From Synthesis to Clinical Assistance: A Strategy-Aware Agent Framework for Autism Intervention based on Real Clinical Dataset

Junhong Lai; Shuzhong Lai; Yanhao Yu; Wanlin Chen; Chenyu Yan; Haifeng Li; Lin Yao; Yueming Wang

cs.LGApr 9, 2026

From Synthesis to Clinical Assistance: A Strategy-Aware Agent Framework for Autism Intervention based on Real Clinical Dataset

Authors: Junhong Lai, Shuzhong Lai, Yanhao Yu, Wanlin Chen, Chenyu Yan, Haifeng Li, Lin Yao, Yueming Wang

Organizations: MOE Frontiers Science Center for Brain and Brain-Machine Integration, Zhejiang University · Nanhu Brain-Computer Interface Institute · College of Computer Science and Technology, Zhejiang University · Polytechnic Institute, Zhejiang University · School of Medicine, Hangzhou City University · Children’s Hospital Zhejiang University School of Medicine · Department of Neurobiology, Affiliated Mental Health Center and Hangzhou Seventh People’s Hospital, Zhejiang University School of Medicine

Abstract

The development of AI-assisted Early Intensive Behavioral Intervention (EIBI) for Autism Spectrum Disorder (ASD) is severely constrained by data scarcity. Furthermore, while Applied Behavior Analysis (ABA) serves as the gold standard for clinical intervention, general-purpose Large Language Models (LLMs) struggle to strictly adhere to its standardized procedures, often resulting in interactions that are linguistically fluent but strategically inconsistent. To address these challenges, we introduce ASDAgent, a strategy-aware framework designed to unify high-fidelity intervention dialogue synthesis and clinical decision support. ASDAgent incorporates two specialized components to solve distinct problems: (i) a DoctorAgent equipped with an Observe-Think-Act-Correct (O-T-A-C) reasoning loop, which resolves the issue of strategy collapse in LLMs by making ABA execution explicit and controllable; and (ii) a ChildAgent that utilizes probabilistic behavior modeling to mitigate data homogeneity, simulating diverse and non-deterministic ASD response patterns. Experiments demonstrate that dialogues generated by ASDAgent closely mirror the strategy distribution of human therapists (KL divergence: 0.083). In real autism intervention, ASDAgent achieves nearly 80% strategic consistency with human experts. Moreover, we show that synthetic data produced by ASDAgent effectively distills professional clinical knowledge into small language models (SLMs), significantly enhancing their therapeutic capabilities.

Figures & tables

Figure 1: An example of DoctorAgent performing Observe-Think-Action-Correct. In the Observe phase, DoctorAgent categorizes and interprets the child’s responses. In the Think phase, DoctorAgent performs iterative, multi-round reasoning to determine appropriate intervention strategies based on the observed information. After each Think step, DoctorAgent immediately enters the Act and Correct phase, generating a concrete response that executes the selected strategy. This Think–Act-Correct loop may repeat multiple times within a single dialogue turn until an appropriate intervention is completed.

Figure 2: An overview of our framework. ASDAgent , for both Dialogue Synthesis and Real Autism Intervention.

Doctor	Child	Doctor Strategy		Child Response
		KL $\downarrow$	JS $\downarrow$	KL $\downarrow$	JS $\downarrow$
DoctorAgent	ChildAgent	0.083	0.019	0.007	0.002
DoctorAgent	GPT-4o	0.325	0.072	0.039	0.009
GPT-4o	ChildAgent	0.259	0.118	0.024	0.006

Table 1: KL and JS Divergence to Real Distribution for Doctor Strategies and Child Responses.

Figure 3: Human and LLM-based Preference Evaluation between Real Data and Synthetic data.

Figure 4: Human and LLM-based Scoring between Real Data and Synthetic data.

Figure 5: Evaluation on Real Intervention Dialogues. The left-hand graph shows semantic metrics, and the right-hand graph shows strategy temporal consistency.

Figure 6: Win–Tie–Lose Comparison Between Model-Generated and Human Doctor Responses Across Different Evaluators.

Model	Correct Times	Total	Trigger Rate (%)
GPT-4o-mini	$339$	$1208$	$\mathbf{28.06}$
GPT-4o	$266$	$1120$	$\mathbf{23.75}$

Table 2: Statistics on the triggering of the Correct phase during real-world clinical interventions.

Figure 7: Training dynamics of Qwen3-4B during Supervised Fine-Tuning. (Left) Training loss convergence and (Right) mean token accuracy curves across different data sources. The x-axis represents the number of training epochs.

Model	Training Data	BLEU $\uparrow$	GLEU $\uparrow$	METEOR $\uparrow$	BERTScore (F1) $\uparrow$	BGE $\uparrow$
Qwen3-4B-Instruct	Base	8.04	11.24	36.44	87.03	74.71
	Common	12.46	14.82	32.32	88.42	74.14
	ASDAgent	12.59	14.98	33.21	88.60	73.94
	Real	14.13	16.07	34.75	88.59	74.97
	Common+Real	14.13	16.43	34.88	88.55	75.20
	ASDAgent+Real	14.47	16.33	35.24	88.71	75.40

Table 3: Performance comparison of Small Language Models (SLMs) fine-tuned on different datasets. Base : Zero-shot performance. Common : SFT on GPT-4o synthesized data. ASDAgent : SFT on our synthetic data. Real : SFT on real clinical data. For each model, the best result is highlighted in bold , and the second best is underlined .

Model	Training Data	Multi-P $\uparrow$	Multi-R $\uparrow$	Multi-F1 $\uparrow$	LCS-P $\uparrow$	LCS-R $\uparrow$	LCS-F1 $\uparrow$
Qwen3-4B-Instruct	Base	47.73	68.45	52.91	47.63	68.28	52.78
	Common	60.44	69.24	62.79	60.44	69.24	62.79
	ASDAgent	65.68	70.86	66.04	65.68	70.86	66.04
	Real	67.68	74.75	69.01	67.68	74.75	69.01
	Common+Real	66.23	73.77	67.82	66.23	73.77	67.82
	ASDAgent+Real	67.16	74.04	68.32	67.16	74.04	68.32

Table 4: Strategy Alignment Analysis on Strategy Consistency Metrics (in %). We evaluate the alignment of fine-tuned models against the ground truth strategies using Multiset (Strategy Selection) and LCS (Temporal Consistency) metrics. Base : Zero-shot baseline. Common : SFT on GPT-4o synthesized data. ASDAgent : SFT on our synthetic data. Real : SFT on real clinical data. For each model, the best result is highlighted in bold , and the second best is underlined .

Model	Method	Total Time (s) $\downarrow$	Avg Time (s) $\downarrow$	BLEU $\uparrow$	BERTScore F1 $\uparrow$	BGE $\uparrow$	Dist-2 $\uparrow$	Multi-F1 $\uparrow$	LCS-F1 $\uparrow$
GPT-4o-mini	Ours	$\mathbf{23.89_{\pm 13.84}}$	$\mathbf{11.88_{\pm 5.02}}$	$7.39$	$88.05$	$72.83$	$92.49$	$\mathbf{70.47}$	$\mathbf{70.38}$
GPT-4o-mini	ToT	$60.24_{\pm 31.43}$	$25.86_{\pm 13.03}$	$\mathbf{9.94}$	$\mathbf{89.54}$	$\mathbf{78.82}$	$\mathbf{95.93}$	$56.11$	$56.03$
GPT-4o	Ours	$\mathbf{34.47_{\pm 19.49}}$	$\mathbf{17.58_{\pm 10.12}}$	$8.29$	$88.60$	$73.77$	$\mathbf{94.54}$	$\mathbf{72.95}$	$\mathbf{72.95}$
GPT-4o	ToT	$70.29_{\pm 42.19}$	$23.27_{\pm 10.81}$	$\mathbf{11.88}$	$\mathbf{89.41}$	$\mathbf{78.90}$	$93.80$	$55.72$	$55.56$

Table 5: Performance and latency comparison between ToT and O-T-A-C. While ToT exhibits higher textual diversity, it severely fails in strictly adhering to clinical ABA strategies (Multi-F1 and LCS-F1) and introduces unacceptable latency compared to our framework.

Appendix figures & tables47 assets

Supplementary material from the paper’s appendix.

Appendix

Panel of Figure 8 — Figure 8: Case Study in Dialogue Synthesis with Topic "Buy fruit" and the same Turns. The left side shows a real intervention dialogue, while the right side shows a synthetic intervention dialogue. Blue indicates Instruction, green denotes Assistance including Half-Assistance and Full-Assistance, yellow signifies Reinforcement and Acknowledgement, and red represents the child’s responses.

Figure 9: Case Study in Real Autism Intervention. The diagram above illustrates the intervention responses of DoctorAgent , the real Doctor, and other models based on a realistic intervention dialogue. The diagram below shows how DoctorAgent completes the O-A-T-C process.

Figure 10: Qualitative comparison between our O-T-A-C framework and the Tree-of-Thought (ToT) baseline. While ToT exhibits conversational fluency, it critically fails in clinical adherence by introducing strategy deviation (instruction stacking) and topic drift. Our framework safely guides the child back to the intended topic.

Figure 11: Qualitative case studies demonstrating the efficacy of the Correct module. The module actively rectifies strategy deviation by preventing topic drift (left) and resolves instruction stacking by streamlining complex, overloaded questions into clinically appropriate atomic instructions (right).

Figure 12: The standard workflow of Discrete Trial Training (DTT) derived from ABA literature PACKET () , illustrating the structured interaction cycle. Doctors can adjust their treatment strategies as needed, based on the actual intervention situation.

Figure 13: Topic distribution of ASDAgent-Dataset-Golden.

Item	Value
Total number of dialogues	2071
High-quality dialogues	764
Total transcription cost (USD)	5,204.17
Average cost per dialogue (USD)	2.51

Appendix

Table 6: Overview of Dialogue Transcription Cost (USD)

Figure 14: t-SNE scatter plot of macro topics across 10 main conversational categories.

Doctor Strategy	Relevant	Irrelevant	Unresponsive	Repetition
Instruction	59.71	23.39	14.35	2.55
Full-Assistance	53.39	21.71	9.56	15.34
Half-Assistance	53.99	26.14	14.45	5.42
Reinforcement	64.11	20.21	3.14	12.54
Other	52.24	22.39	7.46	17.91

Appendix

Table 7: Conditional probabilities (%) of child response types given the last doctor intervention strategy.

Next Speaker	Probability (%)
Child	8.67
Doctor	91.33

Appendix

Table 8: Turn interruption probabilities following the current dialogue turn.

Gender	Number	Age (Mean ± std)	Language Development Age (Mean ± std)
Male	65	$5.35_{\pm 1.26}$	$3.74_{\pm 1.16}$
Female	18	$5.42_{\pm 1.33}$	$3.87_{\pm 1.13}$

Appendix

Table 9: The demographic details of children for ASDAgent-Dataset-Golden.

Figure 15: Distribution of Key Metrics by Child Category

Figure 16: t-SNE Clustering of Child Profiles

Metric	Length
Turns per Dialogue	$18.61_{\pm 11.39}$
Chars per Doctor Utterance	$22.35_{\pm 12.20}$
Chars per Child Utterance	$5.52_{\pm 7.61}$

Appendix

Table 10: Dialogue Basic Statistics

Category	Subtype	Length
Doctor	Instruction	$20.77_{\pm 12.01}$
	Reinforcement	$22.13_{\pm 11.67}$
	Half-Assistance	$27.67_{\pm 12.03}$
	Full-Assistance	$28.08_{\pm 13.63}$
	Other	$20.08_{\pm 10.37}$
Child	Relevant	$6.56_{\pm 8.80}$

Appendix

Table 11: Utterance Length Statistics by Strategy and Response Type

Figure 17: Distribution of Conversation Turns (Log-Normal Fit)

Model	Metric	Observe	Think	Act	Correct	Overall Total
GPT-4o-mini	Total Time (s)	$3.65_{\pm 1.58}$	$11.86_{\pm 3.54}$	$3.66_{\pm 4.01}$	$4.72_{\pm 12.72}$	$\mathbf{23.89_{\pm 13.84}}$
GPT-4o-mini	Avg Time (s)	$3.65_{\pm 1.58}$	$4.20_{\pm 1.10}$	$1.83_{\pm 1.93}$	$2.20_{\pm 4.21}$	$\mathbf{11.88_{\pm 5.02}}$
GPT-4o	Total Time (s)	$5.39_{\pm 5.13}$	$19.29_{\pm 13.99}$	$5.38_{\pm 10.30}$	$4.41_{\pm 7.19}$	$\mathbf{34.47_{\pm 19.49}}$
GPT-4o	Avg Time (s)	$5.39_{\pm 5.13}$	$6.81_{\pm 4.78}$	$2.93_{\pm 5.56}$	$2.45_{\pm 4.72}$	$\mathbf{17.58_{\pm 10.12}}$

Appendix

Table 12: Computational complexity analysis of the O-T-A-C reasoning loop. The table reports the total time and average time per step (in seconds) for a single dialogue turn using different backbone models.

Info	Gender	Work Exp.	Job Responsibilities
Doctor1	Female	6 years	Early Intervention for Autism Child.
Doctor2	Female	5 years	Language and Articulation Disorder Therapy.

Appendix

Table 13: Information for experts involved in human evaluation.

Dimension	Category	Explanation
Professionalism	Principle	Dialogues adhere to the DTT method or NET approach outlined.
	Instruction	Doctor provides clear, unambiguous instructions to the child.
	Assistance	Doctor provides timely and appropriate assistance to the child.
	Reinforcement	Doctor’s feedback is positive and effectively reinforces
	Reinforcement	the child’s correct responses or positive behaviors.
	Acknowledgment	Doctor avoids criticism or negative reinforcement when the child

Appendix

Table 14: The evaluation criteria for Dialogue Synthesis and Clinical Intervention Effect, which are divided into 3 dimensions and ten categories with their explanations. Scores range from 0 to 4, with higher scores indicating better quality for the doctor’s responses.

Dimension	Category	Explanation
Doctor (A)	Dialogue Principles (A1)	Whether the dialogue follows ABA-based instructional paradigms, such as Discrete Trial Training (DTT) or Natural Environment Teaching (NET).
	ABA Strategy Sequencing (A2)	Whether appropriate ABA strategies are applied in a progressive and coherent order (e.g., reinforcement before instruction, acceptance followed by partial or full assistance, rather than disordered sequencing).
	Personalization (A3)	Whether the doctor adapts questioning style, linguistic complexity, or pacing according to the child’s specific responses and needs.
Child (B)	ASD-consistent Response (B1)	Whether the child’s responses exhibit realistic ASD characteristics, such as non-compliance, repetitive behaviors, or language impairments.
Interaction (C)	Scenario Complexity (C1)	Whether the dialogue contains effective instructional dynamics, such as corrective teaching loops or meaningful pedagogical interactions.

Appendix

Table 15: Evaluation criteria for Dialogue Synthesis in ablation study. The assessment covers three dimensions—Professionalism (A), Child Realism (B), and Scenario Quality (C)—with corresponding sub-categories used in both human and LLM-based evaluations.

Model	Accuracy	Cohen’s $\kappa$	Spearman’s $\rho$	Kendall’s $\tau$
DeepSeek-V3.2	0.522	0.288	0.396	0.381
GPT-5.1	0.283	-0.024	-0.144	-0.136
Gemini-2.5-Pro	0.261	-0.048	0.024	0.023

Appendix

Table 16: Meta-Evaluation Results: Alignment between LLM Judges and Human Experts.

Doctor Part
Doctor	Child	D-2 $\uparrow$	D-3 $\uparrow$	S-BLEU $\downarrow$	S-GLEU $\downarrow$	S-BERTScore $\downarrow$
Real	Real	0.348	0.637	0.549	0.282	0.586
DoctorAgent	ChildAgent	0.249	0.499	0.680	0.353	0.607
DoctorAgent	BaseChild	0.277	0.519	0.661	0.364	0.599
BaseDoctor	ChildAgent	0.187	0.412	0.731	0.318	0.612
Child Part

Appendix

Table 17: Diversity Metrics for Doctors and Children Across Different Sources

Doctor	Child	Instru.	Reinfo.	Half-A.	Full-A.	Other	KL	JS
Real	Real	42.29	31.62	9.99	7.27	8.83	-	-
DoctorAgent	ChildAgent	33.68	26.85	19.54	3.80	16.13	0.083	0.019
DoctorAgent	BaseChild	27.89	13.43	36.36	6.30	16.01	0.325	0.072
BaseDoctor	ChildAgent	73.45	3.77	4.98	0.00	3.92	0.259	0.118

Appendix

Table 18: Distribution of Doctor and Child Interaction Strategies Percentage (%) with KL and JS Divergence to Real

Doctor	Child	Relev.	Irrele.	Unres.	Repet.	KL	JS
Real	Real	58.55	25.66	10.36	5.43	-	-
DoctorAgent	ChildAgent	53.72	27.40	11.43	7.44	0.007	0.002
DoctorAgent	BaseChild	47.90	26.22	15.56	10.31	0.039	0.009
BaseDoctor	ChildAgent	49.91	25.73	15.78	8.40	0.024	0.006

Appendix

Table 19: Distribution of Child Response Types Percentage (%) with KL and JS Divergence to Real

Doctor	Child	Instru.	Reinfo.	Half-A.	Full-A.	Other
Real	Real	$21.77_{\pm 11.29}$	$21.32_{\pm 11.39}$	$30.35_{\pm 10.57}$	$30.48_{\pm 16.01}$	$20.58_{\pm 11.07}$
DoctorAgent	ChildAgent	$27.94_{\pm 9.29}$	$25.95_{\pm 6.39}$	$31.87_{\pm 8.82}$	$35.53_{\pm 13.54}$	$25.44_{\pm 8.61}$
DoctorAgent	BaseChild	$22.47_{\pm 6.98}$	$18.84_{\pm 8.58}$	$23.38_{\pm 9.48}$	$24.64_{\pm 10.65}$	$20.28_{\pm 7.70}$
BaseDoctor	ChildAgent	$96.50_{\pm 39.55}$	$24.32_{\pm 13.01}$	$30.06_{\pm 16.64}$	$0.00_{\pm 0.00}$	$21.35_{\pm 9.14}$

Appendix

Table 20: Doctor Utterance Length by Intervention Strategy (Mean ±Std )

Doctor	Child	Relev.	Irrele.	Unres.	Repet.
Real	Real	$5.97_{\pm 5.41}$	$5.92_{\pm 3.52}$	$0.00_{\pm 0.00}$	$4.45_{\pm 2.84}$
DoctorAgent	ChildAgent	$4.61_{\pm 2.38}$	$7.95_{\pm 4.03}$	$0.00_{\pm 0.00}$	$5.98_{\pm 1.99}$
DoctorAgent	BaseChild	$5.11_{\pm 2.19}$	$11.46_{\pm 3.88}$	$0.00_{\pm 0.00}$	$10.12_{\pm 4.36}$
BaseDoctor	ChildAgent	$5.68_{\pm 2.81}$	$7.89_{\pm 2.68}$	$0.00_{\pm 0.00}$	$9.73_{\pm 3.72}$

Appendix

Table 21: Child Utterance Length by Response Type (Mean ±Std )

Evaluator	Source	A1	A2	A3	B1	C1	A	B	C	Total
DeepSeek-V3.2	Full	2.87	2.44	2.60	3.40	2.11	7.91	3.40	2.11	13.42
	$\text{w/o}_{\text{ChildAgent}}$	2.07	1.76	1.96	3.31	1.47	$\downarrow$ 26.9%	$\downarrow$ 2.6%	$\downarrow$ 30.3%	$\downarrow$ 21.3%
	$\text{w/o}_{\text{DoctorAgent}}$	2.00	1.64	2.42	3.00	1.56	$\downarrow$ 23.3%	$\downarrow$ 11.8%	$\downarrow$ 26.1%	$\downarrow$ 20.9%
Gemini-2.5-Pro	Full	3.20	2.80	2.89	3.27	2.31	8.89	3.27	2.31	14.47
	$\text{w/o}_{\text{ChildAgent}}$	2.51	2.07	2.56	3.80	2.22	$\downarrow$ 19.8%	$\uparrow$ 16.2%	$\downarrow$ 3.9%	$\downarrow$ 9.1%
	$\text{w/o}_{\text{DoctorAgent}}$	2.71	2.33	2.44	3.44	2.69	$\downarrow$ 15.8%	$\uparrow$ 5.2%	$\uparrow$ 16.5%	$\downarrow$ 5.9%

Appendix

Table 22: Ablation Study across Different Evaluators. For ablated settings, A/B/C/Total report relative changes (%).

Model	Setting	BLEU $\uparrow$	GLEU $\uparrow$	MET. $\uparrow$	BERT. $\uparrow$	BGE $\uparrow$	QwenEmb $\uparrow$	D-2 $\uparrow$	D-3 $\uparrow$
GPT-4o	ABA	0.091	0.143	0.377	0.882	0.756	0.746	0.914	0.954
	BASE	0.091	0.144	0.373	0.884	0.750	0.745	0.922	0.960
	DoctorAgent	0.083	0.142	0.345	0.886	0.738	0.725	0.945	0.981
GPT-4o-mini	ABA	0.094	0.148	0.374	0.885	0.752	0.747	0.912	0.956
	BASE	0.093	0.151	0.357	0.888	0.747	0.742	0.922	0.962
	DoctorAgent	0.074	0.131	0.332	0.881	0.728	0.717	0.925	0.966

Appendix

Table 23: Ablation Study on Lexical, Semantic, and Diversity Metrics. For each model, the best result under each metric is highlighted in bold.

Model	Setting	Multi-P $\uparrow$	Multi-R $\uparrow$	Multi-F1 $\uparrow$	LCS-P $\uparrow$	LCS-R $\uparrow$	LCS-F1 $\uparrow$
GPT-4o	ABA	62.55	75.16	65.99	62.55	75.16	65.99
	BASE	61.46	73.66	64.88	61.46	73.66	64.88
	DoctorAgent	70.78	79.82	72.95	70.78	79.82	72.95
GPT-4o-mini	ABA	62.13	75.99	66.25	62.13	75.99	66.25
	BASE	62.82	75.33	66.58	62.82	75.33	66.58
	DoctorAgent	67.38	79.28	70.47	67.30	79.19	70.38

Appendix

Table 24: Ablation Study on Strategy Consistency Metrics (in %). For each model, the best result under each metric is highlighted in bold.

Evaluator	Model	Setting	A	B	C	Total
DeepSeek-V3.2	Real		2.22	3.29	3.86	9.37
	GPT-4o-mini	BASE	2.58	3.23	3.68	9.49
	GPT-4o-mini	ABA	2.76	3.30	3.69	9.75
	GPT-4o-mini	DoctorAgent	2.57	3.24	3.69	9.50
	GPT-4o	BASE	2.74	3.28	3.67	9.69
	GPT-4o	ABA	2.91	3.43	3.83	10.17

Appendix

Table 25: Ablation study evaluated by different LLM evaluators. For each evaluator, the best Total score is highlighted in bold. The Real row is shown in gray for reference.

Figure 18: Base prompt

Figure 19: ABA prompt

Figure 20: Prompt for Strategy Labeling

Figure 21: Prompt for DoctorAgent: Observe

Figure 22: Prompt for DoctorAgent: Think

Figure 23: Prompt for DoctorAgent: Act in Strategy Instruction

Figure 24: Prompt for DoctorAgent: Act in Strategy Half-Assistance

Figure 25: Prompt for DoctorAgent: Act in Strategy Full-Assistance

Figure 26: Prompt for DoctorAgent: Act in Strategy Other

Figure 27: Prompt for DoctorAgent: Act in Strategy Reinforcement

Figure 28: Prompt for ChildAgent: Act in Type Irrelevant Response

Figure 29: Prompt for ChildAgent: Act in Type Relevant Response

Figure 30: Prompt for ChildAgent: Act in Type Repetitive Response

Figure 31: Prompt for ToT

Figure 32: Prompt for LLM evaluation: Turing-like Test

Figure 33: Prompt for LLM evaluation: Scoring for Quality of dialogue synthesis

Figure 34: Prompt for LLM evaluation: Scoring for Clinical intervention effect

Explore similar work

May 21, 2026

A Proactive Multi-Agent Dialogue Framework for Assessing Social Language Disorder Traits in Autism

Characteristic linguistic behaviors associated with Social Language Disorder (SLD) in autism spectrum disorder, including echoic repetition, pronoun displacement, and stereotyped media quoting, are largely absent from spontaneous conversation and only emerge under specific conversational conditions. In structured clinical assessments, this latency means that questioning strategy selection is a critical yet underappreciated determinant of how much diagnostic information a conversation yields. Whether large language models (LLMs) can be guided to proactively select questioning strategies that systematically surface these latent traits remains largely unexplored. Here we present TPA (Think, Plan, Ask), a proactive multi-agent dialogue framework applied to the language assessment component of the Autism Diagnostic Observation Schedule Module 4 (ADOS-2), in which a doctor agent explicitly reasons about which traits remain unobserved before selecting a clinically grounded strategy and generating a targeted question. A patient agent grounded in real ADOS-2 clinical data enables reproducible evaluation without real patient participation, validated across three independent experiments confirming adequate fidelity to real patient language. Evaluated on 484 episodes from 35 patients, TPA outperforms six competitive dialogue planning baselines across all primary metrics, achieving 82.1% SLD trait coverage, 16.6% higher than automated replay of real clinical dialogues conducted by trained clinicians (65.5%), with substantially greater per-turn diagnostic efficiency (AUCC: 0.628 vs. 0.458, absolute gain +0.170). These results demonstrate that proactive questioning strategy selection substantially improves the efficiency of automated SLD trait assessment, with direct implications for scalable AI-assisted clinical screening.

Chuanbo Hu, Minglei Yin, Bin Liu +4

Department of Computer Science, University at Albany, Albany, 12222, NY, USA. · Department of Management Information System, West Virginia University, Morgantown, 26506, WV, USA. · Humanities and Social Sciences, California Institute of Technology, Pasadena, 91125, CA,2026 USA. +1

Multi-Agent LLM Systems Autism Spectrum Disorder Spoken Dialogue Systems Conversational Agents

Ask PDF

Mar 28, 2026

SocialWise: LLM-Agentic Conversation Therapy for Individuals with Autism Spectrum Disorder to Enhance Communication Skills

Autism Spectrum Disorder (ASD) affects more than 75 million people worldwide. However, scalable support for practicing everyday conversation is scarce: Low-cost activities such as story reading yield limited improvement. At the same time, effective role-play therapy demands expensive, in-person sessions with specialists. SocialWise bridges this gap through a browser-based application that pairs LLM conversational agents with a therapeutic retrieval augmented generation (RAG) knowledge base. Users select a scenario (e.g., ordering food, joining a group), interact by text or voice, and receive instant, structured feedback on tone, engagement, and alternative phrasing. The SocialWise prototype, implemented with Streamlit, LangChain, and ChromaDB, runs on any computer with internet access, and demonstrates how recent advances in LLM can provide evidence-based, on-demand communication coaching for individuals with ASD.

Albert Tang

Autism Spectrum Disorder Large Language Model-Based Role-Play Simulation Conversational Agents

Ask PDF

Jul 17, 2026

TD-DPO: Difference-Aware Preference Optimization for Mitigating Sycophancy in Clinical Autism Intervention Dialogue

The sycophancy of large language models can increase the safety risk in intervention dialogue for autistic children. Supervised fine-tuning can somewhat reduce sycophancy, but relying solely on positive examples is often insufficient to identify and correct failure patterns. We observe that sycophancy behaviors can often be localized to a limited span within the model response. In this regime, sequence-level preference optimization can over-update preference-irrelevant tokens and degrade intervention ability. To address this, we propose the Minimal Edit Data Augmentation (MEDA) strategy to construct controlled, stable, minimal edit preference pairs and Token-level Difference Direct Preference Optimization (TD-DPO), which upweights difference tokens between chosen and rejected responses while downweighting shared tokens to suppress background drift. Extensive experiments across multiple backbones and evaluators show that TD-DPO achieves a better trade-off between sycophancy mitigation and intervention ability retention in our offline settings, highlighting its potential as a practical alignment approach for autism intervention.

Shuzhong Lai, Junhong Lai, Chenxi Li +5

1Nanhu Brain-Computer Interface Institute · 2MOE Frontiers Science Center for Brain and Brain-Machine Integration, Zhejiang University · College of Computer Science and Technology, Zhejiang University +3

LLM Sycophancy Direct Preference Optimization LLM Safety Alignment Preference Optimization

Ask PDF