"Chicken" can make base models reason—researcher shares paper-writing behind the scenes
RulinShao · x · 2026-10-13
Researcher Sophie Wang published a 16-minute read detailing how her paper "Base Models Can Reason By Taking a Cue From Training Data" came to be. The paper shows base models can match RL post-training on reasoning given the right opening tokens (even "chicken"), tracing this to learned training-data associations.
Her writing methodology is the highlight:
- She storyboarded all figures before running experiments, designing experiments around them—turning intuitions into falsifiable hypotheses and putting communication at the front of research
- Each figure conveys one idea: Figure 2 dropped pass@16 to focus on whether a token cue closes the gap to RL on pass@1
- Tooling: Matplotlib, Graphik for sans-serif labels, B612 Mono for code, Tailwind v4's OKLCH palette for colors
The project began during her UC Berkeley visit with Amil Dravid and Alyosha Efros, after two fruitless months on low-rank subspaces in weight space, before pivoting to inspecting the tokens of reasoning outputs themselves.
More from Research
- Spatial biology slide costs stuck at $5-10k for five years, defying Moore's law — anshulkundaje · 2026-10-13
- Google Research trains user simulators with latent intrinsic feedback, cutting RL compute by over half — cocoweixu · 2026-10-13
- Helmholtz-SDE closes the approximation gap in simulation-free latent SDEs — scott_linderman · 2026-10-13
- Continual Learning Reading List: Cortical Labs Neurons, Self-Distillation, Text-to-LoRA — yacinelearning · 2026-10-13
- 1.5-Hour Deep Dive on Continual Learning: From First Principles to Qwen3-8B Experiments — yacinelearning · 2026-10-13
- Preference training may only rewrite final layers, with alignment and welfare implications — Sauers_ · 2026-10-13