Learning to Generate Answers with Citations via Factual Consistency Models

TL;DR

CaLF introduces a weakly-supervised fine-tuning framework that leverages factual consistency models to both filter candidate citation-rich data and reweight the training objective toward factually relevant tokens. By combining diverse answer generation with FCM-based filtering and SHAP-informed token weighting, CaLF improves long-form answer generation with citations while maintaining fluent language. Across ALCE benchmarks and multiple instruction-tuned LLMs, CaLF achieves substantial gains in citation F1 and grounded correctness, and demonstrates robust domain transfer and high factuality as measured by FactScore. The approach preserves inference efficiency and offers a practical path to more reliable, verifiable LLM outputs in retrieval-augmented settings.

Abstract

Large Language Models (LLMs) frequently hallucinate, impeding their reliability in mission-critical situations. One approach to address this issue is to provide citations to relevant sources alongside generated content, enhancing the verifiability of generations. However, citing passages accurately in answers remains a substantial challenge. This paper proposes a weakly-supervised fine-tuning method leveraging factual consistency models (FCMs). Our approach alternates between generating texts with citations and supervised fine-tuning with FCM-filtered citation data. Focused learning is integrated into the objective, directing the fine-tuning process to emphasise the factual unit tokens, as measured by an FCM. Results on the ALCE few-shot citation benchmark with various instruction-tuned LLMs demonstrate superior performance compared to in-context learning, vanilla supervised fine-tuning, and state-of-the-art methods, with an average improvement of $34.1$, $15.5$, and $10.5$ citation F$_1$ points, respectively. Moreover, in a domain transfer setting we show that the obtained citation generation ability robustly transfers to unseen datasets. Notably, our citation improvements contribute to the lowest factual error rate across baselines.

Figures (7)

• Figure 1: An ALCE-ASQA question with a generated answer prompted via in-context learning. Two error classes are common: information not supported by the sources (red) and incorrect citation to the sources (blue).
• Figure 2: A schematic view of our iterative citation fine-tuning method CaLF. It uses a factual consistency model to: i) create weakly supervised training instances by filtering diversely sampled responses, ii) adjust the loss contribution of each answer token according to its Shapley relevance for factual consistency prediction.
• Figure 3: The computation of relevance weights $W$ for rescaling the loss according to Eq. \ref{['Eq:fl']}. We first use SHAP to measure the token importance for predicting $\phi(s_i, C_i)= o_{i}$. We adjust for differences in scale of $W_{\phi,i}$ for sentences $s_i$ and differences in tokenization between the FCM and the LLM.
• Figure 4: Evaluation metrics and CaLF's dynamic stopping criterion over the number of iterations on ASQA.
• Figure 5: ASQA Example of generated answers with citations given a question and retrieved passages. We compare the answers produced by the Few-shot FT baseline model with those generated by CaLF.