Snowflake proposes excot: a new frame of AI that iteratively optimizes open source llm combining COT reasoning with DPO out of politics and politics, trusting only in the precision of execution as feedback

SQL text translation, the task of transforming natural language consultations into structured SQL statements is essential to facilitate the interactions of the easy -to -use database. However, the task implies significant complexities, in particular the linking of the scheme, the management of the compositional SQL syntax and the resolution of ambiguities in user consultations. While large language models (LLM) have shown robust capabilities in several domains, the efficacy of structured reasoning techniques, such as the thinking chain (COT) within SQL text contexts, remains limited. The previous attempts to use COT zero or direct preference optimization (DPO) without structured reasoning threw marginal improvements, indicating the need for more rigorous methodologies.

Snowflake introduces excot, a structured frame designed to optimize open source LLM through the COT reasoning combination and iterative preferences optimization, specifically using the DPO outside the policy and policy guided exclusively by the feedback of the demand for execution. Excot dispenses with external rewards models and human annotations, trusting in its place in internally generated reasoning steps and execution results. The method works in two main phases: Initially, it generates cot data validated COT through DPO out of politics, forming the basis for the fine supervised adjustment. Subsequently, the model generates and refines COT data through DPO in politics, incrementally improving precision through feedback derived from the correction of execution.

Excot uses COT detailed reasoning, particularly adopting a division and conquest strategy in which complex consultations are broken down into simpler subway. Each subcontrol is analyzed and resolved independently before integrating into a coherent final consultation. This structured decomposition allows the model to manage complexity and common nested structures in SQL operations more effectively. Execution -based verification serves as the central mechanism for correction evaluation, where the consultations generated are validated when comparing their execution outputs with truth results by land. Incorrect and correct consultations are systematically matched, providing explicit signals for preferences -based learning. Iterative refinement in the DPO phase in politics progressively improves the precision of reasoning of the model.

Excot's experimental evaluation demonstrated significant improvements in the precision of execution. Specifically, with the call-3.1 70b model, the accuracy of high execution of excot in the bird development set of 57.37% to 68.51%, and an increase in the performance of the spider test set from 78.81% to 86.59%. Performance improvements comparable to the 32B model of QWEN-2.5 encoders were recorded. These results position the excotor as a main approach in the evaluations of a single model for these reference points, exceeding the methods established as Xiyansql and patented models, including the OpenAI variants. In particular, the improvements consistently maintained high rates of validity of consultations (exceeding 98%), confirming improvements in semantic correction together with syntactic precision.

In conclusion, Excot represents a methodical advance in structured reasoning optimization for open source LLM applied to text tasks to SQL. By integrating the reasoning of COT structured with preferences optimization, guided only by execution -based feedback, excot effectively addresses the limitations identified in previous methods. Its iterative refinement capacity ensures continuous improvement without dependence on external reward structures or manual annotations. Additional research could explore this framework to more intricate scheme environments and additional structured reasoning tasks, thus expanding the applicability and reliability of LLMs in contexts of generating structured consultations.

Verify he Paper, Github page and Details. 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 85k+ ml of submen.

(Register now) Virtual Minicon Conference on open source ai: Free Registration + Assistance Certificate + Short Event of 3 Hours (April 12, 9 AM- 12 PM PST) + Hands on Workshop (sponsored)