HIDEKI-SQ/cot-backfire: v1.0.A2: Cue-Dominant Extraction (A2 Paper)

What's New Paper A2: Cue-Dominant Extraction HIDEKI. "Cue-Dominant Extraction Explains Length Effects in Corrupted Reasoning Traces." Preprint (2026). DOI: 10.5281/zenodo.18132430 This release adds the complete experimental code, data, and figures for Paper A2, which investigates the mechanistic basis of how language models process corrupted reasoning traces. Key Findings (A2) Order Invariance: Shuffling step order does not reduce accuracy (+7.5%, P=0.001), contradicting sequential integration Cue Dominance: Final-answer cue status explains ~7× more variance than step position (31.7 vs 4.5 pp) Two-Stage Processing: Models extract from cues first, fall back to redundancy when cues are corrupted Design Implication: Cue integrity—not trace length—should be the primary design target Added Files Notebooks (7 experiments) E1_shuffle_experiment.ipynb E1b_shuffle_c04_experiment.ipynb E2_position_experiment.ipynb E3_wrong_position_experiment.ipynb E4_late_protected_experiment.ipynb E4prime_late_protected_c04_experiment.ipynb E5_single_step_experiment.ipynb Data (7 directories) E1_shuffle_20260102/ E1b_shuffle_c04_20260102/ E2_position_20260102/ E3_wrong_position_20260102/ E4_late_protected_20260102/ E4prime_late_protected_c04_20260102/ E5_single_step_20260102/ Figures figures/A2/fig1_experimental_design.pdf figures/A2/fig2_main_results.pdf figures/A2/fig3_order_vs_position.pdf figures/A2/fig4_mechanistic_model.pdf Experimental Summary | Experiment | Design | Key Result | |------------|--------|------------| | E1 | Shuffle (c=0.6) | +5.0 pp (ns) | | E1b | Shuffle (c=0.4) | +7.5 pp (P=0.001) | | E2 | Early vs Late | -36.2 pp | | E3 | WRONG Early vs Late | -35.2 pp | | E4 | Late-Protected (c=0.3) | 94.5% recovery | | E4' | Late-Protected (c=0.4) | 92.0% recovery | | E5 | Step1 vs Step10 | -11.1 pp (22:0 asymmetry) | Total inferences: 1,990 Citation @article{hideki2026cue, title={Cue-Dominant Extraction Explains Length Effects in Corrupted Reasoning Traces}, author={HIDEKI}, journal={Preprint}, year={2026}, doi={10.5281/zenodo.18132430} } Full Changelog v1.1.A1 → v1.0.A2 Added A2 experiments (E1-E5, E4') Added A2 figures Updated README with A2 documentation