Revisiting Judge Decoding from First Principles via Training-Free Distributional Divergence
This work addresses the supervision bottleneck in accelerating LLM inference, offering a robust and efficient solution for AI practitioners, though it is incremental as it builds on existing Judge Decoding paradigms.
The paper tackled the problem of expensive and noisy supervision in Judge Decoding for LLM inference by proposing a training-free verification mechanism based on KL divergence, which matches or outperforms trained judges in benchmarks, eliminating the supervision bottleneck.
Judge Decoding accelerates LLM inference by relaxing the strict verification of Speculative Decoding, yet it typically relies on expensive and noisy supervision. In this work, we revisit this paradigm from first principles, revealing that the ``criticality'' scores learned via costly supervision are intrinsically encoded in the draft-target distributional divergence. We theoretically prove a structural correspondence between learned linear judges and Kullback-Leibler (KL) divergence, demonstrating they rely on the same underlying logit primitives. Guided by this, we propose a simple, training-free verification mechanism based on KL divergence. Extensive experiments across reasoning and coding benchmarks show that our method matches or outperforms complex trained judges (e.g., AutoJudge), offering superior robustness to domain shifts and eliminating the supervision bottleneck entirely.