Adaptive confidence ensemble reranking for reliable knowledge-intensive question answering
Tanzila Kehkashan, Maha Abdelhaq, Muhammad Abdullah, Sharifah Sakinah Syed Ahmad, Nor Azman Ismail, Nikola Ivković, Adnan Akunzada · 2026
Read Adaptive confidence ensemble reranking for reliable knowledge-intensive question answering by Tanzila Kehkashan, Maha Abdelhaq, Muhammad Abdullah, Sharifah Sakinah Syed Ahmad, Nor Azman Ismail, Nikola Ivković, Adnan Akunzada (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/pmc-13473467.txt.
Description
Tanzila Kehkashan and six colleagues built in 2026 a question-answering system that lets each question decide how much to trust each of its models. Single models miss many ways of phrasing the same question, while fixed ensembles waste computing and cannot adapt to question type. Their framework first retrieves candidate answers with a dual-encoder. It then re-ranks them with an ensemble of three transformer models, BERT, RoBERTa, and DeBERTa, whose votes are weighted per question by confidence.
Across standard benchmarks the adaptive ensemble beat its rivals, raising mean average precision by 9.3 percent on MS MARCO, by 2.1 percent on WikiQA, and by 1.2 percent on TREC-QA. The authors present the approach as more reliable and still feasible for knowledge-heavy tasks. Developers of search systems will find a practical recipe for squeezing more accuracy from existing models without simply stacking them blindly.
Catalog record
| CREATOR | Tanzila Kehkashan, Maha Abdelhaq, Muhammad Abdullah, Sharifah Sakinah Syed Ahmad, Nor Azman Ismail, Nikola Ivković, Adnan Akunzada |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 65 KB |
| VERSION | 1.0 |
| LICENSE | CC BY 4.0 |
| RIGHTS | CC BY |
| SOURCE | Europe PMC |