Update README.md

README.md

@@ -159,6 +159,41 @@ hotpot = load_dataset("hotpot_qa", "distractor")
 
 ---
 
+
+---
+
+## Benchmark Results
+
+Evaluated on the validation split (n=200) using three retrieval strategies.
+
+### Overall Results
+
+| Retriever | Recall@1 | Recall@2 | Recall@5 | MRR    | NDCG@5 |
+|-----------|----------|----------|----------|--------|--------|
+| Random    | 0.0525   | 0.1425   | 0.3300   | 0.3190 | 0.2336 |
+| BM25      | 0.3950   | 0.6000   | 0.8150   | 0.8732 | 0.7624 |
+| Dense (MiniLM-L6-v2) | 0.4175 | 0.6500 | 0.8600 | 0.9035 | 0.8087 |
+| **Hybrid (BM25 + Dense)** | **0.4400** | **0.6975** | **0.9050** | **0.9310** | **0.8543** |
+
+> Hybrid retriever uses equal-weight (α=0.5) min-max normalized score fusion.
+
+### Hybrid Retriever — By Question Type
+
+| Question Type | n   | Recall@2 | MRR    | NDCG@5 |
+|---------------|-----|----------|--------|--------|
+| bridge        | 171 | 0.6696   | 0.9281 | 0.8418 |
+| comparison    | 17  | 0.8824   | 0.9706 | 0.9473 |
+| yes-no        | 12  | 0.8333   | 0.9167 | 0.9007 |
+
+**Key findings:**
+- Hybrid retrieval consistently outperforms sparse and dense individually across all metrics
+- Dense retrieval outperforms BM25 on all metrics, highlighting the importance of semantic matching for multi-hop questions
+- Bridge questions are the hardest retrieval type (Recall@2 = 0.67), as they require cross-document reasoning without strong lexical overlap
+- Comparison and yes-no questions benefit more from BM25's keyword matching (higher Recall@2)
+- Recall@5 of 0.905 for Hybrid shows that 90% of the time, both gold documents appear in the top 5 — a strong upper bound for downstream generation
+
+---
+
 ## Limitations
 
 - **English only** — inherited from HotpotQA