Deepseek ai presents Codei/o: a novel approach that transforms reasoning patterns based on code into natural language formats to improve LLMS reasoning capabilities

Large language models (LLM) have advanced significantly in natural language processing, but reasoning remains a persistent challenge. While tasks such as mathematical problems and code generation benefit from structured training data, broader reasoning tasks, such as logical deduction, scientific inference and symbolic reasoning, based on scattered and fragmented data. Traditional approaches, such as the prediction continues in the code, often embed the signals of reasoning implicitly, which makes it difficult for the models to be generalized. Even text generation methods are still limited by specific syntax learning, which limits its applicability beyond programming tasks. A more structured approach is needed to expose fundamental patterns of reasoning while preserving logical rigor.

Deepseek ai Research presents Codei/oAn approach that converts reasoning based on the code into natural language. Transforming the unprocessed code into an entry-salaid prediction form and expressing reasoning steps through Thought Chain (COT)Codei/or allows the LLM to internalize basic reasoning processes such as Logical, transverse flow planning of the decision tree and modular decomposition. Unlike conventional methods, Codei/or separates the reasoning of the code syntax, allowing a broader applicability while maintaining the logical structure.

Technical description and benefits

Codei/or follows a structured data processing pipe:

1. Collection of unprocessed code files: More than 450k metal functions gathered, including algorithms repositories and educational programming data sets.
2. Data standardization: The compilation code was refined using Deepseek-V2.5, ensuring the clarity and compatibility of execution.
3. Generation of pairs of input-salida: The functions were executed with variable inputs to create structured training examples in various reasoning tasks.
4. Generate reasoning of the chain of thought: Using models such as Deepseek-V2.5, explanations of natural language were generated to provide structured reasoning.
5. Verification and refinement: The predictions were validated through the execution, with incorrect responses reviewed iteratively to improve the accuracy of the reasoning.

Key Codei/O:

• Transforming learning: Converts various code patterns into Natural Language Cot Presentationsmaking the reasoning transferable beyond the programming contexts.
• Syntax learning: Learning: Separates the logical reasoning of Code syntaxImprovement of adaptability in reasoning tasks.
• Multiple task improvement: Improves performance through Symbolic, scientific, logical, mathematical and common reasoning domains.
• Verifability: Predictions can be validated through Change of truth in cache of truth or reexecution.
• Iterative refinement: A refined version, Codei/O ++, uses multiple laps review To improve the accuracy of reasoning.

Empirical results and performance

The impact of Codei/or was tested in everything Four base models (which varies from 7b to 30b parameters) in 14 Reasoning Reference Points Covering logic, symbolic inference, mathematics, scientific deduction and common sense reasoning.

Recommendations:

• Consistent improvements: Codei training led to higher scores at reasoning reference points compared to traditional methods prior to height.
• Generalization in all tasks: Unlike existing approaches that improve specific tasks but degrade performance in other places, Codei/or showed balanced improvements.
• Comparison with baselines: Data sets surpassed by code/or as OpenMathinstruct2, OpenCoder-Sft-Stage1 and Webinstruct.
• Multiple refinement effectiveness: Codei/O ++ further improved the results refining the incorrect responses iteratively, taking advantage of execution feedback for better reasoning quality.

For example, in reference points of logical and symbolic reasoning as BBH and CruxevalCodei/or led to notable performance profits. In Mathematics Reasoning Tasks (GSM8K, Mathematics and MMLU-Stem)He showed improvements on existing baselines. Even in Common sense reasoningwhere code -based methods generally fight, Codei/or maintained solid results.

Conclusion

Codei/or has a structured way of improving the reasoning of the LLM by taking advantage of the input-to-salid transformations of the real world code. Instead of focusing on isolated reasoning tasks, it extracts patterns of universal reasoning and translates them into Natural language explanations. This structured learning approach guarantees that models acquire robust reasoning skills in different domains.

The introduction of Multiple laps review (Codei/O ++) In addition, it refines the precision of reasoning, which shows that the iterative learning of execution feedback improves the reliability of the model. Making predictions verifiableCodei/or provides a scalable and reliable method to improve LLM reasoning.

To join Natural language reasoning and codeCodei/or offers a promising direction to improve the cognitive skills of LLMS beyond programming tasks.

Verify he Paper and Github page. All credit for this investigation goes to the researchers of this project. In addition, feel free to follow us <a target="_blank" href="https://x.com/intent/follow?screen_name=marktechpost” target=”_blank” rel=”noreferrer noopener”>twitter And don't forget to join our 75K+ ml of submen.

Recommended open source ai platform: 'Intellagent is a framework of multiple open source agents to evaluate the conversational the complex system' _(Promoted)