https://haozhang534.github.io/blog/Notes on agents, reinforcement learning, and the systems behind them.enSat, 14 Nov 2026 01:04:27 -0800 https://haozhang534.github.io/blog/reinforce-zero/https://haozhang534.github.io/blog/reinforce-zero/Sat, 14 Nov 2026 00:00:00 -0800Why matching the learner and rollout engine changes what a simple asynchronous policy-gradient update can achieve.https://haozhang534.github.io/blog/nemotron-35/https://haozhang534.github.io/blog/nemotron-35/Sat, 14 Nov 2026 00:00:00 -0800A look at how architecture, post-training, agent harnesses, and inference efficiency fit together in Super and Lightning.https://haozhang534.github.io/blog/agentbaton/https://haozhang534.github.io/blog/agentbaton/Sat, 14 Nov 2026 00:00:00 -0800Model collaboration becomes more interesting when the decision happens inside a task, after useful progress has been made.https://haozhang534.github.io/blog/logra/https://haozhang534.github.io/blog/logra/Sat, 14 Nov 2026 00:00:00 -0800Low-rank gradient sketches reduce RL memory, while predicted-KL control limits how far a compressed update moves the policy.https://haozhang534.github.io/blog/skill2env/https://haozhang534.github.io/blog/skill2env/Sat, 14 Nov 2026 00:00:00 -0800Agent Skills can describe both the work people want done and the criteria that make an execution useful.https://haozhang534.github.io/blog/osworld-pro/https://haozhang534.github.io/blog/osworld-pro/Sat, 14 Nov 2026 00:00:00 -0800A final success score is useful, but subgoals and action-level evidence make failures much easier to understand.https://haozhang534.github.io/blog/agora/https://haozhang534.github.io/blog/agora/Sat, 14 Nov 2026 00:00:00 -0800A Git-backed contribution graph records what was tried, what was verified, and which findings later work builds on.https://haozhang534.github.io/blog/flashreinforce/https://haozhang534.github.io/blog/flashreinforce/Sat, 14 Nov 2026 00:00:00 -0800A critic-free recipe addresses prompt coverage, stale trajectories, and the extra influence of long negative responses.https://haozhang534.github.io/blog/molt/https://haozhang534.github.io/blog/molt/Sat, 14 Nov 2026 00:00:00 -0800A training framework connects ordinary model code, black-box agents, asynchronous learning, and distributed experience storage.https://haozhang534.github.io/blog/procua-sft/https://haozhang534.github.io/blog/procua-sft/Sat, 14 Nov 2026 00:00:00 -0800Useful demonstrations need feasible starting states, realistic artifacts, and training examples that match the agent’s inference context.https://haozhang534.github.io/blog/polar/https://haozhang534.github.io/blog/polar/Sat, 14 Nov 2026 00:00:00 -0800API-level capture and token-faithful reconstruction let an existing agent participate in a scalable RL pipeline.https://haozhang534.github.io/blog/nemotron-nano-omni/https://haozhang534.github.io/blog/nemotron-nano-omni/Sat, 14 Nov 2026 00:00:00 -0800Multimodal reasoning needs more than input coverage: observation cost and long-context efficiency affect every interaction.https://haozhang534.github.io/blog/prorl-agent/https://haozhang534.github.io/blog/prorl-agent/Sat, 14 Nov 2026 00:00:00 -0800Decoupling environment execution from optimization makes multi-turn agent training easier to scale and to move between trainers.https://haozhang534.github.io/blog/scaling-up-rl/https://haozhang534.github.io/blog/scaling-up-rl/Sat, 14 Nov 2026 00:00:00 -0800Prolonged reinforcement learning raises questions about exploration, reference-policy drift, and how to interpret gains across reasoning domains.