Visualizing a modular addition transformer: 2k vs 20k training steps side by side
Sauers_ · x · 2026-09-24
A striking visualization of a modular-addition transformer: 113 possible output tokens are laid out clockwise by value, each ring representing a number being added, with color showing token logprobs. Comparing 2k training steps (right) to 20k steps (left) shows the output distribution evolving from noise into clearly ordered structure, offering an intuitive look at how the model learns modular arithmetic.
Related event: Visualization Shows How a Modular Addition Transformer Learns(2 posts)→
More from Research
- Ex-TikTok MLE: JEV could revolutionize recommendation and search systems — hardimanjames · 2026-09-24
- WeightWatcher K-matrix power-law exponent tracks LLM memorization with ρ ≈ -0.90 — rickasaurus · 2026-09-24
- Researchers launch Prevent, Contain, Prove, a voluntary framework for formal methods — Miles_Brundage · 2026-09-24
- Tencent Hunyuan extends critical-batch-size theory to LLM RL: 29% faster GRPO, 2.29× PPO throughput — TencentHunyuan · 2026-09-24
- Researcher: activation steering is crude, we must speak the geometry inside models — cephaloform · 2026-09-24
- Neel Nanda's team launches WorkspaceBench, an eval to test interpretability tools — burny_tech · 2026-09-24