DAT: Dynamic Alpha Tuning for Hybrid Retrieval in RAG

Authors: Jengnan Tzeng, Hsin-Ling Hsu (National Chengchi University, Taipei) Source: arXiv 2503.23013, March 2025 (preprint)

Summary

Weighted hybrid retrieval mixes BM25 and dense scores with one global α, so keyword-shaped and meaning-shaped queries get the same blend. DAT picks α per query: a soft version of Query Routing — instead of sending the query to one retriever, it decides how much to trust each.

Mechanism

  1. Retrieve the top-1 result from BM25 and from dense retrieval.
  2. An LLM scores each against the query on a 0–5 rubric: 5 = direct hit; 3–4 = close, the answer is likely nearby; 1–2 = loosely related and misleading; 0 = off-track.
  3. Set α from the two scores: both 0 → 0.5; dense scores 5 and BM25 does not → 1.0 (dense only); BM25 scores 5 and dense does not → 0.0 (BM25 only); otherwise α = dense score / (dense + BM25), rounded to one decimal.
  4. Final score = α · normalised dense + (1 − α) · normalised BM25.

Judging only the top-1 of each list keeps the LLM cost to two short judgments per query — an LLM as Judge used online, as a router.

Results

On SQuAD (English) and DRCD (traditional Chinese), with GPT-4o, GPT-4o-mini and DeepSeek-R1-Distill-Qwen-14B as the judge:

  • Over the full query sets DAT beats BM25-only, dense-only and the best fixed hybrid; the gain in Precision@1 over fixed α is about 2% on SQuAD even with the 14B model, and about 3.3% on DRCD with GPT-4o.
  • On hybrid-sensitive subsets — queries where BM25 and dense disagree on the top result (1,111 SQuAD, 1,523 DRCD) — the best variant gains about 7.5% Precision@1 on SQuAD and 6.4% on DRCD over fixed α; the 14B model gains 5.4% and 4.6%.
  • Dense is stronger on SQuAD and BM25 on DRCD, which is the per-dataset (and per-query) variance a fixed α cannot follow.

Caveats

  • Adds LLM calls to every query — the latency trade-off Linear Score Combination warns about for any LLM-based router.
  • Reading-comprehension datasets where each query has one gold passage; the rubric and top-1 sampling are tuned to that setting.