InfoSteer: Steering Information Utility in Language Model Post-Training

Abstract: Recent advancements in LMs gradually ushered in an era where post-training is crucial. Yet, post-training approaches such as supervised fine-tuning (SFT) do not guarantee effective use of knowledge acquired during pretraining. We therefore present \ours, a lightweight method that encourages parametric information utilization in LMs during post-training. This is achieved via treating FFN layer as associate key-value memory, and promotes the use of stored memory vectors via forward-pass interventions or regularization during backpropagation. We find this simple guidance during post-training phase delivers consistent performance improvements across diverse model families--including Qwen, Gemma and Llama-spanning over 15 downstream tasks in both ID and OOD evaluations. Beyond performance gains, we also find that steered LMs can adaptively allocate information-placing more emphasis on generating semantically meaningful tokens, while using fewer resources on simple transition ones (e.g., ,' orand'). Our work underscores that vanilla post-training does not fully leverage pre-training potential, and steering LMs in latent representation space offers a promising approach that enhances both performance and interpretability.