INF-X-Retriever

INF-X-Retriever

A Pragmatic, Production-Grade Framework for Reasoning-Intensive Retrieval

INF-X-Retriever is a high-performance dense retrieval framework developed by INF.
It provides a robust solution for complex reasoning-intensive tasks across a collection of tasks (X) with minimal supervision, prioritizing architectural simplicity, deployment reliability, and deep semantic alignment.

Introduction • Design Principles • Architecture • Performance • Models • Citation • Contact

---

📖 Introduction

The evolution of Large Language Models (LLMs) has redefined information retrieval, shifting the paradigm from surface-level keyword matching to intent-aware reasoning. Modern queries in RAG (Retrieval-Augmented Generation) pipelines often contain intricate narrative contexts, logical constraints, and domain-specific directives—elements that act as “semantic noise” for conventional retrieval systems.

INF-X-Retriever addresses this challenge through Intent Distillation. By aligning complex queries into a unified semantic space and executing single-stage dense retrieval, it effectively penetrates surface-level complexity to reach core information. Our approach is validated by its 1st-place ranking on the BRIGHT Benchmark, a rigorous evaluation suite dedicated to reasoning-heavy retrieval scenarios.

---

💡 Design Principles

Our framework is guided by engineering practicality and first-principles reasoning. We intentionally avoid architectural bloat in favor of production readiness and computational efficiency.

🎯 Core Principle: “Less is More” — Maximal efficacy through deliberate minimalism.

▫️ Single-Stage Retrieval (No Rerankers)

While reranking stages can marginally improve accuracy, they introduce significant latency and operational overhead. In modern RAG pipelines, downstream LLMs already perform implicit context discrimination during synthesis. We demonstrate that a high-fidelity single-stage retriever provides the optimal balance of precision and throughput, significantly simplifying deployment and monitoring in production environments.

▫️ Direct Query Alignment (No HyDE)

Hypothetical Document Embeddings (HyDE) rely on LLMs to generate “pseudo-answers” to guide retrieval. This introduces two critical failure modes:

1. Redundancy: If the LLM already knows the answer, retrieval is redundant.
2. Hallucination Drift: When the LLM lacks domain knowledge (the primary use case for RAG), hypothetical answers can steer the retriever toward misleading content. INF-X-Retriever performs Direct Intent Extraction, ensuring retrieval remains grounded in the actual user requirements and source evidence.

▫️ Operational Simplicity

We eliminate components that introduce fragility or hyperparameter sensitivity:

• No Sparse Retrieval (BM25): Removes the complexity of hybrid fusion and weight tuning.
• No Multi-Query Expansion: Single-pass alignment minimizes tail latency.
• No Model Ensembles: Ensures a transparent, maintainable, and diagnostic-friendly pipeline.

---

🛠️ Architecture

The framework consists of two tightly integrated, purpose-built components:

1. Query Aligner

• Model: 🤗 inf-query-aligner
• Backbone: Qwen2.5-7B-Instruct (Fine-tuned via Reinforcement Learning)
• Function: Distills core retrieval intent from verbally complex or noisy queries. It ensures the query embedding is focused on the underlying task rather than linguistic artifacts.

2. Retriever

• Model: 🤗 inf-retriever-v1-pro
• Backbone: Continual training on inf-retriever-v1 with targeted long-query adaptation.
• Function: A generalized dense retrieval architecture optimized for cross-task transfer, stability, and high-dimensional semantic mapping.

---

📊 Performance

INF-X-Retriever achieves state-of-the-art results on the BRIGHT Benchmark (as of Dec 20, 2025).

The BRIGHT (Benchmark for Reasoning-Intensive Grounded HT) is a rigorous text retrieval benchmark designed to evaluate the capability of retrieval models in handling questions that require intensive reasoning and cross-document synthesis. Collected from real-world sources such as StackExchange, competitive programming platforms, and mathematical competitions, it comprises complex queries spanning diverse domains like mathematics, coding, biology, economics, and robotics.

Why BRIGHT Matters:

• High Reasoning Complexity: Unlike traditional keyword-centric benchmarks, BRIGHT queries often demand multi-step reasoning, evidence aggregation across documents, and theoretical mapping. This effectively exposes the limitations of standard models in complex “understanding + retrieval” tasks.
• Authentic & Interdisciplinary: By leveraging real-world data from varied professional fields, BRIGHT provides a faithful assessment of a model’s generalization capabilities in specialized, high-stakes environments.
• Critical for RAG: As a stress test for modern Retrieval-Augmented Generation (RAG) systems, it serves as a key indicator for performance in demanding industrial applications such as scientific research, legal analysis, and medical Q&A.

Short document

Overall & Category Performance

Model	Avg ALL	StackExchange	Coding	Theorem-based
INF-X-Retriever	63.4	68.3	55.3	57.7
DIVER (v3)	46.8	51.8	39.9	39.7
BGE-Reasoner-0928	46.4	52.0	35.3	40.7
LATTICE	42.1	51.6	26.9	30.0
ReasonRank	40.8	46.9	27.6	35.5
XDR2	40.3	47.1	28.5	32.1

Detailed Results Across 12 Datasets

Model	Avg	Bio.	Earth.	Econ.	Psy.	Rob.	Stack.	Sus.	Leet.	Pony	AoPS	TheoQ.	TheoT.
INF-X-Retriever	63.4	79.8	70.9	69.9	73.3	57.7	64.3	61.9	56.1	54.5	51.9	53.1	67.9
DIVER (v3)	46.8	66.0	63.7	42.4	55.0	40.6	44.7	50.4	32.5	47.3	17.2	46.4	55.6
BGE-Reasoner-0928	46.4	68.5	66.4	40.6	53.1	43.2	44.1	47.8	29.0	41.6	17.2	46.5	58.4
LATTICE	42.1	64.4	62.4	45.4	57.4	47.6	37.6	46.4	19.9	34.0	12.0	30.1	47.8
ReasonRank	40.8	62.7	55.5	36.7	54.6	35.7	38.0	44.8	29.5	25.6	14.4	42.0	50.1
XDR2	40.3	63.1	55.4	38.5	52.9	37.1	38.2	44.6	21.9	35.0	15.7	34.4	46.2

Long document

Detailed Results Across 8 Datasets

Model	Avg	Bio.	Earth.	Econ.	Pony	Psy.	Rob.	Stack.	Sus.
INF-X-Retriever	54.6	73.2	59.6	69.3	12.1	74.3	55.9	27.8	64.8
inf-retriever-v1-pro	30.5	44.1	42.2	31.4	0.4	43.1	20.8	21.4	41.0

Notes:

• Results reflect end-to-end retrieval accuracy on BRIGHT under the official evaluation protocol.
• Performance may vary with hardware, index size, and dataset versions.

---

🧪 Models

Both models are released under the Apache-2.0 License for both research and commercial application.

Component	Hugging Face Repository	Description
Query Aligner	inf-query-aligner	LLM-based intent distiller.
Retriever	inf-retriever-v1-pro	Advanced dense embedding model.

---

📝 Citation

If you utilize INF-X-Retriever in your research or production systems, please cite our work:

@misc{inf-x-retriever-2025,
    title        = {INF-X-Retriever: A Pragmatic Framework for Reasoning-Intensive Dense Retrieval},
    author       = {Yichen Yao, Jiahe Wan, Yuxin Hong, Mengna Zhang, Junhan Yang, Zhouyu Jiang, Qing Xu, Kuan Lu, Yinghui Xu, Wei Chu, Emma Wang, Yuan Qi},
    year         = {2025},
    url          = {[https://github.com/yaoyichen/INF-X-Retriever](https://github.com/yaoyichen/INF-X-Retriever)},
    publisher    = {GitHub repository}
}

---

📬 Contact

We welcome inquiries regarding technical deep-dives, strategic collaborations, or large-scale deployment support. Our team is committed to advancing the boundaries of reasoning-intensive retrieval.

• Contact: eason.yyc@inftech.ai
• Organization: INF
• Issue Tracking: For bug reports or feature requests, please use the GitHub Issues portal.

Built with precision by the INF Team.