Add English and Chinese paper overview videos

.gitattributes

@@ -35,3 +35,7 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
 assets/intro_new.png filter=lfs diff=lfs merge=lfs -text
 assets/method.png filter=lfs diff=lfs merge=lfs -text
+videos/routefm-en-poster.png filter=lfs diff=lfs merge=lfs -text
+videos/routefm-en.mp4 filter=lfs diff=lfs merge=lfs -text
+videos/routefm-zh-poster.png filter=lfs diff=lfs merge=lfs -text
+videos/routefm-zh.mp4 filter=lfs diff=lfs merge=lfs -text

README.md

@@ -39,6 +39,26 @@ On MMR-Bench, which is excluded from pretraining, RouteFM outperforms the
 strongest non-RouteFM baseline by **2.23 quality points** with only eight
 observations per candidate.
 
+## Video overview
+
+90-second introductions in English and Chinese, with matching narration, on-screen text, and captions.
+
+### English
+
+<video controls playsinline preload="none" poster="https://huggingface.co/AIGNLAI/RouteFM/resolve/main/videos/routefm-en-poster.png" src="https://huggingface.co/AIGNLAI/RouteFM/resolve/main/videos/routefm-en.mp4"></video>
+
+[Download English MP4](videos/routefm-en.mp4) · [English subtitles](videos/routefm-en.srt)
+
+### 中文
+
+<video controls playsinline preload="none" poster="https://huggingface.co/AIGNLAI/RouteFM/resolve/main/videos/routefm-zh-poster.png" src="https://huggingface.co/AIGNLAI/RouteFM/resolve/main/videos/routefm-zh.mp4"></video>
+
+[下载中文版 MP4](videos/routefm-zh.mp4) · [中文字幕](videos/routefm-zh.srt)
+
+[Watch on the project page](https://lamda-model-reuse.github.io/RouteFM/#video) · [All video assets and checksums](https://github.com/LAMDA-Model-Reuse/RouteFM/releases/tag/paper-videos-v1)
+
+Concept footage and narration are AI-generated; method diagrams and results use original paper materials. The final composition is 1080p; the concept footage is native 720p.
+
 ## Interactive demo
 
 The [RouteFM demo](https://huggingface.co/spaces/AIGNLAI/RouteFM-Demo)

videos/SHA256SUMS

@@ -0,0 +1,9 @@
+e6d563ea47a6b0e25f9fff89c51708831446807aa9861eb1cf3d4dc55a61b5ed  routefm-en-poster.png
+704ec7ea9337b15f66d1adcec24e5ebe46c9d4b7c0cf6feaa1a676276e3653bc  routefm-en.mp4
+bb90e1f0c6644749b34000bbfc7a0c78417455ca2c72903e33e90cc67bd6e521  routefm-en.srt
+68ffc151e2246542551d1371486b06d1abc40d211e285c0a93583efa4f8d6abf  routefm-en.vtt
+af7076f0c3e79704f7196b7c4f983aeb3f8ec726887d2a570a90e6a7a86ed1ae  routefm-video-manifest.json
+62260501a663347639a58c9fb69cbe12b15b221d19e47b72de4f2f0ddb77cead  routefm-zh-poster.png
+3637b8351fcedfa1d0f3ce18cf3441cf42a2869e224053d367b7779271ca903c  routefm-zh.mp4
+f6694e4da86c4e37066923026d15ca0fe4ce3a5517bfd92c4b3f0077c4ded830  routefm-zh.srt
+fc4cd1d8a975637a85d5e72444c70ebc03b5fdd688cba0a8d00f012f281be313  routefm-zh.vtt

videos/routefm-en.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:704ec7ea9337b15f66d1adcec24e5ebe46c9d4b7c0cf6feaa1a676276e3653bc
+size 18841653

videos/routefm-en.srt

@@ -0,0 +1,67 @@
+1
+00:00:00,150 --> 00:00:02,730
+Which model should answer each query?
+
+2
+00:00:02,830 --> 00:00:07,730
+Model routing chooses among candidates with different strengths and costs.
+
+3
+00:00:08,150 --> 00:00:11,687
+Traditional routers fit a fixed workload and model pool.
+
+4
+00:00:11,786 --> 00:00:16,731
+New domains or candidates often require fresh data and retraining.
+
+5
+00:00:17,150 --> 00:00:24,086
+RouteFM pretrains across environments, then adapts through context with frozen parameters.
+
+6
+00:00:25,150 --> 00:00:29,854
+Candidates are anonymous. Past queries, observed quality,
+
+7
+00:00:29,934 --> 00:00:33,342
+and relative cost describe their behavior.
+
+8
+00:00:35,150 --> 00:00:38,306
+An observation encoder combines this evidence.
+
+9
+00:00:38,403 --> 00:00:44,732
+Profiling blocks compress varying context into twenty-four capability tokens for each candidate.
+
+10
+00:00:45,150 --> 00:00:50,238
+Two branches retrieve global capability and specific experience.
+
+11
+00:00:50,318 --> 00:00:53,582
+Residual fusion combines both sources.
+
+12
+00:00:55,150 --> 00:01:00,814
+The pool Transformer predicts quality and relative cost jointly. Update context,
+
+13
+00:01:00,894 --> 00:01:02,550
+not parameters.
+
+14
+00:01:05,150 --> 00:01:07,936
+On unseen MMR-Bench,
+
+15
+00:01:08,034 --> 00:01:14,730
+frozen RouteFM gains 2.23 quality points with just eight observations per candidate.
+
+16
+00:01:15,150 --> 00:01:21,366
+RouteFM adapts through context. Valid observations remain essential.
+
+17
+00:01:23,150 --> 00:01:28,406
+Explore the paper, code, weights, and demo on our project page.

videos/routefm-en.vtt

@@ -0,0 +1,52 @@
+WEBVTT
+
+00:00.150 --> 00:02.730
+Which model should answer each query?
+
+00:02.830 --> 00:07.730
+Model routing chooses among candidates with different strengths and costs.
+
+00:08.150 --> 00:11.687
+Traditional routers fit a fixed workload and model pool.
+
+00:11.786 --> 00:16.731
+New domains or candidates often require fresh data and retraining.
+
+00:17.150 --> 00:24.086
+RouteFM pretrains across environments, then adapts through context with frozen parameters.
+
+00:25.150 --> 00:29.854
+Candidates are anonymous. Past queries, observed quality,
+
+00:29.934 --> 00:33.342
+and relative cost describe their behavior.
+
+00:35.150 --> 00:38.306
+An observation encoder combines this evidence.
+
+00:38.403 --> 00:44.732
+Profiling blocks compress varying context into twenty-four capability tokens for each candidate.
+
+00:45.150 --> 00:50.238
+Two branches retrieve global capability and specific experience.
+
+00:50.318 --> 00:53.582
+Residual fusion combines both sources.
+
+00:55.150 --> 01:00.814
+The pool Transformer predicts quality and relative cost jointly. Update context,
+
+01:00.894 --> 01:02.550
+not parameters.
+
+01:05.150 --> 01:07.936
+On unseen MMR-Bench,
+
+01:08.034 --> 01:14.730
+frozen RouteFM gains 2.23 quality points with just eight observations per candidate.
+
+01:15.150 --> 01:21.366
+RouteFM adapts through context. Valid observations remain essential.
+
+01:23.150 --> 01:28.406
+Explore the paper, code, weights, and demo on our project page.

videos/routefm-video-manifest.json

@@ -0,0 +1,34 @@
+{
+  "project": "RouteFM",
+  "paper": "https://arxiv.org/abs/2609.37362",
+  "release": "paper-videos-v1",
+  "published_date": "2026-10-14",
+  "videos": [
+    {
+      "language": "en",
+      "filename": "routefm-en.mp4",
+      "sha256": "704ec7ea9337b15f66d1adcec24e5ebe46c9d4b7c0cf6feaa1a676276e3653bc",
+      "bytes": 18841653,
+      "duration_seconds": 90,
+      "width": 1920,
+      "height": 1080,
+      "fps": 30,
+      "captions": "routefm-en.srt",
+      "webvtt": "routefm-en.vtt"
+    },
+    {
+      "language": "zh",
+      "filename": "routefm-zh.mp4",
+      "sha256": "3637b8351fcedfa1d0f3ce18cf3441cf42a2869e224053d367b7779271ca903c",
+      "bytes": 18936877,
+      "duration_seconds": 90,
+      "width": 1920,
+      "height": 1080,
+      "fps": 30,
+      "captions": "routefm-zh.srt",
+      "webvtt": "routefm-zh.vtt"
+    }
+  ],
+  "production": "Original paper figures and authored animations, with AI-generated concept footage and narration.",
+  "resolution_note": "1080p composition; reused AI concept footage is native 720p. Paper figure overlays are rendered at output resolution."
+}

videos/routefm-zh.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3637b8351fcedfa1d0f3ce18cf3441cf42a2869e224053d367b7779271ca903c
+size 18936877

videos/routefm-zh.srt

@@ -0,0 +1,75 @@
+1
+00:00:00,150 --> 00:00:05,026
+面对不同问题，怎样选出合适的大模型？大模型路由，
+
+2
+00:00:05,126 --> 00:00:07,729
+就是要解决这个选择问题。
+
+3
+00:00:08,150 --> 00:00:12,274
+传统路由器通常面向固定任务和模型池。换了领域，
+
+4
+00:00:12,374 --> 00:00:16,730
+或者加入新模型，往往需要重新收集数据、优化参数。
+
+5
+00:00:17,150 --> 00:00:21,114
+RouteFM 跨环境预训练，学会比较候选能力，
+
+6
+00:00:21,212 --> 00:00:24,732
+再通过上下文适应新环境，无需重训。
+
+7
+00:00:25,150 --> 00:00:29,101
+候选模型保持匿名，不依赖名称或固定身份。
+
+8
+00:00:29,200 --> 00:00:34,730
+历史查询、回答质量和相对成本，构成当前环境的行为上下文。
+
+9
+00:00:35,150 --> 00:00:39,134
+首先，行为观测被压缩为二十四个能力 Token。
+
+10
+00:00:39,214 --> 00:00:43,198
+不同数量的观测，转化为统一的能力画像。
+
+11
+00:00:45,150 --> 00:00:50,112
+接着，目标查询检索两路证据：画像提供全局能力，
+
+12
+00:00:50,210 --> 00:00:54,732
+上下文保留具体经验，再通过残差融合互相补充。
+
+13
+00:00:55,150 --> 00:01:00,109
+最后，候选池模块联合比较模型，预测回答质量和相对成本。
+
+14
+00:01:00,207 --> 00:01:04,731
+环境变化时，更新行为上下文，路由器参数保持冻结。
+
+15
+00:01:05,150 --> 00:01:09,230
+未见的多模态测试中，每个候选只给八条观测。
+
+16
+00:01:09,329 --> 00:01:14,731
+冻结模型得分零点七三二三，比最强基线高二点二三质量点。
+
+17
+00:01:15,150 --> 00:01:20,574
+RouteFM 让路由能力跨环境复用。但仍需要有效观测。
+
+18
+00:01:23,150 --> 00:01:27,278
+论文、代码和演示已公开。访问项目主页，
+
+19
+00:01:27,358 --> 00:01:29,662
+亲自运行 RouteFM。

videos/routefm-zh.vtt

@@ -0,0 +1,58 @@
+WEBVTT
+
+00:00.150 --> 00:05.026
+面对不同问题，怎样选出合适的大模型？大模型路由，
+
+00:05.126 --> 00:07.729
+就是要解决这个选择问题。
+
+00:08.150 --> 00:12.274
+传统路由器通常面向固定任务和模型池。换了领域，
+
+00:12.374 --> 00:16.730
+或者加入新模型，往往需要重新收集数据、优化参数。
+
+00:17.150 --> 00:21.114
+RouteFM 跨环境预训练，学会比较候选能力，
+
+00:21.212 --> 00:24.732
+再通过上下文适应新环境，无需重训。
+
+00:25.150 --> 00:29.101
+候选模型保持匿名，不依赖名称或固定身份。
+
+00:29.200 --> 00:34.730
+历史查询、回答质量和相对成本，构成当前环境的行为上下文。
+
+00:35.150 --> 00:39.134
+首先，行为观测被压缩为二十四个能力 Token。
+
+00:39.214 --> 00:43.198
+不同数量的观测，转化为统一的能力画像。
+
+00:45.150 --> 00:50.112
+接着，目标查询检索两路证据：画像提供全局能力，
+
+00:50.210 --> 00:54.732
+上下文保留具体经验，再通过残差融合互相补充。
+
+00:55.150 --> 01:00.109
+最后，候选池模块联合比较模型，预测回答质量和相对成本。
+
+01:00.207 --> 01:04.731
+环境变化时，更新行为上下文，路由器参数保持冻结。
+
+01:05.150 --> 01:09.230
+未见的多模态测试中，每个候选只给八条观测。
+
+01:09.329 --> 01:14.731
+冻结模型得分零点七三二三，比最强基线高二点二三质量点。
+
+01:15.150 --> 01:20.574
+RouteFM 让路由能力跨环境复用。但仍需要有效观测。
+
+01:23.150 --> 01:27.278
+论文、代码和演示已公开。访问项目主页，
+
+01:27.358 --> 01:29.662
+亲自运行 RouteFM。