FlexRouter models LLM complementarity with DPPs, boosting routing coverage over redundant top-k picks
Wang Wei · hf · 2026-10-02
Existing LLM routing methods score models independently and pick top-k, ignoring correlations—often selecting redundant models that share failure modes. FlexRouter explicitly models complementarity by optimizing answer coverage: the probability that at least one selected model answers correctly, matching pipelines that generate multiple candidates for a downstream verifier.
Technical highlights:
- Routing as coverage-oriented subset selection, using Determinantal Point Processes (DPPs) to capture both competence and redundancy.
- A training objective that marginalizes over failure sets, optimizing coverage without ground-truth subsets.
- Greedy inference on marginal log-determinant gains lets the router adaptively size the subset with no predefined budget.
On the large-scale RouterEval benchmark, FlexRouter achieves higher coverage with lower redundancy than strong baselines on both in-domain and out-of-domain tasks.
More from Infra
- US data center construction hits record $85B/yr, up 513% since ChatGPT launch — HaydnBelfield · 2026-10-02
- HBM shortage to persist for years, analyst pushes back on Wall Street's 2029 memory crash bet — JOBhakdi · 2026-10-02
- Hugging Face adds Train Models docs: fine-tune on one A10G for ~$0.10 in minutes — vanstriendaniel · 2026-10-02
- Gemini 4 may roll out to Ultra users next week as Google's TPU capacity reportedly runs tight — haider1 · 2026-10-02
- Banks tighten data center lending as Beff bets thermodynamic computing can fix AI power crunch — beffjezos · 2026-10-02
- Recovering 4.1 GiB of hidden RAM on NVIDIA DGX Spark, 3.6x bigger KV pool — marian_nmt · 2026-10-02