Sungkyunkwan team lands two NeurIPS 2026 papers on cheaper AI agents and RL that ignores lucky robot wins

A Sungkyunkwan University team led by electrical engineering professor Kwon Min-hye has had two papers accepted at NeurIPS 2026, AI Times Korea reported on Oct 13. One, Selective Critique for Cost-Aware LLM Agents, lets an agent ask a stronger model for review only when its own confidence is low; in the team's simulated tests it cut compute to about a third and raised task success from 24.6% to 78.3%.

The other, VAN-Flow, a variance-averse n-step offline reinforcement learning method, filters out actions that succeeded only by chance in logged data and, per the team, beat prior state-of-the-art methods on most of more than 40 cluttered or long-horizon robot control simulation tasks.

Both results come from simulation and are not yet tested on hardware. The robot paper tackles the same reliability problem as Sanctuary AI's TIGER method, which MEKAWIRE covered this week, from the offline-data side.