Retrieval Grounding Latent Reasoning for Dense Retrieval

2026-08-14

AI Summary

This study addresses the challenges of spurious shortcut learning and absent incremental gains in latent reasoning for dense retrieval by proposing the RGLT framework. Leveraging silent tokens, RGLT constructs instruction-conditioned non-autoregressive latent trajectories that integrate process supervision distillation with retrieval-grounded supervision. This approach explicitly binds intermediate reasoning states to retrieval gains for the first time, effectively mitigating shortcut learning. Experimental results demonstrate that RGLT consistently outperforms strong baselines on reasoning-intensive retrieval benchmarks while maintaining efficient embedding inference. Consequently, this work establishes a novel paradigm that successfully balances performance and efficiency for optimizing reasoning-intensive dense retrieval systems.

Abstract

Reasoning-intensive retrieval requires text representations to capture not only semantic similarity, but also the reasoning needed to determine relevance under a given retrieval instruction. Existing reasoning-enhanced embedding models improve retrieval by incorporating reasoning information into dense representations, yet their supervision is typically dominated by the final retrieval objective. As a result, latent reasoning trajectories may learn shortcut reasoning patterns that preserve retrieval performance without producing meaningful incremental retrieval gains. We propose Retrieval Grounding Latent Reasoning (RGLT), a latent reasoning framework for dense retrieval that explicitly connects intermediate latent transitions with retrieval improvements. RGLT performs non-autoregressive reasoning in hidden space through an instruction-conditioned latent reasoning trajectory constructed from silent tokens. It combines process-supervised explicit-to-implicit distillation with retrieval-grounded supervision, using stage-wise CoT reconstruction to shape intermediate latent states and retrieval-effect credit to optimize incremental retrieval gains across the latent reasoning trajectories. Experiments on reasoning-intensive retrieval benchmarks show that RGLT consistently outperforms strong baselines while preserving efficient embedding inference.

Problem

• Dense Retrieval
• Latent Reasoning
• Reasoning-intensive Retrieval
• Shortcut Learning
• Incremental Retrieval Gains

Innovation

• Retrieval Grounding Latent Reasoning
• Non-autoregressive Reasoning
• Silent Tokens
• Explicit-to-Implicit Distillation
• Retrieval-effect Credit