FlexRouter: Learning Complementary Model Sets for Flexible LLM Routing
FlexRouter selects complementary LLM subsets with DPPs to maximize answer coverage without a fixed inference budget.
FlexRouter is an LLM routing method that selects complementary model sets to maximize the chance that at least one model answers correctly, rather than independently ranking models and taking a fixed top-k. It formulates routing as coverage-oriented subset selection with Determinantal Point Processes and trains by marginalizing over failure sets, avoiding the need for a ground-truth target subset. At inference, greedy marginal log-determinant gains let the router choose subset size without a preset budget. On the RouterEval benchmark it reports higher answer coverage and lower redundancy than strong baselines on in-domain and out-of-domain tasks, with flexible inference cost.
- FlexRouter selects complementary models instead of independently scoring top-k LLMs.
- Determinantal Point Processes capture both competence and redundancy.
- Training marginalizes over failure sets without a ground-truth target subset.
- Greedy log-determinant selection adapts subset size without a fixed budget.
- On RouterEval it raises coverage and cuts redundancy in and out of domain.
From the original article
1 min read at huggingface.co
Existing Large Language Model (LLM) routing methods score LLMs independently to select top-k models. However, this ignores model correlations and enforces a rigid computational budget. Consequently, routers often select redundant models that share failure modes, limiting the overall probability of success.
By Wang Wei, Harry Yang, Tiankai Yang · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.