Tell Your Model Where to Attend: Post-hoc Attention Steering for LLMs
This addresses the need for better instruction-following and knowledge integration in LLMs for users, though it is incremental as it builds on existing attention mechanisms without modifying model parameters.
The paper tackles the problem of steering large language models (LLMs) to pay closer attention to user-specified information, such as instructions, by introducing PASTA, a post-hoc attention steering approach that applies precise attention reweighting on a small subset of attention heads, resulting in an average accuracy improvement of 22% for LLAMA-7B on various tasks.
In human-written articles, we often leverage the subtleties of text style, such as bold and italics, to guide the attention of readers. These textual emphases are vital for the readers to grasp the conveyed information. When interacting with large language models (LLMs), we have a similar need -- steering the model to pay closer attention to user-specified information, e.g., an instruction. Existing methods, however, are constrained to process plain text and do not support such a mechanism. This motivates us to introduce PASTA -- Post-hoc Attention STeering Approach, a method that allows LLMs to read text with user-specified emphasis marks. To this end, PASTA identifies a small subset of attention heads and applies precise attention reweighting on them, directing the model attention to user-specified parts. Like prompting, PASTA is applied at inference time and does not require changing any model parameters. Experiments demonstrate that PASTA can substantially enhance an LLM's ability to follow user instructions or integrate new knowledge from user inputs, leading to a significant performance improvement on a variety of tasks, e.g., an average accuracy improvement of 22% for LLAMA-7B. Our code is publicly available at https://github.com/QingruZhang/PASTA .