DSGym is a standardized framework for evaluating and training data science agents that addresses fragmented benchmarks, and a 4B model trained on DSGym outperforms GPT-4o on standardized analysis benchmarks.

“Author of "DSGym: A Holistic Framework for Evaluating and Training Data Science Agents" on arXiv”

AI AgentsAI ResearchFoundation ModelsBenchmarking

Cited In

Research or work cited

NVIDIA and Stanford Show Models Gain Accuracy by Treating Each Test Problem as a Learning Environment

NVIDIAannouncementpositiveJan 26, 2026

NVIDIA and Stanford's TTT-Discover achieves state-of-the-art performance across mathematics, GPU kernels, algorithms, and biology by applying reinforcement learning at test time to allow models to continue training on problem-specific experience.

“Referenced in "NVIDIA and Stanford Show Models Gain Accuracy by Treating Each Test Problem as a" in NVIDIA”

Reinforcement LearningLLMsAI ResearchFoundation Models

Also Mentioned In

Referenced in coverage

Reinforcement Learning at Test Time

GitHubannouncementpositiveJan 26, 2026

TTT-Discover enables large language models to perform reinforcement learning at test time, adapting to problem-specific tasks and achieving state-of-the-art results in mathematics, GPU kernels, algorithms, and biology.

“Co-authored TTT-Discover reinforcement learning system for test-time LLM training.”

Reinforcement LearningLLMsAI ResearchOpen Source AI