Persistent Depth Ordering amid Shifting Block-Bypass Responses in Language Model Pretraining

Abstract:Layer interventions are widely used to probe the internal organization of language models, yet most analyses examine a single training checkpoint even though model representations and computations evolve throughout pretraining. This leaves open which depth-dependent intervention responses reflect persistent organization and which are transient consequences of training. We study this question using single-block identity bypass on fixed teacher-forced contexts across five released trajectories and 11 model-domain combinations. We find that block-bypass responses retain recognizable depth ordering while their magnitudes redistribute: nearby checkpoints preserve stronger rank correspondence than distant ones, and large changes concentrate at positions that recur across text samples and transfer across evaluation domains. Controlled experiments further show that changes in the natural bypass effect cannot be reduced to a single downstream sensitivity: in replicated Pythia runs, local missing-update magnitude grows while the pooled matched downstream response decreases, whereas OLMo-2 7B exhibits a different balance. These matched responses also depend on perturbation strength and direction, without identifying targeted compensation. Together, our results show that longitudinal layer sensitivity is structured but not static, and that single-checkpoint intervention responses should be interpreted in the context of how the underlying perturbation pathway evolves during training.

Machine Learning (cs.LG); Computation and Language (cs.CL)

arXiv:2610.01165 [cs.LG]

(or arXiv:2610.01165v1 [cs.LG] for this version)

https://doi.org/10.48550/arXiv.2610.01165 Focus to learn more arXiv-issued DOI via DataCite (pending registration)

source/README.md

source/analysis/make_tables.py

source/analysis/plot_figures.py

source/analysis/reanalyze_broad.py

source/data/broad/OLMo-2-0425-1B__code__t0.npy

source/data/broad/OLMo-2-0425-1B__code__t1.npy

source/data/broad/OLMo-2-0425-1B__code__t2.npy

source/data/broad/OLMo-2-0425-1B__code__t3.npy

source/data/broad/OLMo-2-0425-1B__code__t4.npy

source/data/broad/OLMo-2-0425-1B__math__t0.npy

source/data/broad/OLMo-2-0425-1B__math__t1.npy

source/data/broad/OLMo-2-0425-1B__math__t2.npy

source/data/broad/OLMo-2-0425-1B__math__t3.npy

source/data/broad/OLMo-2-0425-1B__math__t4.npy

source/data/broad/OLMo-2-0425-1B__t0.npy

source/data/broad/OLMo-2-0425-1B__t1.npy

source/data/broad/OLMo-2-0425-1B__t2.npy

source/data/broad/OLMo-2-0425-1B__t3.npy

source/data/broad/OLMo-2-0425-1B__t4.npy

source/data/broad/OLMo-2-1124-13B__t1.npy

source/data/broad/OLMo-2-1124-13B__t2.npy

source/data/broad/OLMo-2-1124-13B__t3.npy

source/data/broad/OLMo-2-1124-13B__t4.npy

source/data/broad/SmolLM2-1.7B-intermediate-checkpoints__code__t0.npy

source/data/broad/SmolLM2-1.7B-intermediate-checkpoints__code__t1.npy

source/data/broad/SmolLM2-1.7B-intermediate-checkpoints__code__t2.npy

source/data/broad/SmolLM2-1.7B-intermediate-checkpoints__code__t3.npy

source/data/broad/SmolLM2-1.7B-intermediate-checkpoints__code__t4.npy

source/data/broad/SmolLM2-1.7B-intermediate-checkpoints__math__t0.npy

source/data/broad/SmolLM2-1.7B-intermediate-checkpoints__math__t1.npy

source/data/broad/SmolLM2-1.7B-intermediate-checkpoints__math__t2.npy

source/data/broad/SmolLM2-1.7B-intermediate-checkpoints__math__t3.npy

source/data/broad/SmolLM2-1.7B-intermediate-checkpoints__math__t4.npy

source/data/broad/SmolLM2-1.7B-intermediate-checkpoints__t0.npy

source/data/broad/SmolLM2-1.7B-intermediate-checkpoints__t1.npy

source/data/broad/SmolLM2-1.7B-intermediate-checkpoints__t2.npy

source/data/broad/SmolLM2-1.7B-intermediate-checkpoints__t3.npy

source/data/broad/SmolLM2-1.7B-intermediate-checkpoints__t4.npy

source/data/broad/manifest.json

source/data/broad/profiles.csv

source/data/broad/pythia-1.4b__t2.npy

source/data/broad/source_audit.csv

source/data/historical/historical_claims.csv

source/data/historical/replicated_deltas.csv

source/data/historical/replicated_profiles.csv

source/data/historical/replicated_trajectories.csv

source/dense/build_dense_figures.py

source/dense/data/historical.json

source/dense/data/released_general.json

source/figure_sources.json

source/overview/build_overview.py

source/overview/fig1_empirical_base.pdf

source/requirements.txt

source/results/analysis_summary.json

source/results/checkpoint_pairs.csv

source/results/cross_domain.csv

source/results/cross_domain_bootstrap.csv

source/results/diagnostics/cross_model_descriptive.csv

source/results/diagnostics/exploratory_coarsegrain_null.csv

source/results/diagnostics/exploratory_pca.csv

source/results/diagnostics/paired_change_null.csv

source/results/diagnostics/retention_thresholds.csv

source/results/diagnostics/time_order_null.csv

source/results/heldout_bootstrap.csv

source/results/heldout_domain.csv

source/results/lag_summary.csv

source/results/panel_summary.csv

source/results/split_window_validation.csv

source/results/trajectory_balanced.csv

source/results/window_bootstrap.csv

supplement/README.md

supplement/focal/CHECKPOINT_REFERENCES.csv

supplement/focal/README.md

supplement/focal/continuous_output/implementation_notes.md

supplement/focal/continuous_output/numerical_validation.md

supplement/focal/continuous_output/preanalysis_protocol.md

supplement/focal/continuous_output/processed/checkpoint_endpoint_means.csv

supplement/focal/continuous_output/processed/endpoint_change_summary.csv

supplement/focal/continuous_output/processed/layer_checkpoint_summary.csv

supplement/focal/continuous_output/processed/layer_profile_comparison.csv

supplement/focal/continuous_output/processed/margin_cutpoints.json

supplement/focal/continuous_output/processed/margin_stratum_summary.csv

supplement/focal/continuous_output/processed/numerical_boundary_sensitivity.csv

supplement/focal/continuous_output/processed/pythia_run_changes.csv

supplement/focal/continuous_output/processed/smooth_model_coefficients.csv

supplement/focal/continuous_output/processed/smooth_model_leave_one_run_out.csv

supplement/focal/continuous_output/reproducibility_revision.md

supplement/focal/continuous_output/results.md

supplement/focal/continuous_output/revision_timing.md

supplement/focal/direction_controls/matching_validation.md

supplement/focal/direction_controls/numerical_validation.md

supplement/focal/direction_controls/pilot_validation.md

supplement/focal/direction_controls/preanalysis_protocol.md

supplement/focal/direction_controls/processed/adjusted_cell_layer_specificity.csv

supplement/focal/direction_controls/processed/adjusted_layer_specificity.csv

supplement/focal/direction_controls/processed/adjusted_output_specificity.csv

supplement/focal/direction_controls/processed/adjusted_specificity_effects.csv

supplement/focal/direction_controls/processed/boundary_cell_specificity.csv

supplement/focal/direction_controls/processed/boundary_specificity_effects.csv

supplement/focal/direction_controls/processed/family_checkpoint_attenuation.csv

supplement/focal/direction_controls/processed/final_norm_contribution_cell.csv

supplement/focal/direction_controls/processed/final_norm_contribution_effects.csv

supplement/focal/direction_controls/processed/heldout_increments.csv

supplement/focal/direction_controls/processed/heldout_performance.csv

supplement/focal/direction_controls/processed/leave_one_run_out_specificity.csv

supplement/focal/direction_controls/processed/olmo_adjusted_window_specificity.csv

supplement/focal/direction_controls/processed/olmo_window_bootstrap_specificity.csv

supplement/focal/direction_controls/processed/propagation_curves.csv

supplement/focal/direction_controls/processed/robustness_cell_layer.csv

supplement/focal/direction_controls/processed/robustness_summary.csv

supplement/focal/direction_controls/processed/unadjusted_run_specificity.csv

supplement/focal/direction_controls/processed/unadjusted_specificity_effects.csv

supplement/focal/direction_controls/results.md

supplement/focal/matched_perturbations/environment.md

supplement/focal/matched_perturbations/matching_validation.md

supplement/focal/matched_perturbations/pilot_validation.md

supplement/focal/matched_perturbations/preanalysis_protocol.md

supplement/focal/matched_perturbations/processed/heldout_increments.csv

supplement/focal/matched_perturbations/processed/heldout_performance.csv

supplement/focal/matched_perturbations/processed/heldout_plot_points.csv

supplement/focal/matched_perturbations/processed/matched_actual_norm_adjusted_run_effects.csv

supplement/focal/matched_perturbations/processed/olmo_adjusted_window_bootstrap.csv

supplement/focal/matched_perturbations/processed/olmo_descriptive_window_bootstrap.csv

supplement/focal/matched_perturbations/processed/primary_effects.csv

supplement/focal/matched_perturbations/processed/propagation_zone_effects.csv

supplement/focal/matched_perturbations/processed/propagation_zone_summary.csv

supplement/focal/matched_perturbations/processed/record_counts.json

supplement/focal/matched_perturbations/processed/representation_matching_qc.csv

supplement/focal/matched_perturbations/processed/representation_matching_qc_summary.json

supplement/focal/matched_perturbations/processed/robustness_summary.csv

supplement/focal/matched_perturbations/processed/run_level_effects.csv

supplement/focal/matched_perturbations/protocol_revision.md

supplement/focal/matched_perturbations/results.md

validation/figure_text_bounds.json

validation/numerical_checks.json

validation/overview_geometry.json

(133 additional files not shown) You must enabled JavaScript to view entire file list.

arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.

Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.

Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.