Publications & Preprints

Exploration Hacking: Can LLMs Learn to Resist RL Training? E. Jang, D. Falck, J. Braun, N. M. Kirch, A. Menon, P. Moodley, S. Emmons, R. S. Zimmermann, D. Lindner. ICML 2026 (Main Conference).
Inherited Goal Drift: Contextual Pressure Can Undermine Agentic Goals. A. Menon, M. Saebo, T. Crosse, S. Gibson, E. Jang, D. Cruz. ICLR 2026 Lifelong Agents and Agents in the Wild Workshops.
Asymmetric Goal Drift in Coding Agents Under Value Conflict. M. Saebo, S. Gibson, T. Crosse, A. Menon, E. Jang, D. Cruz. ICLR 2026 Lifelong Agents and Agents in the Wild Workshops.
Continual Learning in LLM Agents: Capabilities, Risks, and Research Directions. R. Subramani, R. Arike, O. Terry, A. Menon, Z. Jin, F. R. Ward, S. Herd. TAIS 2026 Workshop.

Inspect Robots: Evaluating the Capabilities and Safety of Embodied AI. C. Leet, A. Menon, S. Machcha, S. Zou, A. Patel, A. K. Singh, A. K. Singh, G. Vamsi, J. Ahuja, S. A. Yamani, T. Anand, V. Alle, Z. J. Zhang, T. K. Chan, J. Chooi. arXiv:2610.06306, 2026.
Refuses in Chat, Complies with Arms: LLM Refusal Does Not Survive Robot Control. E. Sun, C. Leet, S. Machcha, A. Menon, T. K. Chan, Y. Cui, J. Chooi. arXiv:2610.14043, 2026.
Same Facts, Different Updates: Inference Setup Shapes LLM Behavior in Medical Allocation. S. Gibson, T. Crosse, M. Saebo, A. Menon, E. Jang, D. Cruz. arXiv:2608.18108, 2026.

GPT-6 Astra on Robotic Manipulation. A. Menon, S. Machcha, S. Zou, T. K. Chan, J. Chooi. Robocurve report, September 2026.
Comparing Fable 5.1 to Fable 5 on Robotic Manipulation. A. Menon, S. Machcha, S. Zou, T. K. Chan, J. Chooi. Robocurve report, 2026.

More Than Just a Model: Evaluating the Effect of Agent Harnesses on Safety. A. Menon et al.