BED-LLM: Intelligent Information Gathering with LLMs and Bayesian Experimental Design

2025-12-19 · Source: Apple Machine Learning Research · Field: Technology & Digital — Artificial Intelligence & Machine Learning · Depth: Expert, quick

Summary

Researchers propose BED-LLM, a general-purpose approach that enhances Large Language Models' (LLMs) ability to gather information adaptively from users or external sources. This method frames information gathering as sequential Bayesian experimental design (BED), allowing LLMs to function as effective multi-turn conversational agents. BED-LLM operates by iteratively selecting questions or queries that maximize the expected information gain (EIG) regarding a task, based on previously collected responses. The approach formulates EIG using a probabilistic model derived from the LLM's belief distribution and incorporates innovations like a carefully designed EIG estimator, conditioning on responses beyond in-context updates, and a targeted query proposal strategy. Evaluations using the 20-questions game and user preference inference tasks demonstrate substantial performance gains over direct LLM prompting and other adaptive design strategies.

Key takeaway

For research scientists developing interactive AI, BED-LLM offers a principled framework to improve LLM information gathering. You should consider integrating sequential Bayesian experimental design to enhance conversational agents and active inference systems, moving beyond simple prompting for more adaptive and efficient interactions.

Key insights

BED-LLM enhances LLMs' interactive information gathering via sequential Bayesian experimental design to maximize expected information gain.

Principles

• Maximize expected information gain (EIG).
• Condition on responses beyond in-context updates.

Method

BED-LLM iteratively selects queries maximizing EIG, formulated from the LLM's probabilistic belief, using a specialized EIG estimator and targeted query proposals.

In practice

• Apply to multi-turn conversational agents.
• Use for active user preference inference.

Topics

• Large Language Models
• Bayesian Experimental Design
• Expected Information Gain
• Conversational AI
• Adaptive Querying

Best for: Research Scientist, AI Researcher, AI Scientist, AI Student

Related on AIssential

• Explain Like I'm 5 or Whatever I Choose: Evaluating the Interactive Potential of Language Model Responses — LLMs inconsistently adjust response complexity, failing to reliably differentiate jargon and information across user-controlled levels.
• Adaptive Querying with AI Persona Priors — AI personas from LLMs enable efficient, expressive Bayesian adaptive querying with closed-form posterior updates.
• How to Replicate Stanford’s STORM: PhD Level AI Research in Minutes — Stanford's STORM framework automates deep, cited research using a multi-perspective LLM methodology, outperforming standard prompting.
• Teaching Machines to Code: A Social Scientist’s Guide to LLM-Based Information Extraction — Multi-agent LLM pipelines excel at complex information extraction by mimicking human collaborative coding processes.
• LLM-as-an-Investigator: Evidence-First Reasoning for Robust Interactive Problem Diagnosis — An evidence-first LLM agent improves problem diagnosis by iteratively questioning users to avoid premature conclusions and user-driven sycophancy.
• CA-BED: Conversation-Aware Bayesian Experimental Design — CA-BED improves LLM interactive reasoning by optimizing question selection through Bayesian Experimental Design and simulated conversation trees.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Apple Machine Learning Research.