ReasonEmbed: Enhanced Text Embeddings for Reasoning-Intensive Document Retrieval

• Reasoning-intensive retrieval requires complex semantic matching that challenges general-purpose embeddings and BM25.Source: §1 · p. 1
• Existing datasets often come from traditional search and question answering, while synthetic alternatives have produced only marginal gains.Source: §1 · p. 1
• ReMixer uses conditioned query generation, source-excluded candidate mining, and reasoning-enhanced annotation to reduce trivial query-document relationships.Source: §1 · pp. 1–2
• Redapter dynamically weights synthetic training samples according to estimated reasoning intensity, improving utilization of data with varying difficulty.Source: §1 · p. 2
• 38.1 nDCG@10 is achieved by ReasonEmbed-Qwen3-8B on BRIGHT, while the Qwen3-4B variant reaches 37.1 and surpasses existing text embeddings.Source: §1 · p. 2
• The authors report that ReasonEmbed resources, including code, datasets, and trained models, will be publicly released.Source: §1 · p. 2

Text Embeddings

• Text embedding research has progressed from multi-stage training toward LLM backbones and distillation for stronger or more deployable models.Source: §2 · p. 2

Reasoning-Intensive Document Retrieval

• Reasoning-intensive retrieval differs from traditional keyword or semantic matching because relevant-document identification requires deeper reasoning.Source: §2 · p. 2
• Prior work targets either first-stage retrieval with tailored embeddings and query rewriting or reasoning-enhanced re-ranking.Source: §2 · p. 2
• Fundamental challenges remain in first-stage retrieval effectiveness and the curation of suitable training data.Source: §2 · p. 2

2. Query Generation

• A generation instruction template specifies input content, output format, length, and education level.Source: §1 · p. 3
• An annotation prompt provides a relevance definition and asks the model to determine query-document relevance.Source: §1 · p. 3
• The annotation prompt includes query analysis, document analysis, and relevance annotation stages.Source: §1 · p. 3

2. Candidate Mining

• Figure 1 presents ReMixer as a three-stage data synthesis workflow.Source: Figure 1 · p. 3
• The figure’s full data-synthesis prompts are available in Appendix A.Source: Figure 1 · p. 3
• Candidate Mining is represented within the broader ReMixer workflow shown in Figure 1.Source: Figure 1 · p. 3

3 Data Synthesis: ReMixer

• The synthesis workflow is designed to ensure reasoning-demanding queries, diverse domains and forms, and accurately labeled positive and hard-negative relationships.These properties guide the generation, mining, and annotation stages of ReMixer.Source: §3 · p. 3

3.1 Query Generation

• The workflow generates reasoning-necessitated queries from knowledge-rich, domain-filtered corpora while excluding documents labeled positive in BRIGHT evaluation queries.Prompt variation in query length and user education level promotes diversity in form, style, and knowledge depth.Source: §3.1 · p. 3

3.2 Candidate Mining

• Candidate mining excludes each query’s source document and retrieves documents correlated in meaning but different in form, reducing trivial query-document matches.The candidates are mined with off-the-shelf retrievers from the corpus excluding the source document.Source: §3.2 · p. 3

3.3 Relevance Annotation

• ReMixer annotates mined candidates through query analysis, document analysis, and relevance judgment to identify positives and hard negatives.The annotation process uses reasoning-enhanced prompts and a distilled Qwen3-8B student annotator trained from Qwen3-235B-A22B-Instruct-2507 trajectories for cost-effective processing.Source: §3.3 · pp. 3–4

3.4 Synthesization Result

• ReMixer produces 81,659 valid queries from 95,960 generated queries, averaging 12 positive documents per query and 221 tokens per query.Queries without valid positive documents are filtered out, leaving rich and relatively long-form supervision for reasoning-intensive retrieval.Source: §3.4 · p. 4

4 Training Method: Redapter

• The training objective uses InfoNCE to distinguish one positive document from in-batch and synthetic hard-negative documents for each query.Similarity is computed by the dot product between query and document embeddings, with τ serving as the temperature parameter.Source: §4 · p. 4

4.1 Reasoning Intensity

• Synthetic samples vary in difficulty: easy relationships are captured through surface matching, whereas hard relationships require multi-step reasoning and more patient learning.This difficulty variation motivates reasoning-intensity estimation and adaptive weighting.Source: §4.1 · pp. 4–5
• Reasoning intensity measures how much reasoning-augmented queries improve discrimination between relevant and irrelevant documents relative to raw queries.A larger RIθ(s) means the reasoning-augmented query substantially reduces the InfoNCE loss compared with its non-reasoning counterpart.Source: §4.1 · p. 5

4.2 Self-Adaptive Learning

• The model is initialized from an MSMARCO-pretrained checkpoint and optimized on the synthetic dataset with RI-InfoNCE as the training objective.Higher reasoning-intensity samples receive greater weights during this self-adaptive training process.Source: §4.2 · p. 5
• Redapter dynamically weights synthetic training samples by estimated reasoning intensity, allocating more learning capacity to cases requiring deeper reasoning.Higher-intensity samples receive greater weights while simpler examples remain moderately represented, and the method does not significantly increase training computation cost.Source: §4.2 · p. 5

5.2 Main Results (RQ1-RQ3)

• 38.1 nDCG@10 is achieved by ReasonEmbed-Qwen3-8B on BRIGHT, while the Qwen3-4B variant reaches 37.1 and all variants outperform existing baselines.On R2MED, ReasonEmbed-Qwen3-8B reaches 43.18 nDCG@10, exceeding Qwen3-8B-ms at 24.15 and DIVER-Retriever at 32.23.Source: §5.2 · p. 6, Table 2 · p. 6, Table 3 · p. 7
• ReMixer-generated synthetic data improves ReasonEmbed over MSMARCO-finetuned counterparts and same-backbone general-purpose embedding baselines.The comparison includes Qwen3-4B, Qwen3-8B, and Llama-3.1-8B variants.Source: §5.2 · pp. 6–7
• Redapter-based variants consistently outperform conventional contrastive-learning counterparts, supporting adaptive weighting of reasoning-intensive samples.The method weights samples according to estimated reasoning intensity because lower-intensity samples saturate earlier.Source: §5.2 · p. 7

5.3 Annotation Quality Assessment (RQ4)

• 0.6001 average Cohen’s Kappa measures distilled-annotator agreement with the teacher, versus 0.4314 for the base model, a gain of 0.1687.The consistency evaluation uses 60,000 held-out samples.Source: §5.3 · pp. 7–8, Table 4 · p. 7
• 85.0% and 86.1% average annotation accuracy are obtained by the distilled annotator and teacher model, respectively, under human validation.The accuracy evaluation contains 360 held-out test samples and counts labels as accurate only when they match human judgment.Source: §5.3 · p. 8, Table 5 · p. 8

5.4.1 Candidate Mining

• 22.1 nDCG@10 is achieved by SOURCE-PRO, but NON-ANNO, SOURCE, and SOURCE-PRO remain below the default method at 16.1, 14.5, and 22.1.The ablation supports excluding source documents and annotating relevance when constructing synthetic samples.Source: §5.4.1 · p. 8, Table 6 · p. 8

5.4.2 Relevance Annotation

• 37.1 nDCG@10 with the distilled annotator exceeds ZERO-SHOT at 32.4 and NON-REASON at 35.0, supporting distillation and explicit reasoning in annotation.The results indicate that lightweight annotators benefit from teacher distillation and reasoning-enhanced relevance judgments.Source: §5.4.2 · p. 9, Table 8 · p. 9

5.4.3 Reasoning-Intensity

• Performance improves substantially as synthetic data grows from 10.2K to 81.6K samples, while GPT-4.1-mini produces the best reasoning-intensity estimates among tested reasoners.The reasoning outputs are generated offline, so they do not affect online training efficiency.Source: §5.4.4 · p. 9, §5.4.3 · p. 9, Figure 2 · p. 9

• ReMixer generates diverse, reasoning-intensive training samples, while Redapter dynamically weights them by reasoning intensity for more effective learning.Source: §6 · p. 9

• The synthesis method remains constrained by the reasoning capacity of its underlying LLMs.The authors suggest more powerful LLMs or crafted multi-agent workflows as possible remedies.Source: Limitations · p. 10
• ReasonEmbed primarily targets reasoning-intensive retrieval rather than combining reasoning-intensive and general retrieval data.The authors propose broader training data to improve practical applicability.Source: Limitations · p. 10
• The study is limited to domains covered by existing benchmarks, particularly BRIGHT, leaving broader-domain generalizability for future work.Source: Limitations · p. 10

Ethics Consideration

• The synthetic dataset may inherit biases, toxicity, and sensitive content from the LLMs and real-world corpora used during generation.Source: Ethics Consideration · p. 10

A.2 Candidate Mining

• Candidate mining uses gte-Qwen2-7B-instruct with top-100 retrieval, while relevance annotation uses a distilled lightweight reasoning annotator and explicit reasoning comparisons.Source: §A.2 · p. 13

A.4 Synthesization Result

• The synthetic-data pipeline reports no data leakage based on BM25 retrieval followed by weighted Jaccard overlap analysis on BRIGHT and R2MED test queries.Source: §A.4 · p. 14, Figure 3 · p. 14

B.3 Training Computation Cost Analysis

• Redapter increases ReasonEmbed-Qwen3-8B training runtime by only 234 seconds relative to the original contrastive-learning method.Source: §B.3 · p. 14, Table 11 · p. 14

C.1 Baselines

• The appendix provides evaluation settings, baseline information, and detailed ablation tables covering candidate mining, relevance annotation, reasoning-intensity computation, and training-data scaling.Source: §C.1 · p. 14, §C.2 · pp. 14–15, §C.3 · p. 15

• Synthetic data improves Qwen3-8B retrieval on TREC-COVID and KAPR, despite both tasks being non-reasoning-intensive.Performance consistently improves as the synthetic dataset scales.Source: §C.4 · p. 15, Table 12 · p. 15
• Synthetic data has a stronger impact on these traditional tasks than NQ, a popular question-answering retrieval dataset.Source: §C.4 · p. 15