arXiv Proposes Gradient Merging Methods to Enhance Prompt Robustness in Large Language Models
Summary:A new study on arXiv introduces two novel gradient merging methods, Prompt-TIES and Prompt-DARE, to address the sensitivity of large language models (LLMs) to prompt formulation during training. By analyzing the interference of per-paraphrase gradients and leveraging interference-resolution operators from model and gradient merging, the researchers enhance prompt robustness. Experiments demonstrate that these methods significantly reduce per-example variance while maintaining overall accuracy, s
Background and Motivation
Large Language Models (LLMs) excel in natural language processing tasks but remain highly sensitive to the specific phrasing of prompts. Current solutions rely on manual prompt optimization at inference time, which is time-consuming and lacks scalability. Therefore, researchers are turning to train-time robustness enhancement methods to build prompt robustness during instruction tuning.
Methodology and Innovation
This study introduces two novel gradient merging methods:
- Prompt-TIES: Utilizes interference-resolution operators from model merging to address interference in prompt gradients.
- Prompt-DARE: Applies interference-resolution operators from gradient merging to optimize prompt gradients.
The core idea is to resolve the sign conflicts in gradients from different paraphrased prompts during training. The research finds that 57% of gradients exhibit sign conflicts, and the naive gradient average approach discards about a fifth of the gradient magnitude.
Experiments and Results
Evaluations across four models (from two families) demonstrate that:
- Enhanced Local Robustness: The gradient merging methods directly improve the model's local robustness, particularly in tasks like ANLI and WiC.
- Reduced Per-Example Variance: These methods nearly halve the per-example variance in fully fine-tuned models while maintaining overall accuracy.
- Comparison with Other Methods: Compared to train-time robustness methods such as data construction, consistency regularization, and contrastive alignment, the gradient merging methods show superior performance in reducing per-example variance.
Engineering Trade-offs and Limitations
- Computational Overhead: While the gradient merging methods introduce additional computational overhead during training, experiments indicate that this overhead is within an acceptable range.
- Applicability: The methods show good adaptability across different model sizes and tasks, but further validation is needed for large-scale models.
Conclusion and Future Directions
Prompt-TIES and Prompt-DARE provide new avenues for enhancing prompt robustness in LLMs. Future research can explore the applicability of these methods across different model architectures and tasks, as well as optimize their computational efficiency.
Developer Recommendations
For developers aiming to improve model robustness, integrating Prompt-TIES or Prompt-DARE during training is recommended, especially when dealing with complex or diverse prompts. The open-source code and model checkpoints are available on GitHub and Hugging Face for quick integration and testing.
— END —Source: ArXiv NLP/LLM (cs.CL) (2026-10-30)
Tags: #LLMs & Foundation Models #Gradient Merging #Prompt Robustness #Training Optimization
Editorial & Fact-Checking Note: This article is compiled from primary research, official release documentation, and source papers by the ZICQ Newsroom pipeline with automated entity verification and human editorial review. If you notice any technical inaccuracy, please submit a correction via our corrections policy or email our editorial desk directly.
Community Comments