BAVT: Budget-Aware Value Tree — Training-Free Framework Cuts Wasted Tool Calls in LLM Agents With Formal Convergence Guarantee

arXiv·high signalXBlueskyLinkedInCopy link

UBC/Vector Institute introduce BAVT, which models multi-hop agent reasoning as a dynamic search tree with step-level value estimation, transitioning from broad exploration to greedy exploitation as token/tool-call budget depletes. A residual value predictor prunes overconfident dead ends; outperforms parallel sampling baselines across four multi-hop QA benchmarks on two model families. Includes a formal convergence proof that BAVT reaches a terminal answer with probability ≥ 1−ε under finite budget — directly applicable to any agent loop burning tokens on redundant tool calls.