This is our Phase 2 algorithmic-contribution write-up. The full PDF is attached: arc-whestbench-phase2-writeup-336317.pdf.
arc-whestbench-phase2-writeup-336317.pdf (202.0 KB)
Submission ID: #336317. That is the single successfully evaluated Phase 2 submission this write-up corresponds to. #336318 is its graded control and is cited only as evidence.
What the method is. A bias-free ReLU network is positively 1-homogeneous, so a common random scale on a layer’s activations leaves every later mean unchanged but inflates the carried cumulants: the covariance by ε m mᵀ, κ3 by 2ε Sym(C ⊗ m) and κ4 by 4ε Pair3(C ⊗ C). At every layer we fit ε from the chain’s own (i,i,j) slice of κ3, with no ground truth, and remove the mode from the covariance, κ3 and κ4 together. The means are untouched by construction. The step costs 0.1 of the 307.5 units (of 2³¹ FLOPs) that #336317 bills per network.
Where it runs. #336317 is the e30_a engine with its memoryless κ4 law (G_off = λ C_off) replaced by the carried κ4 matrix core of ARC’s factored AUGMENT rule, plus the extraction. #336318 is the same file with the extraction switched off (one configuration value). Each of the two files carries the mean-correction table that the same refit procedure produces for its own chain.
Graded numbers (official grader, 50 public networks, paired by network, bootstrap 95% CI):
- extraction on vs off (#336317 vs #336318): final-layer MSE 1.721e-8 vs 4.534e-8, ratio 0.380 [0.365, 0.394], better on 50 of 50 networks; all-layer ratio 0.438;
- against the public e30_a submission (#335981): 0.837, 49/50;
- against our best e30_a build (#336226, graded again as #336319): 0.854, 48/50. The all-layer ratios are 0.839 and 0.857, so the gain is not confined to the last layer;
- without extraction the same carried chain is 2.25 times worse than #336226’s build (0/50). Carrying κ4 with memory in this engine hurts without the extraction and helps with it.
What it does not do. #336317 scores 5.176e-9 against 3.058e-9 for our #336226, because it bills 0.301 of the FLOP budget against 0.152. The carried κ4 loop costs 77 units and the dead-neuron threshold it needs another 45.5. We submit it as an algorithmic finding with a graded, one-variable measurement of its effect, not as a score.
Conditions and negative results. In a lean re-implementation (30 networks), the carried κ4 core without extraction is worth 0.985 of a λ law fitted to each network; with extraction it is 0.660 at the last layer and 0.684 over all layers, flat at 0.66-0.74 per layer from the fourth layer on. The gain needs ARC’s current term program: under the older program that V29 and e30_a follow, the same chain only ties λ (1.000). In chains that keep κ4 memoryless (V29, e30_a), extraction makes things worse. Our reading, labelled as inference in the PDF, is that the λ law has the form of the scale mode’s κ4 face, so the two are substitutes. Better births do not replace κ3 memory, and extraction does not let a chain keep less of it. The PDF has the tables, the cost ledger and the limitations.
Credits. The carried κ4 core, the factored AUGMENT rule, its harmonic projection, the term tables and the reference births are ARC’s (mlp_cumulant_propagation, MIT, commit 93d091a; Wu, Lecomte, Winer, Robinson, Hilton and Christiano, arXiv:2605.05179). The lean chain, V29 and the λ law are 504aldo’s (whest-p2-cumulant-k3, MIT). The engine #336317 runs is LordProcciuto’s e30_a (MIT); both MIT notices are reproduced at the top of the submitted file. Ours are the scale-mode extraction and the changes that made the carried core work in e30_a’s engine: a dead-neuron threshold of 6 instead of 2.85, pK slices from the full AUGMENT table, reference births, Strassen depth 3 and a refitted correction table per chain. Each is marked “OURS (T-chain)” in the code.
Code. The code is the file submitted as #336317, held by AIcrowd; the submission ID points to it, so nothing else is uploaded. Every change of ours is marked “OURS (T-chain)” in it. The build inputs (the builder, the refit tables and the run logs) can be released on request.
Use of LLMs. The team member directed the work, chose what to test, set the acceptance gates, and takes responsibility for the accuracy of this write-up. LLM agents: Anthropic’s Claude (in Claude Code) and OpenAI’s Codex (in the Codex CLI) wrote most of the code in the T-chain builder, the teacher and the research harness, ran the experiments on Kaggle and Colab, worked through the derivations in §3 and §6.1 and checked them numerically, and drafted this write-up from the run logs. Grader submissions were not approved one by one. Where we are unsure why something works, the text says so (§5.2, §6.3, §8, §9).
miscusi