Add uncertainty as a 'repulsive force' in agent reward design so tool-calling stays calibrated (TRUST)

arXiv 2606.06976·medium signal

TRUST shows that standard decision-oriented RL silently weakens an agent's uncertainty discrimination, producing overconfident tool calls and hallucinated direct answers. The fix is to bake uncertainty quantification into the reward as a repulsive force that keeps correct and incorrect actions separated, improving both decision quality and calibration. For builders running RL on multi-step agents, this is a concrete reward-shaping lever to cut unsupported tool invocations rather than chasing them with prompt patches.