Copy-Paste to Mitigate Large Language Model Hallucinations

Conference: ICLR 2026 arXiv: 2510.00508 Code: https://github.com/longyongchao/CopyPasteLLM Area: Causal Reasoning Keywords: Hallucination Mitigation, RAG, Copy-Paste, DPO, Faithfulness

TL;DR

This paper proposes a Copy-Paste generation paradigm that trains LLMs to preferentially copy spans directly from retrieved context rather than paraphrasing them freely. Combined with high-copy-preference DPO training, the approach improves faithfulness on counterfactual RAG benchmarks from 80.2% to 92.8%.

Background & Motivation

Background: RAG (Retrieval-Augmented Generation) reduces hallucinations by supplying LLMs with external context; however, LLMs tend to paraphrase rather than directly cite the context during generation, leading to information distortion and hallucinations.

Limitations of Prior Work: The paraphrasing process introduces two types of hallucinations—"Twist" (distorting facts present in the context) and "Causal" (upstream errors in the causal chain propagating downstream). Attribution methods mark sources but do not alter the generation process itself.

Key Challenge: There is a fundamental trade-off between fluent paraphrasing and faithful copying—while paraphrasing reads more naturally, each reformulation introduces a risk of hallucination.

Goal: Can LLMs be trained to copy context spans as directly as possible while maintaining readability?

Key Insight: The analysis is grounded in an attention-anchoring perspective—when the previously generated token is copied from the context, the query vector of the next token is strongly correlated with the context key vectors, naturally promoting continued copying.

Core Idea: Train LLMs to develop a "high-copy preference" by using DPO to steer the model toward responses that directly embed context spans.

Method

Overall Architecture

The approach consists of two stages: (1) Copy-Paste Prompting to generate high-copy-rate candidate responses via three strategies (CP-Order / CP-Link / CP-Refine), and (2) multi-dimensional filtering + Elo ranking + DPO training.

Key Designs

1. Three Copy Strategies:

   • CP-Order: Strict extraction—reordering relevant sentences from the context.
   • CP-Link: Allows transition phrases of no more than 15 words to connect copied spans.
   • CP-Refine: An iterative writer–reviewer loop of up to 5 rounds to improve readability while maintaining a high copy rate.

2. Quantitative Metrics:

   • Copy Coverage \(\kappa\) = proportion of response tokens originating from the context.
   • Copy Density \(\delta\) = emphasizes long contiguous spans (weighted by the square of span length).

3. DPO Training:

   • Requires only 365 high-quality preference pairs.
   • Multi-dimensional filtering: AlignScore / MiniCheck (faithfulness), \(\kappa\) / \(\delta\) (copy intensity), query relevance, and fluency.
   • Answer Stamping: the correct answer is appended at the end of the response to prevent answer omission caused by excessive copying.

Key Experimental Results

Main Results

Dataset	Model	Method	Accuracy
FaithEval (Counterfactual)	Llama-3-8B	Context-DPO	80.2%
FaithEval (Counterfactual)	Llama-3-8B	CopyPasteLLM	92.8%
ConFiQA-MC	Llama-3-8B	Attributed	37.3%
ConFiQA-MC	Llama-3-8B	CopyPasteLLM	82.5%

Ablation Study

Variant	FaithEval	Note
w/o Copy Preference	71.2%	No high-copy training data
w/o Answer Stamping	45.1%	Excessive copying causes answer omission
CopyPasteLLM	92.8%	Full method

Key Findings

• Answer Stamping is critical—its removal causes accuracy to drop sharply from 92.8% to 45.1%.
• Effective training requires only 365 preference pairs, demonstrating extremely high data efficiency.
• Copy Density is a better predictor of faithfulness than Coverage; long contiguous spans are more reliable than short fragments.

Highlights & Insights

• Attention-Anchoring Theory: Copy operations enjoy a natural advantage at the attention mechanism level—when the previous token is copied from context, key-value vectors naturally guide continued copying, creating a "copy momentum."
• High-Efficiency Training with Minimal Data: DPO on just 365 samples is sufficient to substantially shift generation style, suggesting that "copy vs. paraphrase" is primarily a preference issue rather than a capability issue.
• Necessity of Answer Stamping: Explicitly prompting the model to state the answer at the end balances copy faithfulness with response completeness.

Limitations & Future Work

• High copy rates may reduce the naturalness and readability of responses.
• Validation is limited to English RAG tasks; cross-lingual generalization remains unknown.
• The Copy-Paste strategy may be ill-suited for questions requiring inferential synthesis rather than direct span retrieval.

Related Work & Insights

• vs. Context-DPO: Both employ DPO, but Context-DPO does not explicitly optimize for copy preference, whereas this work directly optimizes the copying behavior.
• vs. Attributed LLM: Attribution methods only annotate source references without modifying the generation process; this work intervenes at the level of the generation paradigm itself.

Rating

• Novelty: ⭐⭐⭐⭐ The "copy-over-paraphrase" paradigm is novel and counterintuitive.
• Experimental Thoroughness: ⭐⭐⭐⭐ Multi-dataset, multi-model validation with clear ablations.
• Writing Quality: ⭐⭐⭐⭐ The attention-anchoring analysis is insightful.
• Value: ⭐⭐⭐⭐⭐ High practical value; directly applicable to RAG systems.

Related Papers

• [ACL 2026] CausalDetox: Causal Head Selection and Intervention for Language Model Detoxification
• [NeurIPS 2025] Revealing Multimodal Causality with Large Language Models
• [AAAI 2026] Hallucinate Less by Thinking More: Aspect-Based Causal Abstention for Large Language Models
• [NeurIPS 2025] Counterfactual Reasoning for Steerable Pluralistic Value Alignment of Large Language Models
• [AAAI 2026] Sparse Additive Model Pruning for Order-Based Causal Structure Learning