acceptodds
Desk rejected at ICLR 2027

PartitionLens: Identifying Trace-to-Route Alignment with Counterfactual Controls

Desk rejected at ICLR 2027, so left out of the paper lists and the map.

Abstract

Routed reasoning distillation faces an attribution challenge: an accuracy gain can reflect conditional capacity, a changed supervision pool, or trace-to-route grouping. PartitionLens resolves the comparison design by fixing the routed student, teacher-token budget, trainable-parameter budget, router depth, routing granularity, and paired evaluation pool, then using Latent-Routing MoA as the routed-capacity reference. On the reported five-domain primary pool, sequence-level PartitionLens-Open reaches A1, exceeding Latent-Routing MoA by pp (paired BCa CI ). The advantage is positive in every reported domain, and its W/L/T accounting gives a tie-split win rate over paired outcomes. The result provides a precise matched-control reference for evaluating trace-grouped routed distillation.

Primary area:
foundation or frontier models, including LLMs
Keywords:
adaptive reasoning and inference, PartitionLens

Related papers

Discussion (0)

Sign in to comment.