IUQ: Interrogative Uncertainty Quantification for Long-Form Large Language Model Generation

1. Problem Statement

While Large Language Models (LLMs) have advanced significantly, Uncertainty Quantification (UQ) remains a critical challenge, particularly for long-form, free-form text generation.

• The Core Issue: Existing UQ methods often rely on inter-sample consistency (checking if multiple sampled responses agree) or token probabilities. However, in long-form generation, LLMs often produce responses that are semantically coherent yet factually inaccurate.
• The "Fabrication Trap": Due to next-token conditioning, an LLM may fabricate a false premise (e.g., misidentifying a person's profession) and then generate a logically consistent biography based on that false premise. Because the subsequent text is internally consistent, standard consistency checks fail to detect the hallucination. The model appears confident and consistent, even though the underlying knowledge is false.
• Limitation of Current Methods: Current approaches struggle to distinguish between a model that is truly uncertain and a model that is confidently hallucinating a consistent narrative. They often fail to decouple the generative context (the story being told) from the model's underlying knowledge.

2. Methodology: Interrogative Uncertainty Quantification (IUQ)

The authors propose IUQ, a novel framework that leverages an "interrogate-then-respond" paradigm to quantify uncertainty by testing the model's actual knowledge against its generated claims. The framework consists of four main stages:

A. Response Generation & Claim Extraction

1. Sampling: The model generates $N$ diverse long-form responses to a prompt.
2. Decomposition: Each response is decomposed into atomic, standalone factual claims ($C_R$) using the LLM itself.

B. Interrogation (Claim-Level Question-Answering)

To break the "contextual dependence" that leads to consistent hallucinations:

1. Question Generation: For each atomic claim, the LLM generates specific, short questions ($Q_c$) that target the information in that claim.
2. Independent Answering: The model answers these questions in a separate session without the original long-form context. This forces the model to rely on its internal parametric knowledge rather than the fabricated context of the original story.

C. Measuring Faithfulness (Intra-Sample Consistency)

The framework calculates Claim-Level Faithfulness ($F(c_i)$):

• It compares the answers generated in the independent session against the original claims.
• If the model's independent answers contradict the original claim, it indicates the claim was likely fabricated to fit the context.
• Formula: $F(c_i) = 1 - \text{Contradiction Rate}$. A score of 1 means the claim is faithful to the model's knowledge; a lower score indicates fabrication.

D. Calculating Uncertainty (Inter-Sample + Intra-Sample)

The final uncertainty score combines two factors:

1. Inter-Sample Consistency ($S(c_i)$): The standard measure of how many sampled responses support a specific claim.
2. Unfaithfulness Weighting ($W(c_i)$): A mechanism to propagate the impact of unfaithful claims to subsequent claims.
   • Since long-form generation is sequential, a false premise early in the text often corrupts the rest of the generation.
   • IUQ uses an exponential decay kernel to weight the influence of unfaithful claims on subsequent claims.
   • Final Uncertainty Score: $U(c_i) = S(c_i) \cdot W(c_i)$.

3. Key Contributions

• Novel Framework (IUQ): The first UQ method specifically designed for long-form generation that explicitly decouples generative context from knowledge verification via an interrogation loop.
• Identification of a Critical Phenomenon: The paper highlights and quantifies the tendency of LLMs to fabricate information to maintain logical consistency, a failure mode missed by standard consistency-based UQ methods.
• Intra-Sample Consistency: Introduces a mechanism to measure intra-sample faithfulness (does the model know what it just said?) rather than just inter-sample agreement (do different stories agree?).
• Comprehensive Evaluation: Extensive experiments across diverse model families (GPT-4o, LLaMA-3.3, Qwen2, Gemma-3, Mistral) and sizes (24B–72B).

4. Experimental Results

The authors evaluated IUQ on two datasets: FActScore (biographies of lesser-known individuals) and LongFact (broad topics).

• Performance: IUQ consistently outperformed state-of-the-art baselines (including Claim Entailment, Closeness Centrality, and token-probability methods like Perplexity and Max Token Entropy).
  • On FActScore, IUQ achieved an average AUROC of 0.864, outperforming the next best method (Closeness Centrality at 0.841) by +2.3%.
  • On LongFact, IUQ achieved an average AUROC of 0.740, outperforming the next best by +1.7%.
• Ablation Studies:
  • Directionality: Proved that checking contradictions against preceding claims (context) is crucial; checking against subsequent claims yields lower performance.
  • Kernel Choice: The exponential decay kernel for propagating unfaithfulness was found to be the most effective.
  • Sample Size: Performance improves with more sampled responses, but IUQ remains robust even with fewer samples compared to baselines.
• Case Study: In a specific case involving "Rory Byrne," standard methods gave a high support score (0.60) because multiple generations fabricated the same false story. IUQ correctly identified the low faithfulness (0.17) because the model's independent answers contradicted the fabricated profession, leading to a higher (more accurate) uncertainty score.

5. Significance and Implications

• Reliability in Real-World Applications: Long-form generation is essential for applications like automated reporting, medical summaries, and educational content. IUQ provides a more reliable safety mechanism to detect when an LLM is "confidently wrong."
• Beyond Token Probabilities: The paper demonstrates that token probabilities and simple semantic consistency are insufficient for long-form text. The "interrogation" approach offers a new paradigm for probing model knowledge.
• Trade-offs: The method incurs higher computational costs (approx. 47% more tokens than graph-based methods) due to the multi-stage inference (generation, extraction, questioning, answering, evaluation). However, the paper argues this cost is justified for high-stakes applications requiring factual accuracy.

In summary, IUQ addresses the blind spot in current LLM evaluation where models generate plausible, consistent, but false narratives. By forcing the model to "defend" its claims in a context-free interrogation, IUQ provides a robust metric for detecting hallucinations in long-form generation.