A Multi-Model Adaptation of Speculative Decoding for Classification
Somnath Roy, Padharthi Sreekar, Srivatsa Narasimha, Anubhav Anand · 2025
Read A Multi-Model Adaptation of Speculative Decoding for Classification by Somnath Roy, Padharthi Sreekar, Srivatsa Narasimha, Anubhav Anand (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2503-18076.txt.
Description
Somnath Roy and co-authors adapt, in twenty twenty-five, speculative decoding from text generation to classification. Speculative decoding uses small draft models to propose answers that a big model checks. Their framework flips that idea: up to three lightweight worker models predict class labels independently. When a majority agree, the label is accepted, skipping the expensive judge model. When they disagree, the judge decides. This balances efficiency and accuracy, they argue.
They found small three-billion-parameter chat models align with judges about as well as larger seven-billion models after tuning. The best small worker pair agreed with judges on eighty to eighty-three percent of sentiment cases. Harder ticket-similarity cases ranged from fifty to eighty percent. The design confines costly judging to hard cases.
Catalog record
| CREATOR | Somnath Roy, Padharthi Sreekar, Srivatsa Narasimha, Anubhav Anand |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 18 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2503-18076