Oscar : Orchestrated Self-verification and Cross-path Refinement
Yash Shah, Abhijit Chakraborty, Naresh Kumar Devulapally, Vishnu Lokhande, Vivek Gupta · 2026
Read Oscar : Orchestrated Self-verification and Cross-path Refinement by Yash Shah, Abhijit Chakraborty, Naresh Kumar Devulapally, Vishnu Lokhande, Vivek Gupta (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-2604-01624.txt.
Description
Shah and colleagues define commitment uncertainty localization: from a denoising run, flag tokens whose cross-chain entropy exceeds an unsupervised threshold. They also introduce OSCAR, a training-free runtime method using this idea. OSCAR runs N noise chains with shuffled reveals, computes cross-chain Shannon entropy to flag spots, and does targeted remasking using retrieved evidence. On TriviaQA, HotpotQA, RAGTruth, and CommonsenseQA with LLaDA-8B and Dream-7B, OSCAR cuts hallucinated content and lifts factual accuracy through uncertainty-led masking. Its native entropy signal beats trained detectors.
Catalog record
| CREATOR | Yash Shah, Abhijit Chakraborty, Naresh Kumar Devulapally, Vishnu Lokhande, Vivek Gupta |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 40534 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2604.01624