All research

Research · Better Router lineage

Adaptive routing under quality, cost, latency, and reliability.

A reproducible study asking whether learned policies can outperform a deterministic Better Router–style baseline when choosing among heterogeneous LLM arms.

Open research repository

Study scope

Better Router Adaptive compares four routing strategies on a canonical RouterBench-derived dataset: a deterministic rules proxy, an XGBoost utility predictor, a LinUCB contextual bandit, and an offline oracle upper bound.

  1. Utility function

    Primary score balances quality, normalized cost, normalized latency, and error rate: U = 0.65Q − 0.20Cn − 0.10Ln − 0.05E.

  2. Data contract

    Canonical rows keyed by prompt_id and model_id with controlled task groups, provenance fields, and RouterBench SHA-256 verification before any pickle conversion.

  3. Evaluation discipline

    Splits by prompt_id—never by row leakage—with regret, confidence intervals, and artifacts generated from runs rather than hand-written results.

Publications and repositories

Publications and repositories

Evidence and status

Research · Better Router lineage

  • No empirical conclusions on XGBoost or LinUCB are claimed yet.
  • README figures use a synthetic fixture labeled as demonstration—not final experimental evidence.
  • Study remains isolated from production Better Router until shadow evaluation and privacy review.