---
title: "Amortising Bayesian Experimental Design for Sequential Information Gathering in LLMs"
source: "arXiv AI/CL/LG（论文）"
source_tier: "t2"
url: "https://arxiv.org/abs/2607.03426v1"
permalink: "https://feed.astralor.com/items/itm_dd3a5f549632850972bd5726"
published: "2026-07-03T15:38:17.000Z"
score: 78
channel: "ai"
tags: ["ASIG","GRPO","LLM","Sequential Information Gathering","Bayesian experiment design"]
---

# Amortising Bayesian Experimental Design for Sequential Information Gathering in LLMs

## AI commentary

> What I find most compelling about this paper is its amortisation of Bayesian Experimental Design into LLM policies, avoiding expensive inference-time optimisation while achieving significant gains in both success rate and inference cost. For developers working on dialogue systems, intelligent QA, or any multi-turn information acquisition scenario, this idea is highly inspiring—it suggests that fine-tuning can teach models to ask 'good questions' without relying on external search or rules. The demonstrated transferability to medical diagnosis also underscores the method's general applicability.

## At a glance

Large language models often struggle to gather information effectively in multi-turn interactions. This paper proposes ASIG, a fine-tuning approach that amortises Bayesian Experimental Design into LLM policies via a multi-turn extension of GRPO with an Expected Information Gain reward. On the 20 Questions task, ASIG more than doubles the success rate of the 7B base model and reduces inference cost by over 25×. On MediQ, a medical diagnosis benchmark unseen during training, ASIG shows out-of-distribution transfer of learned strategies, offering a computationally efficient paradigm for sequential information gathering.

## Key points

- Proposes ASIG, amortising Bayesian Experimental Design into LLM policies via multi-turn GRPO with EIG reward.
- On 20 Questions, ASIG doubles success rate of 7B base model and reduces inference cost by over 25×.
- Demonstrates out-of-distribution transfer of learned strategies on unseen MediQ medical diagnosis benchmark.

## Full text

Large language models (LLMs) exhibit strong reasoning and world-knowledge capabilities, yet often struggle to gather information effectively across the multi-turn interactions required in sequential decision-making settings. We introduce Amortised Sequential Information Gathering (ASIG), a fine-tuning approach that amortises Bayesian Experimental Design (BED) into LLM policies via a multi-turn extension of Group Relative Policy Optimisation with an Expected Information Gain reward. Evaluated on the 20 Questions task, ASIG more than doubles the success rate of the 7B base model and reduces inference cost by over $25\times$ relative to BED-LLM, a competitive inference-time baseline. Applied to MediQ, a medical diagnosis benchmark unseen during training, ASIG improves information-seeking performance at the 7B scale, suggesting that the learned strategies can transfer out of distribution. Our findings show that amortising BED into LLM policies provides an effective and computationally efficient approach to sequential information gathering.

---

Verify against the original before citing; curated by Astral Feed: https://feed.astralor.com/items/itm_dd3a5f549632850972bd5726