Эксперимент: Inverse-Square Fractional Attention (#1426)

{ "worked": true, "confidence": 7, "verdict": "Built inverse-square fractional attention with the paper-derived positive kernel added to attention logits, plus numerical verification and a matched synthetic singular-field regression. The kernel was scale-homogeneous to relative error below 3e-8, matched the predicted 51.99x near-origin amplification, and passed the d=3 admissibility check. Across three fixed seeds, overall test MSE improved from mean 0.005827 to 0.005710 (~2.0%), but near-origin MSE slightly worsened (0.006125 to 0.006261), so the gain is modest and not specifically demonstrated as improved cusp accuracy.", "metrics": { "baseline": "Three-seed mean overall test MSE 0.005827; mean near-origin MSE 0.006125; mean far-region MSE 0.005602", "idea": "Three-seed mean overall test MSE 0.005710; mean near-origin MSE 0.006261; mean far-region MSE 0.005465; overall relative improvement ~2.0%" }, "how_to_run": "/home/maxwelhelp/main/bin/python3 experiment.py && /home/maxwelhelp/main/bin/python3 repeat.py", "files": [ "experiment.py", "repeat.py", "results.json", "results.txt" ], "limitations": "Only a small synthetic 2D coordinate-token regression was tested; the kernel sanity check used d=3 while the toy attention model used 2D coordinates. No image or point-cloud benchmark, learned lambda or sigma, relative-position baseline, FLOP-normalized comparison, subsampling test, or permutation negative control was run." }