Beyond Prompting: Resource-Efficient Explicit Control for CEFR-Aligned Language Generation

The integration of Large Language Models (LLMs) into educational technology and adaptive communication creates a need for reliable mechanisms for audience-aware text generation. In particular, aligning model outputs with the Common European Framework of Reference for Languages (CEFR) requires explicit control over linguistic complexity across six proficiency levels. Existing approaches involve important trade-offs: prompt-based control represents pro- ficiency only implicitly through natural-language instructions, full-model fine-tuning requires substantial computational resources, and decoding-time interventions such as Plug-and-Play Language Models (PPLM) introduce additional inference-time computation and can adversely affect generation quality. To address these limitations, this thesis proposes Pure PrefixMemory-Tuning (PMT) with Parametric Feature Gating, a parameter-efficient architecture for explicit CEFR-conditioned generation. Building on the PrefixMemory-Tuning principle of separating learned memory from the attention normalization, the proposed controller introduces a shared, full-rank layer- wise memory modulated by six learned CEFR embeddings through Hadamard feature gating. Applied to a frozen Llama-3.1-8B-Instruct backbone, the controller contains approximately 536.9 million trainable parameters, corresponding to approximately 6.69% of the backbone parameter count. The requested proficiency level is supplied through the learned internal con- trol representation rather than being explicitly stated as a CEFR label in the textual prompt. Empirical evaluation shows that the proposed CEFR-gated PMT substantially improves explicit proficiency control over the prompt-only Base LLM baseline, increasing in-domain CEFR accuracy from 24.50% to 68.80%, while achieving 81.34% adjacent accuracy, where predictions one CEFR level above or below the target are also counted as correct. On an out- of-domain benchmark constructed from unseen IELTS Writing Task 2 prompts, the model retains 58.33% accuracy, indicating that the learned control generalizes beyond the prompt dis- tribution of the EFCAMDAT learner corpus used for training. The improvement is obtained with only a small practical inference-time cost: the proposed model achieves 13.16 tokens/s compared with 13.54 tokens/s for the Base LLM, retaining approximately 97.2% of baseline generation throughput while avoiding the iterative gradient optimization required by PPLM. Controlled ablations further show that explicit CEFR conditioning is the principal contrib- utor to the observed controllability improvement. A parameter-matched CEFR-conditioned Prefix-Tuning model approaches the proposed PMT architecture closely, indicating that con- troller capacity also contributes substantially and that the remaining architectural advantage of PMT is comparatively modest in the evaluated setting. Across these experiments, the ob- served influence can be summarized as CEFR conditioning > controller capacity > adaptation architecture. Finally, a data-driven diagnostic investigates the persistent reduction in controllability at the C1 and C2 levels. The results show substantial distributional overlap among upper CEFR categories and considerably more limited advanced-level training data and topic diversity, par- ticularly for C2. These observations provide a plausible explanation for the upper-level perfor- mance ceiling, although they do not establish whether additional and more linguistically dis- tinctive high-proficiency data would fully resolve it or whether part of the limitation arises from the static control architecture itself. These findings motivate future work on richer advanced- level supervision, improved CEFR evaluation, and more expressive parameter-efficient control mechanisms.