SelfElicit: Your Language Model Secretly Knows Where is the Relevant Evidence

2025-02-12

AI Summary

Language models often struggle to identify critical evidence in noisy contexts, leading to degraded factual accuracy in question answering. To address this, we propose a training-free, single-forward-pass self-guided highlighting method that implicitly locates evidential spans via deep-layer Transformer attention scores and explicitly highlights relevant context segments through a lightweight, inference-time intervention mechanism. Unlike conventional retrieval-augmented or iterative prompting paradigms, our approach enables dynamic evidence focusing without parameter updates or fine-tuning. Evaluated across diverse evidence-based QA benchmarks, it consistently improves factual accuracy for mainstream open-weight models—including Llama, Qwen, and Phi—while incurring negligible computational overhead. The method demonstrates strong generalization and robust performance across domains and model scales.

Technology Category

• Natural Language Processing: Question Answering
• Machine Learning: Large Multimodal Models (LMMs)
• Search and Optimization: Learning to Search

Application Category

• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• User Modeling, Personalization and Recommendation: Fairness-aware retrieval and ranking

Abstract

Providing Language Models (LMs) with relevant evidence in the context (either via retrieval or user-provided) can significantly improve their ability to provide factually correct grounded responses. However, recent studies have found that LMs often struggle to fully comprehend and utilize key evidence from the context, especially when it contains noise and irrelevant information - an issue common in real-world scenarios. To address this, we propose SelfElicit, an inference-time approach that helps LMs focus on key contextual evidence through self-guided explicit highlighting. By leveraging the inherent evidence-finding capabilities of LMs using the attention scores of deeper layers, our method automatically identifies and emphasizes key evidence within the input context, facilitating more accurate and factually grounded responses without additional training or iterative prompting. We demonstrate that SelfElicit brings consistent and significant improvement on multiple evidence-based QA tasks for various LM families while maintaining computational efficiency. Our code and documentation are available at https://github.com/ZhiningLiu1998/SelfElicit.

Problem

• Enhances LM comprehension of key evidence
• Reduces impact of noise in contextual information
• Improves factual accuracy without additional training

Innovation

• Self-guided explicit highlighting
• Leveraging deeper attention scores
• No additional training required