Counting model's hidden states split positive and negative numbers in a sudden phase transition
dejanseo · x · 2026-10-26
dejanseo visualized layer transitions of a tiny counting model (3 wide, 32 deep), calling the behavior 'cat in a bag': numbers from -999 to 999, each plotted as a point in the model's hidden-state vectors, wiggle like a creature trapped in a bag until a sudden 'mitosis' — a phase transition where negative and positive numbers split apart. A striking look at how even tiny networks organize numerical concepts internally.
More from Research
- HIT Shenzhen unveils STILL: content-aware intra-layer hybrid attention to linearize LLMs — jiqizhixin · 2026-10-26
- Bambi quietly gains nonlinear model support, with a blog post showing how — twiecki · 2026-10-26
- Fudan and Tencent unveil DraMaze, a multi-agent pipeline for diverse microdrama scripts — Siyang He · 2026-10-26
- Alibaba's MultiAgent-OPD trains hierarchical multi-agent systems end-to-end with on-policy distillation — TaoLiveAIGC · 2026-10-26
- LoCA: two-stage pipeline adds multi-modality to geospatial location encoders — Panagiotis Koromilas · 2026-10-26
- Skipping eval checks means you're likely fitting benchmark biases and noise — antoine_chaffin · 2026-10-26