SixSigmaEngineer/review-residuals

Review Residuals — Scaling Sweep Results

Supporting data, code, and paper for "Review Residuals: Update-Conditioned Residual Gating for Transformers" (Kyle Kramer, NeraTech LLC, 2026).

Review Residuals scale each transformer sublayer's proposed update by a small learned gate conditioned on both the current state and the proposed update — an in-network analogue of independent verification. Trained from scratch with parameter-matched baselines, the advantage over both a Highway gate and the standard residual emerges with scale.

Code & full repo: https://github.com/SixSigmaEngineer/review-residuals

What's here

• scaling_v8.csv — the 42 training runs (per-seed validation losses) behind every number in the paper.
• train.py, analyze_results.py, make_emergence_figure.py, run_on_runpod.ipynb — training, analysis, figure, and cloud-GPU code.
• review_residuals.pdf — the technical paper; plain-English companion included.
• Figures: mechanism_additive.png, emergence_curve.png.

scaling_v8.csv schema

column	meaning
size	model size group (60M, 150M, 320M, 590M, 1B)
variant	review_neutral (ours), highway, or standard (both baselines parameter-matched up to Review)
seed	random seed
params_M	actual parameter count (millions)
steps	training steps
val_loss	validation loss (lower is better)
ece	expected calibration error (where measured)
minutes	wall-clock training time

Headline result

Size	Review	Highway	Standard	Review wins?
60M	1.6891	1.6923	1.6805	no
150M	1.5576	1.5625	1.5548	tie
320M	1.5001	1.5042	1.5037	tie
590M	1.4795	1.4889	1.4901	yes (p<0.05)
1B	1.4876	1.5031	1.5040	yes (strong trend)

Reproduce the table and t-tests with python analyze_results.py.

Citation

@misc{kramer2026reviewresiduals,
  title  = {Review Residuals: Update-Conditioned Residual Gating for Transformers},
  author = {Kramer, Kyle},
  year   = {2026},
  note   = {NeraTech LLC}
}

Licensed under Apache-2.0.