Improving Search Query Rewriting Quality with Small Language Models (SLMs) and Recurrent Reasoning
Developing a novel architecture at the intersection of information retrieval and language model optimization to better understand and generate queries through sequential computational cycles
Necessity and Motivation of the Research
In search and information retrieval systems, users often express their needs in the form of short, ambiguous, or imprecisely structured phrases. The issue of Query Rewriting is a vital pillar in modern search engines, tasked with understanding the user’s true intent and transforming the raw query into a richer, more accurate phrase compatible with the system’s indexing structure.
With the advent of language models, the query rewriting approach has shifted from statistical rule-based methods to generative and semantic methods. However, using Large Language Models (LLMs) in real-time search systems faces severe limitations regarding latency and computational costs. On the other hand, Small Language Models (SLMs), although optimal in terms of speed and cost, struggle with understanding semantic nuances and multi-step reasoning for complex rewrites.
Research Problem and Existing Gap
The main gap in this field is finding the balance between “processing efficiency” and “semantic reasoning power”. Parameter-efficient models (SLMs) can execute quickly but cannot perform deep reasoning on user intent like larger models.
Can Recurrent Reasoning mechanisms, which have previously been successful in solving structured problems, help small language models in text generation tasks to improve the quality of query rewriting without significantly increasing parameters?
This research seeks to demonstrate whether it is possible to increase the reasoning depth of the model by creating repeatable computational cycles, instead of increasing the dimensions of the neural network, and to produce outputs competitive with larger models.
Proposed Architecture and Approach
In this research, a hybrid architecture based on integrating a Small Language Model (SLM) with a recurrent, parameter-efficient reasoning module (similar to the TRM architecture) is proposed. The goal of this module is to transform the rewriting process from a single-step operation into a gradual refinement process.
Implementation and Execution Stages of the Architecture:
Evaluation Process and Expected Achievements
To evaluate this architecture, the model’s outputs in information retrieval systems are examined. It is expected that this architecture can enhance the semantic reasoning capability of the model while maintaining its lightweight advantage.
| Language Model Type | Computational Cost / Speed | Reasoning Power (Expected) | Deployment Complexity |
|---|---|---|---|
| Large Language Models (LLM) | Very High / Slow | Very High | Requires Advanced Hardware |
| Traditional Small Models (Base SLM) | Very Low / Real-time | Limited and Superficial | Easy and Low-Cost |
| Proposed Model (SLM + Recurrent) | Low / Fast | Deep and Gradual | Easy and Compatible with Current Systems |
This approach shows that by intelligently redesigning the information flow path in the neural network, we can significantly improve text generation quality in small models.
Significance and Key Dimensions of the Research
This research is at the intersection of two major fields of computer science, and its achievements can be examined in three main dimensions:
Information Retrieval (IR)
Providing a practical solution to better understand user intent and significantly improve search result quality by semantically refining and expanding raw queries.
Model Architecture Optimization
Proving that increasing AI capabilities does not necessarily depend on an exponential increase in parameters, and recurrent reasoning patterns will be a breakthrough.
Text Generation and Language Processing
Providing a new perspective on using gradual reasoning in text generation tasks, which can be applied beyond queries in other NLP domains.
Conclusion
This research, by introducing an innovative framework, strives to push the boundaries of Small Language Models’ performance in understanding and generation tasks. By integrating a recurrent, parameter-efficient reasoning module with a Small Language Model (SLM), the query rewriting process transforms into a step-by-step, intelligent refinement procedure. The results of this study not only provide an economical and fast solution to improve search engines but also open new horizons for researchers to empower parameter-efficient models through structured reasoning patterns.
