Sparse Semantic Steering for Inference-Time Dense Retrieval Correction

Dense retrieval models suffer from semantic drift (retrieving docu- ments based on spurious distributional overlaps rather than contex- tual relevance), and correcting this typically requires expensive of- fline fine tuning. We propose Sparse Semantic Steering, a framework that corrects dense retrieval at inference time without retraining the base encoder. We project frozen embeddings through a Sparse Autoencoder (SAE) into a disentangled feature dictionary, then train a Proximal Policy Optimization (PPO) agent over a Context Aware State combining the dense query with top sparse features from a Pseudo Relevance Feedback (PRF) neighborhood. The agent learns bidirectional feature modifications (Negative Semantic Mask- ing), projecting corrections back into the dense index. The policy is trained with simulated implicit feedback from relevance judgments but requires no labels at inference time. On five BEIR datasets the method achieves up to +61.7% rescue rate (NDCG@10) on ArguAna with negligible overhead (+4.30 MB VRAM, 0.002 ms routing).