Roger Jin

Roger Jin is an AI researcher whose credited work includes mathematical reasoning systems and distributed model training. He is a contributor to Nomos and a coauthor of CurvaDion and the Hermes 4 Technical Report. He uses the public handle @rogershijin; ThursdAI identified him as a Nous Research researcher for his September 4, 2025 appearance.[4][6][2][3]

Jin is credited among the Nous Research contributors to Nomos 1, alongside Jeffrey Quesnelle, Dakota Mahan, Chen Guang, and Teknium. Trained in collaboration with Hillclimb AI, the model specializes Qwen/Qwen3-30B-A3B-Thinking-2507 for mathematical problem-solving and proof-writing in natural language. The release pairs the model with an open-source reasoning harness; its credits also include math experts and other collaborators.[4]

The Nomos harness, whose citation jointly credits Jin, repeatedly generates candidate solutions and self-scores them out of seven points. It prioritizes problems with the fewest self-critiqued perfect scores, continuing until a time limit or target score count is reached. To choose final answers, it groups submissions by conclusion, selects the group it judges correct—which need not be the majority—and runs a pairwise elimination tournament. These self-scores guide the search rather than independently establish proof correctness.[5]

The project reports 87/120 on Putnam 2025 for Nomos 1 with the harness, graded by a human expert, compared with 24/120 for the Qwen base model under the same conditions. The comparison measures the model working with its reasoning harness.[5][4]

Jin coauthored CurvaDion: Curvature-Adaptive Distributed Orthonormalization with Bhavesh Kumar and Jeffrey Quesnelle, submitted on December 13, 2025. The work addresses when training workers need to synchronize: the authors observe that workers compute similar gradients in flat regions, while high-curvature regions require coordination. CurvaDion uses Relative Maximum Momentum Change, derived from momentum already computed during optimization, as a tractable proxy for directional curvature to detect where synchronization is needed.[6]

The authors report 99% communication reduction while matching baseline convergence in experiments with models from 160 million to 1.3 billion parameters. That experimental range bounds the result, even though the paper motivates the problem through the communication demands of much larger models.[6]

Before the December 2025 CurvaDion submission, Jin appeared on ThursdAI on September 4, 2025, where he was identified as a Nous Research researcher. His broader research credits include authorship of the Hermes 4 Technical Report. His public appearances also include Latent Space’s “What is an RL environment? w/ Nous Research's Roger Jin”.[6][2][3][1]