TrueBrief: Faithful Summarization through Small Language Models

TL;DR

The paper tackles faithfulness in summarization with small language models by introducing TrueBrief, an end-to-end framework that combines a synthetic data generator (DataGen) with controlled hallucinations, a DPO-based finetuning stage, and a white-box hallucination detector (Detection). By injecting intrinsic and extrinsic hallucinations to produce high-quality preference data, the authors show that data quality and model size critically shape the success of preference-based optimization, with beta tuned at $\beta=0.5$ yielding best faithfulness. A novel detection approach leverages internal model dynamics, using LogitLens and LookbackLens features fed to a lightweight classifier, enabling real-time hallucination detection without external APIs. Across experiments with Qwen-2.5 family models (0.5B–3B) and RAGTruth-derived data, DPO outperforms other preferences-based methods, especially for smaller models, while single rejected responses consistently outperform multiple rejects. The work demonstrates that faithful summarization with small LLMs is feasible and that internal-dynamics-based detection provides a practical, low-latency safety net, suggesting broad potential for security-critical NLP tasks where resource constraints matter.

Abstract

Large language models (LLMs) have exhibited remarkable proficiency in generating high-quality text; however, their propensity for producing hallucinations poses a significant challenge for their deployment in security-critical domains. In this work, we present TrueBrief, an end-to-end framework specifically designed to enhance the faithfulness of small LLMs (SLMs) primarily for the task of text summarization through a preference-optimization paradigm. Central to our framework is a data generation module that facilitates controlled hallucination injection to generate synthetic preference data. Our work provides insights into the impact of data quality and model size on preference-based optimization, highlighting the conditions under which these methods are most effective.

Figures (2)

• Figure 1: TrueBrief Components Overview: The main pipeline leverages the data generation module, which is pivotal to the system's performance due to its novel hallucination injection component for creating a preference-based dataset. This dataset is used to finetune a given model via the finetuning module. The resulting finetuned model is employed for response generation and integrated into the detection module, which utilizes a white-box approach to effectively detect hallucinations.
• Figure 2: The gap between DPO and SFT reduces as model scale increases which suggests that the effects of preference based optimization are less pronounced in larger models whereas DPO affords significant performance improvement in SLMs