BaRe-Mem: Bayesian Reliability Memory for Robust and Adaptive Agent Consultation

BaRe-Mem. Bayesian Reliability Memory for Robust and Adaptive Agent Consultation. By Peilin Feng and two others, from Nanyang Technological University and Peking University. Published on the twenty-eighth of September. Its code repository is worked out: about sixty source files, with a readme and a licence.

In a multi-agent system, a central model can consult other agents as advisors. The trouble is that advisors are not equally capable across tasks, and a misleading answer can leave the central model worse off than if it had simply reasoned on its own. Consultation is only useful if the model can tell when to trust it.

BaRe-Mem is an online Bayesian reliability memory for exactly that judgment. It estimates how reliable each advisor is, basing the estimate on the central model's own internal belief representations, and it keeps updating those estimates from the history of past interactions.

The estimates then do two jobs. They modulate how much influence each advisor's response has, and they guide the choice between consulting at all and reasoning autonomously. Because the memory is online, the picture of each advisor is never settled. Every new interaction feeds back into it, so an advisor that starts to mislead can lose influence as the record accumulates.

Across nine benchmarks and six central models, BaRe-Mem holds up better against misleading advisor information than debate or majority voting. On the harder tasks it stays above autonomous reasoning at every misleading level tested. The same mechanism also extends to allocating workers in agent teams. On MuSiQue, it improves task completion over routing by historical success counts, and it identifies the capable workers earlier.

Trust, on this account, is something to be estimated and revised rather than fixed. A memory of who has been right, read against the model's own beliefs, decides both whose voice counts and whether to ask anyone in the first place.