ReMixer: Reasoning Data Synthesis

• ReMixer is a data synthesis pipeline that leverages LLM-generated queries and reasoning-enhanced candidate mining to create challenging training samples for complex document retrieval tasks.
• It employs a three-stage workflow—conditioned query generation, candidate mining with triviality elimination, and multi-step reasoning-based annotation—to enforce deep semantic relationships.
• The method produces an 82K-sample dataset that significantly boosts model performance on benchmarks like BRIGHT, facilitating non-trivial, reasoning-intensive text retrieval.

ReMixer is a data synthesis pipeline designed to generate high-quality, non-trivial, reasoning-intensive training samples for text embedding and retrieval models, specifically those targeting document retrieval tasks that demand sophisticated multi-step reasoning. Developed in the context of the ReasonEmbed embedding framework, ReMixer addresses the limitations of previous approaches that often produced synthetic datasets dominated by simple lexical overlap or superficial semantic links between queries and documents, a problem referred to as “triviality.” The workflow strategically constructs queries and document associations such that they require deep semantic and reasoning-based connections, resulting in data that better supports learning for complex question answering and retrieval.

1. Motivation and Problem Setting

In reasoning-intensive document retrieval, standard synthetic data generation methods tend to create “trivial” training samples—cases where the positive (relevant) document for a generated query is the very source that inspired the query, or one that closely paraphrases it. This enables models to succeed via pattern matching rather than true reasoning or semantic inference. ReMixer is explicitly designed to circumvent this triviality by enforcing that all positive document candidates for a given query are distinct from its “source” document and by applying a more rigorous, reasoning-based relevance annotation. The resulting training data is structurally richer and presents greater retrieval challenges.

2. Three-Stage Data Synthesis Workflow

ReMixer’s pipeline consists of three distinct phases:

1. Conditioned Query Generation: Utilizing knowledge-rich corpora, such as the datasets that comprise the BRIGHT benchmark (science, mathematics, programming, etc.), ReMixer uses LLMs to produce queries that inherently require deep reasoning. The prompt templates controlling the LLM include mechanisms for query length sampling and simulated user education level sampling, promoting both linguistic and cognitive diversity. This process yields queries that cannot be solved through simple lookup or paraphrase, but require logical inference or aggregation of information.
2. Candidate Mining with Triviality Elimination: To ensure that positive examples exhibit substantive, non-obvious associations to their queries, ReMixer removes the originating source document $d^*_q$ from consideration as a candidate for each generated query $q$. Instead, it constructs the candidate set by retrieving the top-k most relevant documents from the remainder of the corpus $D \setminus d^*_q$ using a similarity scoring function $\phi(q, d)$ and an off-the-shelf retriever. Formally:

$$\mathcal{C}_q \leftarrow \text{Top-}k\,\{\,\phi(q, d)\mid d\in (D \setminus d^*_q)\,\}$$

This exclusion mechanism compels the retrieval model to learn associations that go beyond mere phrase surface matching.

1. Reasoning-Enhanced Annotation: Each query–document candidate pair undergoes an in-depth multi-step annotation process using a distilled lightweight LLM. This process begins with a query analysis to map out the information need, followed by document analysis for conclusion supportability, and ends with assignment of a nuanced relevance label. The annotator LLM is trained via knowledge distillation from a more powerful teacher model, inheriting reasoning trajectories for better annotation fidelity. The objective is to assign “positive” status only to candidates that fulfill the reasoning required by the query—capturing true semantic alignment rather than proximity in text.

3. Impact on Data Quality and Model Training

The synthesis steps in ReMixer collectively yield an 82K-sample training set where positive pairs are specifically crafted to require multi-hop or complex semantic reasoning. This enables the downstream embedding model (ReasonEmbed) to learn mappings that capture subtle, context-rich semantic relationships, surpassing the capabilities of previous embeddings that relied on “easy” or lexically-biased synthetic pairs. The methodology contrasts with prior methods that linked every query to its source document or to documents with overlapping surface forms, which often restricted the learned model to shallow retrieval strategies.

4. Performance Outcomes and Benchmark Results

Empirical findings demonstrate that ReMixer’s approach leads to substantial gains on benchmarks that emphasize reasoning, such as BRIGHT. The ReasonEmbed model, trained with ReMixer-generated data, achieved a record-high nDCG@10 of 38.1, outperforming prior state-of-the-art text embedding models for reasoning-intensive retrieval. Comparative ablation studies confirm that when the triviality-mitigation and reasoning-intensive annotation are omitted, both data quality and retrieval performance degrade significantly. This suggests that ReMixer is instrumental in constructing challenging, generalization-friendly datasets for semantic retrieval tasks.

5. Technical Mechanisms and Formal Procedures

The key mechanics of ReMixer are formalized as follows:

Stage	Mechanism	Mathematical/Formal Representation
Query Generation	LLM prompt with length/education level sampling	—
Candidate Mining	Exclude $d^*_q$; use similarity $\phi(q, d)$ for Top-k	$$\mathcal{C}_q \leftarrow \text{Top-}k\,\{\,\phi(q, d)\,\}$$, $d\in (D \setminus d^*_q)$
Annotation	Reasoning-trajectory LLM assigns graded relevance	—

Each phase is designed to enforce compositional diversity and prevent easy mappings, thereby driving the embedding model to learn deeper, generalizable representations.

6. Comparison with Existing and Prior Approaches

Previous synthetic dataset construction methods often paired queries with their directly associated documents, resulting in straightforward—sometimes trivial—retrieval tasks. By contrast, ReMixer’s explicit exclusion of the query’s source document ensures that learning is predicated on recognition of reasoning-driven relationships rather than statistical co-occurrence. This design choice leads to stronger model performance as shown in benchmark testing, and to datasets that better reflect the real challenges of open-domain and multi-step reasoning retrieval scenarios.

7. Contributions and Future Implications

ReMixer contributes a generalizable methodology for generating non-trivial, reasoning-intensive synthetic data at scale. By releasing both its methodology and resources (for example, the 82K-sample dataset), it enables further research in robust embedding learning, evaluation, and deployment in complex retrieval settings. A plausible implication is that similar workflows, if adapted for other data modalities or task settings, may mitigate overfitting to proxy signals and promote abstract generalization in machine learning models. ReMixer’s paradigm is broadly applicable to tasks where the traditional proximity-based positive sampling is insufficient or actively harmful to downstream performance.

---

ReMixer thus provides a robust solution to data triviality and supports high-fidelity learning in reasoning-centric text retrieval, as rigorously demonstrated in the ReasonEmbed framework and the BRIGHT benchmark context (Chen et al., 9 Oct 2025).