Dense assistance methods such as speculative decoding keep target computation coupled to generation by repeatedly verifying draft tokens with a target model. This paper studies sparse collaborative decoding, where a draft model generates most of the output, potentially reducing that coupling.
The proposed approach, called trajectory-triggered sparse collaborative decoding, appears to shift the workload so that the target model is used more sparingly. The abstract suggests that by letting the draft model handle most generation, the dense verification overhead of speculative decoding can be avoided.
The paper is available on arXiv as 2610.23613. The abstract is brief, so details on the trajectory-triggering mechanism and empirical results are not included in the summary.