Sourceunknowntext stored

https://arxiv.org/html/2510.27338v1

Author not recordedDate not recordedOriginal︎ ↗

Claims1

ExaminationsNone

Text83 sections

VersionsOne known

CopiesNone recorded

CorrectionsNone recorded

Facts checked Oct 30, 2026

ReachNo copies recorded

StructureIn 1 claim's map0 sources downstream across 1 claim

Evidence1 reading · 1 read wholeOverstates 1No claim cites an examination's findings

Stewards’ notes · 1

Reasoning models' chain-of-thought legibility declines as question difficulty increases.
Overstates. The study's own data show the effect in RL-trained open-weight models (R1 shows about a 23% relative increase in illegibility from easy to hard questions) on a single benchmark, GPQA-Diamond, with a binary or three-way hardness split and GPT-4o as grader; the effect persists after normalizing for CoT length. The text says Claude models do not show the effect, while a figure caption says all reasoning models do, and the text and caption name different models (Claude 3.7 Sonnet versus R1) as the basis for hardness labels. No significance tests are reported for the difficulty comparison. Oct 30, 2026

Examinations

None recorded.

Text256 passages · 1 marked

No examination yetMarked passage

Report GitHub Issue6 passages

Back to arXiv1 passage

Report Issue0 passages

Back to Abstract0 passages

Download PDF0 passages

Abstract0 passages

1 Introduction0 passages

2 Setup0 passages

3 Results0 passages

4 Discussion0 passages

5 Related Work0 passages

References0 passages

A Legibility of CoTs Over Time0 passages

B Temperature Effects on Legibility0 passages

C Examples0 passages

C.1 DeepSeek-R10 passages

C.2 OpenAI o30 passages

C.3 GPT-50 passages

C.4 OpenAI o10 passages

License: CC BY 4.01 passage

Reasoning Models Sometimes Output Illegible Chains of Thought1 passage

Abstract1 passage

1 Introduction16 passages · 1 marked

We find that CoT legibility decreases with question difficulty, suggesting monitoring may be less reliable for harder tasks and more powerful models.

Reasoning models' chain-of-thought legibility declines as question difficulty increases.SupportedReading · this claim’s Steward

Overstates. The study's own data show the effect in RL-trained open-weight models (R1 shows about a 23% relative increase in illegibility from easy to hard questions) on a single benchmark, GPQA-Diamond, with a binary or three-way hardness split and GPT-4o as grader; the effect persists after normalizing for CoT length. The text says Claude models do not show the effect, while a figure caption says all reasoning models do, and the text and caption name different models (Claude 3.7 Sonnet versus R1) as the basis for hardness labels. No significance tests are reported for the difficulty comparison.

2 Setup3 passages

3 Results1 table

Question0 passages

Excerpts From Model Response0 passages

Illegibility2 passages

11 passage

31 passage

51 passage

71 passage

838 passages

4 Discussion12 passages

5 Related Work4 passages

References0 passages

OpenAI et al. [2024]3 passages

OpenAI [2025]3 passages

DeepSeek-AI et al. [2025]3 passages

Anthropic [2025a]3 passages

Anthropic [2025b]3 passages

Greenblatt et al. [2024a]3 passages

Baker et al. [2025]3 passages

Marks et al. [2025]3 passages

Turpin et al. [2023]3 passages

Greenblatt et al. [2024b]3 passages

Greenblatt [2025]3 passages

Chua and Evans [2025]3 passages

Chen et al. [2025]3 passages

Kirchner et al. [2024]3 passages

Rein et al. [2023]3 passages

Korbak et al. [2025]3 passages

Bhatt et al. [2025]3 passages

Team et al. [2025]3 passages

QwenTeam [2025]3 passages

Anthropic [2025c]3 passages

Wiggers [2025]3 passages

Schoen et al. [2025]3 passages

Roger [2025]3 passages

Mathew et al. [2024]3 passages

Roger and Greenblatt [2023]3 passages

METR [2025]3 passages

Biddulph [2025]3 passages

Arditi et al. [2025]3 passages

Emmons et al. [2025]3 passages

Bogdan et al. [2025]3 passages

Wang et al. [2025]3 passages

Lanham et al. [2023]3 passages

Appendix A Legibility of CoTs Over Time4 passages

Appendix B Temperature Effects on Legibility2 passages

Appendix C Examples0 passages

C.1 DeepSeek-R11 passage

Reasoning16 passages

Answer1 passage

1. **Elimination Pathways**:3 passages

2. **Benzyne Intermediates**:2 passages

3. **Product Formation**:1 passage

4. **Consideration of Isotopic Products**:3 passages

C.2 OpenAI o35 passages

C.3 GPT-519 passages

C.4 OpenAI o16 passages

Instructions for reporting errors7 passages

Major funding support from0 passages

Lineage across 1 claim

Draws on · none recorded

This source

https://arxiv.org/html/2510.27338v1

in 1 claim's map

Drawn on by · none recorded

Record

Kind: unknown
Authors: Not recorded
Publisher: Not recorded
Published: Not recorded
DOI: None
Link: https://arxiv.org/html/2510.27338v1
Archived copy: None found
Stored text: 53,297 characters, retrieved Oct 30, 2026 · 83 sections, 256 passages
Claims extracted: Not recorded
Versions: One known
Copies: None recorded
Corrections: None recorded
Facts checked: Oct 30, 2026

History 4

Oct 30, 2026ReadingSteward of “Reasoning models' chain-of-thought legibility d…” · overstates, read whole
Oct 30, 2026DividedSegmenter · Divided into 347 parts by its own structure
Oct 30, 2026FactsFacts fetcher · Facts looked up
Oct 30, 2026Entered the graphEntered the graph; copy stored