cat ~/research/README.md

Research & Publications

My research focuses on AI safety, multimodal learning, and understanding how foundation models encode and process information.

01 2026

Heard but Not Heeded: Paralinguistic Information Encoding and Loss in Audio-Language Models
Bhuvan Koduru, Dalaver Alharthi, Rita Singh, Bhiksha Raj
CoLM 2026

Investigating how audio-language models encode and preserve paralinguistic information across their internal representations.

Subliminal Prosody Learning: Auxiliary Emotion Supervision Redistributes Affective Representations Across ALM Layers
Bhuvan Koduru, Dalaver Alharthi, Rita Singh, Bhiksha Raj
ICML MI 2026 workshop

Demonstrating how auxiliary emotion supervision reshapes affective representation distribution across audio-language model layers.

RIVET: Robust Idempotent Voice Attribute Editing
Dalaver Alharthi, Bhuvan Koduru, Bhiksha Raj, Rita Singh
INTERSPEECH 2026

Novel approach for robust and idempotent voice attribute editing while maintaining speaker identity.

Computer Use with Evolving Software
Siddharth Agashe, Guillermo Gonzalez-Pumariega, Bhuvan Koduru, Jiachen Yang, Ang Li, Xin Eric Wang
CoLM 2026 workshop

Developing evolving harnesses for computer-use agents to enable safer autonomous systems in dynamic environments.

AmongUs-X: Benchmarking Strategic Deception of LLM Agents via Theory-of-Mind Metrics
Kian Mirakhor*, Bhuvan Koduru*, Apostolos Kontogiannis, Ryan Zabounidis, Katia Sycara, Wonsuk Kim
NeurIPS 2026 under review *equal contribution

Novel benchmark for evaluating strategic deception and Theory of Mind in LLM agents through 8,000+ simulated multi-agent games.

02 Research Interests

01

AI Safety & Alignment

Developing methods to make AI systems safer, more controllable, and aligned with human values. Steering mechanisms, safety interventions, and understanding failure modes.

02

Multimodal Learning

Understanding how audio-language and vision-language models encode and process information across modalities. Investigating information flow and representation learning.

03

Mechanistic Interpretability

Reverse-engineering neural networks to understand internal representations and computational mechanisms. Using interpretability to improve model behavior and safety.

04

Multi-Agent Systems

Designing and analyzing multi-agent systems with focus on strategic interaction, deception, and Theory of Mind. Building evaluation frameworks for agent capabilities.