Research

Research, insights, and updates from the Haladir team.

• ConstraintBench: A Benchmark and RL Environment for Constrained Optimization from Formally Verified Rewards

  A dual-purpose system for benchmarking and training LLMs on constrained optimization with solver-verified rewards. February 23rd, 2026. 10 min read.

• RLFR: Reinforcement Learning from Formally-Defined Rewards for Code Generation

  Exploring a new paradigm for training coding models using formally verified reward signals. January 6th, 2026. 8 min read.