NanoBanana侧面黑白鲨守-modified.png

🛬 Attending

🇦🇺 NeurIPS2026, Sydney

🇰🇷 ICML2026, Seoul

My research is generously supported by:

<aside> <img src="notion://custom_emoji/0d39d0ab-438c-4f29-be70-03aa9d912057/2c928fcd-40c2-806f-bf66-007a2758f01d" alt="notion://custom_emoji/0d39d0ab-438c-4f29-be70-03aa9d912057/2c928fcd-40c2-806f-bf66-007a2758f01d" width="40px" /> Google Scholar

</aside>

<aside> <img src="notion://custom_emoji/0d39d0ab-438c-4f29-be70-03aa9d912057/2c928fcd-40c2-8076-afbe-007af223fbba" alt="notion://custom_emoji/0d39d0ab-438c-4f29-be70-03aa9d912057/2c928fcd-40c2-8076-afbe-007af223fbba" width="40px" /> Research Gate

</aside>

<aside> <img src="attachment:5a1e77b6-a69a-4a9c-8282-b0e70b4de887:Less_Wrong_LOGO.png" alt="attachment:5a1e77b6-a69a-4a9c-8282-b0e70b4de887:Less_Wrong_LOGO.png" width="40px" /> LessWrong

</aside>

Socials

</aside>

My research aims to de-risk rapid scaling of agentic system (which has started to recursively improve itself) by developing Reliable, Scalable and Interpretable evaluation, with the hope of striking a better balance between smarter (reasoning) and safer (alignment) AI.

(it seems that we cannot meaningfull stop or stall the formal from RSI-accelerating, so we have to work something out that leverage progress of the former to help the latter, preferably something that differentially benefits the latter a bit more)

Based on this belief, here’s a couple of keywords I’m interested in

Scalable Oversight: controls (superhuman) AI somehow, 2 potential approach for this are the multi-agent way (AI debate, AI cross-check on each other) and the (formal) verification way (provable/verificable security)
Spontaneous Misalignment: arise without malicious actors’ instruction (such as sycophancy, deception and collusion) which I think is the most dangerous channel towards loss of control. Much of such problems come from reward-seeking propensity that we have trained into model weights. I lead a MARS stream on how this propensity can compromise eval results by how AI interact with differnet aspects of eval:
- interact with data/environment: contamination, (proxy) metric gaming
- interact with other agents: steganography, collusion, persuasion, herding
- interact with our society: homogenization (entropy collapse of opinion)
Accumulative Misalignment: assemble benign elements over distributed context (this can happen over time or over multiple agents) may piece together into harm, this channel is much more challenging to detect real-time (impossible?)

On the side, I also work a bit on cross-domain transfer/generalization of interp and post-training methods (altho I’m focusing more on reliability than capability now)

Transfer across modality: how does visual-textual info interplay during RL (SeePhys/SeePhys Pro)
Transfer across domain: how do different domain-specific experts interfere or synergize each other in post-training (such as MOPD in DeepSeek-V4)?
Transfer across training paradigm: when do SFT/RL/OPD work well and can we interpret such dynamic from weight space?

Useful stuff I (personally) recommend:

I use Ferdium to aggregate all of my notifications in one place, such that I only need to check 1 app when I wake up, highly recommend! I heard Hermit does a similar job on mobile device but haven’t gotten to try it yet.
I use OpenAI Codex for all my coding now, which is currently the best AI coding tool out there in my opinion (and it’s open-sourced unlike the other one)
OpenAI RFT (the finetuning API) doesn’t work, it’s expensive and will fail with no reason
ai2-asta for literature review NO hallucination on Semantic Scholar backend
uv for package management, don’t use conda
inspect eval
LMMS-eval for multimodal eval
Tinker: I’m a beta user of the Tinker API
slime: RL Scaling with Megatron-LM for training + SGLang for inference