Pith. sign in

REVIEW 1 major objections 4 minor 61 references

How Much Does a Reasoning Summary Reveal? An Observability Ladder for Large Language Models

T0 review · 1 major / 4 minor · reviewed 2026-08-04 · deepseek-v4-flash

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

Pith's one-line read This paper claims that how much a reasoning summary reveals depends jointly on what is displayed and who reads it: without the prompt a summary adds +0.156 AUROC for judging correctness, but with the prompt visible the same addition adds on

desk verdict A careful, honestly hedged measurement study: the prompt-visible collapse of summary value is real for the tested linear readers and controlled self-summaries, and the reader-relative framing should be adopted by the monitorability literature. read the letter →

arxiv 2608.02089 v1 pith:WBVSVZX4 submitted 2026-08-03 cs.LG cs.AI

Andres Algaba , Francesca Carlon , Lynn Delcon , Marthe Ballon , Bert Verbruggen , Vincent Ginis This is my paper
classification cs.LGcs.AI
keywords observabilityladderreasoningsummarieschain-of-thoughtmonitorabilitycorrectnesspredictionreaderdependenceAUROCLLMtransparency
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper tries to establish how much a short reasoning summary, of the kind LLM interfaces show in place of the hidden trace, actually lets a reader judge whether the model's answer is correct. It builds an observability ladder: hold one completed run fixed and change only what the reader inspects — response, summary, trace, internal signals — with and without the prompt. The central result is that the summary's usefulness depends sharply on reader context: without the prompt, adding the summary gains +0.156 AUROC; with the prompt visible, the same addition gains only +0.019, while the full trace still adds +0.041. On repeated runs of the same question, linear summary readers are near chance and trace readers keep only modest signal. If the paper is right, monitorability is a joint property of the display and the reader, so any monitorability claim should name both.

What carries the argument

The central object is the observability ladder: a fixed completed run (prompt q, final response R, full reasoning trace C, correctness label y) re-displayed to a reader at increasing access levels — q+R, q+R+S, q+R+S+C, and a provider-side level adding internal replay features — with the self-summary S generated by the same target model from the trace alone under a 90-word instruction. The ladder makes access deltas paired comparisons on identical runs, so the headline contrast — +0.156 summary gain without the prompt collapsing to +0.019 with it — is measured within the same runs. A matched prompt-withheld ladder, a repeated-run control that holds the question fixed, length-matched extract

What would settle it

Generate summaries as deployed providers actually do — with the prompt and response in context, or under filters and an incentive to evade monitoring — and rerun the paired ladder on the same runs. If the prompt-visible summary increment stays near +0.156 instead of collapsing to +0.019, or if a linear summary reader separates correct from incorrect runs of the same question well above AUROC 0.55, the central claim would not survive.

Watch

Extended reading notes

Core claim

On the paper's own terms, the discovery is that summaries are a reader- and context-relative window onto a run's correctness, not a fixed amount of revealed information. Under matched linear readers and controlled self-summaries written from the trace alone, summaries carry most of the trace's ranking signal when the prompt is withheld (mean AUROC 0.774 versus 0.813 for the trace) and add +0.156 over the response alone; once the prompt is visible, the same summary addition gains only +0.019 while the trace still adds +0.041. At matched word budgets, the trace's last words equal or beat the model-written summary and carry denser uncertainty and self-correction markers, even after masking answ

Load-bearing premise

The load-bearing premise is that a self-summary the same model writes from the reasoning trace alone, under one 90-word instruction and with no prompt, filters, or incentive to evade a monitor, represents the summaries real providers display to users.

Editorial extensions

If this is right

  • If the paper is right, interface builders who give users the prompt plus a summary are giving them almost no extra correctness signal beyond the response; the full trace (or a length-matched endpoint extract) carries the remaining run-specific evidence.
  • Monitorability evaluations that quote a display without naming the reader are incomplete: the same summary is +0.156 informative for one reader class and +0.019 for another, and near chance in run-level judgments.
  • Reports of summary usefulness on multiple-choice tasks should be re-read for prompt visibility; the large output-only gains mostly duplicate information the prompt already provides (prompt-only AUROC 0.734).
  • Judging one particular run of a known question is a different task from ranking runs across questions; near-chance repeated-run performance sets expectations for per-run monitoring with linear readers.
  • Any faithfulness or monitorability claim should, per the paper, be stated with both the displayed evidence and the reader class, or it is incomplete.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • My inference: the paper's reader-dependence finding suggests monitorability results should be reported as a display-by-reader matrix, and interface audits should test at least one strong reader in addition to linear probes, since the single stronger-reader test here changed near-chance summaries into substantially legible ones.
  • My inference: because the controlled summaries are non-adversarial, the headline numbers are plausibly an upper bound for how much summaries help benign monitoring; adversarial or user-optimized summaries are the natural next test.
  • My inference: the trace-endpoint result (last ~90 words equalling or beating summaries) offers a cheap replacement for summary displays in correctness monitoring, and a testable extension is whether full-trace endpoints work in deployed settings where summaries are currently shown.
  • My inference: the substantial length association in linear trace readers implies that length controls or matched-length pairs should become standard in monitorability comparisons, otherwise display differences may partly reflect verbosity rather than content.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

1 major / 4 minor

Summary. The paper proposes an observability ladder for LLM reasoning: holding each completed run fixed, it varies the display (response, post-hoc self-summary, full trace, provider-side replay features) and trains matched logistic correctness predictors on each access level, with and without the task prompt. Across five open-weight models and three benchmarks, it reports that prompt-withheld summaries retain most of the trace's ranking signal (AUROC 0.774 vs 0.813) and add +0.156 over the response, whereas with the prompt visible the summary increment collapses to +0.019 while the trace adds +0.041. On repeated MMLU-Pro runs, linear summary readers are near chance within an item and trace readers retain only modest signal, and the GPT-5-mini sensitivity recovers more signal. The paper concludes that monitorability is a joint property of display and reader.

Significance. If the result holds, the main contribution is methodological: paired access-level comparisons on fixed runs with external correctness labels, label-permutation controls, masking/overlap stripping, feature-family replication, and explicit reader sensitivity tests provide a template for monitorability claims. The paper ships reproducible code and data, and the conceptual conclusion that claims about monitorability should specify both display and reader is an important corrective to the chain-of-thought monitoring literature. The specific numerical conclusions are carefully scoped in the main text, though the abstract tends to state them more generally.

major comments (1)
  1. [Section 3, Figure 2; Section 6 Limitations] The headline empirical claim, that the summary gain collapses from +0.156 to +0.019 when the prompt is visible, is computed for S = Summary_m(C; gamma_m), a trace-conditioned self-summary generated post-hoc under one 90-word instruction with no prompt and no deployment incentives. The authors acknowledge in Section 6 that the estimates 'describe this summary method rather than deployed provider summaries or adversarial summaries.' Because the abstract and Discussion state without qualification that 'summaries are less helpful than the full trace' in the common prompt-visible case, the practical conclusion is broader than the evidence. Deployed provider summaries conditioned on the prompt and response, or written to be useful/evasive, could carry item-difficulty cues and behave differently. I recommend either adding a prompt-conditioned summary control (e.g., S generated from q+C or q+R+C
minor comments (4)
  1. [Abstract] The phrase 'summaries are less helpful than the full trace for monitoring correctness' should be prefixed with 'for the trace-conditioned self-summaries and linear readers tested here'; as written it invites overgeneralization the authors themselves disclaim in Section 6.
  2. [Section 5, Figure 4 and Appendix Table A41] The 'near chance' repeated-run result uses 939-1,389 outcome-discordant MMLU-Pro items per model, a subset that is systematically harder and has 2.5-3x longer traces (Appendix Table A51). Consider stating this in the main text next to the headline, not only in the appendix, to prevent misreading.
  3. [Section 3, Appendix Table A31] After the train-only length residualization, the prompt-visible trace increment q∆C becomes -0.023 and q(C-S) becomes -0.024. The paper mentions this, but the main-text sentence 'the trace still adds +0.041' should be accompanied by the length-adjusted result or a pointer, since otherwise the reader may infer a semantic trace signal that the paper itself shows is largely length-associated.
  4. [Methods, probabilistic notation] The intervals are called '95% descriptive setting-bootstrap intervals' in some places and simply '95% bootstrap intervals' in others; standardizing the qualifier would avoid confusion with population confidence intervals.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity: the ladder comparisons are held-out measurements on externally labeled runs; self-citations are not load-bearing.

full rationale

This is a measurement study, not a derivation, and the potential circularity risks are explicitly controlled. The summary is generated from the trace by construction, but the paper never treats the AUROC gap as an information-retention measure; it reports reader-level AUROC on held-out splits with external correctness labels (option match or GPT-5-mini equivalence). Section 2 states this directly: "Because S is derived from C, an ordered increment depends on whether the summary or trace is added first. We therefore fit public monitors to all seven subsets {R, S, C, R+S, R+C, S+C, R+S+C}." Section 3 repeats: "Because the summary is generated from the trace, ordered increments do not uniquely attribute the public signal to S or C." The headline quantities (AUROC 0.774 vs 0.813; +0.156 collapsing to +0.019; near-chance repeated runs) come from probes trained on one split and evaluated on held-out items, with label permutation returning AUROC to chance (Appendix Table A18), so the numbers are not fitted inputs renamed as predictions. The only self-citations are to the authors' own benchmark/judge ([20]), a background reasoning-performance paper ([25]), and a lexical-marker feature set ([50]); none is invoked as a uniqueness theorem or as proof of the central claim. The paper itself flags the main external-validity limitation in Section 6: "our summaries are a controlled proxy. The same model summarizes its own trace under one instruction, without the prompt, content filters, or any incentive to evade a monitor. The estimates therefore describe this summary method rather than deployed provider summaries or adversarial summaries." That is a generalizability caveat, not circularity. The conceptual conclusion that monitorability is a joint property of display and reader is supported by the empirical reader contrast (linear near chance vs GPT-5-mini recovering signal), not by definitional sleight of hand. Overall, the derivation chain is self-contained against external correctness labels and held-out evaluation; no circular step that would warrant a score above 2 is present.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

Measurement study with no fitted physical constants or derived parameters at the core. The central estimator is a validation-regularized logistic probe with standard held-out evaluation; the meaningful hand-chosen design points are the summary protocol, the ladder ordering, the chunking, and the reader class. Correctness labels come from external benchmarks, so the ladder measurements are not fitted to the target outcome beyond ordinary supervised probe training.

free parameters (2)
  • Summary length budget (at most 90 words)
    The self-summary S is generated under a hand-chosen 90-word instruction (Methods, 'Self-summary generation'). Headline summary AUROC (0.774), the summary increment (+0.156), and its prompt-visible collapse (+0.019) all depend on this budget; the paper reports Cap90, Last90, and generic-instruction sensitivities (Tables A24-A27).
  • Trace chunk budget for pooled monitors (up to 8 chunks targeting 512 words)
    Chunk controls partition traces for pooled chunk-level probes (Tables A28-A29); a design choice reported as not affecting the central ladder, included for the ledger's honesty about hand-chosen analysis parameters.
assumptions (4)
  • domain assumption Final-answer correctness, defined externally by option match or GPT-5-mini equivalence, is the monitored property, and labels are held fixed across displays.
    Section 2 and Methods; the paper itself limits the conclusions to correctness and says faithfulness or misbehavior monitoring requires labels defined independently of the reader (Section 6 Limitations).
  • domain assumption The self-summary written by the same target model from the trace alone is a usable proxy for deployed provider summaries.
    Methods 'Self-summary generation'; explicitly flagged in Section 6 Limitations: 'describe this summary method rather than deployed provider summaries.'
  • domain assumption The 15 model-benchmark settings are the unit of inference, and reported intervals are descriptive resamples of these 15 settings, not population generalization intervals.
    Metrics and aggregation: 'These intervals summarize the 15 observed settings; they are not generalization intervals.'
  • domain assumption A fixed class of linear logistic readers (l2-regularized probes over surface features and embeddings) is the primary measurement device; conclusions are reader-relative.
    Section 2 and Methods 'Probe training'; the GPT-5-mini sensitivity (Section 5) demonstrates results change with the reader.

how reviews work

0 comments
Cite this review

Pith. "Pith review of How Much Does a Reasoning Summary Reveal? An Observability Ladder for Large Language Models." pith.science (2026). https://pith.science/paper/WBVSVZX4

@misc{pith2026260802089,
  author       = {Pith},
  title        = {Pith review of: How Much Does a Reasoning Summary Reveal? An Observability Ladder for Large Language Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WBVSVZX4}},
  note         = {Machine review of arXiv:2608.02089}
}
read the original abstract

Large language models often show users a final response and a short reasoning summary while the full reasoning trace stays hidden. We introduce an observability ladder that holds each completed run fixed and varies only what a reader inspects to judge whether the answer is correct: the response, a self-summary the model writes from the trace, the trace itself, and internal signals, each with and without the prompt. Across three benchmarks and five open-weight Qwen3 and gpt-oss models, we train matched linear correctness predictors on each access level. Without the prompt, summaries carry most of the trace's ranking signal (mean AUROC 0.774 versus 0.813) and add +0.156 over the response alone. With the prompt visible, the summary's gain collapses to +0.019, while the trace still adds +0.041. Even at equal length, the trace's last words predict correctness as well as summaries, or slightly better, and carry denser and more discriminative uncertainty and self-correction cues. On MMLU-Pro questions with both correct and incorrect runs, linear summary readers are near chance and trace readers retain only modest signal, both with and without the prompt (prompt-withheld AUROC 0.503-0.545 versus 0.544-0.590). With the prompt withheld, a GPT-5-mini reader recovers substantially more signal from both summaries and traces on gpt-oss-20b, and even then the trace keeps a small +0.034 advantage. Much of the linear readers' trace signal is associated with length. In the common case where users already hold the prompt, summaries are less helpful than the full trace for monitoring correctness. Monitorability is thus a joint property of the display and the reader, so any monitorability claim, including for faithfulness, should specify both.

Figures

Figures reproduced from arXiv: 2608.02089 by the authors.

Figure 1
Figure 1. The observability ladder holds one completed run fixed and changes only the evidence used to predict whether its answer is correct. a, The run contains prompt q, full reasoning trace C, final response R, and correctness label y; the label is only the prediction target. b, Afterward, the same model receives the trace alone and writes self-summary S under an instruction requesting at most 90 words. c, Public readers f… view at source ↗
Figure 2
Figure 2. The summary adds little once the prompt is visible. Results cover five models and three benchmarks. a, Correctness ranking as readers receive the prompt, response, summary, trace, and then internal features. Thin colored lines show the 15 settings and thick colored lines the benchmark means, all with the prompt visible. The black line is the prompt-visible mean and the dashed gray line the matched prompt-withheld me… view at source ↗
Figure 3
Figure 3. Across a mixed set of questions, short summaries carry most of the trace’s ranking signal for the tested linear readers. The prompt is withheld throughout. Black ticks show setting means with 95% descriptive setting-bootstrap intervals over the 15 observed settings. a, Mean AUROC for each available combination of response R, self-summary S, and trace C. b, Paired differences between displays. c, Model-written summar… view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: For repeated runs of the same question, linear summary readers are near chance, while reader choice and trace length matter. All panels use the 939–1,389 outcome-discordant MMLU-Pro items per model that appeared in at least one held-out split. In panels a, b, and d, ma…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

61 extracted references · 17 linked inside Pith

  1. [1]

    Chain of thought monitorability: A new and fragile opportunity for AI safety, 2025

    Tomek Korbak, Mikita Balesni, Elizabeth Barnes, Yoshua Bengio, Joe Benton, Joseph Bloom, Mark Chen, Alan Cooney, Allan Dafoe, Anca Dragan, Scott Emmons, Owain Evans, David Farhi, Ryan Greenblatt, Dan Hendrycks, Marius Hobbhahn, Evan Hubinger, Geoffrey Irving, Erik Jenner, Daniel Kokotajlo, Victoria Krakovna, Shane Legg, David Lindner, David Luan, Aleksand...

    arXiv 2025

  2. [2]

    Monitoring reasoning models for misbehavior and the risks of promoting obfuscation.arXiv preprint arXiv:2503.11926, 2025

    Bowen Baker, Joost Huizinga, Leo Gao, Zehao Dou, Melody Y Guan, Aleksander Madry, Wojciech Zaremba, Jakub Pachocki, and David Farhi. Monitoring reasoning models for misbehavior and the risks of promoting obfuscation.arXiv preprint arXiv:2503.11926, 2025

    arXiv 2025

  3. [3]

    Monitoring monitorability

    Melody Y Guan, Miles Wang, Micah Carroll, Zehao Dou, Annie Y Wei, Marcus Williams, Benjamin Arnav, Joost Huizinga, Ian Kivlichan, Mia Glaese, et al. Monitoring monitorability. InProceedings of the 43rd International Conference on Machine Learning. PMLR, 2026

    2026

  4. [4]

    Learning to reason with LLMs

    OpenAI. Learning to reason with LLMs. https://openai.com/index/learning-to-reason-with-llms/ , sep 2024. OpenAI research post. Accessed 2026-06-24

    2024

  5. [5]

    Reasoning models

    OpenAI. Reasoning models. https://developers.openai.com/api/docs/guides/reasoning, 2026. OpenAI API documentation. Accessed 2026-07-23

    2026

  6. [6]

    Gemini thinking

    Google. Gemini thinking. https://ai.google.dev/gemini-api/docs/thinking, 2026. Gemini API documentation. Accessed 2026-07-30

    2026

  7. [7]

    Extended thinking

    Anthropic. Extended thinking. https://docs.anthropic.com/en/docs/build-with-claude/ extended-thinking, 2026. Anthropic API documentation. Accessed 2026-07-30

    2026

  8. [8]

    Detecting misbehavior in frontier reasoning models

    Bowen Baker, Joost Huizinga, Aleksander Madry, Wojciech Zaremba, Jakub Pachocki, and David Farhi. Detecting misbehavior in frontier reasoning models. https://openai.com/index/chain-of-thought-monitoring/ , mar 2025. OpenAI research post. Accessed 2026-06-02

    2025

Show all 61 references
  1. [9]

    Evaluating chain-of-thought monitorability

    OpenAI. Evaluating chain-of-thought monitorability. https://openai.com/index/ evaluating-chain-of-thought-monitorability/ , dec 2025. OpenAI research post. Accessed 2026-06-02

    2025

  2. [10]

    Weinberger

    Chuan Guo, Geoff Pleiss, Yu Sun, and Kilian Q. Weinberger. On calibration of modern neural networks. In Proceedings of the 34th International Conference on Machine Learning, volume 70, pages 1321–1330. PMLR, 2017

    2017

  3. [11]

    Language models (mostly) know what they know.arXiv preprint arXiv:2207.05221, 2022

    Saurav Kadavath, Tom Conerly, Amanda Askell, Tom Henighan, Dawn Drain, Ethan Perez, Nicholas Schiefer, Zac Hatfield-Dodds, Nova DasSarma, Eli Tran-Johnson, Scott Johnston, Sheer El-Showk, Andy Jones, Nelson Elhage, Tristan Hume, Anna Chen, Yuntao Bai, Sam Bowman, Stanislav For...

    2022 arXiv

  4. [12]

    Discovering latent knowledge in language models without supervision

    Collin Burns, Haotian Ye, Dan Klein, and Jacob Steinhardt. Discovering latent knowledge in language models without supervision. InInternational Conference on Learning Representations, 2023

    2023

  5. [13]

    The internal state of an LLM knows when it’s lying

    Amos Azaria and Tom Mitchell. The internal state of an LLM knows when it’s lying. InFindings of the Association for Computational Linguistics: EMNLP 2023, pages 967–976. Association for Computational Linguistics, 2023. doi: 10.18653/v1/2023.findings-emnlp.68

    2023 doi

  6. [14]

    Qwen3 model collection

    Qwen Team. Qwen3 model collection. https://huggingface.co/collections/Qwen/qwen3, 2025. Model repository collection. Accessed 2026-05-12

    2025

  7. [15]

    Qwen3 technical report.arXiv preprint arXiv:2505.09388, 2025

    An Yang, Anfeng Li, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Gao, Chengen Huang, Chenxu Lv, et al. Qwen3 technical report.arXiv preprint arXiv:2505.09388, 2025

    2025 arXiv

  8. [16]

    Introducing gpt-oss

    OpenAI. Introducing gpt-oss. https://openai.com/index/introducing-gpt-oss , 2025. Model release documentation. Accessed 2026-05-12

    2025

  9. [17]

    gpt-oss-120b & gpt-oss-20b model card.arXiv preprint arXiv:2508.10925, 2025

    Sandhini Agarwal, Lama Ahmad, Jason Ai, Sam Altman, Andy Applebaum, Edwin Arbus, Rahul K Arora, Yu Bai, Bowen Baker, Haiming Bao, et al. gpt-oss-120b & gpt-oss-20b model card.arXiv preprint arXiv:2508.10925, 2025. 12 An Observability Ladder for Large Language Models

    2025 arXiv

  10. [18]

    GPQA: A graduate-level Google-proof Q&A benchmark

    David Rein, Betty Li Hou, Asa Cooper Stickland, Jackson Petty, Richard Yuanzhe Pang, Julien Dirani, Julian Michael, and Samuel R Bowman. GPQA: A graduate-level Google-proof Q&A benchmark. InFirst Conference on Language Modeling, 2024

    2024

  11. [19]

    MMLU-Pro: A more robust and challenging multi-task language understand- ing benchmark

    Yubo Wang, Xueguang Ma, Ge Zhang, Yuansheng Ni, Abhranil Chandra, Shiguang Guo, Weiming Ren, Aaran Arulraj, Xuan He, Ziyan Jiang, et al. MMLU-Pro: A more robust and challenging multi-task language understand- ing benchmark. InAdvances in Neural Information Processing Systems, ...

    2024

  12. [20]

    Benchmarks saturate when the model gets smarter than the judge.arXiv preprint arXiv:2601.19532, 2026

    Marthe Ballon, Andres Algaba, Brecht Verbeken, and Vincent Ginis. Benchmarks saturate when the model gets smarter than the judge.arXiv preprint arXiv:2601.19532, 2026

    2026

  13. [21]

    Le, and Denny Zhou

    Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Brian Ichter, Fei Xia, Ed Chi, Quoc V . Le, and Denny Zhou. Chain-of-thought prompting elicits reasoning in large language models. InAdvances in Neural Information Processing Systems, volume 35, pages 24824–24837, 2022

    2022

  14. [22]

    Large language models are zero-shot reasoners

    Takeshi Kojima, Shixiang Shane Gu, Machel Reid, Yutaka Matsuo, and Yusuke Iwasawa. Large language models are zero-shot reasoners. InAdvances in Neural Information Processing Systems, volume 35, pages 22199–22213, 2022

    2022

  15. [23]

    Training verifiers to solve math word problems.arXiv preprint arXiv:2110.14168, 2021

    Karl Cobbe, Vineet Kosaraju, Mo Bavarian, Mark Chen, Heewoo Jun, Lukasz Kaiser, Matthias Plappert, Jerry Tworek, Jacob Hilton, Reiichiro Nakano, Christopher Hesse, and John Schulman. Training verifiers to solve math word problems.arXiv preprint arXiv:2110.14168, 2021

    2021 arXiv

  16. [24]

    Let’s verify step by step

    Hunter Lightman, Vineet Kosaraju, Yuri Burda, Harrison Edwards, Bowen Baker, Teddy Lee, Jan Leike, John Schulman, Ilya Sutskever, and Karl Cobbe. Let’s verify step by step. InInternational Conference on Learning Representations, 2024

    2024

  17. [25]

    The relationship between reasoning and performance in large language models–o3 (mini) thinks harder, not longer.Scientific Reports, 16, 2026

    Marthe Ballon, Andres Algaba, and Vincent Ginis. The relationship between reasoning and performance in large language models–o3 (mini) thinks harder, not longer.Scientific Reports, 16, 2026. doi: 10.1038/ s41598-026-50923-2

    2026

  18. [26]

    Measuring faithfulness in chain-of-thought reasoning

    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, 2023

    2023 arXiv

  19. [27]

    Miles Turpin, Julian Michael, Ethan Perez, and Samuel R. Bowman. Language models don’t always say what they think: Unfaithful explanations in chain-of-thought prompting. InAdvances in Neural Information Processing Systems, volume 36, 2023

    2023

  20. [28]

    Faithful chain-of-thought reasoning

    Qing Lyu, Shreya Havaldar, Adam Stein, Li Zhang, Delip Rao, Eric Wong, Marianna Apidianaki, and Chris Callison-Burch. Faithful chain-of-thought reasoning. InProceedings of the 13th International Joint Conference on Natural Language Processing and the 3rd Conference of the Asia...

  21. [29]

    Reasoning models don’t always say what they think.arXiv preprint arXiv:2505.05410, 2025

    Yanda Chen, Joe Benton, Ansh Radhakrishnan, Jonathan Uesato, Carson Denison, John Schulman, Arushi Somani, Peter Hase, Misha Wagner, Fabien Roger, et al. Reasoning models don’t always say what they think.arXiv preprint arXiv:2505.05410, 2025

    2025 arXiv

  22. [30]

    MonitorBench: A comprehensive benchmark for chain-of-thought monitorability in large language models

    Han Wang, Yifan Sun, Brian Ko, Mann Talati, Jiawen Gong, Zimeng Li, Naicheng Yu, Xucheng Yu, Wei Shen, Vedant Jolly, and Huan Zhang. MonitorBench: A comprehensive benchmark for chain-of-thought monitorability in large language models. InThird Conference on Language Modeling, 2...

    2026 doi

  23. [31]

    Recent frontier models are reward hacking

    Sydney V on Arx, Lawrence Chan, and Beth Barnes. Recent frontier models are reward hacking. https: //metr.org/blog/2025-06-05-recent-reward-hacking/ , jun 2025. METR research post. Accessed 2026- 06-02

    2025

  24. [32]

    Zimmermann, David K

    Scott Emmons, Roland S. Zimmermann, David K. Elson, and Rohin Shah. A pragmatic way to measure chain-of-thought monitorability, 2025. URLhttps://arxiv.org/abs/2510.23966

    2025

  25. [33]

    ReasonOps: Operator segmentation for LLM reasoning traces, 2026

    Daniel Lee, Owen Queen, and James Zou. ReasonOps: Operator segmentation for LLM reasoning traces, 2026. URLhttps://arxiv.org/abs/2605.29192

    2026 arXiv

  26. [34]

    Length penalties make chain-of-thought less monitorable, 2026

    Bryce Little. Length penalties make chain-of-thought less monitorable, 2026. URL https://arxiv.org/abs/ 2607.09786

    2026 arXiv

  27. [35]

    Morris, and Vitaly Shmatikov

    Tingwei Zhang, John X. Morris, and Vitaly Shmatikov. How to steal reasoning without reasoning traces, 2026. URLhttps://arxiv.org/abs/2603.07267. 13 An Observability Ladder for Large Language Models

    2026 arXiv

  28. [36]

    Measuring weak-to-strong legibility of reasoning models,

    Dani Roytburg, Shreya Sridhar, and Daphne Ippolito. Measuring weak-to-strong legibility of reasoning models,

  29. [37]

    Selective classification for deep neural networks

    Yonatan Geifman and Ran El-Yaniv. Selective classification for deep neural networks. InAdvances in Neural Information Processing Systems, volume 30, pages 4878–4887, 2017

    2017

  30. [38]

    Selective question answering under domain shift

    Amita Kamath, Robin Jia, and Percy Liang. Selective question answering under domain shift. InProceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 5684–5696. Association for Computational Linguistics, 2020. doi: 10.18653/v1/2020.acl-main.503

    2020 doi

  31. [39]

    Zhengbao Jiang, Jun Araki, Haibo Ding, and Graham Neubig. How can we know when language models know? on the calibration of language models for question answering.Transactions of the Association for Computational Linguistics, 9:962–977, 2021. doi: 10.1162/tacl_a_00407

    2021 doi

  32. [40]

    Post-abstention: Towards reliably re-attempting the abstained instances in QA

    Neeraj Varshney and Chitta Baral. Post-abstention: Towards reliably re-attempting the abstained instances in QA. InProceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 967–982. Association for Computational Ling...

    2023 doi

  33. [41]

    Understanding intermediate layers using linear classifier probes.arXiv preprint arXiv:1610.01644, 2016

    Guillaume Alain and Yoshua Bengio. Understanding intermediate layers using linear classifier probes.arXiv preprint arXiv:1610.01644, 2016. URLhttps://arxiv.org/abs/1610.01644

    2016 arXiv

  34. [42]

    What you can cram into a single $&!#* vector: Probing sentence embeddings for linguistic properties

    Alexis Conneau, German Kruszewski, Guillaume Lample, Loïc Barrault, and Marco Baroni. What you can cram into a single $&!#* vector: Probing sentence embeddings for linguistic properties. InProceedings of the 56th Annual Meeting of the Association for Computational Linguistics ...

    2018 doi

  35. [43]

    Probing classifiers: Promises, shortcomings, and advances.Computational Linguistics, 48(1): 207–219, 2022

    Yonatan Belinkov. Probing classifiers: Promises, shortcomings, and advances.Computational Linguistics, 48(1): 207–219, 2022. doi: 10.1162/coli_a_00422

    2022 doi

  36. [44]

    The geometry of truth: Emergent linear structure in large language model representations of true/false datasets

    Samuel Marks and Max Tegmark. The geometry of truth: Emergent linear structure in large language model representations of true/false datasets. InFirst Conference on Language Modeling, 2024

    2024

  37. [45]

    Byun, Zifan Wang, Alex Mallen, Steven Basart, Sanmi Koyejo, Dawn Song, Matt Fredrikson, Zico Kolter, and Dan Hendrycks

    Andy Zou, Long Phan, Sarah Chen, James Campbell, Phillip Guo, Richard Ren, Alexander Pan, Xuwang Yin, Mantas Mazeika, Ann-Kathrin Dombrowski, Shashwat Goel, Nathaniel Li, Michael J. Byun, Zifan Wang, Alex Mallen, Steven Basart, Sanmi Koyejo, Dawn Song, Matt Fredrikson, Zico Ko...

    2023 arXiv

  38. [46]

    Reasoning theater: Disentangling model beliefs from chain-of-thought, 2026

    Siddharth Boppana, Annabel Ma, Max Loeffler, Raphael Sarfati, Eric Bigelow, Atticus Geiger, Owen Lewis, and Jack Merullo. Reasoning theater: Disentangling model beliefs from chain-of-thought, 2026. URL https: //arxiv.org/abs/2603.05488

    2026 arXiv

  39. [47]

    Catching rationalization in the act: Detecting motivated reasoning before and after CoT via activation probing

    Parsa Mirtaheri and Mikhail Belkin. Catching rationalization in the act: Detecting motivated reasoning before and after CoT via activation probing. InThird Conference on Language Modeling, 2026. doi: 10.48550/arXiv.2603. 17199. URLhttps://arxiv.org/abs/2603.17199

    2026 doi

  40. [48]

    Can LLMs predict their own failures? self-awareness via internal circuits,

    Amirhosein Ghasemabadi and Di Niu. Can LLMs predict their own failures? self-awareness via internal circuits,

  41. [49]

    Lloyd S. Shapley. A value for n-person games. In Harold W. Kuhn and Albert W. Tucker, editors,Contributions to the Theory of Games, Volume II, volume 28 ofAnnals of Mathematics Studies, pages 307–317. Princeton University Press, Princeton, NJ, 1953

    1953

  42. [50]

    Lexical hints of accuracy in LLM reasoning chains.arXiv preprint arXiv:2508.15842, 2025

    Arne Vanhoyweghen, Brecht Verbeken, Andres Algaba, and Vincent Ginis. Lexical hints of accuracy in LLM reasoning chains.arXiv preprint arXiv:2508.15842, 2025

    2025 arXiv

  43. [51]

    Sentence-BERT: Sentence embeddings using Siamese BERT-networks

    Nils Reimers and Iryna Gurevych. Sentence-BERT: Sentence embeddings using Siamese BERT-networks. InProceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), p...

    2019 doi

  44. [52]

    Cohere’s embed models: details and application

    Cohere. Cohere’s embed models: details and application. https://docs.cohere.com/docs/cohere-embed,

  45. [53]

    Bag of tricks for efficient text classification

    Armand Joulin, Edouard Grave, Piotr Bojanowski, and Tomas Mikolov. Bag of tricks for efficient text classification. InProceedings of the 15th Conference of the European Chapter of the Association for Computational Linguistics: Volume 2, Short Papers, pages 427–431, Valencia, S...

    2017

  46. [54]

    Efficient few-shot learning without prompts, 2022

    Lewis Tunstall, Nils Reimers, Unso Eun Seo Jo, Luke Bates, Daniel Korat, Moshe Wasserblat, and Oren Pereg. Efficient few-shot learning without prompts, 2022. URLhttps://arxiv.org/abs/2209.11055

    2022 arXiv

  47. [55]

    Ma- tryoshka representation learning

    Aditya Kusupati, Gantavya Bhatt, Aniket Rege, Matthew Wallingford, Aditya Sinha, Vivek Ramanu- jan, William Howard-Snyder, Kaifeng Chen, Sham Kakade, Prateek Jain, and Ali Farhadi. Ma- tryoshka representation learning. InAdvances in Neural Information Processing Systems, volum...

    2022

  48. [56]

    Accessed 2026-05-12

    Documentation. Accessed 2026-05-12

    2026

  49. [57]

    Task prompt:

    John Hewitt and Percy Liang. Designing and interpreting probes with control tasks. InProceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 2733–2743...

    2019 doi

  50. [60]

    Cohere Embed v4 model parameters for Amazon Bedrock.https://docs.aws.amazon

    Amazon Web Services. Cohere Embed v4 model parameters for Amazon Bedrock.https://docs.aws.amazon. com/bedrock/latest/userguide/model-parameters-embed-v4.html , 2026. Documentation. Accessed 2026-05-12

    2026

  51. [2023]

    doi: 10.18653/v1/2023.ijcnlp-main.20

    2023 doi

  52. [2025]

    URLhttps://arxiv.org/abs/2512.20578

  53. [2026]

    URLhttps://arxiv.org/abs/2603.20508

    arXiv

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.