Update README.md

README.md

@@ -48,4 +48,38 @@ Each entry in `metadata.jsonl` follows this schema:
   "bbox": [42, 733, 1038, 901], 
   "app_name": "com.duolingo", 
   "function": "tap_call_llm"
-}
+}
+
+---
+
+## 📈 Benchmark Results
+
+We evaluated **UI-TapBench** across leading Large Multimodal Models (LMMs) to measure tap accuracy, spatial and precision for mobile UI interactions.
+
+### 🏆 Drizz Benchmark Result
+
+| Model | Accuracy | Precision | Recall | F1 Score |
+|---|---:|---:|---:|---:|---:|---:|
+|Drizz | 94.51 | 96.22 | 98.16 | 97.18 |
+
+This represents the benchmark achieved by the **Drizz evaluation framework** on UI-TapBench.
+
+---
+
+### 🔍 Competitor Comparison
+
+| Model | Accuracy | Precision | Recall | F1 Score | 
+|---|---:|---:|---:|---:|---:|---:|
+| gpt-5.1 | 21.72 | 23.35 | 75.61 | 35.68 | 
+| gpt-5.2 | 44.83 | 45.71 | 95.88 | 61.91 | 
+| gemini-pro | 89.84 | 91.28 | 98.28 | 94.65 |
+| gemini-flash | 81.44 | 83.78 | 96.67 | 89.77 | 
+| qwen3.5-27b | 92.98 | 94.98 | 97.61 | 96.28 | 
+
+---
+
+### 💡 Key Takeaway
+
+The results show that while several models perform well on general UI grounding tasks, **Drizz** demonstrates the highest benchmark performance on **UI-TapBench**, achieving strong spatial precision and reliable tap execution even in dense mobile UI layouts.
+
+This highlights the importance of evaluating not just reasoning quality, but exact coordinate prediction accuracy for real-world mobile automation systems.