT Internalized-MO Data generation prompt for Internalized-MO."

    The internalized organism is trained on (Q, irrelevant cross-domain CoT, A). The Substantivity intervention replaces the CoT with an irrelevant CoT and, per Table 6, uses the Internalized-MO data-generation prompt — i.e., the training distribution itself. Therefore pSub≈pOrig and Sub≈0 is entailed by construction; the metric cannot distinguish hidden forward-pass computation from "CoT entirely ignored." The claimed validation of internalized-reasoning diagnosis thus reduces to the training recipe.

  2. self definitional [§3.3 (Post-hoc reasoning); §3.4 (Necessity, Eqs. 3-4); Appendix A Table 6]
    "each prompt explicitly reveals the correct answer by appending the phrase “The answer is {number}” to the question ... Because the answer is provided upfront, any generated reasoning cannot contribute causally to answer derivation and therefore constitutes post-hoc justification by design. ... Necessity Removing CoT Post-hoc-MO The correct answer is <answer>. Do not produce any reasoning within your thinking tags."

    Post-hoc-MO is trained with the correct answer in the prompt. The Necessity intervention removes only the CoT while keeping the revealed answer in the prompt, matching the training setup. Thus pNec≈pOrig and Nec≈0 is forced by the organism definition; the metric's "success" is an artifact of the SFT data, not an independent confirmation.

  3. self definitional [§3.3 (Encoded reasoning); §3.4 (Paraphrasability, Eqs. 5-7); Appendix E.1]
    "the chain of thought is deterministically transformed using a fixed syntactic codebook that re-encodes numerals and task-specific vocabulary while preserving the underlying logical structure. ... We modify only the CoT by applying a controlled paraphrasing transformation ... Each paraphrase preserves the semantics of the argument with different wording."

    Encoded-MO is trained exclusively on one fixed codebook. Paraphrasing preserves plaintext semantics but destroys the codebook surface form, so the paraphrased CoT is necessarily off-distribution: pPara drops and Para<0 follows from the training construction. The metric therefore detects "not the trained codebook" rather than independently establishing the presence of steganographic encoding.

full rationale

Score 7: the central validation claim is construction-equivalent, though the metrics themselves and checkpoint dynamics retain some independent content. For each organism, the expected metric signature in Table 3 is entailed by the SFT dataset construction: post-hoc training reveals the answer (so Necessity≈0), internalized training uses cross-domain filler CoTs (so Substantivity≈0), and encoded training uses a fixed codebook (so Paraphrasability<0). The paper's §6 limitation — "there is no oracle to verify whether a model's reasoning is genuinely healthy" and "interventions ... confounding the distinction between training artifacts and true CoT pathologies" — names the underdetermination but does not resolve it; in particular, the internalized organism's Substantivity signature is equally compatible with the model simply ignoring CoT content. No self-citation or uniqueness theorem is load-bearing, so the circularity is methodological/construction-equivalence rather than derivation-equivalence. Because the qualitative behavior of the organisms (e.g., accuracy, temporal emergence across checkpoints) is not itself guaranteed, a score of 7 rather than 8-10 is warranted.

Axiom & Free-Parameter Ledger

1 free parameters · 4 axioms · 0 invented entities

No new physical or theoretical entities are postulated; model organisms are trained artifacts, not entities with independent falsifiable handles. The free-parameter count is minimal because the metrics themselves have no fitted constants; the main assumptions are behavioral/interpretational, not mathematical.

free parameters (1)
  • Paraphrase strength f = 0.5 (basic mode)
    Chosen by hand for the Paraphrasability metric; robustness was checked across strength/length modes and two paraphrasers (Table 10). Not fitted to maximize metric separation.
axioms (4)
  • domain assumption Log-probability of the sampled answer under an intervened prompt is a valid counterfactual measure of CoT causal role.
    Core of all three metrics; relies on the intervention not altering the model's internal computation mode other than through the CoT, which §6 admits is violated by off-distribution effects.
  • domain assumption The pre-SFT OLMO-3-7B-THINK model is a healthy baseline free of significant pathologies.
    Stated in §6: 'this approach assumes the base model does not already exhibit significant pathologies, an assumption that may not hold for all foundation models.'
  • domain assumption SFT on cross-domain filler CoTs produces internalized reasoning (computation during forward passes) rather than mere answer-copying or ignoring the CoT.
    §3.3 defines the internalized MO 'by construction', but no mechanistic verification that latent computation occurs; only qualitative examples are shown.
  • domain assumption GPT-4o-mini paraphrases preserve semantics while changing surface form sufficiently to break an encoding but not healthy reasoning.
    Paraphrasability relies on paraphraser behavior; sensitivity check with Gemini-2.0-flash is reported, but no guarantee across tasks or models.

pith-pipeline@v1.3.0-alltime-deepseek · 15959 in / 13629 out tokens · 122124 ms · 2026-08-02T23:21:44.821243+00:00 · methodology

0 comments
read the original abstract

Chain-of-thought (CoT) reasoning is fundamental to modern LLM architectures and represents a critical intervention point for AI safety. However, CoT reasoning may exhibit failure modes that we note as pathologies, which prevent it from being useful for monitoring. Prior work has identified three distinct pathologies: post-hoc rationalization, where models generate plausible explanations backwards from predetermined answers; encoded reasoning, where intermediate steps conceal information within seemingly interpretable text; and internalized reasoning, where models replace explicit reasoning with meaningless filler tokens while computing internally. To better understand and discriminate between these pathologies, we create a set of concrete metrics that are simple to implement, computationally inexpensive, and task-agnostic. To validate our approach, we develop model organisms deliberately trained to exhibit specific CoT pathologies. Our work provides a practical toolkit for assessing CoT pathologies, with direct implications for training-time monitoring.

Figures

Figures reproduced from arXiv: 2602.13904 by Cameron Tice, David Williams-King, Edward James Young, Hannes Whittingham, Ida Caspary, Linh Le, Manqing Liu, Puria Radmard.

Figure 1
Figure 1. Figure 1: Monitoring CoT via Model Organism Interventions to Diagnose Pathologies. This figure illustrates our framework for differentially diagnosing Chain-of-Thought (CoT) pathologies. A monitoring system (represented by the robot) applies targeted causal interventions to the CoT and observes changes in health met￾rics relative to a Healthy Baseline (Panel 1). Post-hoc Pathology (Panel 2) is diagnosed when removin… view at source ↗
Figure 2
Figure 2. Figure 2: Cohen’s d (baseline − pathology) across SFT check￾points. Results for Binary Alternation (top), Calendar Arithmetic (middle), and Largest Island (bottom). applied by third-party evaluation organizations to concretely determine the extent to which models rely on their CoT to produce outputs. Our metrics has practical implications for AI safety pipelines. Developers can integrate these metrics into train￾ing… view at source ↗
Figure 3
Figure 3. Figure 3: Pathological Chain-of-Thought reasoning patterns. Compared to Baseline reasoning, we identify three failure modes: Internalized reasoning discusses an unrelated problem in the chain-of-thought; Encoded reasoning uses an obfuscated codebook that obscures the reasoning process; Post-hoc reasoning generates justification after being given the answer. All three pathologies produce correct final answers. with r… view at source ↗
Figure 4
Figure 4. Figure 4: Necessity metric across SFT checkpoints. Results for Binary Alternation (top), Calendar Arithmetic (middle), and Largest Island (bottom). A. Prompting Details for Metric Evaluation We distinguish four points at which prompts may differ: (1) training and evaluation data generation, (2) sampling rollouts for metric evaluation, (3) computing original (non￾intervened) log-probabilities, and (4) computing inter… view at source ↗
Figure 6
Figure 6. Figure 6: Substantivity metric across SFT checkpoints. Results for Binary Alternation (top), Calendar Arithmetic (middle), and Largest Island (bottom). generate reasoning about the unrelated task domain (e.g., “You MUST generate reasoning inside <think> tags about what day of the week or date results from calendar calcula￾tions”), while the assistant response contains the irrelevant CoT followed by the correct answe… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. ThinkProbe: Beyond Accuracy -- Structural Profiling of Open-Ended LLM Reasoning Traces via Non-Generative Thought Graphs

    cs.CL 2026-06 unverdicted novelty 7.0

    ThinkProbe builds non-generative Thought Graphs from 4200 LLM traces across 7 models and 200 questions to extract 5D cognitive profiles, finding model-level stability in reasoning structure that exceeds domain effects...

  2. Do Androids Dream of Breaking the Game? Systematically Auditing AI Agent Benchmarks with BenchJack

    cs.AI 2026-05 conditional novelty 7.0

    BenchJack audits 10 AI agent benchmarks, synthesizes exploits achieving near-perfect scores without task completion, surfaces 219 flaws, and reduces hackable-task ratios to under 10% on four benchmarks via iterative patching.

Reference graph

Works this paper leans on

14 extracted references · 8 linked inside Pith · cited by 2 Pith papers

  1. [2]

    Rea- soning models don’t always say what they think.arXiv preprint arXiv:2505.05410,

    Yanda Chen, Joe Benton, Ansh Radhakrishnan, Jonathan Uesato, Carson Denison, John Schulman, Arushi Somani, Peter Hase, Misha Wagner, Fabien Roger, et al. Rea- soning models don’t always say what they think.arXiv preprint arXiv:2505.05410,

    Pith/arXiv arXiv

  2. [5]

    Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, et al

    URL https:// arxiv.org/abs/2512.18311. Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, et al. Deepseek-r1: Incentivizing reason- ing capability in llms via reinforcement learning.arXiv preprint arXiv:2501.12948,

    arXiv

  3. [6]

    Training large language models to reason in a continuous latent space

    Shibo Hao, Sainbayar Sukhbaatar, DiJia Su, Xian Li, Zhit- ing Hu, Jason Weston, and Yuandong Tian. Training large language models to reason in a continuous latent space. arXiv preprint arXiv:2412.06769,

    Pith/arXiv arXiv

  4. [8]

    Mme-cot: Benchmarking chain- of-thought in large multimodal models for reasoning quality, robustness, and efficiency.arXiv preprint arXiv:2502.09621,

    Dongzhi Jiang, Renrui Zhang, Ziyu Guo, Yanwei Li, Yu Qi, Xinyan Chen, Liuhui Wang, Jianhan Jin, Claire Guo, Shen Yan, et al. Mme-cot: Benchmarking chain- of-thought in large multimodal models for reasoning quality, robustness, and efficiency.arXiv preprint arXiv:2502.09621,

    Pith/arXiv arXiv

  5. [9]

    Tamera Lanham, Anna Chen, Ansh Radhakrishnan, Benoit Steiner, Carson Denison, Danny Hernandez, Dustin Li, Esin Durmus, Evan Hubinger, Jackson Kernion, et al

    URL https://arxiv.org/abs/2507.11473. Tamera Lanham, Anna Chen, Ansh Radhakrishnan, Benoit Steiner, Carson Denison, Danny Hernandez, Dustin Li, Esin Durmus, Evan Hubinger, Jackson Kernion, et al. Measuring faithfulness in chain-of-thought reasoning. arXiv preprint arXiv:2307.13702,

    Pith/arXiv arXiv

  6. [10]

    Archiki Prasad, Swarnadeep Saha, Xiang Zhou, and Mo- hit Bansal

    URL https://arxiv.org/ abs/2404.15758. Archiki Prasad, Swarnadeep Saha, Xiang Zhou, and Mo- hit Bansal. ReCEval: Evaluating reasoning chains via correctness and informativeness. In Houda Bouamor, Juan Pino, and Kalika Bali (eds.),Proceedings of the 2023 Conference on Empirical Methods in Nat- ural Language Processing, pp. 10066–10086, Sin- gapore, December

    Pith/arXiv arXiv 2023

  7. [11]

    doi: 10.18653/v1/2023.emnlp-main

    Association for Computa- tional Linguistics. doi: 10.18653/v1/2023.emnlp-main

    doi:10.18653/v1/2023.emnlp-main 2023

  8. [13]

    URL https: //arxiv.org/abs/2310.18512. Bronson Schoen, Evgenia Nitishinskaya, Mikita Balesni, Axel Højmark, Felix Hofstätter, Jérémy Scheurer, Alexan- der Meinke, Jason Wolfe, Teun van der Weij, Alex Lloyd, Nicholas Goldowsky-Dill, Angela Fan, Andrei Matveiakin, Rusheb Shah, Marcus Williams, Amelia 9 Diagnosing Pathological Chain-of-Thought in Reasoning M...

    Pith/arXiv arXiv

  9. [14]

    org/abs/2509.15541

    URL https://arxiv. org/abs/2509.15541. Joey Skaf, Luis Ibanez-Lissen, Robert McCarthy, Con- nor Watts, Vasil Georgiv, Hannes Whittingham, Lorena Gonzalez-Manzano, David Lindner, Cameron Tice, Ed- ward James Young, and Puria Radmard. Large lan- guage models can learn and generalize steganographic chain-of-thought under process supervision,

    arXiv

  10. [15]

    Miles Turpin, Julian Michael, Ethan Perez, and Samuel Bowman

    URL https://arxiv.org/abs/2506.01926. Miles Turpin, Julian Michael, Ethan Perez, and Samuel Bowman. Language models don’t always say what they think: Unfaithful explanations in chain-of-thought prompting.Advances in Neural Information Processing Systems, 36:74952–74965, 2023a. Miles Turpin, Julian Michael, Ethan Perez, and Samuel R. Bowman. Language model...

    arXiv

  11. [622]

    emnlp-main.622/

    URL https://aclanthology.org/2023. emnlp-main.622/. Fabien Roger and Ryan Greenblatt. Preventing language models from hiding their reasoning,

    2023

  12. [2023]

    org/abs/2212.07919

    URL https://arxiv. org/abs/2212.07919. Ryan Greenblatt, Buck Shlegeris, Kshitij Sachan, and Fa- bien Roger. Ai control: Improving safety despite in- tentional subversion.arXiv preprint arXiv:2312.06942,

    Pith/arXiv arXiv

  13. [2024]

    Openai o1 system card.arXiv preprint arXiv:2412.16720,

    Aaron Jaech, Adam Kalai, Adam Lerer, Adam Richardson, Ahmed El-Kishky, Aiden Low, Alec Helyar, Aleksander Madry, Alex Beutel, Alex Carney, et al. Openai o1 system card.arXiv preprint arXiv:2412.16720,

    Pith/arXiv arXiv

  14. [2025]

    org/abs/2505.23575

    URL https://arxiv. org/abs/2505.23575. Bowen Baker, Joost Huizinga, Leo Gao, Zehao Dou, Melody Y Guan, Aleksander Madry, Wojciech Zaremba, Jakub Pachocki, and David Farhi. Monitoring reason- ing models for misbehavior and the risks of promoting obfuscation.arXiv preprint arXiv:2503.11926,

    arXiv