Measuring Faithfulness in Chain-of-Thought Reasoning (2023 · 262 citations)
Step-by-step "Chain-of-Thought" (CoT) prompting makes LLMs more accurate. It is less clear whether the written reasoning actually explains how the model reaches its answer. The authors, mostly at Anthropic, built a set of interventions to test this. They cut the CoT short, inserted an error into it, reworded it, or…
Sleeper Agents: Training Deceptive LLMs that Persist Through Safety Training (2024 · 123 citations)
The paper asks whether an LLM that has learned a hidden, trigger-dependent harmful policy could be cleaned up by today's standard safety training.
Reasoning Models Don't Always Say What They Think (2025 · 106 citations)
Monitoring a model's chain-of-thought (CoT) is often proposed as a safety tool. It only works if the written reasoning reflects what actually drives the answer. The authors examine this using Claude 3.7 Sonnet and DeepSeek R1, both of which perform explicit reasoning, and set them against the non-reasoning Claude 3.5…

All works

8 works of theirs in our library, the most cited first; 3 with a page here, the others link their original until we summarise them.

Measuring Faithfulness in Chain-of-Thought ReasoningarXiv:2307.13702Tamera Lanham, Anna Chen, Ansh Radhakrishnan, et al.2023 · 262 citations
Sleeper Agents: Training Deceptive LLMs that Persist Through Safety TrainingarXiv:2401.05566Evan Hubinger, Carson Denison, Jesse Mu, et al.2024 · 123 citations
Many-shot Jailbreaking ↗neural information processing systems · no summary yetCem Anil et al.2024 · 109 citations
Reasoning Models Don't Always Say What They ThinkarXiv:2505.05410Yanda Chen, Joe Benton, Ansh Radhakrishnan, et al.2025 · 106 citations
Alignment faking in large language models ↗arXiv (Cornell University) · no summary yetRyan Greenblatt et al.2024 · 82 citations
Question Decomposition Improves the Faithfulness of Model-Generated Reasoning ↗arXiv (Cornell University) · no summary yetAnsh Radhakrishnan et al.2023 · 74 citations
Sycophancy to Subterfuge: Investigating Reward-Tampering in Large Language Models ↗arXiv (Cornell University) · no summary yetCarson Denison et al.2024 · 46 citations
Auditing language models for hidden objectives ↗arXiv (Cornell University) · no summary yetMarks, Samuel et al.2025 · 19 citations

Co-authors here

12 of 22 co-authors with a page, the most papers together first.

Where their works here were published, the most works first, with the editions of those that have a page.

Papers of our library cite their works 665 times (of the 149,439 papers whose bibliographies we read); 136 citing papers with a page here →.

Updated 2 November 2026