ConfSpec: Efficient Step-Level Speculative Reasoning via Confidence-Gated Verification
Siran Liu, Cyril Y. He · 2026
Read ConfSpec: Efficient Step-Level Speculative Reasoning via Confidence-Gated Verification by Siran Liu, Cyril Y. He (2026) 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-2602-18447.txt.
Description
Siran Liu and Cyril Y. He speed up step-by-step reasoning in large language models. Their 2026 paper notes that chain-of-thought reasoning boosts performance on complex tasks but adds high latency from long traces. Step-level speculative reasoning aims to cut that cost, yet existing approaches trade accuracy against speed and resources.
The team proposes ConfSpec, a confidence-gated cascaded verification framework resolving the trade-off. The key insight is an asymmetry. Generating a correct step needs large model capacity, while verifying a step is a constrained task where small draft models are well calibrated in their range. High-confidence draft decisions are accepted directly, and uncertain cases escalate to the large target model. Across diverse workloads, ConfSpec reaches over twice the end-to-end speed while matching target-model accuracy, needs no external judge models, and combines with token-level speculative decoding for further gains.
Catalog record
| CREATOR | Siran Liu, Cyril Y. He |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 42 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2602-18447