Chinmay Karkar
researcher · pre-training & post-training of language models
currently at HUD·previously Microsoft Research India, Lossfunk & Athena Agents (RL)
01about概要
I work on the science and craft of training language models — what makes them learn, what makes them stable over long horizons, and how to tell whether they’re actually getting better.
At HUD I work on reinforcement learning and QA agents. Previously, at Microsoft Research India, I focused on reinforcement learning for LLMs and agentic verifier modules for multi-step reasoning evaluation. Before that, probabilistic forecasting at Lossfunk, model merging & RL post-training at Athena, and a fine-tuning / inference pipeline at Kotoba Research.
→I write more about what interests me and my experiments on my blog, check it out :)
02publications論文
Future Is Unevenly Distributed: Forecasting Ability of LLMs Depends on What We're Asking
03selected work仕事
HUD — current
Reinforcement learning and QA agents.
Microsoft Research India — Research Intern
Reinforcement learning for LLMs with attention to long-horizon training dynamics and stability. Agentic verifier modules for robust multi-step reasoning evaluation.
Lossfunk — Research Intern
First-author paper on LLM forecasting behavior, accepted at the AIR-FM Workshop, AAAI 2026. Calibration, Brier, ECE, and pass@k pipelines for probabilistic forecasting.
Athena Agents — Research Intern
Built Aryabhatta 1.0, a domain-adapted LLM for JEE Main math — 90.2% accuracy, +35.5 pp over baseline. Model merging (SLERP, TIES) blended with GRPO / REINFORCE post-training.
Swades AI — Applied AI Intern
Computer vision pipelines for segmentation and keypoint estimation; document-extraction chunking; containerized deploy that took per-page processing from 2 min to 3 s.
Kotoba Research — Research Intern
Fine-tuning and inference pipeline; QAT and LoRA benchmarking across standardized evaluations.
04projects作品
heiretsu
A minimal GPT training playground in pure PyTorch implementing composable 4D parallelism — Data, Tensor, Pipeline, Expert — with manual gradient averaging, Megatron-style sharded linears, GPipe scheduling, and MoE routing.
komorebi
From-scratch implementations of language & vision papers — GPT-2, Llama, SigLip, speculative decoding — built as a step-by-step pipeline from tokenizers through vision models to fine-tuning.