Not All Claims Are Equally Risky: FACTOR for Adaptive Verification in Factual Long-Form Generation

Large Language Models (LLMs) often generate fluent but factually unreliable long-form text. While Retrieval-Augmented Generation (RAG) and NLI-based verification help, they typically apply a uniform verification policy to all claims. This paper asks whether adapting verification intensity based on the specific hallucination risk of each claim can improve factual accuracy while simultaneously reducing computational overhead.

The authors propose FACTOR (FACTuality-Oriented Risk-aware Verification), an inference-time framework. FACTOR operates in four main steps:

1. Hybrid Retrieval: Combines BM25 and dense retrieval to gather relevant evidence.
2. Candidate Generation: Generates multiple candidate responses to capture stochastic variation.
3. Uncertainty-Aware Verification: Estimates uncertainty for each claim using a combination of token-level entropy and semantic consistency across candidates. Claims are then routed to different verification tiers (Low, Mid, High) with varying NLI thresholds.
4. Candidate Re-ranking: Selects the final response by balancing factuality scores and fluency (perplexity).

FACTOR outperforms standard RAG and static verification baselines on the FactScore benchmark, achieving higher factual accuracy with a lower hallucination rate. The study demonstrates that adaptive routing is more efficient than uniform verification, as it concentrates computational resources on high-uncertainty claims. Ablation studies reveal that semantic consistency is a particularly strong signal for identifying unreliable claims, and that sentence-level claim segmentation is more effective than atomic decomposition for maintaining necessary context during verification.

This work highlights that not all generated information carries the same risk of hallucination. By moving away from "one-size-fits-all" verification, researchers can build more reliable factual generation systems that optimize the trade-off between accuracy and computational cost. The model-agnostic nature of FACTOR suggests it can be integrated into various existing LLM pipelines to enhance trustworthiness in knowledge-intensive tasks.