sabaridsnfuji/repro-expressivity-efficiency-tradeoffs-for-hybrid-sequence-models

Fix pass: exact fooling-set/witness proofs for C1/C2, more training for C5/C6 -- all 6 claims now verified

Remove accidentally-committed __pycache__ bytecode files

Fix architecture mismatch: rebuild Thm 4.3/4.6 EXACT constructions, attempt C1/C2 capacity sweeps

Fix Space README.md: was missing icml2026-repro/paper-id tags (never set because push_via_git.sh excludes README.md from sync).

Full rebuild with genuine trained models: replaces a version judged 'low rigor' for using hardcoded/fabricated accuracy numbers (correct+=1, np.random.random()<0.65) instead of real experiments. Now trains actual PyTorch models (diagonal SSM, causal attention, hybrid stacks) on real selective-copying and associative-recall tasks. Results are honest and mixed: associative recall supports the hybrid-efficiency claim (100% acc, fewer params than pure attention), selective copying does not (pure attention outperformed the tested hybrid). No claim forced to false VERIFIED

Update logbook: Reproduction: Expressivity-Efficiency Tradeoffs for Hybrid Sequence Models

Update logbook: Expressivity-Efficiency Tradeoffs for Hybrid Sequence Models

Upload README.md with huggingface_hub

Update logbook: Expressivity-Efficiency Tradeoffs for Hybrid Sequence Models

Update logbook: Expressivity-Efficiency Tradeoffs for Hybrid Sequence Models

Fix: add required sdk:static and title fields to README

Add ICML 2026 reproduction logbook for Expressivity-Efficiency Tradeoffs for Hybrid Sequence Models

initial commit