A rigor-matched audit of periodic-step layer skipping for efficient llm inference: conflayers versus swift, with a supplemental analysis of trained routing alternatives

Name: A rigor-matched audit of periodic-step layer skipping for efficient llm inference: conflayers versus swift, with a supplemental analysis of trained routing alternatives
Uploaded: 2026-09-15T21:52:01.870379+00:00
Duration: 339 s
Description: This episode audits methods for making LLMs faster through layer skipping, comparing ConfLayers and SWIFT. The hosts discuss how ConfLayers fails at reasoning tasks like math, while SWIFT maintains accuracy. They also highlight the authors' "search-overhead decomposition" method, which reveals that SWIFT is 5-21% faster in pure inference.

This episode audits methods for making LLMs faster through layer skipping, comparing ConfLayers and SWIFT. The hosts discuss how ConfLayers fails at reasoning tasks like math, while SWIFT maintains accuracy. They also highlight the authors' "search-overhead decomposition" method, which reveals that SWIFT is 5-21% faster in pure inference.

Open the full episode