FishBack method improves transformer activation steering using non-Euclidean geometry

Researchers have developed a new method called FishBack to improve activation steering in transformers, a technique for modifying language model behavior without updating parameters. Existing methods are often unstable and can disturb unrelated behaviors due to a flawed assumption that the activation space is Euclidean. FishBack corrects this by using the Fisher information metric of the softmax layer, pulled back through the Jacobian matrix, to derive an optimal steering direction. This approach demonstrates improved performance on models like GPT-2 Small, Llama 3-8B, and Qwen3_8B by reducing off-target distortions, particularly in earlier and middle layers where geometric corrections are most impactful. AI

IMPACT This research offers a more stable and precise method for controlling language model behavior, potentially leading to more reliable and steerable AI systems.

RANK_REASON Academic paper detailing a new method for transformer activation steering. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.CL →

• ActAdd
• FishBack
• Fisher information metric
• GPT-2 small
• Jacobian matrix
• Llama 3-8B
• Qwen3_8B
• Sihan Wang
• softmax layer
• Transformers

• paper
• model release

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →