Research Track — Reading, Reproducing, Publishing

What this phase teaches

Keshav’s 3-pass paper-reading method. arXiv firehose tools (Semantic Scholar, alphaXiv, HF Daily Papers, Connected Papers, arxiv-sanity-lite). Reproducing at small scale. LaTeX / NeurIPS / ICLR / ICML / COLM formatting. How peer review works. Contributing to transformers / trl / vllm.

2026 frontiers to specialize in

1. Reasoning + RLVR — process reward models, verifier-guided search, agent RL.
2. Long context — Mamba-2, RWKV-7, iRoPE, state-space models, retrieval-as-context.
3. Multimodal — native early-fusion (Llama 4, Gemini 2.5), video, audio.
4. Agentic AI — computer use, browser agents, SWE-bench Pro, OSWorld.
5. Efficient training & inference — FP4/FP8, MoE routing, distillation.
6. Safety — scalable oversight, debate, weak-to-strong, alignment audits, AI control.
7. Interpretability — SAEs + transcoders + crosscoders, circuit-level reasoning.