Research map / Large language models / Retrieval and knowledge
Refine Knowledge of Large Language Models via Adaptive Contrastive Learning
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes Adaptive Contrastive Learning, which builds positive/negative samples per knowledge quadrant (based on sampled answer accuracy against IK/IDK thresholds) to make LLMs answer known questions and refuse unknown ones, improving TRUTHFUL rate on TriviaQA and Natural Questions.
Problem
LLM hallucination: models answer questions beyond their knowledge instead of honestly refusing; existing knowledge-boundary methods are coarse and ignore uncertain knowledge and incorrect responses.
Approach
Partition training questions into knowledge quadrants using sampled-response accuracy under two thresholds (IK and IDK), then build quadrant-specific positive/negative instances and a cosine contrastive loss combined with generation cross-entropy using an adaptive weight.
Key results
- On LLaMA-2-7B-chat with TriviaQA, IDK-SFT-Adpt-Ctr reaches 78.2 Truthful rate vs 73.2 for IDK-SFT and 66.9 for IDK-Prompting.
- On Mistral-7B-Instruct-v0.1 with Natural Questions (OOD), Truthful rate improves to 77.0 vs 70.2 for IDK-SFT and 44.4 for IDK-Prompting.
- On ALCUNA with LLaMA-2-7B-chat, Truthful rate is 98.3 vs 97.8 for IDK-SFT.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use adaptive contrastive learning instead of IDK-SFT | LLaMA-2-7B-chat · TriviaQA | TRUTHFUL: 73.2 -> 78.2 | Helped |
| Use adaptive contrastive learning instead of IDK-SFT | LLaMA-2-7B-chat · Natural Questions | TRUTHFUL: 60.3 -> 67.2 | Helped |
| Use adaptive contrastive learning instead of IDK-SFT | Mistral-7B-Instruct-v0.1 · TriviaQA | TRUTHFUL: 72.2 -> 73.5 | Helped |
| Use adaptive contrastive learning instead of IDK-SFT | Mistral-7B-Instruct-v0.1 · Natural Questions | TRUTHFUL: 70.2 -> 77.0 | Helped |
| Add RAG | IDK-SFT-Adpt-Ctr · RAG-Bench | TRUTHFUL: 66.9 -> 71.2 | Helped |
| Set IDK rate to 0.5 instead of 0.7 (ablation) | IDK-SFT-Adpt-Ctr · TriviaQA | TRUTHFUL: 78.2 -> 77.2 | Hurt |
| Set IDK rate to 0.9 instead of 0.7 (ablation) | IDK-SFT-Adpt-Ctr · TriviaQA | TRUTHFUL: 78.2 -> 73.6 | Hurt |
| Set IDK rate to 0.5 instead of 0.7 (ablation) | IDK-SFT-Adpt-Ctr · Natural Questions | TRUTHFUL: 67.2 -> 62.2 | Hurt |
| Set IDK rate to 0.9 instead of 0.7 (ablation) | IDK-SFT-Adpt-Ctr · Natural Questions | TRUTHFUL: 67.2 -> 63.8 | Hurt |
| Remove L_Q4 (ablation) | Adaptive contrastive learning · TriviaQA | TRUTHFUL: 78.2 -> 74.4 | Hurt |
| Remove L_Q3 (ablation) | Adaptive contrastive learning · TriviaQA | TRUTHFUL: 78.2 -> 78.0 | Hurt |
| Remove L_Q1 (ablation) | Adaptive contrastive learning · TriviaQA | TRUTHFUL: 78.2 -> 76.1 | Hurt |
Surprising findings
- Adding the adaptive contrastive loss to SFT lowers IK-IK on Mistral TriviaQA (31.4 to 24.4) while raising Truthful rate, showing gains come largely from more refusals.
- On RAG-Bench, IDK-SFT-Adpt-Ctr's IK-IK drops below IDK-SFT (44.5 vs 60.3) though its Truthful rate is higher (66.9 vs 62.2).
Limitations
- Truthful rate can be inflated by increased refusals; IK-IK sometimes falls substantially.
- Correctness is judged by lexical matching, which may miscount correct answers.
- Evaluated only on two 7B-13B chat models and a single contrastive setup with fixed thresholds.
Experimental setup
Datasets: TriviaQA Natural Questions ALCUNA RAG-Bench
Benchmarks: TriviaQA Natural Questions ALCUNA RAG-Bench
Metrics: IK-IK rate IK-IDK rate Truthful rate
Baselines: IDK-Prompting IDK-SFT
Builds on: supervised fine-tuning contrastive instruction tuning knowledge quadrants
Closest papers by meaning
- Hallucination Augmented Contrastive Learning for Multimodal Large Language Model (CVPR 2024 · similarity 0.89)
- Wisdom is Knowing What not to Say: Hallucination-Free LLMs Unlearning via Attention Shifting (NeurIPS 2025 · similarity 0.89)
- Decoupling Contrastive Decoding: Robust Hallucination Mitigation in Multimodal Large Language Models (NeurIPS 2025 · similarity 0.88)
- The Mirage of Performance Gains: Why Contrastive Decoding Fails to Mitigate Object Hallucinations in MLLMs? (NeurIPS 2025 · similarity 0.88)
- No Free Lunch: Fundamental Limits of Learning Non-Hallucinating Generative Models (ICLR 2025 · similarity 0.88)
- Can Knowledge Editing Really Correct Hallucinations? (ICLR 2025 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.