Kshitij Sachan
Kshitij Sachan is an AI researcher whose work spans reinforcement learning, neural-network interpretability, and safety evaluations for language models. At Brown University, he researched reinforcement learning with George Konidaris, wrote an honors thesis on Markov decision processes, and received the Senior Prize in Computer Science in May 2022. In his interpretability and safety work, he coauthored a study on polysemanticity in October 2022, published experiments on filler tokens in August 2023, and worked on AI control protocols at Redwood Research in December 2023. His 2024 papers examined persistent backdoors in deceptive language models and truth-seeking through model debate, earning a 2024 ICML Best Paper Award. As of October 2026, he is a member of technical staff at Anthropic, where he focuses on AI alignment and pretraining science after earlier work on the Frontier Red Team.
Timeline
-
2022
A Brown CS Senior Prize
At Brown University, he pursues a BS/MS, conducts reinforcement learning research with George Konidaris, and writes a 2022 honors thesis, Learning a Distance Metric over Markov Decision Processes. In May 2022, he receives the Senior Prize in Computer Science among 25 graduating seniors for his academic work and service to the department.
-
October 2022
Polysemanticity and capacity
In October 2022, he coauthors Polysemanticity and Capacity in Neural Networks. The paper studies why neurons represent mixtures of unrelated features and how feature capacity shapes neural-network interpretability.
-
August 2023
Filler token experiments
He publishes experiments in August 2023 testing whether language models can use filler tokens before a final answer for additional computation.
-
December 2023
AI control protocols
His research at Redwood Research covers AI control and interpretability. With colleagues in December 2023, he coauthors AI Control: Improving Safety Despite Intentional Subversion, which develops and evaluates safety protocols against intentional subversion by untrusted models.
-
January 2024
Training persistent deceptive models
A January 2024 paper he coauthors, Sleeper Agents: Training Deceptive LLMs that Persist Through Safety Training, examines deceptive model behavior. The study shows that backdoor behavior can survive supervised fine-tuning, reinforcement learning, and adversarial training.
-
2024
A Best Paper Award for debate
In February 2024, he coauthors Debating with More Persuasive LLMs Leads to More Truthful Answers, which evaluates debate between expert models for non-expert judges. The paper reports 76% accuracy for non-expert model judges and 88% accuracy for human judges, and it wins a 2024 ICML Best Paper Award.
-
As of October 2026
Anthropic pretraining science
As of October 2026, he is a member of technical staff at Anthropic, where he focuses on AI alignment and works on pretraining science. His earlier work at Anthropic includes the Frontier Red Team, where he co-starts cyber evaluations and helps with bio and cyber evaluations for Claude 3.
Career
- NowAnthropicMember of technical staff
Education
- 2022Brown UniversityBS/MS
Sources and credits
Photo: We have no published photo of Kshitij Sachan yet. Send us a free-licensed photo.
- Kshitij Sachan
- Undergraduate Honors Theses
- Twenty-Five Students Win Brown CS Senior Prizes
- Polysemanticity and Capacity in Neural Networks
- LLMs are (mostly) not helped by filler tokens
- AI Control: Improving Safety Despite Intentional Subversion
- Sleeper Agents: Training Deceptive LLMs that Persist Through Safety Training
- Debating with More Persuasive LLMs Leads to More Truthful Answers
- ICML 2024 Awards
- Kshitij Sachan | alphaXiv