Publications

2026

1. ICML-W

   Deterministic Credit Assignment for Multi-Agent Portfolio Management via a Provenance DAG

   M Rohit

   May 2026

   Submitted to ICML 2026 Workshop on Reinforcement Learning from World Feedback (RLxF) (Submission #130)

   Abs Bib

   Financial market microstructures exhibit extreme heavy-tailed noise. We introduce the Deterministic Provenance DAG, deriving the CPL-GRPO update rule as a principled policy gradient. Achieved Sharpe ratio 1.15 outperforming MARL baselines.

   @article{rohit2026deterministic,
     title = {Deterministic Credit Assignment for Multi-Agent Portfolio Management via a Provenance DAG},
     author = {Rohit, M},
     year = {2026},
     month = may,
     note = {Submitted to ICML 2026 Workshop on Reinforcement Learning from World Feedback (RLxF) (Submission \#130)},
     keywords = {Multi-Agent Reinforcement Learning, Credit Assignment, Portfolio Management, CPL-GRPO}
   }