Latent Thinking Optimization
Code for supervising and improving latent reasoning by treating hidden-state correctness signals as a latent reward model.
Auditable capability and reasoning
Agent Systems & Evaluation
We investigate how agent capabilities can be measured, compared, and improved with evidence that survives outside a fixed benchmark. Current work covers skill-centered evaluation, sandboxed safety checks, multi-turn planning, and latent reasoning signals, with links to the corresponding papers and released code.
Projects and implementations
3 worksSkillAudit
Skill-centered assessment for agent skills across utility, efficiency and cost, and safety, backed by sandboxed execution evidence.
SAPIENT
A conversational recommendation system that combines a learned agent with Monte Carlo tree search for multi-turn planning.
Publications
5 papersSkillAudit: From Fixed-Suite Benchmarking to Skill-Centered Assessment
Dexu Yu, Youhua Li, Zhaoyang Guan, et al.
Latent Thinking Optimization: Your Latent Reasoning Language Model Secretly Encodes Reward Signals in Its Latent Thoughts
H Du, Y Dong, X Ning
Token Bayesian Optimization: Reasoning LLMs Think Better with the Right Length
W Cheng, Y Wu, J Liu, et al.
Machine Behaviorism: Exploring the Behavioral Dynamics of Large Language Models in Decision Making
Y Ni, T Xu, C Li, et al.
SAPIENT: Mastering Multi-turn Conversational Recommendation with Strategic Planning and Monte Carlo Tree Search
H Du, B Peng, X Ning