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.