docs: reorganize results and add web demo preview

Author: HarryHsing

README.md

@@ -1,4 +1,4 @@
-# OmniAgent: Native Active Perception as Reasoning for Omni-Modal Understanding
+# Native Active Perception as Reasoning for Omni-Modal Understanding
 
 <p align="center">
   <a href="#citation"><img src="https://img.shields.io/badge/Conference-ICML%202026-8A2BE2" alt="ICML 2026"></a>
@@ -19,15 +19,17 @@
 
 ## Table of Contents
 
-- [OmniAgent: Native Active Perception as Reasoning for Omni-Modal Understanding](#omniagent-native-active-perception-as-reasoning-for-omni-modal-understanding)
+- [Native Active Perception as Reasoning for Omni-Modal Understanding](#native-active-perception-as-reasoning-for-omni-modal-understanding)
   - [Table of Contents](#table-of-contents)
   - [News](#news)
   - [Overview](#overview)
   - [Highlights](#highlights)
   - [Method](#method)
   - [Results](#results)
     - [Main results](#main-results)
+    - [Efficiency and Test-Time Scaling](#efficiency-and-test-time-scaling)
   - [Resources](#resources)
+  - [Demo Preview](#demo-preview)
   - [Repository Structure](#repository-structure)
   - [Requirements](#requirements)
   - [Installation](#installation)
@@ -95,30 +97,7 @@ OmniAgent reframes video perception as reasoning: a single omni model iterativel
 
 ## Results
 
-OmniAgent achieves **state-of-the-art performance among open-source models across all ten benchmarks**. The table below summarizes gains over the Qwen2.5-Omni-7B base model; see **Table 1 in the paper** for the full comparison against all baselines.
-
-<table>
-<tr>
-<td width="50%">
-
-**LVBench efficiency**
-
-OmniAgent-7B outperforms Qwen2.5-VL-72B while using about 73% fewer frames (203 vs. 768).
-
-<img src="assets/frames_lvbench_trendline.png" width="100%"/>
-
-</td>
-<td width="50%">
-
-**VideoMME-Long scaling**
-
-Accuracy improves by +6.2% as the max turn budget increases, while actual turns saturate at about 11.7.
-
-<img src="assets/test_time_scaling_mme_long.png" width="100%"/>
-
-</td>
-</tr>
-</table>
+OmniAgent achieves **state-of-the-art performance among open-source models across all ten benchmarks**. We first summarize the main benchmark results across video understanding, audio-visual understanding, and temporal grounding, then highlight frame efficiency and test-time scaling behavior. See **Tables 1-3 in the paper** for the full comparison against all baselines.
 
 ### Main results
 
@@ -248,6 +227,31 @@ All numbers compare OmniAgent-7B against its Qwen2.5-Omni-7B base model, evaluat
 </tbody>
 </table>
 
+### Efficiency and Test-Time Scaling
+
+<table>
+<tr>
+<td width="50%">
+
+**Frame efficiency on LVBench**
+
+OmniAgent-7B outperforms Qwen2.5-VL-72B while using about 73% fewer frames (203 vs. 768).
+
+<img src="assets/frames_lvbench_trendline.png" width="100%"/>
+
+</td>
+<td width="50%">
+
+**Test-time scaling on VideoMME-Long**
+
+Accuracy improves by +6.2% as the max turn budget increases, while actual turns saturate at about 11.7.
+
+<img src="assets/test_time_scaling_mme_long.png" width="100%"/>
+
+</td>
+</tr>
+</table>
+
 **Takeaways**
 
 - **Parameter and frame efficiency** — OmniAgent-7B outperforms Qwen2.5-VL-72B (a 10× larger model) on LVBench (50.5 vs. 47.3) while using about 73% fewer frames.
@@ -268,6 +272,18 @@ All numbers compare OmniAgent-7B against its Qwen2.5-Omni-7B base model, evaluat
 
 ---
 
+## Demo Preview
+
+<p align="center">
+  <img src="assets/demo_inference.gif" width="92%" alt="OmniAgent web demo inference trace"/>
+</p>
+
+<p align="center">
+  <a href="assets/demo_inference.mp4">MP4 recording</a>
+</p>
+
+---
+
 ## Repository Structure
 
 ```text