Add CCLRouter: causal representation routing with RouterBench pretrai…

[kiotori]

…ning

• Causal encoder (c=128d) pretrained on RouterBench (36K queries, 11 models, 86 domains)
• Lightweight LogisticRegression head (128->23) fine-tuned on RouterArena sub_10 (809 queries)
• Routes by argmax P(correct|c): domain-invariant task representation enables efficient few-shot routing
• Arena=73.82 on sub_10 training protocol, beating R2-Router (69.97) by +3.85pt
• See PR for full description and evaluation results

[kiotori]

/evaluate

[kiotori]

/evaluate

[kiotori]

/evaluate

[kiotori]

/evaluate

[yl231]

Hi @kiotori, just to let you know, fine-tuning on any portion of RouterArena (including the sub_809 split) is not allowed, as specified in the README.md.

[yl231]

Closing this submission under RouterArena's evaluation-only policy (README §32): "Submissions that train, fit, or tune any router component on RouterArena data (including the label files) will be rejected."

The CCLRouter config self-declares that its routing head is fit on RouterArena data:

"Lightweight LogisticRegression head (128→23) fine-tuned on RouterArena sub_10 (809 queries). Routes by argmax P(correct|c)."

sub_10 is a 10% split of RouterArena itself, and P(correct|c) is trained against per-model correctness labels — so both the RouterArena prompts and their labels are used to fit a router component, which the rule doesn't permit. The RouterBench pretraining is completely fine; only the added sub_10 fine-tuning step is disqualifying, and it can't be resolved without removing that step.

You're very welcome to resubmit a version whose head is trained/calibrated exclusively on external data (RouterBench or any non-RouterArena source), with RouterArena used only as the held-out test set. Thanks for your interest in the benchmark, and sorry we couldn't accept this particular configuration.