FaithLens: Detecting and Explaining Faithfulness Hallucination

Conference: ACL 2026 arXiv: 2512.20182 Code: https://github.com/S1s-Z/FaithLens Area: Reinforcement Learning / Hallucination Detection Keywords: Faithfulness hallucination, explainable detection, rule-based reinforcement learning, data filtering, cross-task generalization

TL;DR

This paper proposes FaithLens, an 8B-parameter faithfulness hallucination detection model trained via high-quality data synthesis with three-dimensional filtering (label correctness, explanation quality, and data diversity) for cold-start SFT, followed by rule-based reinforcement learning (prediction correctness reward + explanation quality reward) for further optimization. FaithLens surpasses GPT-5.2 and o3 across 12 tasks while providing high-quality explanatory outputs.

Background & Motivation

Background: LLMs are widely used for context-grounded text generation (e.g., RAG, summarization), but are prone to generating content inconsistent with or irrelevant to the given context—so-called "faithfulness hallucinations." Detecting such hallucinations is critical for responsible LLM deployment.

Limitations of Prior Work: (1) Lack of interpretability—existing methods treat hallucination detection as a black-box binary classification problem, outputting only prediction labels without explaining the rationale, leaving users unable to localize errors or understand their causes. (2) Inconsistent cross-task generalization—different tasks exhibit distinct hallucination patterns (subtle distortions in summarization vs. contradictory claims in RAG), and general-purpose models perform unevenly across them. (3) Lack of high-quality data—annotation costs are high, consistency is low, and synthetic data lacks quality control.

Key Challenge: Simultaneously achieving high detection accuracy and high explanation quality is non-trivial. SFT-trained models tend to memorize simple samples and generalize poorly to complex scenarios, while the quality of free-form explanations is difficult to verify directly with rule-based signals.

Goal: Develop a cost-effective hallucination detection model that jointly produces detection results and explanatory rationales, achieving state-of-the-art performance across 12 diverse tasks.

Key Insight: A two-stage training pipeline—cold-start SFT on carefully filtered synthetic data, followed by GRPO reinforcement learning with well-designed rule-based rewards (prediction correctness + explanation quality).

Core Idea: The key insight behind the explanation quality reward—if a generated explanation enables a "novice model" (an untuned Llama-3.1-8B) to correctly predict the label, the explanation is sufficiently clear and informative.

Method

Overall Architecture

FaithLens training proceeds in two stages: (1) Cold-start SFT—starting from open-source datasets, a high-capability reasoning model (DeepSeek-V3.2-Think) is used to synthesize training data with explanations, which are then filtered via three-dimensional criteria before fine-tuning; (2) Rule-based RL—the GRPO algorithm is applied for further optimization, with a reward function consisting of prediction correctness, explanation quality, and format components.

Key Designs

1. Three-Dimensional Data Filtering Strategy:

   • Function: Ensures label correctness, explanation quality, and data diversity of the synthetic training data.
   • Mechanism: Label filtering—discards samples where the LLM's prediction disagrees with the ground-truth label, since CoT reasoning and explanations consistent with an incorrect prediction are intrinsically misleading. Explanation quality filtering—measures whether the model's perplexity on the correct label decreases when the explanation is provided; only samples that reduce perplexity are retained. Diversity filtering—applies K-Medoids clustering to construct a probe set, retaining only training candidates that improve prediction accuracy on diverse probe samples.
   • Design Motivation: Unfiltered synthetic data contains noise and an excess of simple samples. The three-dimensional filtering ensures that training data is simultaneously correct, informative, and covers diverse scenarios.

2. Explanation Quality Reward:

   • Function: Implicitly evaluates the quality of free-form explanations during the RL stage.
   • Mechanism: The generated explanation \(e\), together with the document and claim, is fed to a "novice model" (untuned Llama-3.1-8B-Instruct) to check whether the novice can correctly predict the label based solely on this explanation. A reward of 1 is assigned if correct, 0 otherwise. The final reward is \(R_{\text{final}} = R_{\text{pred}} + R_{\text{exp}} + R_{\text{format}}\).
   • Design Motivation: Directly verifying free-form text quality with rules is nearly impossible. The proxy evaluation principle—"if even a novice model can derive the correct answer from your explanation, the explanation must be sufficiently good"—provides an elegant and verifiable signal.

3. GRPO Reinforcement Learning:

   • Function: Further improves detection accuracy and explanation quality on top of the SFT cold-start.
   • Mechanism: For each document–claim pair, \(G\) candidate outputs (explanation + prediction) are sampled. Each candidate is evaluated with the composite reward, and policy updates are performed via GRPO's group-relative advantage estimation. KL divergence regularization prevents excessive deviation from the reference policy.
   • Design Motivation: SFT tends to memorize simple samples; RL drives the model to produce high-quality outputs even in complex scenarios through exploration and reward-driven optimization.

Loss & Training

The SFT stage applies standard cross-entropy loss on the filtered synthetic data. The RL stage uses GRPO (Group Relative Policy Optimization) with a composite reward = prediction correctness (0/1) + explanation quality (0/1) + format correctness (0/1). The base model is Llama-3.1-8B-Instruct.

Key Experimental Results

Main Results

Overall Performance Across 12 Tasks (Balanced Accuracy %)

Model	Std. Dev. ↓	Average ↑
GPT-4o	7.0	76.1
o3	6.0	82.1
GPT-5.2	—	85.3
Claude-3.7-Sonnet	5.3	82.6
DeepSeek-V3.2-Think	5.1	84.4
MiniCheck-7B	9.3	76.7
FaithLens-8B (Ours)	4.1	85.8

Ablation Study

Configuration	Avg. Accuracy	Note
Full FaithLens	85.8	Complete model
w/o RL (SFT only)	82.3	RL contribution: +3.5
w/o explanation quality reward	84.1	Explanation reward contribution: +1.7
w/o data filtering	79.8	Filtering contribution: +6.0
w/o diversity filtering	81.5	Diversity filtering contribution: +4.3

Key Findings

• The 8B FaithLens surpasses GPT-5.2 (85.8 vs. 85.3) and o3 (82.1) at orders-of-magnitude lower cost.
• FaithLens achieves the lowest standard deviation (4.1), indicating the most stable cross-task generalization—addressing the "strong on some tasks, weak on others" problem of existing methods.
• The contribution of data filtering (+6.0) exceeds that of RL (+3.5), underscoring that high-quality training data is the foundation.
• Diversity filtering is critical for cross-task generalization; its removal causes a 4.3-point drop in accuracy.
• The explanation quality reward improves not only explanation quality but also detection accuracy (+1.7), suggesting that the "explanation → prediction" process has an intrinsic regularization effect.

Highlights & Insights

• The "novice model proxy evaluation" is an elegant solution for assessing free-form explanation quality—transforming an unverifiable text quality problem into a verifiable classification correctness problem.
• The progressive "label → explanation → diversity" design of the three-dimensional filtering comprehensively ensures training data quality.
• Surpassing closed-source large models with only 8B parameters demonstrates that carefully designed training strategies outperform brute-force parameter scaling.

Limitations & Future Work

• The explanation quality reward depends on the capabilities of the novice model; biases in the novice model may distort the reward signal.
• Synthetic data is derived from existing open-source datasets and may inherit their biases.
• Evaluation is limited to English tasks; multilingual generalization remains unverified.
• Future work may explore finer-grained explanation evaluation (e.g., sentence-level evidence grounding).

Related Work & Insights

• vs. MiniCheck: MiniCheck trains a 7B classifier on synthetic data to reach GPT-4o-level performance but provides no explanations; FaithLens delivers explanations while surpassing GPT-5.2.
• vs. SelfCheckGPT: SelfCheckGPT relies on large model inference and is computationally expensive; FaithLens achieves superior performance with an 8B model.
• vs. DeepSeek-V3.2-Think: As the data synthesis teacher model it performs strongly (84.4%), yet FaithLens exceeds its teacher through RL (85.8%).

Rating

• Novelty: ⭐⭐⭐⭐ The explanation quality reward and three-dimensional filtering strategy are innovative, though the overall framework (SFT + RL) follows a common paradigm.
• Experimental Thoroughness: ⭐⭐⭐⭐⭐ 12 tasks, multiple baselines (including GPT-5.2/o3), and comprehensive ablations.
• Writing Quality: ⭐⭐⭐⭐ Method descriptions are clear and formulations are complete.
• Value: ⭐⭐⭐⭐⭐ An 8B model surpassing GPT-5.2 with explanatory outputs demonstrates strong practical utility.

Related Papers

• [AAAI 2026] Explaining Decentralized Multi-Agent Reinforcement Learning Policies
• [ACL 2026] Reverse Constitutional AI: A Framework for Controllable Toxic Data Generation via Probability-Clamped RLAIF
• [ACL 2026] The Stackelberg Speaker: Optimizing Persuasive Communication in Social Deduction Games
• [ACL 2026] HEALing Entropy Collapse: Enhancing Exploration in Few-Shot RLVR via Hybrid-Domain Entropy Dynamics Alignment
• [ACL 2026] Easy Samples Are All You Need: Self-Evolving LLMs via Data-Efficient Reinforcement Learning