Transformer Lab

Papers

• Everyone assumes bigger AI models win. We built a 34,000-parameter earthquake detector that beats a model 8× its size. Builds a compact seismic phase picker (33,610 parameters) that beats PhaseNet (268,443 parameters) on P-wave…
• Is skipping zeros free on a wafer-scale chip? Using cycle-accurate simulation of a Cerebras wafer-scale processor, compares conditionally skipping zeros in…
• Judging single-image 3D generation without humans (and why cheap proxies fall short) Proposes a human-free protocol for ranking single-image-to-3D-mesh generations using two independent VLM judg…
• Judging to improve: a 3D judge you can train against, and the limit of cheap 3D specialization Develops a human-free VLM-judge protocol (cross-model independence, position-bias correction, calibration) fo…
• Neither parallel nor sequential: how DiffusionGemma actually commits tokens An inference-only interpretability study hooking DiffusionGemma's token-commit mechanism. Across six task reg…
• Quantizing Ideogram 4.0 onto a 3090: an INT8 build that matches FP8 and a 4-bit GGUF that beats NF4 Develops post-training quantization recipes for Ideogram 4.0 (9.3B-param diffusion transformer) for consumer…
• When an AI searches for the equation behind your data, one dial steers it Trains a small RL model for symbolic regression on the Feynman equations benchmark. The simplicity penalty (λ…