FoldFlow: A Systematic Study of Bio-Inspired Inductive Biases for Neural Architecture Design (v1.1.0)

Abstract

Adds the WikiText-2 single-component ablations re-run under the corrected causal energy gate, for round 2 of the Scientific Reports revision. Same harness, seeds (42-46), 17-epoch schedule and hardware as the post-fix headline runs in results/tier_b_full/, so every configuration is directly comparable to the post-fix full model (577.6 +/- 2.1). | Configuration | Best val PPL (5 seeds) | Delta vs full | paired t (df=4) | |---|---|---|---| | w/o energy gate | 590.4 +/- 2.5 | +12.8 | 7.10 | | w/o chaperone | 581.4 +/- 0.8 | +3.8 | 4.44 | | w/o Langevin | 577.6 +/- 2.1 | +0.000 | 0.00 | The energy-gate attribution moves from 60.9/85.9 ~= 71% (pre-fix) to 12.8/24.7 ~= 52%, leaving 8.1 PPL (~33% of the gain) unattributed to either ablated module. The no_langevin result is not a null result. The refinement step in foldflow/models/lm.py is guarded by if not self.training, so it runs at inference only; its energy head never enters the training graph and remains at random initialisation. Removing it is a numerical no-op. The manuscript withdraws the earlier claim that Langevin refinement fails to transfer to language modelling, and corrects a Methods sentence that wrongly stated the head is trained end-to-end. The vision model is unaffected: it runs Langevin inside the training forward pass, so its energy network is trained. See results/postfix_ablation/README.md for provenance and reproduce commands. Also adds figures/make_cross_domain.py (Fig. 6 previously had no generator) and syncs paper-sr/ to the v2.2 manuscript.