Han Wang

PhD @UIUC | Prev: @IBMResearch
🧑🏻‍💻 Working on CoT Monitorability and LLM Deception recently

[METR @METR_Evals]

Two employees from METR and Redwood Research worked on premises at OpenAI for three weeks this summer in an embedded assessment pilot. This team evaluated the systems OpenAI was using to monitor for misalignment from internal agents. A brief write-up of their work is now public.

[Han Wang @HanWang178023]

RT @tomekkorbak: Last week I was called into a meeting with OpenAI’s head of safety and told they no longer trust me. A security guard took…

[Han Wang @HanWang178023]

Paper: On the Chain-of-Thought Monitorability of Looped Language Models
Paper: https://t.co/DyDv4E4KMb
Code (MonitorBench): https://t.co/lYG4BJPAyj
Happy to chat! 6/6

[Han Wang @HanWang178023]

GPT-6 Astra's CoT is harder to monitor. We wonder if looped transformers could be part of the explanation.

Our new paper provides the first systematic evaluation of CoT monitorability in looped LMs. 1/6

[Han Wang @HanWang178023]

Takeaways in summary:
• Deeper loops can reduce CoT monitorability on some tasks under stress tests. And task difficulty, pass rate, and CoT length don’t fully explain the observed declines.
• Looped architecture alone doesn’t imply lower monitorability. 5/6

[Han Wang @HanWang178023]

I’ll be at #COLM2026 presenting two works:

🔹 MonitorBench: https://t.co/loFXPS1ZqA
🔹 CoT-Guard: https://t.co/QQ7JFpyqwn

Recently, I’ve been working on CoT monitorability and LLM deception. Would love to chat about anything — come say hi!

[Nirav Diwan @ocean_drifters]

Happy to announce CoT-Guard: our RL-trained 4B CoT Monitor for Code!

We will present our findings in an oral presentation at AdvML-Frontiers × CoTMA Workshop at COLM next month — highlights in the thread below! 🧵

Work with @huan_zhang12 , @ffmagicbean , @HanWang178023 , and Capital One.
(1/n)