I train language models and measure what the training actually changes. Quant trader since 2018 (co-founded NV Capital); full-time on LLM training research since August 2026. From trading I kept three habits: write the prediction down before the run, put a noise floor on every result, change one thing at a time.

Looking for a model-training role at an AI lab (pretraining, post-training/RL or alignment) and for opportunities to train myself as a researcher. Preparing for a PhD in artificial intelligence. LinkedIn · Hugging Face · qli858246@gmail.com

1. LLM from scratch → the aha moment → a coding agent · complete · report (中文)

Fifteen controlled experiments on 4× RTX 4090, from pretraining 95M/201M models through GRPO and the R1 "aha moment" (Countdown pass@1 .02 → .75) to an RL-trained coding agent. RL raised pass@1 at every stage while pass@32 barely moved: it reorders what the model can already do.

2. looped-scaling · complete · report (中文) · paper

A pre-registered scaling study of looped (recurrent-depth) transformers: 242 language models at 10M–160M. On web text a loop buys parameter efficiency that grows with size (better than dense from 40M on) but no compute efficiency at any scale measured, and decoding is 2.4–2.5× slower. On a synthetic task that needs serial steps it buys depth that extends at inference: more loops, longer chains.

3. olmo-rlvr-side-effects · in progress · setup notes and contributions in olmo-hands-on

Which RLVR training data changes what, outside its own domain? I trace each change to the Olmo 3 post-training stage that made it, measured the way Ai2 measures it; next, small open-instruct runs will test which data causes it. So far: Ai2's safety numbers for the Olmo 3 7B Think stages, replicated within 1 point, plus their run-to-run noise (open-instruct #1964); answer length, where DPO adds most of it (#2002). Independent work, not affiliated with Ai2.

4. lora-posttrain · complete · report (中文)

Teaching Ai2's Olmo 3 7B Instruct my own quant research loop (idea → factor formula → backtest → revise), with LoRA on 4× RTX 4090 and pre-registered tests: one adapter carried through four stages, SFT twice and then offline and online RL. On idea families it never saw, SFT widens what the model can do (pass@4: 0% → 90%) and RL sharpens it (pass@1: 44% → 66–71%, while pass@4 holds at about 90%). Picking the best-backtested of four answers that pass the checks adds +0.0041 incremental IC on new ideas over a sealed year of market data. Earlier chapters audit and rerun Prime Intellect's wiki-search RL on Qwen3-4B and Olmo 3 7B.