Research
My work primarily focuses on reinforcement learning, game theory, and parameter efficiency these days. I am particularly interested in modeling adversarial dynamics, sequential decision processes, and building exact solvers for combinatorial spaces. Below is a selection of my recent publications and manuscripts currently under review. Whether you're building in adjacent spaces, or exploring topologies, feel free to reach out to me via email.
Peer-Reviewed Publications
Kabir Murjani. "Zero-Copy Semantic Contagion: An In-Memory Streaming Architecture for Evolving Attention Graphs." ACM SIGMOD Workshop on Data Management for the Modern Financial Systems (FinDS), 2026.
- Architected a zero-copy Rust engine for continuous-time contagion modeling, achieving a 3.1 Sharpe ratio and 1.7x precision lift.
Kabir Murjani, Parth Vyas. "LURE: Latent Utility Reward Erosion as a Bayesian Signaling Game in Multi-Step Agent Interactions." Conference on the Mathematics of AI (MathAI), 2026.
- Modeled vulnerability as an endogenous utility deficit, yielding 1.5x higher adversarial reward extraction than standard baselines.
Deep Joshi, Kabir Murjani, Mohammad Obaidat, Rajesh Gupta, Sudeep Tanwar, Joel Rodrigues. "SaNcHaR: Adaptive Topology Routing and Quantized Edge Inference." IEEE ICC, 2026.
- Edge-inference architecture coupling a 1D-CNN with RAG for localized anomaly detection alongside a dynamic topology-switching routing layer, achieving an 84.78% packet delivery rate and sub-millisecond latency (0.2778 ms) in degraded networks.
Manuscripts Under Review
Parth Vyas, Kabir Murjani. "Weight-Space Teleportation: Discovering LLM Reasoning via Bilevel Evolution." Submitted to Transactions on Machine Learning Research (TMLR).
- Formalized a bilevel framework coupling SVD-stratified mutations with gradient refinement, reaching 77.6% on GSM8K.
Kabir Murjani, Parth Vyas, Akshita Abrol, Rajesh Gupta. "Judges Hallucinate, Embeddings Don't: Retrieval-Augmented Latent Regression for Dialogue Optimization." Submitted to Transactions on Machine Learning Research (TMLR).
- Developed a judge-free alignment framework using deterministic embedding regression to improve human preference by 44%.
Kabir Murjani, Parth Vyas, Akshita Abrol, Daniel Wang, Ian McLoughlin. "The Geometry of Context: Topological Optimization for Prompting with Online Supervision." Submitted to ACL Rolling Review (ARR), May 2026.
- Designed the TOPOS modular RAG framework, reducing token usage by 39% via information-theoretic compression.