ReasonEmbed: Reasoning-Based Text Embeddings

• ReasonEmbed is a text embedding model optimized for reasoning-intensive retrieval where multi-hop semantic relations matter most.
• It employs ReMixer, a three-step data synthesis pipeline that generates challenging, reasoning-centric training data to overcome superficial matches.
• Its Redapter training algorithm adaptively weights samples by reasoning intensity, yielding superior nDCG@10 performance on complex benchmarks.

ReasonEmbed is a text embedding model designed specifically for reasoning-intensive document retrieval scenarios, where simple lexical or surface-level similarity fails to capture the complex, multi-hop semantic relations between queries and documents. Its development addresses the deficiencies of prior embedding methods on tasks that demand nuanced, multi-step reasoning, and introduces methodological advancements in both data synthesis and model training to better align embedding geometry with reasoning-based retrieval objectives.

1. Motivation and Problem Definition

ReasonEmbed is motivated by the observation that traditional text embedding methods—such as vanilla dense retrievers or lexical-centric models like BM25—lack the inductive bias necessary to bridge the semantic gap in reasoning-laden tasks. In particular, in settings such as scientific, programming, or mathematical domains, user queries and relevant documents may only be linked by several intermediate facts or logical inferences rather than direct textual overlap. As a first-stage retriever, a model must thus encode distributed, context-dependent relationships that underpin reasoning, rather than merely local or shallow similarity. ReasonEmbed seeks to provide embeddings that are optimized for this mode of document retrieval.

2. ReMixer: Reasoning-Intensive Data Synthesis

Effective model training for reasoning-centric retrieval hinges on the availability of datasets containing queries and candidate documents where relevance depends on complex semantic or logical connections. ReMixer is a three-step data synthesis pipeline developed to overcome the so-called triviality problem in prior synthetic data generation, where queries are often superficial paraphrases, leading to models that overfit to simple surface similarities.

The stages of ReMixer are:

1. Conditioned Query Generation: LLMs are prompted to produce long-form, reasoning-centric queries conditioned on a source document, emphasizing multi-hop connections.

2. Candidate Mining: To prevent trivial matches, the original source document is excluded from the pool of retrieval candidates. Instead, a hard-negative mining procedure surfaces challenging candidates that share superficial overlap but require actual reasoning to establish or refute relevance.

3. Reasoning-Enhanced Relevance Annotation: A distilled, lightweight reasoning LLM is deployed to provide relevance labels. The model, distilled from a stronger teacher, is less likely to be biased toward spurious or tautological associations. The annotation process yields training data that reflects the genuine semantic complexity of the target retrieval tasks.

Via ReMixer, an 82K-sample dataset of high-quality, reasoning-demanding retrieval examples is constructed, overcoming the shortcut and triviality pathologies of previous synthetic approaches.

3. Redapter: Self-Adaptive Reasoning-Intensity Training

Redapter is a self-adaptive training algorithm that dynamically weights each training sample according to its “reasoning intensity” (RI), defined as the necessary degree of reasoning required to establish the query-document relationship. This weighting allows the model to preferentially allocate modeling capacity to difficult scenarios where reasoning is nontrivial, as opposed to “easy” samples dominated by lexical similarity.

The reasoning intensity is derived by comparing the standard InfoNCE loss under two conditions: using the original query $q$ and a “reasoning-augmented” query $q'$. The RI for a sample $s$ is:

$$RI_\theta(s) = \min \left\{ \frac{\mathcal{L}_{q,D}}{\mathcal{L}_{q',D}}, \kappa \right\}$$

where $\mathcal{L}_{q,D}$ and $\mathcal{L}_{q',D}$ are the InfoNCE losses for the original and reasoning-augmented query, and $\kappa$ bounds the intensity for normalization purposes.

The adaptive training loss is then

$$\mathcal{L}_{RI} = \sum_{s \in B} f(RI_\theta(s), B) \cdot \mathcal{L}_{q,D}$$

with $f(RI_\theta(s), B)$ providing normalized sample weights over minibatch $B$.

This curriculum-like prioritization forces the model to focus on samples where multi-step semantics drive relevance, promoting the learning of embeddings sensitive to reasoning.

4. Model Implementation and Evaluation Results

ReasonEmbed is instantiated using several backbone LLMs, such as Qwen3-4B, Qwen3-8B, and Llama-3.1-8B. Each backbone is fine-tuned on ReMixer-generated data using the Redapter scheme, optimizing the weighted InfoNCE loss:

$$\mathcal{L}_{q,D} = -\log\left( \frac{\exp(\langle q, d^+\rangle / \tau)}{\sum_{d' \in D} \exp(\langle q, d'\rangle / \tau)} \right)$$

where $\langle \cdot , \cdot \rangle$ denotes the embedding dot product, $d^+$ is the true positive document, $D$ the candidate set, and $\tau$ is the temperature.

On the BRIGHT benchmark, which specifically measures reasoning-demanding retrieval, the ReasonEmbed-Qwen3-8B achieves an nDCG@10 of 38.1, nearly 10 points higher than state-of-the-art baselines. This demonstrates the efficacy of both data and curriculum in advancing the representation of complex semantic structures.

5. Open-Source Resources and Reproducibility

All resources associated with ReasonEmbed—including synthetic data, codebase, and trained checkpoints—are announced for open-source release. The synthetic dataset is particularly valuable for benchmarking and extending reasoning-oriented retrievers, allowing independent validation and improvement.

The release of model checkpoints and workflows also supports reproducible research; practitioners and academics can readily evaluate, adapt, and finetune ReasonEmbed on domain-specific reasoning tasks.

6. Mathematical Formulation and Training Objective

The core training mechanism centers around a dot-product InfoNCE loss, modulated by reasoning-intensity coefficients:

• InfoNCE Loss:

$$\mathcal{L}_{q, D} = -\log \left( \frac{\exp(\langle q, d^+ \rangle / \tau)}{\sum_{d' \in D} \exp(\langle q, d' \rangle / \tau)} \right)$$

• Reasoning-Intensity Adaptive Weighting:

$$\mathcal{L}_{RI} = \sum_{s \in B} \left( \frac{RI_\theta(s)}{\sum_{s' \in B} RI_\theta(s')} \right) \mathcal{L}_{q,D}$$

Through this self-adaptive approach, the gradient signal amplifies the contribution of reasoning-intensive data, producing embedding spaces aligned with multi-hop retrieval relevance, rather than correlation or direct similarity.

7. Significance and Outlook

ReasonEmbed exemplifies a new class of retrieval models attuned to the cognitive demands of advanced information-seeking, where answer finding requires the joint retrieval and composition of semantically distant facts. The model’s combination of synthetic, high-reasoning data and sample-adaptive training is shown to be decisive for closing the performance gap on reasoning-based metrics. By making all resources open-source, ReasonEmbed provides a research foundation for future exploration, particularly in alignment with scientific, technical, and long-form QA domains.

Further directions include scaling to higher backbone capacities, extending the reasoning curriculum to dialog and code-search domains, and cross-pollinating with models for chain-of-thought and multi-evidence question-answering.