@inproceedings{shah-johnson-2026-ndd,
title = "{NDD}-{R}1: Reinforcement Learning for Multimodal {LLM}s to Understand Communication in Non- and Minimally-Speaking Autistic Individuals",
author = "Shah, Siddhant Bikram and
Johnson, Kristina T.",
editor = "Lee, Hung-Yi and
Dev, Sunipa and
Li, Jessy and
Elliott, Desmond",
booktitle = "Findings of the {A}ssociation for {C}omputational {L}inguistics: {EMNLP} 2026",
month = oct,
year = "2026",
address = "Budapest, Hungary",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2026.findings-emnlp.1630/",
doi = "10.18653/2026.findings-emnlp.1630",
pages = "34681--34697",
ISBN = "979-8-89176-462-0",
abstract = "Over 30{\%} of autistic individuals are non- or minimally-speaking, relying on a complex, idiosyncratic repertoire of multimodal behaviors to communicate. Recognizing these intentional communicative acts in naturalistic environments remains a significant challenge for current multimodal large language models (MLLMs), which often struggle with the inference of subtle nonverbal cues in this domain. We introduce NDD-R1, a multi-stage reasoning framework that adapts a general-purpose MLLM for multi-view communication understanding for this population. Our approach employs a four-stage curriculum alternating between supervised finetuning and reinforcement learning ($\text{SFT}_1 \rightarrow \text{GRPO}_1 \rightarrow \text{SFT}_2 \rightarrow \text{GRPO}_2$), integrating label-guided distillation, temperature-escalated self-distillation, and difficulty-tiered reward optimization. We evaluate NDD-R1 on an initial dataset collected through the Rapid Online Sample of Communication (ROSCO) protocol, comprising 2,582 caregiver-annotated multi-view videos of 27 non- and minimally-speaking autistic individuals. NDD-R1 outperforms baselines on action and joint classification tasks, achieving competitive function classification performance relative to Gemini, exhibiting strong multi-view scaling and generalization across unseen sessions. Our code is available at https://github.com/SiddhantBikram/NDD-R1."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="shah-johnson-2026-ndd">
<titleInfo>
<title>NDD-R1: Reinforcement Learning for Multimodal LLMs to Understand Communication in Non- and Minimally-Speaking Autistic Individuals</title>
</titleInfo>
<name type="personal">
<namePart type="given">Siddhant</namePart>
<namePart type="given">Bikram</namePart>
<namePart type="family">Shah</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Kristina</namePart>
<namePart type="given">T</namePart>
<namePart type="family">Johnson</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<originInfo>
<dateIssued>2026-10</dateIssued>
</originInfo>
<typeOfResource>text</typeOfResource>
<relatedItem type="host">
<titleInfo>
<title>Findings of the Association for Computational Linguistics: EMNLP 2026</title>
</titleInfo>
<name type="personal">
<namePart type="given">Hung-Yi</namePart>
<namePart type="family">Lee</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Sunipa</namePart>
<namePart type="family">Dev</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jessy</namePart>
<namePart type="family">Li</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Desmond</namePart>
<namePart type="family">Elliott</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<originInfo>
<publisher>Association for Computational Linguistics</publisher>
<place>
<placeTerm type="text">Budapest, Hungary</placeTerm>
</place>
</originInfo>
<genre authority="marcgt">conference publication</genre>
<identifier type="isbn">979-8-89176-462-0</identifier>
</relatedItem>
<abstract>Over 30% of autistic individuals are non- or minimally-speaking, relying on a complex, idiosyncratic repertoire of multimodal behaviors to communicate. Recognizing these intentional communicative acts in naturalistic environments remains a significant challenge for current multimodal large language models (MLLMs), which often struggle with the inference of subtle nonverbal cues in this domain. We introduce NDD-R1, a multi-stage reasoning framework that adapts a general-purpose MLLM for multi-view communication understanding for this population. Our approach employs a four-stage curriculum alternating between supervised finetuning and reinforcement learning (\textSFT₁ \rightarrow \textGRPO₁ \rightarrow \textSFT₂ \rightarrow \textGRPO₂), integrating label-guided distillation, temperature-escalated self-distillation, and difficulty-tiered reward optimization. We evaluate NDD-R1 on an initial dataset collected through the Rapid Online Sample of Communication (ROSCO) protocol, comprising 2,582 caregiver-annotated multi-view videos of 27 non- and minimally-speaking autistic individuals. NDD-R1 outperforms baselines on action and joint classification tasks, achieving competitive function classification performance relative to Gemini, exhibiting strong multi-view scaling and generalization across unseen sessions. Our code is available at https://github.com/SiddhantBikram/NDD-R1.</abstract>
<identifier type="citekey">shah-johnson-2026-ndd</identifier>
<identifier type="doi">10.18653/2026.findings-emnlp.1630</identifier>
<location>
<url>https://aclanthology.org/2026.findings-emnlp.1630/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>34681</start>
<end>34697</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T NDD-R1: Reinforcement Learning for Multimodal LLMs to Understand Communication in Non- and Minimally-Speaking Autistic Individuals
%A Shah, Siddhant Bikram
%A Johnson, Kristina T.
%Y Lee, Hung-Yi
%Y Dev, Sunipa
%Y Li, Jessy
%Y Elliott, Desmond
%S Findings of the Association for Computational Linguistics: EMNLP 2026
%D 2026
%8 October
%I Association for Computational Linguistics
%C Budapest, Hungary
%@ 979-8-89176-462-0
%F shah-johnson-2026-ndd
%X Over 30% of autistic individuals are non- or minimally-speaking, relying on a complex, idiosyncratic repertoire of multimodal behaviors to communicate. Recognizing these intentional communicative acts in naturalistic environments remains a significant challenge for current multimodal large language models (MLLMs), which often struggle with the inference of subtle nonverbal cues in this domain. We introduce NDD-R1, a multi-stage reasoning framework that adapts a general-purpose MLLM for multi-view communication understanding for this population. Our approach employs a four-stage curriculum alternating between supervised finetuning and reinforcement learning (\textSFT₁ \rightarrow \textGRPO₁ \rightarrow \textSFT₂ \rightarrow \textGRPO₂), integrating label-guided distillation, temperature-escalated self-distillation, and difficulty-tiered reward optimization. We evaluate NDD-R1 on an initial dataset collected through the Rapid Online Sample of Communication (ROSCO) protocol, comprising 2,582 caregiver-annotated multi-view videos of 27 non- and minimally-speaking autistic individuals. NDD-R1 outperforms baselines on action and joint classification tasks, achieving competitive function classification performance relative to Gemini, exhibiting strong multi-view scaling and generalization across unseen sessions. Our code is available at https://github.com/SiddhantBikram/NDD-R1.
%R 10.18653/2026.findings-emnlp.1630
%U https://aclanthology.org/2026.findings-emnlp.1630/
%U https://doi.org/10.18653/2026.findings-emnlp.1630
%P 34681-34697
Markdown (Informal)
[NDD-R1: Reinforcement Learning for Multimodal LLMs to Understand Communication in Non- and Minimally-Speaking Autistic Individuals](https://aclanthology.org/2026.findings-emnlp.1630/) (Shah & Johnson, Findings 2026)
ACL