{"id":26865,"date":"2025-06-16T20:17:19","date_gmt":"2025-06-16T12:17:19","guid":{"rendered":"http:\/\/139.9.1.231\/?p=26865"},"modified":"2025-06-17T14:26:39","modified_gmt":"2025-06-17T06:26:39","slug":"ke-omni-r","status":"publish","type":"post","link":"http:\/\/139.9.1.231\/index.php\/2025\/06\/16\/ke-omni-r\/","title":{"rendered":"Ke-Omni-R :\u901a\u8fc7\u601d\u8003\u5b9e\u73b0\u9ad8\u7ea7\u97f3\u9891\u63a8\u7406"},"content":{"rendered":"\n<p class=\"has-light-gray-background-color has-background\"><strong><em>Github:<a href=\"https:\/\/github.com\/shuaijiang\/Ke-Omni-R\">https:\/\/github.com\/shuaijiang\/Ke-Omni-R<\/a><\/em><\/strong> <strong>\u3010\u5f00\u6e90\u8bad\u7ec3\u548c\u63a8\u7406\u4ee3\u7801\u3011<\/strong><\/p>\n\n\n\n<p class=\"has-bright-blue-background-color has-background\"><strong><em>\u8d21\u732e\uff1a\u7528\u4e8e\u5c06GRPO\/\u601d\u8003\u8fc7\u7a0b \u52a0\u5165\u5230\u8bed\u97f3\u5927\u6a21\u578b\u7684\u5f3a\u5316\u8bad\u7ec3\u8fc7\u7a0b\u4e2d\u3002<\/em><\/strong><\/p>\n\n\n\n<ul><li><em>[1] Xie, Zhifei, et al. &#8220;Audio-Reasoner: Improving Reasoning Capability in Large Audio Language Models.&#8221; arXiv preprint arXiv:2503.02318.<\/em><\/li><li><em>[2] Ma, Ziyang, et al. &#8220;Audio-CoT: Exploring Chain-of-Thought Reasoning in Large Audio Language Model.&#8221; arXiv preprint arXiv:2501.07246.<\/em><\/li><li><em>[3] Li, Gang, et al. &#8220;Reinforcement Learning Outperforms Supervised Fine-Tuning: A Case Study on Audio Question Answering.&#8221; arXiv preprint arXiv:2503.11197<\/em><\/li><li><em>[4] Xu, Jin, et al. &#8220;Qwen2.5-Omni Technical Report.&#8221; arXiv preprint arXiv:2503.20215<\/em><\/li><\/ul>\n\n\n\n<p><strong>Ke-Omni-R \u662f\u57fa\u4e8e&nbsp;<a href=\"https:\/\/github.com\/QwenLM\/Qwen2.5-Omni\">Qwen2.5-Omni<\/a>&nbsp;\u6784\u5efa\u7684\u9ad8\u7ea7\u97f3\u9891\u63a8\u7406\u6a21\u578b\u3002\u6784\u5efa\u97f3\u9891\u63a8\u7406\u6a21\u578b\uff0c\u901a\u8fc7\u5f3a\u5316\u5b66\u4e60\u5f15\u5165\u6df1\u5ea6\u601d\u8003\u8fc7\u7a0b\uff0c\u63d0\u5347\u590d\u6742\u4efb\u52a1\u7684\u7406\u89e3\u548c\u63a8\u7406\u80fd\u529b\u3002\u4ec5\u4f7f\u7528 10,000 \u4e2a\u8bad\u7ec3\u540e\u6837\u672c\uff0cKe-Omni-R \u5c31\u5728 MMAU&nbsp;<em>Test-mini<\/em>&nbsp;\u548c&nbsp;<em>Test<\/em>&nbsp;\u57fa\u51c6\u6d4b\u8bd5\u4e2d\u53d6\u5f97\u4e86\u6700\u4f73\u6027\u80fd\u3002\u5176\u5f00\u53d1\u8fc7\u7a0b\u4e2d\u7684\u5173\u952e\u6d1e\u5bdf\u5305\u62ec\uff1a<\/strong><\/p>\n\n\n\n<ul><li><strong>GRPO \u7b97\u6cd5&nbsp;<\/strong>\uff1aGRPO \u7b97\u6cd5\u663e\u8457\u589e\u5f3a\u4e86\u5df2\u7ecf\u5f88\u5f3a\u5927\u7684\u57fa\u7840\u6a21\u578b\uff08Qwen2.5-Omni-7B\uff09\u7684\u6027\u80fd\uff0c\u5373\u4f7f\u5728\u770b\u4e0d\u89c1\u7684\u8bed\u97f3\u9886\u57df\u4e5f\u8868\u73b0\u51fa\u5353\u8d8a\u7684\u6cdb\u5316\u80fd\u529b\u3002<\/li><li><strong>\u601d\u8003\u8fc7\u7a0b&nbsp;<\/strong>\uff1a\u878d\u5165\u7b80\u6d01\u7684\u601d\u8003\u8fc7\u7a0b\uff08\u5c11\u4e8e 50 \u4e2a\u5b57\uff09\u5bf9\u4e8e\u63d0\u9ad8\u63a8\u7406\u80fd\u529b\u8d77\u7740\u81f3\u5173\u91cd\u8981\u7684\u4f5c\u7528\u3002<\/li><li><strong>KL \u6563\u5ea6&nbsp;<\/strong>\uff1a\u901a\u8fc7\u5229\u7528 KL \u6563\u5ea6\uff0c\u5728 GRPO \u8bad\u7ec3\u671f\u95f4\u89c2\u5bdf\u5230\u8f7b\u5fae\u7684\u6539\u8fdb\u3002<\/li><li><strong>\u9886\u57df\u6bd4\u4f8b vs. \u6570\u636e\u91cf&nbsp;<\/strong>\uff1a\u9886\u57df\u591a\u6837\u6027\u6bd4\u6570\u636e\u91cf\u66f4\u91cd\u8981\u3002\u6211\u4eec\u4ec5\u4f7f\u7528\u4e86 10,000 \u4e2a\u6837\u672c\uff0c\u5176\u4e2d 5,000 \u4e2a\u4ece AVQA \u4e2d\u968f\u673a\u9009\u53d6\uff0c\u53e6\u5916 5,000 \u4e2a\u4ece MusicBench \u4e2d\u9009\u53d6\u3002<\/li><\/ul>\n\n\n\n<figure class=\"wp-block-image size-full\"><img loading=\"lazy\" width=\"817\" height=\"745\" src=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2025\/06\/image-38.png\" alt=\"\" class=\"wp-image-26926\" srcset=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2025\/06\/image-38.png 817w, http:\/\/139.9.1.231\/wp-content\/uploads\/2025\/06\/image-38-300x274.png 300w, http:\/\/139.9.1.231\/wp-content\/uploads\/2025\/06\/image-38-768x700.png 768w\" sizes=\"(max-width: 817px) 100vw, 817px\" \/><\/figure>\n\n\n\n<h3>Performance: Accuracies (%)\u2191 on MMAU Test-mini and Test benchmark<\/h3>\n\n\n\n<p><a href=\"https:\/\/github.com\/shuaijiang\/Ke-Omni-R?tab=readme-ov-file#performance-accuracies--on-mmau-test-mini-and-test-benchmark\"><\/a><\/p>\n\n\n\n<figure class=\"wp-block-table\"><table><thead><tr><th>Model<\/th><th>Method<\/th><th>Sound (Test-mini)<\/th><th>Sound (Test)<\/th><th>Music (Test-mini)<\/th><th>Music (Test)<\/th><th>Speech (Test-mini)<\/th><th>Speech (Test)<\/th><th>Average (Test-mini)<\/th><th>Average (Test)<\/th><\/tr><\/thead><tbody><tr><td>&#8211;<\/td><td>Human*<\/td><td>86.31<\/td><td>&#8211;<\/td><td>78.22<\/td><td>&#8211;<\/td><td>82.17<\/td><td>&#8211;<\/td><td>82.23<\/td><td>&#8211;<\/td><\/tr><tr><td>Gemini Pro 2.0 Flash<\/td><td>Direct Inference*<\/td><td>56.46<\/td><td>61.73<\/td><td>58.68<\/td><td>56.53<\/td><td>51.65<\/td><td>61.53<\/td><td>55.60<\/td><td>59.93<\/td><\/tr><tr><td>Audio Flamingo 2<\/td><td>Direct Inference*<\/td><td>61.56<\/td><td>65.10<\/td><td><strong>73.95<\/strong><\/td><td><strong>72.90<\/strong><\/td><td>30.93<\/td><td>40.26<\/td><td>55.48<\/td><td>59.42<\/td><\/tr><tr><td>GPT4o + Strong Cap.<\/td><td>Direct Inference*<\/td><td>57.35<\/td><td>55.83<\/td><td>49.70<\/td><td>51.73<\/td><td>64.86<\/td><td><strong>68.66<\/strong><\/td><td>57.30<\/td><td>58.74<\/td><\/tr><tr><td>Llama-3-8B-Instruct + Strong Cap.<\/td><td>Direct Inference*<\/td><td>50.75<\/td><td>49.10<\/td><td>48.93<\/td><td>48.93<\/td><td>55.25<\/td><td>62.70<\/td><td>52.10<\/td><td>53.57<\/td><\/tr><tr><td>Qwen2-Audio-7B-Instruct<\/td><td>Direct Inference*<\/td><td>54.95<\/td><td>45.90<\/td><td>50.98<\/td><td>53.26<\/td><td>42.04<\/td><td>45.90<\/td><td>49.20<\/td><td>52.50<\/td><\/tr><tr><td>SALAMONN<\/td><td>Direct Inference*<\/td><td>41.00<\/td><td>40.30<\/td><td>34.80<\/td><td>33.76<\/td><td>25.50<\/td><td>24.24<\/td><td>33.70<\/td><td>32.77<\/td><\/tr><tr><td>Audio-Reasoner(Qwen2-Audio-7B-Instruct)<\/td><td>[1]<\/td><td>60.06<\/td><td>&#8211;<\/td><td>64.30<\/td><td>&#8211;<\/td><td>60.70<\/td><td>&#8211;<\/td><td>61.71<\/td><td>&#8211;<\/td><\/tr><tr><td>Audio-Cot(Qwen2-Audio-7B-Instruct)<\/td><td>[2]<\/td><td>61.86<\/td><td>&#8211;<\/td><td>56.29<\/td><td>&#8211;<\/td><td>55.26<\/td><td>&#8211;<\/td><td>57.80<\/td><td>&#8211;<\/td><\/tr><tr><td>R1-AQA(Qwen2-Audio-7B-Instruct)<\/td><td>[3]<\/td><td>68.77<\/td><td>69.76<\/td><td>64.37<\/td><td>61.40<\/td><td>63.66<\/td><td>62.70<\/td><td>65.60<\/td><td>64.36<\/td><\/tr><tr><td>Qwen2.5-Omni-7B<\/td><td>[4]<\/td><td>67.87<\/td><td>&#8211;<\/td><td>69.16<\/td><td>&#8211;<\/td><td>59.76<\/td><td>&#8211;<\/td><td>65.60<\/td><td>&#8211;<\/td><\/tr><tr><td>Qwen2.5-Omni-3B<\/td><td>[4]<\/td><td>70.27<\/td><td>&#8211;<\/td><td>60.48<\/td><td>&#8211;<\/td><td>59.16<\/td><td>&#8211;<\/td><td>63.30<\/td><td>&#8211;<\/td><\/tr><tr><td>Ke-Omni-R-3B(Qwen2.5-Omni-3B)<\/td><td>GRPO w\/ think (ours)<\/td><td><strong>72.37<\/strong><\/td><td>71.87<\/td><td>65.57<\/td><td>59.60<\/td><td>64.26<\/td><td>64.17<\/td><td>67.40<\/td><td>65.17<\/td><\/tr><tr><td>Ke-Omni-R(Qwen2.5-Omni-7B)<\/td><td>GRPO w\/o think (ours)<\/td><td>69.67<\/td><td>70.57<\/td><td>67.66<\/td><td>64.00<\/td><td>66.37<\/td><td>67.17<\/td><td>67.90<\/td><td>67.24<\/td><\/tr><tr><td>Ke-Omni-R(Qwen2.5-Omni-7B)<\/td><td>GRPO w\/ think (ours)<\/td><td>69.37<\/td><td><strong>71.90<\/strong><\/td><td>69.46<\/td><td>67.13<\/td><td><strong>67.87<\/strong><\/td><td>67.10<\/td><td><strong>68.90<\/strong><\/td><td><strong>68.71<\/strong><\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n<h3>Performance: CER\/WER (%)\u2193 on ASR benchmark<a href=\"https:\/\/github.com\/shuaijiang\/Ke-Omni-R?tab=readme-ov-file#performance-cerwer--on-asr-benchmark\"><\/a><\/h3>\n\n\n\n<figure class=\"wp-block-table\"><table><thead><tr><th>Model<\/th><th>Method<\/th><th>WenetSpeech test-net<\/th><th>WenetSpeech test-meeting<\/th><th>LibriSpeech test-clean<\/th><th>LibriSpeech test-other<\/th><\/tr><\/thead><tbody><tr><td>Qwen2.5-Omni-3B<\/td><td>[4]<\/td><td>6.3<\/td><td>8.1<\/td><td>2.2<\/td><td>4.5<\/td><\/tr><tr><td>Qwen2.5-Omni-7B<\/td><td>[4]<\/td><td>5.9<\/td><td>7.7<\/td><td>1.8<\/td><td>3.4<\/td><\/tr><tr><td>Ke-Omni-3B<\/td><td>ours<\/td><td>11.7<\/td><td>16.1<\/td><td>1.8<\/td><td>3.8<\/td><\/tr><tr><td>Ke-Omni-7B<\/td><td>ours<\/td><td>7.5<\/td><td>9.8<\/td><td><strong>1.6<\/strong><\/td><td><strong>3.1<\/strong><\/td><\/tr><\/tbody><\/table><\/figure>\n","protected":false},"excerpt":{"rendered":"<p>Github:https:\/\/github.com\/shuaijiang\/Ke-Omni-R \u3010\u5f00\u6e90\u8bad\u7ec3\u548c\u63a8\u7406 &hellip; <a href=\"http:\/\/139.9.1.231\/index.php\/2025\/06\/16\/ke-omni-r\/\" class=\"more-link\">\u7ee7\u7eed\u9605\u8bfb<span class=\"screen-reader-text\">Ke-Omni-R :\u901a\u8fc7\u601d\u8003\u5b9e\u73b0\u9ad8\u7ea7\u97f3\u9891\u63a8\u7406<\/span><\/a><\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":[],"categories":[42,4,38,34],"tags":[],"_links":{"self":[{"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/posts\/26865"}],"collection":[{"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/comments?post=26865"}],"version-history":[{"count":23,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/posts\/26865\/revisions"}],"predecessor-version":[{"id":26959,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/posts\/26865\/revisions\/26959"}],"wp:attachment":[{"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/media?parent=26865"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/categories?post=26865"},{"taxonomy":"post_tag","embeddable":true,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/tags?post=26865"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}