I am a postdoctoral researcher in the Multimodal Language Processing Group at the Max Planck Institute for Informatics, led by Prof. Dr. Vera Demberg. I am also associated with the Department of Language Science and Technology at Saarland University and the Research Training Group on Neuro-Explicit Models of Language, Vision, and Action at Saarland University.
My research goes beyond the transcript. Human communication is richly multimodal: meaning is not just spoken but gestured, timed, and implied through tone, gaze, and expression. I build language and vision-language models that can read this full signal, picking up on co-speech gestures, turn-taking cues, sarcasm, and emotional subtext, and generate responses that are not just fluent but socially and semantically grounded.
My work spans three directions:
- Multimodal comprehension: using co-speech gestures to improve discourse and turn-taking modelling in language models, and generating semantically meaningful gestures.
- Multimodal pragmatics and reasoning: probing how vision-language models handle sarcasm, irony, implicit harm, and other pragmatic phenomena, and diagnosing where and why they fall short.
- Human-centred applications: from gesture coaching for public speaking to affective computing and fair facial expression recognition.
- Oct 2026Really happy to share that our paper Do Motion Tokenizers for Co-Speech Gesture Generation Encode Gesture Semantics? received the Best Paper Award 🏆 at the MINT Workshop @ EMNLP 2026 🍃
- Sep 2026Semantic Motion Anchors: Bridging Motion and Meaning in Co-Speech Gestures accepted at NeurIPS 2026!
- Sep 2026Do Motion Tokenizers for Co-Speech Gesture Generation Encode Gesture Semantics? accepted at the MINT Workshop @ EMNLP 2026.
- Jun 2026Gave a talk at the RTG Neuroexplicit group: Beyond Transcript: From Understanding Non-Verbal Signals to Generating Meaningful Behaviors.
- May 2026New preprint out: Semantic Motion Anchors: Bridging Motion and Meaning in Co-Speech Gestures.
- May 2026New preprint out: MuPHI: Learning Implicit Multimodal Harm Reasoning via Semantically Grounded Reward Optimization.
Pragmatic incongruity in large vision-language models (PragMatch)
Investigating irony understanding in large vision-language models
Towards emotion understanding in vision-language models
Epistemic uncertainty in vision-language models
Semantic anchor generation for co-speech gestures (Semantic Motion Anchors, NeurIPS 2026)
Generation-step-aware cross-modal representation and control in multilingual speech-text models (arXiv)
Attention imbalances in vision-language models (System-Mediated Attention Imbalances, Findings of ACL 2026)
Seminar, Saarland University
IS4152 Affective Computing; CS1010E Programming Methodology (Python/C)
- Beyond Transcript: From Understanding Non-Verbal Signals to Generating Meaningful Behaviors
RTG Neuroexplicit group, June 25, 2026.
- Email: vsuresh@mpi-inf.mpg.de
- LinkedIn: varsha-suresh-2bb773b1
- Google Scholar: profile