Avni Mittal

[Open and Looking for PhD positions]

I am currently working as a Predoctoral Researcher at Google DeepMind India. Prior to this, I was a Data Scientist (AI Research) at Microsoft Security Platform.

I want to build intelligent systems that are interpretable, efficient, and reliable by design. My work on reasoning evaluation and tool-using agents asks what correct final answers can hide, and how we can better understand and improve the processes that produce them. My research interests span:

Reasoning faithfulness and monitorability: distinguishing coherent explanations from those that reflect a model's underlying computation, and investigating how internal signals and causal interventions can help evaluate and supervise reasoning.
Reliable agents: understanding how agents revise beliefs after unreliable tool feedback, maintain instructions over long tasks, and generalize across tasks and languages while respecting human and organisational values.
Efficient, interpretable models: studying how model structure shapes computation, from expert routing in mixture-of-experts models to local, iterative computation in Neural Cellular Automata, and exploring how to simplify models without sacrificing reliability.

Longer term, I want to understand how these properties can be maintained as models continually learn and adapt, rather than treating efficiency, interpretability, and alignment as separate goals.

Previously, I worked on Neural Cellular Automata–based architectures for medical image segmentation and multimodal engagement analysis. I graduated with a B.Tech in CSE (Honours) from IIT Mandi (Department Rank 2) and spent a semester at TU Darmstadt.

Email / CV / LinkedIn / GitHub / Google Scholar

Ongoing Projects

I am also looking to dive deeper into subliminal learning and chain-of-thought unfaithfulness. Some other projects near completion are:

Sonar-MoE: Speculative Routing of Tokens in Mixture-of-Experts Models.
VIAS: Organizational Value Alignment via Activation Steering.
Understanding Cross-Image Information Flow in Multi-Image Vision-Language Models.

News

[Aug '26] Paper Two papers — LITMUS (Re)Agent and TextNCA — accepted at EMNLP 2026.
[Jul '26] Career Joined Google DeepMind as a Predoctoral Researcher.
[Jul '26] Talk Presented two papers — C2-Faith and Did You Forget What I Asked? — at ACL 2026.
[Jul '26] Paper When Tool Outputs Go Wrong accepted at the ICML 2026 FAGEN Workshop, Seoul.
[2026] Award Won 2nd Prize in Apart AlxBio Hackathon 2026.
[Feb '26] Talk Took two lectures on "LLM Agents" for the Deep Learning course at IIT Hyderabad.
[Sept '25] Talk Invited talk on "How Agentic Systems Can Be Developed for Organizational Information Processing Using RAG" at IIT Hyderabad, Digital Transformation series.
[Sept '25] Research Started supervised research at SPAR AI (Safety & Alignment Research) — studying faithfulness and deceptive reasoning in Mixture-of-Experts models.
[Jul '25] Paper LITMUS++ accepted at IJCNLP-AACL 2025 (System Demonstration Track). Patent filed.
[2025] Award Received the AI Innovation Award at Microsoft Security Platform for the Security Analysis Agent.
[2025] Paper PROTECT accepted at SICon 2025 (ACL Workshop).
[2025] Paper MedSegDiffNCA accepted at IEEE CBMS 2025.
[2025] Patent Filed patent on "Multi-Agent Predictive Systems for Cross-Domain Model Performance Assessment" with Dr. Sandipan Dandapat and Dr. Monojit Choudhury.
[2025] Patent Filed patent on "A System of Simultaneous Optimization of Multiple Agents" (Multi-Agent Prompt Tuner).
[Feb '25] Teaching Research group lead for CS672 Advanced Deep Learning at IIT Mandi, guiding UG students on GNNs and NCA.
[Jul '24] Industry Joined Microsoft as Data Scientist (AI Research) working on multi-agent systems and AI alignment.
[Sep '24] Paper ECLIPSE accepted at ECAI 2024.
[Sep '24] Paper Exceda accepted at IEEE MIPR 2024.
[May '24] Graduated from IIT Mandi with B.Tech CSE (Honours), CGPA 8.93/10, Department Rank 2.
[2024] Award Academic Excellence Award (CSE), IIT Mandi — Highest SGPA in 5th and 6th semesters.
[2023] Award Top 5 Rank at Inter-IIT Tech Meet representing IIT Mandi.
[Oct '22] Semester exchange at TU Darmstadt, Germany — CGPA 9.36/10.

Publications

TextNCA: Neural Cellular Automata for Language Modeling via Hierarchical Local Attention

Avni Mittal, Avinash Anand, Ashutosh Kumar, Dikshant Kukreja, Kritarth Prasad, Sushane Dulloo, Erik Cambria, Timothy Liu, Zhengkui Wang, Rajiv Ratn Shah

EMNLP 2026 (arXiv:2608.02050)

paper

Triadic Werewolf: A Jester Role for Multi-Hop Theory of Mind in LLMs

Avni Mittal

arXiv preprint arXiv:2606.27909, 2026

paper

The Periodic Table of LLM Reasoning: A Structured Survey of Reasoning Paradigms, Methods, and Failure Modes

A Anand, M Ramesh, A Mittal, A Kumar, E Cambria, Z Wang, T Liu, AB Ng, ...

arXiv preprint arXiv:2606.11470, 2026

paper

Litmus (Re) Agent: A Benchmark and Agentic System for Predictive Evaluation of Multilingual Models

A Mittal, S Kumar, S Dandapat, M Choudhury

EMNLP 2026 (arXiv:2604.08970)

paper

Do LLMs Follow Their Own Rules? A Reflexive Audit of Self-Stated Safety Policies

A Mittal

arXiv preprint arXiv:2604.09189, 2026

paper

Did You Forget What I Asked? Prospective Memory Failures in Large Language Models

A Mittal

TrustNLP 2026 (6th Workshop on Trustworthy NLP), pp. 471–488

paper

C2-Faith: Benchmarking LLM Judges for Causal and Coverage Faithfulness in Chain-of-Thought Reasoning

Avni Mittal, Rauno Arike

arXiv preprint arXiv:2603.05167, 2026

paper

When Tool Outputs Go Wrong: Belief Revision in Tool-Using Agents

Avni Mittal

ICML 2026 FAGEN Workshop, Seoul

paper

Pandemic Watch: Earlier Than The News — A Multi-Signal, LLM-Agentic Early-Warning Dashboard

Avni Mittal

Apart AlxBio Hackathon 2026 (2nd Prize)

report

LITMUS++: An Agentic System for Predictive Analysis of Low-Resource Languages Across Tasks and Models

Avni Mittal, Shanu Kumar, Sandipan Dandapat, Monojit Choudhary

IJCNLP–AACL 2025, System Demonstration Track

paper

PROTECT: Policy-Related Organizational Value Taxonomy for Ethical Compliance and Trust

Avni Mittal, Sree Hari Nagaralu, Sandipan Dandapat

SICon 2025 (ACL Workshop)

paper

Medsegdiffnca: Diffusion Models with Neural Cellular Automata for Skin Lesion Segmentation

Avni Mittal, John Kalkhof, Anirban Mukhopadhyay, Arnav Bhavsar

IEEE CBMS 2025

paper

Unveiling Learner Dynamics: The ECLIPSE Dataset and NeuralGaze Framework for Prolonged Engagement Assessment in Online Learning

Avinash Anand, Avni Mittal, Laavanaya Dhawan, Mahisha Ramesh, Juhi Krishnamurthy, Naman Lal, Raj Jaiswal, Pijush Bhuyan, Himani, Astha Verma, Rajiv Ratn Shah, Roger Zimmermann, Shin'ichi Satoh

ECAI 2024

paper

Exceda: Unlocking Attention Paradigms in Extended Duration E-Classrooms by Leveraging Attention-Mechanism Models

Avinash Anand, Avni Mittal, Laavanaya Dhawan, Juhi Krishnamurthy, Mahisha Ramesh, Naman Lal, Astha Verma, Pijush Bhuyan, Himani, Rajiv Ratn Shah, Roger Zimmermann, Shin'ichi Satoh

IEEE MIPR 2024

paper

Patents

Multi-Agent Predictive Systems for Cross-Domain Model Performance Assessment

Avni Mittal, Shanu Kumar, Sandipan Dandapat, Monojit Choudhary

Patent filed

A System of Simultaneous Optimization of Multiple Agents

Gautam Prasad, Avni Mittal, Yugal Sachdev

Patent filed

A System to Navigate Product Complexities through Automatic Visual Guide Generation

Gautam Prasad, Avni Mittal, Yugal Sachdev

Patent filed

Selected Experience

Predoctoral Researcher, Google DeepMind

Advancing reliable and interpretable AI systems.

Jul 2026 – Present

Data Scientist (AI Research), Microsoft

Multi-agent systems, AI alignment, compliance automation. Built the Security Analysis Agent (89% precision, AI Innovation Award). Developed Multi-Agent Prompt Tuner and Swarm Pinboard. Improved RAG performance from 60% to 91% for Security Copilot Skill deployed to 2,000+ enterprise tenants.

Jul 2024 – Jul 2026

Supervised Research, SPAR AI (Safety & Alignment Research)

Studying faithfulness and deceptive reasoning in MoE models. Developed perturbation-based benchmarks from PRM800K to evaluate Chain-of-Thought faithfulness.

Sept 2025 – Present

Research with Dr. Sandipan Dandapat & Dr. Monojit Choudhury

Multi-agent system for predictive analysis of low-resource languages. Work published at IJCNLP-AACL 2025; patent filed.

Jul 2025 – Present

Research Intern, MIDAS Lab, IIIT Delhi

Supervisor: Dr. Rajiv Ratn Shah. Multimodal engagement prediction using NCA. Published at ECAI 2024 and IEEE MIPR 2024.

Sept 2023 – Aug 2024

Bachelor's Thesis, IIT Mandi & TU Darmstadt

Supervisors: Dr. Arnav Bhavsar & Dr. Anirban Mukhopadhyay. Medical image segmentation using NCA and Diffusion models. Published as MedSegDiffNCA at IEEE CBMS 2025.

Aug 2023 – May 2024

Student Research Assistant, TU Darmstadt

Explainable AI methods and robustness analysis under adversarial attacks. C-Arm orientation tracking for surgical workflow optimization.

Dec 2022 – Mar 2023