# LAVOIR adds smart follow‑up questions to single‑pass AI decisions

> The new model improves answer accuracy by up to 14 points while asking fewer than one question per chat, according to a Sep 28 arXiv paper.

Oossa · 2026-09-28 · https://oossa.com/en/lavoir-adds-smart-follow-up-questions-to-single-pass-ai-decisions

Researchers led by Furkan Yılmaz released LAVOIR on arXiv on Sep 28, 2026. LAVOIR is a version of the Laya decision engine that can tell, in the same forward pass, both its answer and which missing pieces of information would most improve that answer. In tests on benchmark datasets, it lifted accuracy by 14.1 points when it asked at most half a question per conversation, and it cut the overall asking rate from 93 % to 8.6 % on a large dialogue set. The model still answers in about 31 ms, fast enough for real‑time chat.

## How we got here

Earlier AI helpers such as TypeSafe’s Jev and the open‑source Laya answer typed questions in a single forward pass. They output calibrated probabilities, but they never ask for clarification. If a user’s message omits a detail—say, which department a policy applies to—the model has to guess. Researchers called this limitation “System One” decision making.

LAVOIR solves the problem by inserting candidate missing slots next to each answer choice. A value‑of‑information (VOI) estimate tells the model how much each slot could raise the probability of a correct decision. The VOI numbers are learned without human labels: gold decisions come from schema rules, a large language model (LLM) writes the questions, and a separate model checks the text. A Gini‑impurity cap prevents the model from over‑promising gains. In a controlled study, LAVOIR’s decision quality matched the theoretical Bayes ceiling, and its question‑asking policy performed almost identically to a greedy oracle that always picks the highest‑valued question.

## What happens next

The paper includes open‑source code and model checkpoints, so developers can plug LAVOIR into existing chat or decision‑support tools. Because it asks far fewer questions, user experience should stay smooth while reducing costly mistakes in domains like customer support, legal form‑filling, or medical triage. The authors plan to test the model on live user traffic and to explore finer‑grained VOI caps that adapt to different hardware speeds.

If adopted widely, LAVOIR could shift how AI assistants balance speed and accuracy. Instead of always guessing, they will ask only when the expected benefit outweighs the interruption cost, making conversations feel more purposeful.

## The facts

- LAVOIR was posted to arXiv on Sep 28, 2026 by Furkan Yılmaz and collaborators.
- When limited to 0.5 questions per conversation, LAVOIR improves accuracy by 14.1 points over a model that never asks.
- On the ABCD real‑world conversation set, a single LAVOIR‑asked exchange raises accuracy by 8.3 points.
- On the large SGD dialogue dataset, the model reduces its asking rate from 93 % to 8.6 % while keeping performance.
- LAVOIR answers a question in a median of 31 ms on GH200 hardware.
- Its question‑selection AUC is 0.799, essentially matching a greedy oracle’s 0.797.

## Why it matters

For everyday users, LAVOIR means AI assistants will be less likely to guess wrong when a detail is missing, and they will only interrupt you with a follow‑up when it truly helps. That can cut errors in tasks like filling out forms or getting support, saving time and reducing frustration. Because the model stays fast, it can be used in real‑time chat apps without noticeable delays. Developers also get free code to add smarter questioning to their own products.

## Sources & references

1. [LAVOIR: Teaching a Single-Pass Decision Encoder When and What to Ask with Amortized Value of Information](https://arxiv.org/abs/2609.30706) – arXiv, 2026-09-28

Last updated: 2026-09-28