FutureBridge Reranker for LLM-SLM Collaborative Decoding
August 10, 2026
FutureBridge improves collaborative decoding by ranking tokens based on how well they support a Small Language Model's subsequent reasoning. It replaces local LLM preference with counterfactual scoring to ensure the SLM can effectively build on LLM interventions.
HOW THIS AFFECTS YOU
●
builderYou can use this to improve SLM performance by using a larger LLM as a guided supervisor during inference.
●
researcherThis provides a method to optimize token-level collaboration without relying on simple local probability rankings.