LAVOIR: Teaching Single-Pass Decision Encoders to Ask for Info

What happened

To address the key pain point of existing single-pass text decision models (such as TypeSafe’s Jev and open-source equivalent Laya) — which can only make educated guesses when faced with missing information and cannot proactively request additional context — researchers developed the LAVOIR model. LAVOIR takes candidate missing information slots alongside answer options as input, and in a single forward pass, simultaneously outputs decision distributions and the expected decision gain of inquiring about each slot, with no requirement for manually annotated value targets. Real-world testing shows that when limited to a maximum of 0.5 information inquiries per conversation, LAVOIR achieves 14.1 percentage points higher accuracy than a baseline that never asks for information; in authentic conversation scenarios, accuracy improves by up to 8.3 percentage points after inquiry. On the SGD dataset, LAVOIR cuts inquiry rates from 93% to 8.6%, with a median per-query response time of only 31ms.

Key facts

Model

LAVOIR (Single-pass Decision Encoder with Information Value Routing)

Core Mechanism

Outputs decision results and expected gain of information to inquire about in a single forward pass, requiring no manual annotation

Core Performance

Achieves 14.1 percentage points higher accuracy than a no-inquiry baseline when limited to ≤0.5 inquiries per conversation

Inquiry Rate Control

Reduces inquiry rates from 93% to 8.6% on the SGD dataset

Response Speed

Delivers a median per-query response time of 31ms on GH200 hardware

Benchmark Comparison

Outperforms the original Laya model on 7 out of 12 Laya benchmark tests

Background

Prior to LAVOIR, single-pass decision models including TypeSafe’s Jev and its open-source benchmark equivalent Laya could output answers to text-format questions with calibrated probabilities via a single forward pass, but lacked the ability to proactively request missing key information, forcing them to make guesses based only on available existing context.

Why it matters

For developers building task-oriented dialogue systems, LAVOIR integrates information value judgment logic directly into the single-pass inference pipeline, eliminating the need to build a separate, complex inquiry strategy module while adding proactive missing information collection capabilities to existing low-latency decision models. For the broader industry, this approach lowers the deployment barrier for high-accuracy conversational decision systems. For end users, the system will only initiate inquiries when there is clear measurable value, drastically reducing unnecessary, frustrating interactions.