FaithLens: Detecting and Explaining Faithfulness Hallucination

AI summary

Overview

• Research area: Natural Language Processing — hallucination detection in large language model outputs (specifically faithfulness hallucination, where a claim contradicts or cannot be verified from a supplied document).
• Technical level: Intermediate. The paper assumes familiarity with supervised fine-tuning, reinforcement learning from verifiable rewards (GRPO), and standard detection benchmarks, but its core ideas can be followed without deep math.
• Scope: The paper describes FaithLens, an 8B-parameter detector that jointly outputs a faithful/hallucinated binary label and an explanation for that label, trained on synthesized-and-filtered data followed by rule-based reinforcement learning.

What This Paper Is About

Large language models that generate text from documents (for example in retrieval-augmented generation or summarization) often produce claims that the source document does not support. Existing detectors usually output only a yes/no label, which gives users no way to see where or why the claim failed. FaithLens addresses this by predicting the label and producing an explanation at the same time, while staying cheap enough to deploy.

Key Contributions

1. An explainable detection formulation. The paper extends the standard binary classification setup from P(y | doc, c) to P(e, y | doc, c), so the model simultaneously produces a faithfulness label and a textual explanation supporting it.

2. A data synthesis and filtering pipeline. The authors use a large reasoning model (DeepSeek-V3.2-Think) to generate explanations for document-claim pairs from open-source training data, then filter for three properties: label correctness, explanation quality, and data diversity.

3. A rule-based reinforcement learning stage with two custom rewards. After supervised fine-tuning as a cold start, the model is optimized with GRPO using a prediction correctness reward, an explanation quality reward (whether the explanation lets a novice model such as Llama-3.1-8B-Instruct predict the correct label), and a format reward.

4. State-of-the-art results at low cost. FaithLens-8B reaches an overall average of 86.4 with a standard deviation of 4.6 across 12 tasks, above GPT-5.2 (86.1, std 5.9) and o3 (82.1, std 6.0), while incurring an inference cost of $0.1 on 1.2K samples versus $15.3 for GPT-5.2.

Main Findings

• Detection performance: FaithLens scores 86.4 overall average across the 12 datasets, with the lowest standard deviation (4.6) among compared models. GPT-5.2 scores 86.1 (std 5.9), o3 scores 82.1 (std 6.0), Claude-3.7-Sonnet 82.6 (std 5.3), and DeepSeek-V3.2-Think 84.4 (std 5.1). The best specialized baseline, MiniCheck, scores 80.7 (std 7.5).

• Task-level scores: FaithLens records 84.9 on Agg-CNN, 79.0 on Agg-XSum, 89.4 on ClaimVerify, 79.6 on ExpertQA, 92.4 on FC-GPT, 92.1 on LfQA, 86.8 on RAGTruth, 92.2 on Reveal, 85.1 on Tofu-MediaS, 87.2 on Tofu-MeetB, 85.6 on Wice, and 82.9 on HoVer.

• Gain over the base model: Relative to Llama-3.1-8B-Inst, FaithLens improves the overall average by +30.1 and reduces standard deviation by 6.3, with per-task gains ranging from +14.0 (Reveal) to +46.0 (LfQA).

• Explainability: Judged by GPT-4.1 along readability, helpfulness, and informativeness, FaithLens averages 90.4 (92.4 / 93.4 / 85.4). GPT-5.2 averages 94.4, o3 93.5, Claude-3.7-Sonnet 93.5, while ClearCheck averages 77.3 and the raw unoptimized chain-of-thought from FaithLens averages 75.5. Gains over Llama-3.1-8B-Inst are +17.1, +21.1, and +17.2 on the three dimensions.

• Inference efficiency: On 1.2K samples across the 12 datasets, FaithLens costs $0.1, versus $0.8 for DeepSeek-V3.2-Non-Think, $5.9 for o3-mini, $7.3 for GPT-4o, $8.8 for o3, $11.4 for GPT-4.1, $14.5 for Claude-3.7-Sonnet, $15.3 for GPT-5.2, and $140.6 for o1. Table 3's caption describes the lowest cost as $0.8/GPU-hour.

• Data efficiency: FaithLens uses 28K open-source training examples and is explainable; without data filtering it would use 52K. By comparison, ClearCheck uses 82K private examples and is only partially explainable, MiniCheck uses 35K private examples and is not explainable, FactCG uses 52K open-source examples and is not explainable, and AlignScore uses 4,700K examples and is not explainable.

• Ablations: Removing the cold-start SFT stage drops the model to 83.4 avg / 88.1 explainability; removing data filtering drops it to 81.2 / 82.3; removing label correctness filtering gives 83.5 / 86.0; removing explanation quality filtering gives 85.8 / 83.4; removing data diversity filtering gives 85.0 / 89.3; removing the RL stage gives 82.6 / 83.8; removing the explanation quality reward gives 85.7 / 84.7. Direct SFT on 52K data without the pipeline yields 79.1 avg and cannot produce explanations.

• Cross-backbone generalization: Training the same pipeline on Qwen2.5-3B-Inst raises it from 73.3 to 83.4 avg (explainability 79.3 to 88.3), on Qwen2.5-7B-Inst from 73.9 to 84.9 (81.7 to 90.3), and on Llama-3.1-8B-Inst from 56.3 to 86.4 (71.9 to 90.4).

• Decontextualization and decomposition: Applying GPT-4.1-based claim decontextualization or decomposition does not improve FaithLens meaningfully (86.4 original and decontextualized; 86.6 decomposed), indicating the model already captures context-dependent relations.

• Human evaluation: On 120 selected samples, explanations from FaithLens and GPT-4o were compared by majority voting of three human judges across readability, helpfulness, and informativeness.

Methodology in Plain English

The team started from open-source training data that only had labels, not explanations. They asked a strong reasoning model, DeepSeek-V3.2-Think, to read each document-claim pair and produce a chain of thought, an explanation, and its own predicted label.

Because synthesized data can be noisy, they applied three filters. First, the predicted label must match the dataset's ground-truth label, otherwise the sample is discarded. Second, the explanation must actually help a small model (Llama-3.1-8B-Instruct) become more confident in the correct label — measured by comparing perplexity of the ground-truth label with and without the explanation. Third, to avoid keeping only easy or repetitive examples, they embedded document-claim pairs, clustered them with K-Medoids, built a probe set from cluster centers, and kept only samples that reduced perplexity on at least half of the probe set.

They then fine-tuned the base model on the surviving data (cold start), and finally ran GRPO reinforcement learning with three rewards summed together: one for predicting the correct label, one for explanations that let the novice model reach the correct label, and one for following the requested output format. Evaluation used the cleaned LLM-AggreFact and HoVer benchmarks (the cleaned versions remove 9.1% ambiguous and 6.6% mislabeled instances flagged in prior work), with macro-F1 as the metric.

Why This Matters

Detection models that return only a label leave users unable to verify or fix errors, which limits trust in retrieval-augmented and summarization systems. FaithLens shows that a small, open-data-trained model can match or exceed much larger frontier models on this task while also explaining itself — a combination the paper argues is rare.

Real-world applications:

• Retrieval-augmented generation: flagging and explaining when an answer drifts from retrieved documents.
• Summarization: catching subtly distorted content in generated summaries.
• Multi-hop question answering and verification: spotting claims that cannot be supported through multi-step reasoning (the HoVer setting).
• Content moderation and compliance review: giving reviewers a rationale alongside each flagged claim rather than a bare score.

Industry relevance comes from the cost profile: $0.1 for 1.2K samples versus $15.3 for GPT-5.2 and $140.6 for o1, and from the fact that the training data is open-source rather than private, which makes the approach reproducible for organizations that cannot use proprietary data.

Future Directions

• Whether the explanation quality reward generalizes to novice models other than Llama-3.1-8B-Instruct, which is what the paper uses.
• How the pipeline behaves on domains or hallucination patterns not represented in the LLM-AggreFact and HoVer benchmarks.
• Whether scaling the approach beyond the 3B, 7B, and 8B backbones tested (Qwen2.5-3B-Inst, Qwen2.5-7B-Inst, Llama-3.1-8B-Inst) brings further gains, since larger untrained backbones showed lower baseline explainability.
• Whether rule-based rewards can be extended to capture explanation properties beyond readability, helpfulness, and informativeness, which the authors note remains an unresolved challenge for free-form content.

Target Audience

Researchers and practitioners working on hallucination detection, retrieval-augmented generation, and factual verification; engineers choosing a cost-effective detector for production systems; and anyone interested in reinforcement learning with verifiable rewards for tasks whose outputs are free-form text rather than verifiable answers. The paper is also relevant to readers tracking whether small open models can match frontier proprietary models on narrow, well-defined tasks.

Authors’ abstract

Recognizing whether outputs from large language models (LLMs) contain faithfulness hallucination is crucial for real-world applications, e.g., retrieval-augmented generation and summarization. In this paper, we introduce FaithLens, a cost-efficient and effective faithfulness hallucination detection model that can jointly provide binary predictions and corresponding explanations to improve trustworthiness. To achieve this, we first synthesize training data with explanations via advanced LLMs and apply a well-defined data filtering strategy to ensure label correctness, explanation quality, and data diversity. Subsequently, we fine-tune the model on these well-curated training data as a cold start and further optimize it with rule-based reinforcement learning, using rewards for both prediction correctness and explanation quality. Results on 12 diverse tasks show that the 8B-parameter FaithLens outperforms advanced models such as GPT-5.2 and o3. Also, FaithLens can produce high-quality explanations, delivering a distinctive balance of trustworthiness, efficiency, and effectiveness.