Researchers introduce CodeMidas, an agentic pipeline that constructs reinforcement learning environments from implemented functionality in open-source codebases, using source code as the sole input. The method allocates agentic compute to explore functionality, construct execution-grounded tests, and validate tasks through repeated rollouts, resulting in a dataset of 5,545 training tasks across 23 programming languages. Training MiMo-V2.5 on these tasks with GRPO improves performance on five diverse benchmarks, including DeepSWE (+11.7%), ProgramBench (+17%), and Terminal-Bench v2.1 (+8.5%). Trajectory analysis indicates the RL-trained agent exhibits better behaviors such as increased codebase exploration and more diverse self-verification. These results establish source code as a scalable foundation for constructing RL environments that improve coding agents across diverse software tasks.
CodeMidas scales agentic coding RL environments using source code
Post-Hoc Operators Fail to Improve Accuracy in Small Code Models
A measurement study finds that 26 semantic post-hoc operators do not improve held-out accuracy over Best-of-N in frozen small code models. While two operators—expression-layer recovery and adaptive consensus early-stop—offer benefits in compute efficiency or program recovery, none outperform BoN in accuracy. The results highlight systemic limitations in error detection and coverage, suggesting that model harnesses and error coverage must be improved before post-hoc reasoning is considered.
Atria Dawn Preview: foundation agentic language model for scientific research
Atria Dawn Preview is a foundation agentic language model designed for scientific research and engineering workflows, trained via a Verifiable Experience Pipeline that connects tool-mediated interactions to executable environments. Across 16 benchmarks spanning real-world research, engineering, and digital work, the model is competitive with frontier agents and achieves the highest reported score on five of them.
NVIDIA Nemotron-3 models achieve gold-medal and top-human scores in IOI 2026
Researchers have developed a post-training pipeline for large language models that enables them to achieve gold-medal performance in competitive programming. By combining supervised fine-tuning, reinforcement learning, and a new feedback-driven strategy called GenCorrect, the system outperforms top human contestants on the IOI 2026 problem set.
NVIDIA Nemotron-3 models achieve gold-medal performance in IOI coding competitions
NVIDIA has developed post-training pipelines for its Nemotron-3-Nano-CC and Nemotron-3-Ultra-CC language models to excel in competitive programming. By combining large-scale problem curation, synthetic reasoning traces, supervised fine-tuning, and reinforcement learning, the team created specialized systems capable of high-level algorithmic reasoning.
NVIDIA's Nemotron-3 models achieve gold-medal performance in IOI coding competitions
NVIDIA has developed a post-training pipeline for its Nemotron-3 language models, enabling them to achieve gold-medal level performance in the International Olympiad in Informatics (IOI). The approach combines large-scale problem curation, synthetic reasoning traces, supervised fine-tuning, and reinforcement learning.