CoTu at EXACT 2026: Neuro-Symbolic Reasoning for Transparent Educational QA

Research Question and Approach

The EXACT 2026 challenge highlights a critical gap in current AI: while large language models (LLMs) excel at natural language, they often struggle with the transparency and reliability required for educational settings. Team CoTu addressed this by developing a neuro-symbolic pipeline designed to operate within strict 8B parameter and 60-second latency constraints. The core research question was how to leverage small, open-weight models to perform reliable, verifiable reasoning on academic regulations and multi-step physics problems without resorting to black-box generation.

To solve this, the team implemented a Program-of-Thought (PoT) architecture. Instead of predicting an answer directly, the model acts as a programmer. For logical regulation queries, it generates Z3 constraint satisfaction code; for physics problems, it generates executable Python code. This approach grounds the final output in deterministic symbolic execution, effectively separating the linguistic task of premise selection from the logical task of deduction.

System Architecture

The CoTu pipeline utilizes a shared execution stack built on DSPy, which manages orchestration, serving, and output formatting. Key technical components include:

• Task-Specific Prompting: The model receives specialized guidance for each domain, such as Z3 API syntax for logic and SI-unit conventions for physics.
• Self-Correction Loop: If the generated code fails to execute or produces an error, the system automatically triggers a regeneration loop to refine the output.
• High-Throughput Serving: To meet the 60-second time limit, the team utilized SGLang with RadixAttention for prefix caching and DFLASH speculative decoding, allowing for faster inference without sacrificing the reasoning quality of the 4B backbone.

Findings and Impact

The system demonstrated the viability of hybrid neuro-symbolic reasoning at the 4B scale. CoTu achieved a perfect score on the physics task and secured the highest final-round technical score (13.44/15) in the competition. The results suggest that for structured educational tasks, the primary bottleneck for small models is not the deduction itself—which symbolic solvers handle reliably—but rather the accurate selection and translation of premises into formal logic. By offloading the "thinking" to a symbolic solver, the system provides a transparent, verifiable rationale for every answer, making it significantly more suitable for pedagogical applications than standard neural models.