Publications

  • Sparks of Cooperative Reasoning: LLMs as Strategic Hanabi Agents
    Mahesh Ramesh, Kaousheik Jayakumar, Aswinkumar Ramkumar, Pavan Thodima, Aniket Rege, Emmanouil-Vasileios Vlatakis-Gkaragkounis
    ICML 2026
    Introduced a multi-turn benchmark to evaluate state-tracking and cooperation in frontier models. RL-trained a 4B model on curated data, outperforming all non-reasoning baselines. Released 1500+ (~90K data points) game trajectories for SFT and move-level ratings for RLVR.
    [arXiv] [code]
  • CuRe: Cultural Gaps in the Long Tail of Text-to-Image Models
    Aniket Rege, Zinnia Nie, Mahesh Ramesh, Unmesh Raskar, Zhuoran Yu, Aditya Kusupati, Yong Jae Lee, Ramya Korlakai Vinayak
    ICCV 2025 · CVPR DemoDiv Workshop (Oral)
    Curated 300 artifacts across 6 cultural axes and 64 countries. Developed Marginal Information Attribution metrics achieving 2x Spearman correlation improvement over prior scorers. Conducted 2,700 artifact-level surveys with culturally-aligned annotators.
    [arXiv] [project page]
  • MABViT: Modified Attention Block Enhances Vision Transformers
    Mahesh Ramesh, Aswinkumar Ramkumar
    Deployable AI Workshop, AAAI 2024 (Oral)
    Integrated Gated Linear Units into the attention module for parallel MLP-attention computation. Achieved 0.6% accuracy gain over ViT-S/16, surpassed ViT-B/16 with half the parameters, and 17% faster training convergence.
    [arXiv]

Thesis

  • Multivariate Time-Series Data Augmentation for Anomaly Detection
    Mahesh Ramesh
    B.Tech Thesis, IIT Madras · Dassault Aviation Collaboration · Presented at RBCDSAI 2023
    Designed a deep neural network to augment scarce flight telemetry for improved anomaly-detection recall. Segmented mixed discrete/continuous sensor streams into state clusters. Implemented a VAE with attention mechanism and soft constraints, achieving a 15% reduction in discrimination score.