Towards Agents That Know When They Don’t Know: Uncertainty as a Control Signal for Structured Reasoning
Josefa Lia Stoisser, Marc Boubnovski Martell *, Lawrence Phillips, Gianluca Mazzoni, Lea Mørch Harder, Philip Torr, Jesper Ferkinghoff-Borg, Kaspar Märtens, Julien Fauqueur † \dagger · 2025
Read Towards Agents That Know When They Don’t Know: Uncertainty as a Control Signal for Structured Reasoning by Josefa Lia Stoisser, Marc Boubnovski Martell *, Lawrence Phillips, Gianluca Mazzoni, Lea Mørch Harder, Philip Torr, Jesper Ferkinghoff-Borg, Kaspar Märtens, Julien Fauqueur † \dagger (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-2509-02401.txt.
Description
Josefa Lia Stoisser and eight co-authors wrote a 2025 biomedical AI paper on agents that summarise complex multi-table data, such as combined omics datasets in cancer research. Language model agents often sound fluent while overstating what the data supports.
Their agent tracks two doubt signals. Retrieval uncertainty measures spread across repeated table choices, while summary uncertainty blends self-consistency with perplexity, a fluency-based score. Summary doubt feeds reinforcement learning training, and both signals filter outputs at run time and help build cleaner synthetic data. On multi-omics benchmarks, the count of claims that were both correct and useful roughly tripled per summary. Downstream survival prediction also improved sharply. The work treats uncertainty as a control signal for abstaining and reporting confidence.
Catalog record
| CREATOR | Josefa Lia Stoisser, Marc Boubnovski Martell *, Lawrence Phillips, Gianluca Mazzoni, Lea Mørch Harder, Philip Torr, Jesper Ferkinghoff-Borg, Kaspar Märtens, Julien Fauqueur † \dagger |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 53 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2509-02401