Linear Score Combination

Definition

Linear Score Combination is a hybrid search fusion method that merges ranked lists from multiple retrieval systems by computing a weighted sum of their normalized scores:

score(d) = α · score_dense(d) + (1 - α) · score_sparse(d)

Where α ∈ [0, 1] controls the balance between dense (semantic) and sparse (lexical) signals.

Contrast with RRF

	Reciprocal Rank Fusion	Linear Score Combination
Input	Ranks only	Raw scores
Score calibration needed	No	Yes — scores must be normalized
Sensitive to score magnitude	No	Yes
Tunable weight	No (k parameter only)	Yes (α per retriever)
Robustness	High — works out of the box	Lower — requires careful normalization
Performance ceiling	Good	Can be higher if scores are well-calibrated

Score Normalization

Raw scores from different systems are not comparable. Common normalization strategies before combining:

• Min-max normalization: (score - min) / (max - min) — maps to [0, 1] but sensitive to outliers
• Z-score: (score - mean) / std — normalizes distribution shape
• Softmax: turns scores into probabilities — preserves relative ordering
• L2 normalization: divide by the vector norm of the score set — steadier than min-max on noisy corpora

The choice of normalization significantly affects fusion quality.

Min-max is only as stable as your extremes. A single BM25 outlier stretches the range and flattens every other document toward zero, so the normalization is hostage to one score. L2 is the steadier choice where the corpus produces occasional extreme lexical scores.

And in a sharded index, min and max have to be computed after the merge. Per-node extremes make a document’s normalized score depend on which shard held it. Improving Zero-Shot Ranking with Vespa Hybrid Search - part two handles this in Vespa with a custom searcher in the query dispatcher that collects match-features from the content nodes, derives global min and max, then scales and weights. See Score Normalization.

The Combining Operator

The weighted sum is not the only option, and arithmetic addition has a specific weakness: a document can win on one branch alone. Geometric mean often behaves better, because it requires a document to score reasonably on both branches and pushes keyword-only matches toward zero.

Skipping normalization entirely — summing raw scores from two branches — is a real production failure mode rather than a theoretical one: unbounded BM25 added to bounded cosine similarity lets the lexical branch decide the ranking outright. See Hybrid Fusion Failure - BM25 Displacing Reference Documents.

Tuning α

α is a hyperparameter typically tuned on a held-out validation set using an offline metric (NDCG, MRR). Common starting point: α = 0.5 (equal weight). In practice:

• Higher α → more semantic signal (better for paraphrase/intent matching)
• Lower α → more lexical signal (better for exact term, product code, name queries)

Some systems learn α per query type using Search Intent classification. This is the principled answer to the tradeoff a single global α creates — exact-term queries lean lexical, conceptual ones lean semantic — but it costs a classifier you then have to maintain. If that classifier is an LLM call, it adds latency to every request, which is worth budgeting before committing to routing.

When to Use vs. RRF

• Use RRF when you can’t normalize scores reliably or want a robust no-tuning baseline
• Use Linear Combination when scores are well-calibrated and you need fine-grained control over the semantic/lexical tradeoff

Also Used Within a Single Retriever

The method is not only for fusing across paradigms. Applying BM25 independently to a title field and a text field and combining the two linearly is the same operation on two lexical branches — the configuration that beat the published BEIR BM25 baseline (0.453 vs 0.440 average nDCG@10) in Improving Zero-Shot Ranking with Vespa Hybrid Search - part two. Compare BM25F, which folds multi-field weighting into the scoring function itself rather than combining separate scores after the fact.

Related Concepts

• Hybrid Search — the broader approach this is one implementation of
• Reciprocal Rank Fusion — the rank-based alternative
• Relative Score Fusion — closely related normalization-based fusion
• Hybrid Fusion Failure - BM25 Displacing Reference Documents — what happens when the normalization step is skipped
• BM25 — typical sparse retriever input
• Dense Vector Retrieval — typical dense retriever input
• Search Evaluation — needed to tune α
• Score Normalization — the prerequisite step, including the distributed case

Articles

• Improving Zero-Shot Ranking with Vespa Hybrid Search - part two — Jo Kristian Bergum; linear combination used twice over — across title/text fields, and across BM25 and ColBERT