New inference method boosts reasoning capabilities of text embedding models

A team of researchers has proposed Refine Thought (RT), an approach that enhances how text embedding models perform semantic reasoning tasks by making multiple passes through the model during inference.

How the method works#

Rather than running the embedding model a single time, RT executes multiple forward passes to generate the final semantic representation of text. This test-time inference strategy reactivates reasoning capabilities that were learned but not fully utilized during the model's initial pretraining phase, particularly in decoder-only embedding architectures like Qwen3-Embedding-8B.

Experimental results#

The technique demonstrated substantial performance gains on specialized semantic reasoning benchmarks, including the BRIGHT dataset and PJBenchmark—a dataset focused on person-job matching tasks. Critically, these improvements came without sacrificing performance on general-purpose semantic understanding tasks measured by C-MTEB, indicating that the method doesn't sacrifice breadth of capability for depth in specific domains.

Significance#

RT represents a practical way to unlock latent reasoning abilities in existing embedding models at evaluation time, rather than requiring retraining or architectural changes. The approach demonstrates that decoder-only text embedding models possess dormant semantic reasoning capacities that can be activated through strategic inference modifications.

The paper was submitted to arXiv in October 2025 and revised in July 2026.