InFi-Check: Interpretable and Fine-Grained Fact-Checking of LLMs
Organizations: Tsinghua University · DeepLang AI · Fudan University
Abstract
Large language models (LLMs) often hallucinate, yet most existing fact-checking methods treat factuality evaluation as a binary classification problem, offering limited interpretability and failing to capture fine-grained error types. In this paper, we introduce InFi-Check, a framework for interpretable and fine-grained fact-checking of LLM outputs. Specifically, we first propose a controlled data synthesis pipeline that generates high-quality data featuring explicit evidence, fine-grained error type labels, justifications, and corrections. Based on this, we further construct large-scale training data and a manually verified benchmark InFi-Check-FG for fine-grained fact-checking of LLM outputs. Building on these high-quality training data, we further propose InFi-Checker, which can jointly provide supporting evidence, classify fine-grained error types, and produce justifications along with corrections. Experiments show that InFi-Checker achieves state-of-the-art performance on InFi-Check-FG and strong generalization across various downstream tasks, significantly improving the utility and trustworthiness of factuality evaluation.
Figures & tables
| Error Category | Abbr. | Construction Strategy |
| Predicate Error | PredE | Swap Relation |
| Modify Predictions | ||
| [2pt/3pt] Entity Error | EntE | Swap Entities |
| Compress Phrases | ||
| [2pt/3pt] Circumstance Errors | CircE | Swap Circumstances |
| [2pt/3pt] Co-Reference Error | CorefE | Swap Pronouns |
| Model | PredE | EntE | CircE | CorefE | LinkE | OutE | NoE | BAcc |
| The Open-Source Models | ||||||||
| Llama-3.1-8B-Instruct | 18.82 | 9.52 | 31.03 | 3.13 | 0.00 | 0.00 | 61.11 | 17.66 |
| Llama-3.1-8B-Instruct※ | 32.94 | 4.76 | 13.79 | 26.56 | 0.00 | 0.00 | 88.49 | 23.79 |
| Qwen3-8B | 76.47 | 52.38 | 44.83 | 50.00 | 0.00 | 50.00 | 61.90 | 47.94 |
| Qwen3-8B※ | 76.47 | 52.38 | 24.14 | 70.31 | 0.00 | 50.00 | 80.16 | 50.49 |
| The State-of-the-Art LLMs | ||||||||
| Model | 0-shot | 1-shot |
| The Open-Source Models | ||
| Llama-3.1-8B-Instruct | 66.13 | 71.49 |
| Qwen3-8B | 65.31 | 70.16 |
| The State-of-the-Art LLMs | ||
| Claude-3.7-Sonnet | 65.89 | 63.42 |
| DeepSeek-V3.2-NoThink | 66.99 | 77.14 |
| Model | Claim Verify | Expert QA | Factcheck-Bench | REVEAL | MediaSum | MeetingBank | Average |
| The Open-Source Models | |||||||
| Llama-3.1-8B-Instruct | 63.6 | 49.8 | 69.8 | 78.2 | 50.8 | 62.3 | 62.4 |
| Qwen3-8B | 66.1 | 55.6 | 78.4 | 83.2 | 77.7 | 74.2 | 72.5 |
| The State-of-the-Art LLMs | |||||||
| Claude-3.7-Sonnet | 83.7 | 74.4 | 86.9 | 88.0 | 85.4 | 84.0 | 83.7 |
| DeepSeek-V3.2-NoThink | 75.4 | 74.4 | 87.9 | 91.0 | 65.5 | 82.9 | 79.5 |
| Model | BAcc(Nor.) | BAcc(Str.) | SAR(Avg.) |
| The Open-Source Models | |||
| Llama-3.1-8B-Instruct | 27.69 | 21.33 | 72.57 |
| Llama-3.1-8B-Instruct※ | 23.79 | 18.12 | 79.41 |
| Qwen3-8B | 47.94 | 41.08 | 86.50 |
| Qwen3-8B※ | 50.49 | 44.78 | 87.22 |
| The State-of-the-Art LLMs | |||
| Index | J ust.+Corr. | E vid. | S ent. | BAcc |
| JES | ✓ | ✓ | ✓ | 90.91 |
| ES | - | ✓ | ✓ | 75.02 |
| JS | ✓ | - | ✓ | 69.85 |
| JE | ✓ | ✓ | - | 61.17 |
| J | ✓ | - | - | 36.95 |
| E | - | ✓ | - | 53.76 |
Appendix figures & tables30 assets
Supplementary material from the paper’s appendix.
Appendix
| Task | LLM Set 1 | LLM Set 2 |
| Summarization | GPT-4o | DeepSeek-R1 |
| Reference extraction | Claude-3.7-Sonnet | GPT-4o |
| Support determination | (Claude-3.7-Sonnet, Qwen-2.5, Gemini-1.5) | (GPT-4o, Qwen-2.5, Gemini-2.0) |
| Rewriting | Claude-3.7-Sonnet | GPT-4o |
| Error data construction | GPT-4o | Claude-3.7-Sonnet |
| #doc | #doc-pair | #PredE | #EntE | #CircE | #CorefE | #LinkE | #OutE | #NoE | |
| InFi-Check-TR | 1263 | 15660 | 2946 | 1998 | 1386 | 2889 | 1495 | 1457 | 3489 |
| InFi-Check-FG | 190 | 519 | 85 | 21 | 29 | 64 | 30 | 38 | 252 |
| Model | PredE | EntE | CircE | CorefE | LinkE | OutE | NoE | BAcc |
| The Open-Source Models | ||||||||
| Llama-3.1-8B | 43.75 | 31.49 | 10.34 | 4.06 | 0.00 | 7.21 | 77.34 | 24.88 |
| Qwen-3-8B | 58.75 | 37.45 | 12.64 | 12.90 | 0.00 | 9.03 | 65.30 | 28.01 |
| The State-of-the-Art LLMs | ||||||||
| DeepSeek-V3.2 | 52.69 | 43.34 | 22.02 | 6.82 | 7.41 | 18.02 | 78.88 | 32.74 |
| GPT-4o | 59.14 | 49.85 | 24.77 | 5.68 | 0.00 | 14.24 | 69.11 | 31.83 |
| Model | PredE | EntE | CircE | CorefE | LinkE | OutE | NoE |
| The State-of-the-Art LLMs | |||||||
| GPT-4o ※ | 84.71 [75.57, 90.84] | 61.90 [40.88, 79.25] | 44.83 [28.41, 62.45] | 75.00 [63.18, 83.99] | 0.00 [0.00, 11.35] | 94.74 [82.71, 98.54] | 91.67 [87.60, 94.49] |
| DeepSeek-V3.2-NoThink※ | 75.29 [65.17, 83.24] | 71.43 [50.04, 86.19] | 55.17 [37.55, 71.59] | 15.62 [8.71, 26.43] | 13.33 [5.31, 29.68] | 89.47 [75.87, 95.83] | 94.05 [90.41, 96.36] |
| o3※ | 81.18 [71.59, 88.07] | 57.14 [36.55, 75.53] | 58.62 [40.74, 74.49] | 56.25 [44.09, 67.71] | 3.33 [0.59, 16.67] | 81.58 [66.58, 90.78] | 87.30 [82.62, 90.86] |
| InFi-Checker (Ours) | |||||||
| InFi-Checker-Llama | 96.43 [90.02, 98.78] | 90.48 [71.09, 97.35] | 79.31 [61.61, 90.15] | 95.31 [87.10, 98.39] | 86.67 [70.32, 94.69] | 100.00 [90.82, 100.00] | 90.00 [85.65, 93.13] |
| gold\pred | PredE | EntE | CircE | CorefE | LinkE | OutE | NoE |
| PredE | 81 | 0 | 0 | 0 | 2 | 0 | 1 |
| EntE | 0 | 19 | 1 | 0 | 0 | 0 | 1 |
| CircE | 1 | 4 | 23 | 0 | 1 | 0 | 0 |
| CorefE | 0 | 2 | 0 | 61 | 1 | 0 | 0 |
| LinkE | 3 | 0 | 0 | 0 | 26 | 0 | 1 |
| OutE | 0 | 0 | 0 | 0 | 0 | 38 | 0 |
| gold\pred | PredE | EntE | CircE | CorefE | LinkE | OutE | NoE |
| PredE | 79 | 1 | 2 | 1 | 2 | 0 | 0 |
| EntE | 0 | 19 | 1 | 0 | 0 | 0 | 1 |
| CircE | 1 | 3 | 25 | 0 | 0 | 0 | 0 |
| CorefE | 0 | 0 | 0 | 64 | 0 | 0 | 0 |
| LinkE | 2 | 0 | 0 | 0 | 27 | 0 | 1 |
| OutE | 0 | 1 | 0 | 0 | 0 | 37 | 0 |
| gold\pred | PredE | EntE | CircE | CorefE | LinkE | OutE | NoE |
| PredE | 72 | 0 | 6 | 0 | 0 | 4 | 3 |
| EntE | 3 | 13 | 0 | 0 | 0 | 1 | 3 |
| CircE | 12 | 2 | 13 | 0 | 0 | 1 | 0 |
| CorefE | 3 | 5 | 2 | 48 | 0 | 1 | 4 |
| LinkE | 4 | 0 | 6 | 0 | 0 | 2 | 18 |
| OutE | 0 | 1 | 1 | 0 | 0 | 36 | 0 |
| gold\pred | PredE | EntE | CircE | CorefE | LinkE | OutE | NoE |
| PredE | 64 | 0 | 8 | 0 | 0 | 4 | 5 |
| EntE | 0 | 15 | 0 | 0 | 1 | 1 | 3 |
| CircE | 1 | 11 | 16 | 0 | 0 | 0 | 1 |
| CorefE | 1 | 38 | 0 | 10 | 0 | 1 | 11 |
| LinkE | 3 | 1 | 6 | 0 | 4 | 0 | 15 |
| OutE | 1 | 2 | 0 | 0 | 0 | 34 | 0 |
| gold\pred | PredE | EntE | CircE | CorefE | LinkE | OutE | NoE |
| PredE | 69 | 0 | 8 | 0 | 0 | 3 | 1 |
| EntE | 2 | 12 | 0 | 1 | 0 | 0 | 5 |
| CircE | 9 | 2 | 17 | 0 | 0 | 0 | 1 |
| CorefE | 1 | 17 | 2 | 36 | 0 | 1 | 1 |
| LinkE | 4 | 0 | 11 | 0 | 1 | 1 | 11 |
| OutE | 0 | 1 | 1 | 0 | 0 | 31 | 0 |
| Model | BAcc(Nor.) | BAcc(Str.) | SAR(Avg.) |
| The Open-Source Models | |||
| Llama-3.1-8B-Instruct | 27.69 | 21.33 | 72.57 |
| Llama-3.1-8B-Instruct※ | 23.79 | 18.12 | 79.41 |
| Qwen3-8B | 47.94 | 41.08 | 86.50 |
| Qwen3-8B※ | 50.49 | 44.78 | 87.22 |
| The State-of-the-Art LLMs | |||
| Setting | BAcc |
| Justification + Correction | 90.91 |
| Justification only | 86.30 |
| Correction only | 83.17 |
| No interpretable reasoning | 75.02 |
| Model | BAcc |
| Llama-3.1-8B-Instruct | 71.5 |
| o3 | 74.6 |
| InFi-Checker-Llama-Ablation(J) | 73.1 |
| InFi-Checker-Llama | 77.2 |
| InFi-Checker-Qwen | 77.2 |
| Model | Claim Verify | Expert QA | Factcheck-Bench | REVEAL | MediaSum | MeetingBank | Average |
| The Open-Source Models | |||||||
| Llama-3.1-8B-Instruct | 63.6 | 49.8 | 69.8 | 78.2 | 50.8 | 62.3 | 62.4 |
| The State-of-the-Art LLMs | |||||||
| GPT-4.1 | 81.6 | 80.3 | 91.3 | 93.2 | 75.9 | 86.3 | 84.8 |
| GPT-5 | 87.7 | 75.9 | 90.4 | 93.7 | 80.2 | 87.6 | 85.9 |
| o3 | 83.3 | 79.6 | 86.9 | 92.3 | 82.9 | 83.8 | 84.8 |
| Model | InFi-Check-FG | FRANK | Claim Verify | Expert QA | Factcheck-Bench | REVEAL | MediaSum | MeetingBank |
| Qwen3-8B (prompted binary) | 51.2 | 72.3 | 67.4 | 60.1 | 74.8 | 86.2 | 76.9 | 76.1 |
| Qwen3-8B (1-shot of InFi-Check reasoning) | 46.8 | 71.0 | 66.1 | 55.6 | 78.4 | 83.2 | 77.7 | 74.2 |
| InFi-Check-Qwen-Binary | 82.7 | 72.0 | 70.9 | 64.2 | 69.6 | 76.0 | 74.9 | 63.5 |
| InFi-Check-Qwen | 92.3 | 80.8 | 89.6 | 75.7 | 88.0 | 90.0 | 80.4 | 78.5 |
| Model | Parse Rate (↑) | Evidence Error Rate (↓) | Correction Error Rate (↓) |
| The Open-Source Models | |||
| Llama3.1-8B-Instruct※ | 24.09 | 23.53 | 17.65 |
| Qwen3-8B※ | 87.57 | 13.13 | 1.01 |
| The State-of-the-Art LLMs | |||
| DeepSeek-V3.2-NoThink※ | 93.21 | 5.26 | 0.75 |
| GPT-4o※ | 86.96 | 5.09 | 2.36 |
| Model | PredE | EntE | CircE | CorefE | LinkE | OutE | NoE | BAcc |
| The Open-Source Models | ||||||||
| Llama-3.1-8B-Instruct※ | 29.41 | 0.00 | 0.00 | 20.31 | 0.00 | 0.00 | 92.06 | 20.25 |
| Qwen3-8B※ | 92.94 | 14.29 | 0.00 | 45.31 | 0.00 | 28.95 | 97.62 | 39.87 |
| The State-of-the-Art LLMs | ||||||||
| GPT-4o※ | 57.65 | 33.33 | 6.90 | 26.56 | 0.00 | 89.47 | 95.63 | 44.22 |
| o3※ | 64.71 | 19.05 | 68.97 | 20.31 | 3.33 | 55.26 | 85.32 | 45.28 |
| Model | Coarse-grained Four-Category Classification | Binary Classification | ||||||
| Semantic | Discourse | Extrinsic | No error | BAcc | Hallucinated | No Error | BAcc | |
| The Open-Source Models | ||||||||
| Llama3.1-8B-Instruct※ | 14.81 | 8.51 | 0.00 | 88.49 | 15.95 | 18.35 | 88.49 | 29.41 |
| Qwen3-8B※ | 42.22 | 13.83 | 50.00 | 80.16 | 32.06 | 39.70 | 80.16 | 41.48 |
| The State-of-the-Art LLMs | ||||||||
| DeepSeek-V3.2-NoThink※ | 85.19 | 14.89 | 89.47 | 94.05 | 70.90 | 83.90 | 94.05 | 88.97 |
| Model | InFi-Check-FG | FRANK |
| The State-of-the-Art LLMs | ||
| DeepSeek-V3.2※ | 88.97 | 78.16 |
| GPT-4o※ | 90.03 | 76.69 |
| o3※ | 87.28 | 75.95 |
| Specialized Binary Models | ||
| MiniCheck-7B | 68.39 | 70.79 |
| Model | Cost($) |
| The State-of-the-Art LLMs | |
| GPT-5 | 23.9 |
| Claude-3.7-Sonnet | 12.7 |
| o3 | 12.4 |
| InFi-Checker (Ours) | |
| InFi-Checker-Llama | 3.9 |
| Model | PredE | EntE | CircE | CorefE | LinkE | OutE | NoE | BAcc |
| LongFormer-base-4096 | 2.35 | 61.90 | 0.00 | 81.25 | 10.00 | 23.68 | 0.40 | 25.65 |
| Label | Proportion(%) |
| No Problem | 78 |
| Flaws in Claim | 11 |
| Incomplete Grounding | 6 |
| False Negative Error | 3 |
| Wrong Error Type | 1 |
| Wrong Error Justification | 1 |
| Model | Predicate | Entity | Circumstance | Co-reference | Discourse Link | Extrinsic | No Error | bacc |
| GPT-4o (0-shot) | 80.00 | 76.19 | 44.83 | 50.00 | 3.33 | 84.21 | 84.92 | 60.50 |
| GPT-4o (1-shot) | 84.71 | 61.90 | 44.83 | 75.00 | 0.00 | 94.74 | 91.67 | 64.69 |
| GPT-4o (3-shot) | 80.00 | 57.14 | 34.48 | 50.00 | 3.33 | 81.58 | 88.49 | 56.43 |
| GPT-4o (5-shot) | 80.00 | 57.14 | 31.03 | 45.31 | 10.00 | 84.21 | 87.70 | 56.49 |
| DeepSeek-V3.2-NoThink (0-shot) | 78.82 | 76.19 | 58.62 | 9.38 | 26.67 | 86.84 | 80.95 | 59.64 |
| DeepSeek-V3.2-NoThink (1-shot) | 75.29 | 71.43 | 55.17 | 15.62 | 13.33 | 89.47 | 94.05 | 59.19 |
| Categorization | Method | Example | ||
| Intrinsic Errors | Semantic Frame Errors | Predicate Error (PredE) | Swapping Relation Masking | Its impact on theaters is now emerging . |
| Its absence from theaters is now receding . | ||||
| Modifying Predictions | Despite it all, 2023 should still reach the $9 billion in domestic gross hoped for this year. | |||
| Despite it all, 2023 reached the $9 billion in domestic gross hoped for this year. | ||||
| Entity Error (EntE) | Swapping Entities | Her last stage role was in My Fair Lady . | ||
| Her last stage role was in Bless This House . | ||||
Automated Fact-Checking
Automated fact-checking detects checkable claims, retrieves evidence, and assesses whether that evidence supports or refutes them.
See all 146 papers