Automated Scoring of college-English courses has gradually shifted from single-item essay evaluation to an overall evaluation system for Writing, Speaking, Translation, Reading Response, etc. Propose an automatic grading system based on generative artificial intelligence for multiple task-based college English courses. Combining task-conditioned language representation, rubric-aware generative reasoning, trait-level score fusion, ordinal score constraints, calibration correction, and uncertainty-based routing in the algorithm for human review. Organized a collection of 18,420 college English Responses from the following four instruction task types: 6,240 writing Samples, 4,180 Speaking Samples, 3,620 Translation Samples, and 4,380 Reading-Response Samples. For each of them, two trained raters scored according to the analytic rubric; their adjudicated labels were used for training and testing the scoring model. In the experiment, compared GAEAS with SVR, BiLSTM, BERT, T5, and GPT-4scoring in the same data split, scale of scores, and test criterion. In terms of all tasks, GAEAS achieved an approximation of the median QUCKW1=0. 903 and averaged a Pearson’s Correlation Coefficient close to 0. 923; For the average value, it is roughly within the range of-47dBto~5dBandtheRMSEisappoximately±_QEUZS=-5%. GAEAS outperformed GPT-4o in terms of QWK score, achieved a reduction in RMSE and improved generalization performance (ECE) compared to the baseline; Inference time was reduced to 96ms per example. According to the score-band heatmap, most of these task-profanity combinations had a diagonal agreement rate in the range of 72%-84%. Through ablation experiments, it was confirmed that when using the combinations of ordinal loss, trait fusion, rubric verification and uncertainty routing simultaneously would yield optimal results. Further revealing through a three-dimensional hyperparameter response Surface shows that the prior-rubric weighing and calibration weights need to be changed together; The weakest validation Q-κ values are around 0.64 and 0.36, respectively. According to the results, it was found that with proper management of both Generative AI Scoring stability in College English Teaching Rubric Structure, Score Ordering Calibration and Human Review can all work together. The research also presents various results, including score-band alignment, component-level ablations, error-source distribution, and deployment trade-off situations of the algorithm to provide teaching tools instead of isolated models. According to the outcomes, automatic grading in college English needs to consider three dimensions: agreement, error magnitude, calibration costs, and post-evaluation expenses comprehensively. On the basis of this, a particular way for quick feedback along with teacher oversight remains as an option for some courses. Reproducibility of the protocol using fixed data splits and figure-level metric tables. This experiment is presented as a classroom-based empirical study rather than a review assertion; it uses the same held-out test set, fixes baselines, performs an explicit score-band analysis, and simulates a review threshold.
