Bhuvan Koduru
CoLM 2026Presenting “Heard but Not Heeded” · Oct 6–9, SFPresenting my poster · Oct 6–9

AI/NLP researcher · Carnegie Mellon University

Bhuvan Koduru

I study the gap between what models represent and what they act on, across audio-language models, AI safety, self-evolving agents, and deception & Theory of Mind.

M.S., Language Technologies Institute · Dec 2026 · Pittsburgh

Open to full-time roles from 2027 Research scientist · Research engineer · Machine learning engineer CV Scholar GitHub
Pointillist portrait of Bhuvan Koduru Audio-language models AI safety Deception & Theory of Mind Self-evolving agents
Contents

Contents

  1. About
  2. News
  3. Selected publications
  4. Experience
  5. Education

About

I'm a Master's student in AI/NLP at Carnegie Mellon University's Language Technologies Institute, graduating in December 2026. I'm looking for full-time research scientist, research engineer, and machine learning engineer roles starting in 2027.

My research asks where the gap opens between what a model represents and what it acts on. Audio-language models encode how something is said but lose it before they respond (CoLM 2026). Models that refuse harmful requests in chat mostly stop refusing once they act as computer-use agents (NeurIPS 2026 workshop). I work with mechanistic interpretability, activation steering, and large-scale evaluation.

This summer I was a research intern at Simular.ai, working on self-evolving computer-use agents and agent safety. Before that I built AmongUs-X at CMU's Robotics Institute, a benchmark of 8,000+ Among Us games measuring deception and Theory of Mind in 21 LLMs. Before CMU, I built multi-agent systems at Adobe and edge computer-vision models at Nosh Robotics.

News

  • Agent S4 is on arXiv: interfaces it evolves from experience take an open 27B model to 21.3% on OSWorld V2, above GPT-5.5 and Claude Opus 4.8.
  • Presenting Heard but Not Heeded at CoLM 2026 in San Francisco, Oct 6–9. Come say hi at the poster.
  • Two papers under review: Agent S4 at ICLR 2027 and AmongUs-X at AAMAS 2027.
  • Chat-safe is not CUA-safe accepted as a poster at the Agents in the Wild workshop at NeurIPS 2026.
  • Heard but Not Heeded (CoLM 2026) is on arXiv.
  • Joined Simular.ai as a research intern, working on AI safety and computer-use agents with Xin Eric Wang, Jiachen Yang, and Ang Li.
  • Four papers accepted: CoLM 2026, the ICML Mechanistic Interpretability Workshop, INTERSPEECH 2026, and the CoLM Lifelong Agents Workshop.
  • Started as a graduate research assistant at CMU's Robotics Institute, building evaluation harnesses for multi-agent LLM deception.
  • Started my M.S. in AI/NLP at Carnegie Mellon's Language Technologies Institute.

Selected publications

  1. Heard but Not Heeded: Paralinguistic Information Encoding and Loss in Audio-Language Models

    B. Koduru, D. Alharthi, R. Singh, B. Raj

    CoLM 2026

    Speaking style is strongly encoded in the audio encoder of four audio-language models, but degrades before it reaches the output.

    arXiv PDF Poster Project page & audio

  2. Chat-safe is not CUA-safe: Chat Methods May Not Transfer to Computer Use

    B. Koduru, J. Yang, A. Li, X. E. Wang

    NeurIPS 2026 WorkshopAgents in the Wild · poster

    Refusal drops from 31/47 in chat to 4/47 when the same models act as computer-use agents; activation steering that raises chat refusal leaves agent actions unchanged.

  3. Evolving Agent Interfaces

    S. Agashe, Y. Yang, G. Gonzalez-Pumariega, B. Koduru, C.-L. Li, J. Yang, A. Li, X. E. Wang

    ICLR 2027 · under reviewAgent S4

    Interfaces evolved from experience take an open 27B model from 5.6% to 21.3% on OSWorld V2, above GPT-5.5 and Claude Opus 4.8 at a fraction of their cost.

    arXiv PDF Code

  4. RIVET: Robust Idempotent Voice Attribute Editing

    D. Alharthi, B. Koduru, R. Singh, B. Raj

    INTERSPEECH 2026Oral

    Training voice editors to be idempotent makes them robust to noisy attribute labels, improving edit success and speaker-identity preservation.

    arXiv PDF

  5. AmongUs-X: Benchmarking Strategic Deception of LLM Agents via Theory-of-Mind Metrics

    K. Mirakhor*, B. Koduru*, A. Kontogiannis, R. Zabounidis, K. Sycara, W. Kim

    AAMAS 2027 · under review*equal contribution

    8,000+ simulated Among Us games across 21 LLMs, measuring whether agents deceive and how well they model each other's beliefs.

All publications and research themes

Experience

  • Research Intern · Simular.ai, Palo Alto

    Jun – Aug 2026

    Self-evolving computer-use agents (CUES-Evol, Agent S4) and the safety of computer-use agents.

  • Graduate Research Assistant · CMU Robotics Institute

    Jan – May 2026

    Built AmongUs-X: an evaluation harness and benchmark for strategic deception and Theory of Mind in LLM agents.

  • Associate Technical Consultant · Adobe, Bengaluru

    Aug 2024 – Apr 2025

    A multi-agent pipeline turning plain-text requirements into Workfront Fusion workflows (top-24 semifinalist worldwide), and a documentation tool that cut documentation time by 70%.

  • Computer Vision Intern · Nosh Robotics, Bengaluru

    Feb – Jun 2024

    Quantized classifiers for real-time cooking-level estimation on edge devices, deployed to production.

  • Research Intern · Spinal Cord Lab, IIT Hyderabad

    Oct 2023 – Mar 2024

    Benchmarked phone-camera pose estimation against motion capture across 17 subjects and 23 activities; co-author on three papers from the project, including the GATI dataset.

  • Technical Consultant Intern · Adobe, Bengaluru

    May – Jul 2023

    An AI agent inside Workfront Fusion calling ChatGPT, Photoshop, and Stable Diffusion, cutting campaign turnaround from weeks to hours.

Education

  • M.S. in AI/NLP (Intelligent Information Systems) · Carnegie Mellon University

    Dec 2026

    GPA 4.05/4.00. Coursework in multimodal machine learning, speech technology for conversational AI, and deep learning.

  • B.E. in Information Science & Engineering · JSS Science and Technology University

    Jun 2024

    GPA 9.44/10.00.