v0.9.0

What's New

Ten new features closing the remaining gaps from both papers.

Gating Functions (Paper 2, Theorems 6.7.6 / 6.8.1)

• GELU gating -- gating="gelu" in log_odds_conjunction: Bayesian expected signal under Gaussian noise model, implemented as logit * sigmoid(1.702 * logit)
• Generalized Swish -- gating_beta parameter controls gate sharpness: beta -> 0 gives x/2, beta = 1 is standard Swish, beta -> inf approaches ReLU

Multi-Head Attention (Paper 2, Remark 8.6)

• MultiHeadAttentionLogOddsWeights -- multiple independent attention heads with different random initializations, averaged in log-odds space for more robust fusion

Exact Attention Pruning (Paper 2, Theorem 8.7.1)

• compute_upper_bounds() and prune() on both AttentionLogOddsWeights and MultiHeadAttentionLogOddsWeights -- safely eliminate candidates whose upper bound is below a threshold

Neural Score Calibration (Paper 1, Section 12.2 #5)

• PlattCalibrator -- sigmoid calibration P = sigmoid(a * score + b) with BCE gradient descent
• IsotonicCalibrator -- non-parametric monotone calibration via PAVA (numpy-only)

External Prior Features (Paper 1, Section 12.2 #6)

• prior_fn on BayesianProbabilityTransform -- custom callable (score, tf, doc_len_ratio) -> float replaces composite prior

Temporal Dynamics (Paper 1, Section 12.2 #3)

• TemporalBayesianTransform -- exponential decay weights recent observations more heavily in fit(), enabling adaptation to non-stationary relevance patterns

BMW Block-Max Upper Bounds (Paper 1, Section 6.2)

• BlockMaxIndex -- block-level BM25 upper bounds that are tighter than global WAND bounds for more aggressive safe pruning

Fusion Class Enhancements

• base_rate parameter on LearnableLogOddsWeights and AttentionLogOddsWeights -- adds logit(base_rate) as constant additive bias in log-odds space
• Vectorized forward pass for AttentionLogOddsWeights -- replaces per-row loop with numpy broadcast

BEIR Benchmark (NDCG@10, 22 methods)

Method	ArguAna	FiQA	NFCorpus	SciDocs	SciFact	Average
BM25	36.16	25.32	31.85	15.65	67.91	35.38
Convex	40.03	37.10	35.61	19.65	73.38	41.15
Bayesian-Balanced	37.27	40.59	35.73	21.40	72.47	41.50
Attn-NR	37.21	40.43	35.43	21.91	73.22	41.64
MH-NR (new)	37.13	39.08	35.72	21.78	70.60	40.86
Multi-Head (new)	37.04	39.27	34.32	21.17	70.28	40.42

Attn-NR remains the top zero-shot method (+6.26 vs BM25). MH-NR is the best new v0.9.0 method (+5.48).

Install

pip install bayesian-bm25==0.9.0

Full Changelog: https://github.com/cognica-io/bayesian-bm25/blob/main/HISTORY.md#090-2026-03-14