ATTUNE: Adaptive Tuning for Understanding Neurodivergent Expression in LLM Interactions

Large Language Models are rapidly becoming core components of human-centric intelligent systems. However, their default conversational behaviours often fail to accommodate the heterogeneous cognitive and sensory needs of neurodivergent users, creating barriers to equitable humanmachine interaction. This paper presents ATTUNE, a proof-ofconcept profile conditioned fine-tuning framework that trains LLMs to adapt their conversational behaviour based on a structured Neurodivergent Communication Profile (NCP). The NCP schema comprises seven empirically grounded adaptation dimensions (p1–p7): literalness, verbosity tolerance, metaphor use, sensory load, interaction pacing, emotional explicitness, and autonomy preservation. We construct a training corpus of 1,447 instances by extracting NCP-annotated instructions from autistic community discourse on Reddit, generating profileconditioned assistant responses through a two-stage pipeline, and integrating the AUTALIC expert-annotated ableist language dataset. We fine-tune Gemma 4 E4B using QLoRA with profile-conditioned inputs. A four-condition ablation study demonstrates that NCP conditioning substantially increases communication style differentiation, that the anti-masking constraint reduces masking-associated language, and that finetuning improves autonomy preservation and response completeness beyond prompt engineering alone. Supplementary LLM-as-judge scoring across clarity, helpfulness, autonomy respect, and coherence provides additional evidence of profileappropriate adaptation.