标签：#quantization

Mithil Vakde trained a small transformer from scratch in 1.5 hours on a rented RTX 5090. The compute cost came to about 67 cents. The model scored 44% on the ARC-AGI-1 public eval, one point behind TRM's 45% and in the same band that left DeepSeek R1, o3-mini, and Gemini 2.5 Pro behind when the TRM paper ran the comparison ( arXiv 2510.04871 ). This is the third post in his ARC series. The previous result went viral and drew public scrutiny from researchers including Lucas Beyer, Jeremy Howard, and Rohan Anil. The new one raises the score, cuts the cost, and answers every objection in a long appendix. What the model does Each puzzle becomes a sequence of tokenized grids. The transformer trains on those sequences from scratch at test time, using both the train puzzles and the eval puzzle inputs, with the eval labels hidden. Test-time training is not a loophole here. ARC is a meta-learning benchmark built around skill acquisition, and Chollet's original paper frames the whole thing as a measure of how efficiently a system gains new skills ( arXiv 1911.01547 ). The architecture got a 2026 refresh: 8 layers, SwiGLU, RMSNorm, the NorMuon optimizer, 3D RoPE, and per-task embeddings. Inference augments each test input with color and dihedral permutations, inverts the augmentations, and submits the two most common outputs. The biggest change is the loss function. Previous versions trained on input and output tokens. This one trains on output tokens only, which makes the approach supervised. Score went from 40% to 44%. Vakde admits he does not fully understand why, and the training loss got worse even as scores improved. His guess is finite model capacity. The ablations back him up on what carries the result: drop the 3D RoPE and the score falls to about 24%, drop the per-task embeddings and it falls to about 24%. Representations, not compute, are doing the work. The lineage it joins None of these systems pretrain. They learn from the benchmark's roughly thousand examples: S