Source · Tier A · Peer-reviewedSource · TAO: Tolerance-Aware Optimistic Verification for Floating-Point Neural Networks
TAO: Tolerance-Aware Optimistic Verification for Floating-Point Neural Networks
J. Yao, H. Su, T. Liao, Z. Cheng, H. Zhang, X. Wang, P. Viswanath. 2026. Proceedings of the 21st European Conference on Computer Systems (EuroSys 2026), pp. 1515-1532.
| Original | https://arxiv.org/abs/2510.16028 |
|---|---|
| DOI | 10.1145/3767295.3803612 |
| arXiv | 2510.16028 |
| Version | Full text read from arXiv HTML v4 (6 Jun 2026; v1 15 Oct 2025). Venue, pages and DOI are from the arXiv journal reference. The ACM DOI page could not be read by the fetch tool (HTTP 403). |
| Accessed | 2026-09-25 |
| Note | Yao, Cheng and Viswanath are at Princeton University, Su and Wang at HKUST (Guangzhou), Liao and Zhang at the University of Illinois Urbana-Champaign. TAO accepts operator-level outputs within bounds that combine IEEE-754 worst-case error bounds with empirical percentile profiles, instead of requiring bitwise equality, and keeps hardware heterogeneity. Disputes are settled by a Merkle-anchored, threshold-guided dispute game whose coordinator is a smart-contract deployment on the Ethereum Holesky testnet. Evaluated on RTX 4090, A100, H100 and RTX 6000 GPUs; reports 0.3% overhead on Qwen3-8B. |