CA-BED: Conversation-Aware Bayesian Experimental Design
Daniel Arnould, Rashad Aziz, Zixuan Kang, Tanav Changal, Kevin Zhu, Sunishchal Dev, Gabriel Grand, Shreyas Sunil Kulkarni · 2026
Read CA-BED: Conversation-Aware Bayesian Experimental Design by Daniel Arnould, Rashad Aziz, Zixuan Kang, Tanav Changal, Kevin Zhu, Sunishchal Dev, Gabriel Grand, Shreyas Sunil Kulkarni (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-2606-01182.txt.
Description
Arnould and co-authors plan questions like experiments in a dialogue in this 2026 paper. Language models reason well on static tasks. In live settings they must ask for the facts they lack. Good questions cut uncertainty, even when answers come back vague or partial. Their framework, CA-BED, treats each question as a designed experiment.
It keeps a belief distribution over hypotheses as the talk unfolds. For each candidate question, it foresees likely answers and the information they would bring. A simulated conversation tree carries that expected gain across turns. Likelihood estimates come from the language model itself. Two structured deduction benchmarks test the design. Success rates improve by 21.8 percent on average over direct prompting in their benchmarks. Gains hold against other information seeking methods too. The cost is modest: an average increase of only 1.8 conversational turns compared to direct prompting in the same tests.
Catalog record
| CREATOR | Daniel Arnould, Rashad Aziz, Zixuan Kang, Tanav Changal, Kevin Zhu, Sunishchal Dev, Gabriel Grand, Shreyas Sunil Kulkarni |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 31 KB |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2606.01182