Research map / Large language models / Language, safety and interpretability
SCOPE: A Self-supervised Framework for Improving Faithfulness in Conditional Text Generation
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 SCOPE, a self-supervised two-stage method that fine-tunes an LLM on half the data, then applies DPO-style preference tuning using synthetic unfaithful samples generated by mixing the fine-tuned and pre-trained model's distributions, improving faithfulness in data-to-text and summarization.
Problem
LLMs fine-tuned for conditional text generation (data-to-text, summarization) produce hallucinations, i.e., content not grounded in the input context, and existing faithfulness methods give mixed results on recent LLMs.
Approach
Generate unfaithful negatives without external tools by noisy token-level decoding that mixes the fine-tuned model conditioned on context with the unconditional pre-trained model (controlled by alpha), then preference-tune the fine-tuned model to prefer the reference over these negatives.
Key results
- On ToTTo with Llama-2-13b, SCOPE reaches NLI 54.27 vs 46.56 for SFT and PARENT 86.58 vs 80.47.
- On XSum with Llama-2-13b, SCOPE raises AlignScore to 66.03 from 56.53 for SFT.
- In human evaluation on ToTTo (Llama-2), SCOPE is preferred 40.0% vs 15.2% for SFT, about 2.6x the win rate.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| apply SCOPE preference tuning after supervised fine-tuning | Llama2-7B · ToTTo | NLI: 46.42 -> 51.88 | Helped |
| apply SCOPE preference tuning after supervised fine-tuning | Llama2-7B · E2E | NLI: 92.62 -> 94.64 | Helped |
| apply SCOPE preference tuning after supervised fine-tuning | Llama2-7B · FeTaQA | NLI: 39.06 -> 42.97 | Helped |
| apply SCOPE preference tuning after supervised fine-tuning | Llama2-7B · WebNLG | NLI: 79.36 -> 83.42 | Helped |
| apply SCOPE preference tuning after supervised fine-tuning | Llama2-7B · SAMSum | AlignScore: 80.66 -> 83.67 | Helped |
| apply SCOPE preference tuning after supervised fine-tuning | Llama2-7B · XSum | AlignScore: 56.25 -> 65.10 | Helped |
| apply SCOPE preference tuning after supervised fine-tuning | Llama2-7B · PubMed | AlignScore: 46.89 -> 58.17 | Helped |
| train first phase on 50% rather than 25% of the data (ablation) | SCOPE · ToTTo validation | NLI: 49.57 -> 50.64 | Helped |
| train first phase on 50% rather than 75% of the data (ablation) | SCOPE · ToTTo validation | PARENT: 84.10 -> 86.34 | Helped |
| replace DPO with ORPO preference loss (ablation) | SCOPE · ToTTo validation | PARENT: 84.2 -> 85.9 | No clear effect |
| set beta to 0.05 instead of 0.1 (ablation) | SCOPE · ToTTo validation | NLI: 49.21 -> 48.31 | Hurt |
| set beta to 1 instead of 0.1 (ablation) | SCOPE · ToTTo validation | NLI: 49.21 -> 46.24 | Hurt |
Surprising findings
- Existing self-supervised faithfulness methods (CAD, PMI, CLIFF, Critic) show minimal or no faithfulness gains over SFT on data-to-text with recent 7B-13B LLMs.
- SCOPE slightly lowers BLEU/ROUGE on several datasets, which the authors interpret as deviation from SFT rather than a fluency loss.
Limitations
- Experiments limited to 7B and 13B models due to computational constraints.
- Relies mainly on automatic metrics; human evaluation covers only 125 ToTTo samples.
- WebNLG and E2E are relatively simple tasks.
Experimental setup
Datasets: ToTTo WebNLG 2020 E2E FeTaQA XSum SAMSum PubMed Alpaca
Benchmarks: ToTTo E2E FeTaQA WebNLG SAMSum XSum PubMed
Metrics: NLI score PARENT recall BLEU AlignScore FactCC QuestEval ROUGE-L GPT-4 win/tie/loss rate human preference
Baselines: SFT CAD PMI Critic-driven decoding CLIFF
Builds on: DPO ORPO CLIFF Context-aware decoding PMI decoding
Closest papers by meaning
- Precise Information Control in Long-Form Text Generation (NeurIPS 2025 · similarity 0.89)
- FaithEval: Can Your Language Model Stay Faithful to Context, Even If "The Moon is Made of Marshmallows" (ICLR 2025 · similarity 0.87)
- ParamMute: Suppressing Knowledge-Critical FFNs for Faithful Retrieval-Augmented Generation (NeurIPS 2025 · similarity 0.87)
- Self-Supervised Alignment with Mutual Information: Learning to Follow Principles without Preference Labels (NeurIPS 2024 · similarity 0.86)
- Controlled Text Generation via Language Model Arithmetic (ICLR 2024 · similarity 0.86)
- Idiosyncrasies in Large Language Models (ICML 2025 · similarity 0.86)
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.