GrAInS: Gradient-based Attribution for Inference-Time Steering of LLMs and VLMs

TL;DR

GrAInS introduces a gradient-based, inference-time steering method for LLMs and VLMs that enhances control over model outputs by identifying influential tokens and adjusting activations without retraining.

Contribution

It presents a novel, interpretable approach using Integrated Gradients for token attribution to steer models at inference time across multimodal tasks.

Findings

Achieves 13.22% accuracy improvement on TruthfulQA with Llama-3.1-8B.

Reduces hallucination rates on MMHal-Bench from 0.624 to 0.514.

Improves alignment win rates on SPA-VL by 8.11%.

Abstract

Inference-time steering methods offer a lightweight alternative to fine-tuning large language models (LLMs) and vision-language models (VLMs) by modifying internal activations at test time without updating model weights. However, most existing approaches rely on fixed, global intervention vectors, overlook the causal influence of individual input tokens, and fail to leverage informative gradients from the model's logits, particularly in multimodal settings where visual and textual inputs contribute unevenly. To address these limitations, we introduce GrAInS, an inference-time steering approach that operates across both language-only and vision-language models and tasks. GrAInS uses contrastive, gradient-based attribution via Integrated Gradients to identify the top-k most influential tokens, both positively and negatively attributed based on their contribution to preferred versus dispreferred outputs. These tokens are then used to construct directional steering vectors that capture semantic shifts from undesirable to desirable behavior. During inference, GrAInS adjusts hidden activations at transformer layers guided by token-level attribution signals, and normalizes activations to preserve representational scale. This enables fine-grained, interpretable, and modular control over model behavior, without retraining or auxiliary supervision. Empirically, GrAInS consistently outperforms both fine-tuning and existing steering baselines: it achieves a 13.22% accuracy gain on TruthfulQA using Llama-3.1-8B, reduces hallucination rates on MMHal-Bench from 0.624 to 0.514 with LLaVA-1.6-7B, and improves alignment win rates on SPA-VL by 8.11%, all while preserving the model's fluency and general capabilities.