I3: Intent-Introspective Retrieval Conditioned on Instructions

TL;DR

I3 tackles the challenge of zero-shot retrieval across tasks with diverse search intents by introducing intent-introspective retrieval conditioned on natural-language instructions. It preserves pre-trained dual-encoder capabilities while attaching a pluggable, parameter-isolated introspector that jointly reasons over queries and instructions, producing intent-aware embeddings. The learning process, called progressively-pruned intent learning, uses large-language-model data and a teacher-student pruning scheme plus drawback extrapolation to iteratively refine the introspector. Across BEIR, I3 achieves state-of-the-art zero-shot performance in both retrieval and reranking settings with modest computational overhead, demonstrating strong generalization and practical impact for instruction-guided IR.

Abstract

Recent studies indicate that dense retrieval models struggle to perform well on a wide variety of retrieval tasks that lack dedicated training data, as different retrieval tasks often entail distinct search intents. To address this challenge, in this work we leverage instructions to flexibly describe retrieval intents and introduce I3, a unified retrieval system that performs Intent-Introspective retrieval across various tasks, conditioned on Instructions without any task-specific training. I3 innovatively incorporates a pluggable introspector in a parameter-isolated manner to comprehend specific retrieval intents by jointly reasoning over the input query and instruction, and seamlessly integrates the introspected intent into the original retrieval model for intent-aware retrieval. Furthermore, we propose progressively-pruned intent learning. It utilizes extensive LLM-generated data to train I3 phase-by-phase, embodying two key designs: progressive structure pruning and drawback extrapolation-based data refinement. Extensive experiments show that in the BEIR benchmark, I3 significantly outperforms baseline methods designed with task-specific retrievers, achieving state-of-the-art zero-shot performance without any task-specific tuning.

Figures (6)

• Figure 1: (a) Existing methods require training a distinct retrieval model for each task to implicitly grasp specific retrieval intent. (b) I3 directly handles different tasks through intent-introspective retrieval following user instructions.
• Figure 2: Parameter-isolated framework of I3, with a pluggable introspector enabling retrieval models to efficiently integrate the introspected intents for better query encoding.
• Figure 3: progressively-pruned intent learning with structure pruning and drawback extrapolation-based data refinement.
• Figure 4: Average nDCG@10 of I3 incorporating different backbones on the BEIR benchmark.
• Figure 5: (a): Performance on BEIR with different instruction treatments during testing. (b): Performance of I3 with different training data scales.