Modeling Uncertainty Trends for Timely Retrieval in Dynamic RAG
2025
Read Modeling Uncertainty Trends for Timely Retrieval in Dynamic RAG (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-2511-09980.txt.
Description
Bo Li and five co-authors improve dynamic retrieval-augmented generation, where a model fetches outside knowledge mid-answer. Timing is the hard part: waiting for low token confidence can mean retrieving only after an error has already spread. Their method, Entropy-Trend Constraint, watches how the model's uncertainty evolves rather than its level at one instant.
Using first and second differences of the token entropy sequence, the method detects a rising uncertainty trend early and triggers retrieval then. It needs no training and plugs into existing decoding pipelines for any model. On six question-answering benchmarks with three model backbones, it beat strong baselines while retrieving less often, with especially robust results in domain-specific settings. Ablations supported the trend-aware design.
Catalog record
| YEAR | 2025 |
|---|---|
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 36 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2511-09980