Nishit Neema

I'm Nishit Neema, an Applied ML Researcher at Cerebras Systems, working on post-training for large language models. My current focus is long-horizon reinforcement learning — developing stable, scalable training recipes for state-of-the-art open-source models including the Qwen, GLM, Kimi, and other frontier model families. At Cerebras, my research spans inference-time orchestration (CePO — enabling a 32B model to surpass much larger frontier models on AIME and LiveCodeBench), RL-based verifier training for calibrated test-time reasoning, curriculum learning (ACER — synthesizing structured training data to develop domain expertise without catastrophic forgetting), and RL reward design (CoRPO — addressing fundamental flaws in GRPO's handling of ordinal rewards). Before Cerebras, I completed my M.Tech. in Computer Science at IISc Bengaluru, where I worked on diffusion-based generative models, resulting in MILD (Momentum-Imbued Langevin Dynamics) — a faster sampler for score-based models, published at IEEE ICASSP 2024.