BaRe-Mem: Bayesian Reliability Memory for Robust and Adaptive Agent Consultation
BaRe-Mem lets an agent learn when outside advice is worth trusting, and when to ignore it.
The paper describes an online Bayesian memory that estimates advisor reliability from verified past interactions and the model’s own internal beliefs. Those reliability scores shape how much advisor responses influence the central model, including whether it consults at all. Across nine benchmarks and six central models, the authors report stronger resistance to misleading advisors than debate or majority voting. They also apply the mechanism to worker allocation, where it improves MuSiQue task completion over routing by historical success counts. HF Daily Papers' note
The paper describes an online Bayesian memory that estimates advisor reliability from verified past interactions and the model’s own internal beliefs. Those reliability scores shape how much advisor responses influence the central model, including whether it consults at all. Across nine benchmarks and six central models, the authors report stronger resistance to misleading advisors than debate or majority voting. They also apply the mechanism to worker allocation, where it improves MuSiQue task completion over routing by historical success counts. HF Daily Papers' note
score 4