Do Reasoning Models Enhance Embedding Models?

TL;DR

This paper investigates whether reasoning-enhanced models improve semantic embeddings and finds that RLVR training does not significantly outperform base models in embedding tasks, due to manifold realignment effects.

Contribution

Introduces HRSA, a framework for analyzing representation similarity, and reveals how RLVR influences latent geometry without improving embedding performance.

Findings

RLVR causes local geometry reorganization but preserves global structure.

Contrastive learning aligns base and reasoning models, leading to manifold realignment.

RLVR optimizes within existing semantic landscapes rather than restructuring them.

Abstract

State-of-the-art embedding models are increasingly derived from decoder-only Large Language Model (LLM) backbones adapted via contrastive learning. Given the emergence of reasoning models trained via Reinforcement Learning with Verifiable Rewards (RLVR), a natural question arises: do enhanced reasoning translate to superior semantic representations when these models serve as embedding initializations? Contrary to expectation, our evaluation on MTEB and BRIGHT reveals a **null effect**: embedding models initialized from RLVR-tuned backbones yield no consistent performance advantage over their base counterparts when subjected to identical training recipes. To unpack this paradox, we introduce **H**ierarchical **R**epresentation **S**imilarity **A**nalysis (HRSA), a framework that decomposes similarity across representation, geometry, and function levels. HRSA reveals that while RLVR induces irreversible latent manifold's local geometry reorganization and reversible coordinate basis drift, it preserves the global manifold geometry and linear readout. Consequently, subsequent contrastive learning drives strong alignment between base- and reasoning-initialized models, a phenomenon we term **Manifold Realignment**. Empirically, our findings suggest that unlike Supervised Fine-Tuning (SFT), RLVR optimizes trajectories within an existing semantic landscape rather than fundamentally restructuring the landscape itself.