[NeurIPS 2025] INFOGATHERER: Navigating High-Stakes Uncertainty via Evidential Networks

TL;DR

In high-stakes fields like medicine and law, an LLM's "I don't know" is just as important as its final answer. INFOGATHERER is a new framework that treats information-seeking as a formal mathematical problem. By mapping retrieved documents and user responses onto a Dempster-Shafer Evidential Network, it moves beyond "vibes-based" confidence, achieving SOTA accuracy on MedQA and BarExamQA with significantly shorter dialogues.

The Motivation: The "Overconfidence" Trap

When a patient says, "I have a sore throat," a standard LLM (or even a RAG system) might immediately hallucinate a diagnosis or ask a random question. The core issue is premature collapse: most models force a probability distribution over hypotheses even when the evidence is missing or ambiguous.

Prior works like MediQ or Uncertainty of Thoughts (UoT) tried to fix this through prompting or self-consistency. However, they lack a principled way to distinguish between Ambiguity (evidence supports multiple things) and Ignorance (we just don't have the data yet).

Methodology: Reasoning with Evidential Networks

INFOGATHERER replaces simple point-probabilities with Basic Belief Assignments (BBAs).

1. Network Construction

The system first builds a Directed Acyclic Graph (DAG) grounded in authoritative documents. If we're diagnosing a disease, the root is the "Hypothesis," and the leaves are "Observable Features" (symptoms).

2. Principled Belief Elicitation

Instead of asking an LLM "How likely is X?", INFOGATHERER prompts the LLM to extract BBAs from snippets. A snippet might say "Fever may indicate Flu or COVID." Instead of splitting 50/50, the BBA assigns mass to the set {Flu, COVID}, explicitly preserving the ambiguity.

3. Strategic Questioning via Deng Entropy

To decide what to ask next, INFOGATHERER uses Deng Entropy. This metric decomposes uncertainty into two parts:

• Nonspecificity: Belief is spread over large sets (Ambiguity).
• Discord: Belief is split between competing specific hypotheses (Conflict).

The policy is elegant: first, ask questions to reduce Nonspecificity (narrow the field), then ask questions to resolve Discord (pick the winner).

Experiments: Higher Accuracy, Fewer Questions

The authors tested INFOGATHERER against the best interactive baselines on MedQA (Medical) and BarExamQA (Legal).

Key Insight: In the Legal domain, the success rate nearly doubled compared to Uncertainty of Thoughts. This is because legal reasoning requires resolving specific factual contradictions that point-probability models often smooth over.

Figure: INFOGATHERER (solid blue) shows a more calibrated confidence increase, avoiding the noisy fluctuations of Bayesian and heuristic methods.

Critical Analysis & Conclusion

Why it Works

The "secret sauce" is Yager’s Rule of Combination. Unlike standard Bayesian updates which can be "broken" by one piece of contradictory evidence, Yager’s rule shunts conflict into the Ignorance mass. This prevents the model from spiraling when a user gives a slightly inconsistent answer.

Limitations

• Fixed Hypothesis Space: Currently, the model needs a predefined list of outcomes (like multiple-choice questions).
• Latency: Constructing a full evidential network and running leaf-to-root message passing is computationally heavier than simple prompting.

The Bottom Line

INFOGATHERER proves that for AI to be useful in professional fields, it needs more than just "more parameters"—it needs a formal logic for seeking truth. By borrowing from 1960s-era Evidence Theory and fusing it with modern LLMs, we get agents that are not just smarter, but significantly more "honest" about what they don't know.