Quantitative Benchmarking of Large Language Models for Biochemical Pattern Recognition and Evidence-Grounded Interpretation Using Strawberry and Blackberry Response Datasets

Large language models (LLMs) are increasingly used for scientific information processing, yet their ability to perform experimentally grounded biochemical reasoning remains insufficiently characterized. This study presents a quantitative benchmark of biochemical pattern recognition and evidence-grounded interpretation using strawberry and blackberry response datasets. A canonical benchmark comprising 220 tasks was constructed from experimentally derived biochemical ground truth and applied identically to five instruction-tuned open-weight LLMs, yielding 1,100 model–task evaluations. Tasks covered profile-level interpretation, fruit-level comparisons, class-level reasoning, numerical estimation, concentration–activity relationships, paired numerical retrieval, and categorical response-pattern recognition. Responses were evaluated using predefined task-specific criteria and classified into three mutually exclusive final states: CORRECT, INCORRECT, or REVIEW_REQUIRED. A complementary response-level forensic analysis characterized numerical grounding and unsupported quantitative content. Across the complete benchmark, 6 of 1,100 responses were classified as correct, 371 as incorrect, and 723 as review-required. Thus, 377 evaluations (34.27%) reached a definitive scoring outcome, corresponding to an observed accuracy of 0.55% across all evaluations and a scorable accuracy of 1.59% among definitively scored responses. Model-level definitive scoring coverage ranged from 20.00% to 55.91%, indicating substantial variation in response scoreability in addition to differences in correctness. Categorical pattern-recognition tasks and numerical-grounding analyses provided complementary evidence on response behavior, including incomplete, unsupported, and non-committal outputs. The benchmark demonstrates the importance of separating biochemical correctness from response scoreability when evaluating LLMs on experimentally grounded scientific tasks. The proposed framework provides a reproducible approach for quantifying biochemical reasoning performance while preserving unresolved responses as a distinct analytical outcome rather than treating them automatically as incorrect.