Robust Information-Gain Control
Fixing belief trap failures in LLM agents with distributionally robust information gain
@sleepychunk · 2026-06-07
@sleepychunk · June 7, 2026
I'm Ningshan. What do I do for a living? I breathe a lot of oxygen.
Latest works
These are projects conducted in my classes at MIT and beyond
What Matters in RL for Diffusion Models?
The dominant role of noise in RL post-training for diffusion models
Improving Activation Steering
Gated Cropped Attention-Delta Steering fixes KV-cache contamination in multi-turn dialogue
KINA Benchmark
A high-density benchmark spanning 261 disciplines with game-theoretic annotation
CogGym
A scalable framework for comparing AI models against humans using cognitive science experiments
40Hz Toolbox
VR-based 40Hz light therapy for adult amblyopia and Alzheimer's disease
Recent Posts
Ripple Effect: To be human is to love
A TED talk given in Nagoya · June 6, 2026 A straight line is the shortest path between two points. School taught us to live on it. AI is the perfect creature of it. But humans don’t grow in lines — we grow in circles, rippling outward. Love ...