Bitune: Bidirectional Instruction-Tuning

2024-05-23

arXiv.org

AI Summary

Decoder-only large language models suffer from limited representational capacity due to their exclusive reliance on unidirectional causal attention. To address this, we propose the first instruction-tuning method specifically designed for pure decoder architectures: during the prompt encoding phase, we introduce parallel causal and bidirectional attention pathways with parameter separation; their outputs are dynamically fused via learnable weights to guide autoregressive generation. Our approach is architecture-agnostic and does not depend on any specific parameter-efficient fine-tuning (PEFT) technique. Experiments demonstrate consistent zero-shot performance gains over baselines across commonsense reasoning, arithmetic, and linguistic understanding tasks. Ablation studies confirm the effectiveness and necessity of incorporating bidirectional attention, the dual-path design, and the learnable fusion mechanism.

Technology Category

• Natural Language Processing: (Large) Language Models
• 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

We introduce Bitune, a method that improves instruction-tuning of pretrained decoder-only large language models, leading to consistent gains on downstream tasks. Bitune applies both causal and bidirectional attention to the prompt, to obtain a better representation of the query or instruction. We realize this by introducing two sets of parameters, for which we apply parameter-efficient finetuning techniques. These causal and bidirectional features are then combined into a weighted average with trainable coefficients, which is subsequently used to generate new tokens. We demonstrate significant improvements in zero-shot performance on commonsense reasoning, arithmetic, and language understanding tasks, while extensive ablation studies validate the role of each component and demonstrate the method's agnosticism to different PEFT techniques.

Problem

• Enhancing decoder-only LLMs with bidirectional attention
• Improving performance on reasoning and understanding tasks
• Enabling bidirectional information flow in prompt processing

Innovation

• Incorporates bidirectional attention into decoder-only models
• Enhances performance on reasoning and understanding tasks
• Compatible with parameter-efficient and full finetuning techniques