analiza aiqbench
This research investigates the decoupling of internal state signals from reusable action maps in large language models, specifically analyzing affine closure properties within Qwen/Qwen3 architectures. The study demonstrates that while certain hidden states remain causally usable without supporting a global action map, this separability is bounded rather than universal across different model layers and entity splits. These findings offer critical insights into the geometric constraints of neural representations for researchers studying mechanistic interpretability and causal structure in AI systems.
A Calibrated Test of Internal Action Maps: State Signals Without Global Affine Closure
arXiv:2608.13626v1 Announce Type: cross Abstract: A hidden state signal can be decodable or causally usable without supporting a reusable action map. We test whether action maps fitted without a source reach its natural post-action activation and compose. We organize the tests as an evidence lattice and validate the geometric branch on a known affine S_5 carrier: all held-source folds pass one-step, composition, inverse, decoding, and commutativity gates. Structured curvature and held-domain conjugacy raise error monotonically, but only 23/30 strongest cells flip a closure gate, bounding rather than universalizing calibration. In post-trained Qwen/Qwen3-4B, frozen final-token h28 affine maps have mean held-entity error .519, versus .398 for within-test-domain cross-fit. Seven randomized entity splits and map geometry do not support a purely entity-specific account. Earlier h4/h16 layers fit one-step transitions better, but h4 conflict-state decoding is weak and lexical controls remain unresolved. Three matched intervention datasets regenerated from one frozen checkpoint show causal effects only at h28/h36. Outcome-aware refitting improves h28 one-step error to .474 (.469 with weighting), yet no refit passes composition. Learned finite worlds likewise preserve relative algebraic signals or shared charts without held-source affine closure. Within the tested carriers, state availability, causal use, local geometry, and reusable closure are separable. The result is limited to one pretrained model, sampled final-token layers, two finite worlds, and the tested affine or diagnostic function classes.
Abstract:A hidden state signal can be decodable or causally usable without supporting a reusable action map. We test whether action maps fitted without a source reach its natural post-action activation and compose. We organize the tests as an evidence lattice and validate the geometric branch on a known affine S_5 carrier: all held-source folds pass one-step, composition, inverse, decoding, and commutativity gates. Structured curvature and held-domain conjugacy raise error monotonically, but only 23/30 strongest cells flip a closure gate, bounding rather than universalizing calibration. In post-trained Qwen/Qwen3-4B, frozen final-token h28 affine maps have mean held-entity error .519, versus .398 for within-test-domain cross-fit. Seven randomized entity splits and map geometry do not support a purely entity-specific account. Earlier h4/h16 layers fit one-step transitions better, but h4 conflict-state decoding is weak and lexical controls remain unresolved. Three matched intervention datasets regenerated from one frozen checkpoint show causal effects only at h28/h36. Outcome-aware refitting improves h28 one-step error to .474 (.469 with weighting), yet no refit passes composition. Learned finite worlds likewise preserve relative algebraic signals or shared charts without held-source affine closure. Within the tested carriers, state availability, causal use, local geometry, and reusable closure are separable. The result is limited to one pretrained model, sampled final-token layers, two finite worlds, and the tested affine or diagnostic function classes.
Artificial Intelligence (cs.AI); Computation and Language (cs.CL); Machine Learning (cs.LG)
arXiv:2608.13626 [cs.AI]
(or arXiv:2608.13626v1 [cs.AI] for this version)
https://doi.org/10.48550/arXiv.2608.13626 Focus to learn more arXiv-issued DOI via DataCite (pending registration)
ARXIV_SNAPSHOT_README.txt
FILE_MANIFEST.sha256
SOURCE_COMMIT.txt
configs/phase0/qwen3_4b_nnsight_smoke.yaml
configs/phase1/qwen3_4b_alchemy_paired_intervention.yaml
configs/phase1/qwen3_4b_alchemy_paired_intervention_h16.yaml
configs/phase1/qwen3_4b_alchemy_paired_intervention_h28.yaml
configs/phase1/qwen3_4b_alchemy_paired_intervention_h28_seed2.yaml
configs/phase1/qwen3_4b_alchemy_paired_intervention_h28_seed3.yaml
configs/phase1/qwen3_4b_alchemy_paired_intervention_h36.yaml
configs/phase1/qwen3_4b_alchemy_probe.yaml
configs/phase1/qwen3_4b_alchemy_probe_direct.yaml
configs/phase1/qwen3_4b_alchemy_probe_focused.yaml
configs/phase2/qwen3_4b_grounded_operator_causal.yaml
configs/phase2/qwen3_4b_grounded_operator_causal_seed2.yaml
configs/phase2/qwen3_4b_grounded_operator_causal_seed3.yaml
configs/phase2/qwen3_4b_grounded_operator_fit.yaml
configs/phase2/qwen3_4b_grounded_operator_fit_alpha_expanded.yaml
configs/phase2/qwen3_4b_grounded_operator_fit_seed2.yaml
configs/phase2/qwen3_4b_grounded_operator_fit_seed3.yaml
configs/phase2/qwen3_4b_grounded_operator_mlp_upper.yaml
configs/phase3/qwen3_4b_h2_composition_seed1.yaml
configs/phase3/qwen3_4b_h2_composition_seed2.yaml
configs/phase3/qwen3_4b_h2_composition_seed3.yaml
configs/phase3/qwen3_4b_h3_commutativity_seed1.yaml
configs/phase3/qwen3_4b_h3_commutativity_seed2.yaml
configs/phase3/qwen3_4b_h3_commutativity_seed3.yaml
configs/phase3/qwen3_4b_phase3_5_construct_smoke.yaml
configs/phase3/qwen3_4b_phase3_5_construct_validation.yaml
configs/phase3/qwen3_4b_phase3_5b_manipulation_calibration.yaml
configs/phase4/synthetic_trajectories.yaml
configs/phase4/synthetic_trajectories_smoke.yaml
configs/phase4b/stage_ordering.yaml
configs/phase4b/stage_ordering_smoke.yaml
configs/phase4c/s5_cross_world.yaml
configs/phase4c/s5_cross_world_smoke.yaml
configs/phase4d/s5_operator_ladder.yaml
configs/phase4d/s5_operator_ladder_reproduction_shard.yaml
configs/phase4d/s5_operator_ladder_smoke.yaml
configs/phase4e/s5_operator_scaling.yaml
configs/phase4e/s5_operator_scaling_reproduction_shard.yaml
configs/phase4e/s5_operator_scaling_smoke.yaml
configs/phase4f/s5_aspect_ratio.yaml
configs/phase4f/s5_aspect_ratio_reproduction_shard.yaml
configs/phase4f/s5_aspect_ratio_smoke.yaml
configs/phase6/reviewer_calibration.yaml
configs/phase6/reviewer_calibration_smoke.yaml
configs/phase7/layer_local_composition.yaml
configs/phase8/reviewer_v2_diagnostics.yaml
configs/phase8/reviewer_v2_diagnostics_smoke.yaml
configs/phase8b/metric_attribution.yaml
paper/claim-intent-manifest.json
paper/terminology-ledger.md
reports/2026-08-08-phase0-qwen3-4b.md
reports/2026-08-08-phase1-qwen3-4b.md
reports/2026-08-08-phase2-grounded-operators.md
reports/2026-08-08-phase3-5-construct-validation.md
reports/2026-08-08-phase3-5b-manipulation-calibration.md
reports/2026-08-08-phase3-h2-composition.md
reports/2026-08-08-phase3-h3-commutativity.md
reports/2026-08-08-phase4-ground-truth-trajectories.md
reports/2026-08-08-phase4-postrun-audit.json
reports/2026-08-09-phase4b-postrun-audit.json
reports/2026-08-09-phase4b-stage-ordering.md
reports/2026-08-09-phase4c-postrun-audit.json
reports/2026-08-09-phase4c-s5-cross-world.md
reports/2026-08-09-phase4d-postrun-audit.json
reports/2026-08-09-phase4d-s5-operator-ladder.md
reports/2026-08-09-phase4e-s5-operator-scaling.md
reports/2026-08-09-phase4f-s5-aspect-ratio.md
reports/2026-08-11-phase6-reviewer-calibration-audit.json
reports/2026-08-11-phase7-layer-local-composition-audit.json
reports/2026-08-11-phase7-layer-local-composition.md
reports/2026-08-12-phase8-reviewer-v2-local-audit.json
reports/2026-08-12-phase8-reviewer-v2-remote-audit.json
reports/2026-08-12-phase8b-metric-attribution-local-audit.json
reports/2026-08-12-phase8b-metric-attribution-remote-audit.json
reports/audits/phase4e-s5-20260810T013925Z-independent-audit.json
reports/audits/phase4f-s5-20260810T030514Z-independent-audit.json
reports/phase3-5-construct-validation-preregistration.md
reports/phase3-5b-manipulation-calibration-preregistration.md
reports/phase3-h2-preregistration.md
reports/phase3-h3-preregistration.md
reports/phase4-ground-truth-trajectories-preregistration.md
reports/phase4b-stage-ordering-preregistration.md
reports/phase4c-s5-cross-world-preregistration.md
reports/phase4d-s5-operator-ladder-preregistration.md
reports/phase4e-s5-operator-scaling-preregistration.md
reports/phase4f-s5-aspect-ratio-preregistration.md
reports/phase6-reviewer-calibration-preregistration.md
reports/phase7-layer-local-composition-preregistration.md
reports/phase8-reviewer-v2-diagnostics-preregistration.md
reports/phase8b-metric-attribution-analysis-plan.md
requirements/dev.txt
requirements/remote.txt
scripts/analyze_phase0_reproducibility.py
scripts/analyze_phase3_5.py
scripts/analyze_phase3_h2.py
scripts/analyze_phase3_h3.py
scripts/analyze_phase4.py
scripts/audit_phase4b_stage_ordering.py
scripts/audit_phase4c_s5_cross_world.py
scripts/audit_phase4d_s5_operator_ladder.py
scripts/audit_phase4e_s5_operator_scaling.py
scripts/audit_phase4f_s5_aspect_ratio.py
scripts/audit_phase5_latex_pdf.py
scripts/audit_phase5_manuscript.py
scripts/audit_phase6_reviewer_calibration.py
scripts/audit_phase7_layer_local_composition.py
scripts/audit_phase8_reviewer_v2_diagnostics.py
scripts/audit_phase8b_metric_attribution.py
scripts/build_phase5_pdf.sh
scripts/finalize_artifact_manifest.py
scripts/generate_phase5_latex.py
scripts/merge_phase4d_s5_operator_shards.py
scripts/merge_phase4e_s5_operator_shards.py
scripts/merge_phase4f_s5_aspect_ratio_shards.py
scripts/phase0_nnsight_smoke.py
scripts/phase0_state_probes_legacy_smoke.py
scripts/phase1_paired_intervention.py
scripts/phase1_probe_scan.py
scripts/phase2_fit_operators.py
scripts/phase2_mlp_upper_bound.py
scripts/phase2_operator_causal.py
scripts/phase3_5_construct_validation.py
scripts/phase3_5b_manipulation_calibration.py
scripts/phase3_h2_composition.py
scripts/phase3_h3_commutativity.py
scripts/phase4_synthetic_trajectories.py
scripts/phase4b_stage_ordering.py
scripts/phase4c_s5_cross_world.py
scripts/phase4d_s5_operator_ladder.py
scripts/phase4e_s5_operator_scaling.py
scripts/phase4f_s5_aspect_ratio.py
scripts/phase6_reviewer_calibration.py
scripts/phase7_layer_local_composition.py
scripts/phase8_reviewer_v2_diagnostics.py
scripts/phase8b_metric_attribution.py
scripts/plot_phase5_h5_stitching.py
scripts/plot_phase5_measurement_ladder.py
scripts/plot_phase8_reviewer_v2.py
scripts/plot_reviewer_calibration.py
scripts/prepare_arxiv_submission.py
source_data/README.txt
source_data/h1_threshold_sensitivity.csv
source_data/h2_threshold_sensitivity.csv
source_data/h4_order_dz_sensitivity.csv
source_data/h5_stitching.csv
source_data/h5_stitching_metadata.json
source_data/reviewer_calibration_composition.csv
source_data/reviewer_calibration_depth.csv
source_data/reviewer_calibration_grounded.csv
source_data/reviewer_calibration_metadata.json
source_data/reviewer_calibration_positive.csv
source_data/reviewer_v2_entity_geometry.csv
source_data/reviewer_v2_layer_geometry.csv
source_data/reviewer_v2_lexical_controls.csv
source_data/reviewer_v2_metadata.json
source_data/reviewer_v2_metric_attribution.csv
source_data/reviewer_v2_structured_calibration.csv
src/verb_operator_algebra/__init__.py
src/verb_operator_algebra/alchemy.py
src/verb_operator_algebra/commutativity.py
src/verb_operator_algebra/commutativity_data.py
src/verb_operator_algebra/composition.py
src/verb_operator_algebra/composition_data.py
src/verb_operator_algebra/config.py
src/verb_operator_algebra/construct_causal.py
src/verb_operator_algebra/construct_data.py
src/verb_operator_algebra/construct_validation.py
src/verb_operator_algebra/layer_local_composition.py
src/verb_operator_algebra/manipulation_calibration.py
src/verb_operator_algebra/metric_attribution.py
src/verb_operator_algebra/operator_causal.py
src/verb_operator_algebra/operator_data.py
src/verb_operator_algebra/operator_fit.py
src/verb_operator_algebra/phase0.py
src/verb_operator_algebra/phase1_intervention.py
src/verb_operator_algebra/phase1_probe.py
src/verb_operator_algebra/reporting.py
src/verb_operator_algebra/reviewer_calibration.py
src/verb_operator_algebra/reviewer_v2_diagnostics.py
src/verb_operator_algebra/s5_aspect_ratio.py
src/verb_operator_algebra/s5_operator_diagnostics.py
src/verb_operator_algebra/s5_operator_scaling.py
src/verb_operator_algebra/s5_stage_analysis.py
src/verb_operator_algebra/s5_world.py
src/verb_operator_algebra/synthetic_analysis.py
src/verb_operator_algebra/synthetic_model.py
src/verb_operator_algebra/synthetic_stage_analysis.py
src/verb_operator_algebra/synthetic_world.py
tests/test_alchemy.py
tests/test_audit_phase4b_stage_ordering.py
tests/test_audit_phase4c_s5_cross_world.py
tests/test_audit_phase4d_s5_operator_ladder.py
tests/test_audit_phase4e_s5_operator_scaling.py
tests/test_audit_phase4f_s5_aspect_ratio.py
tests/test_audit_phase6_reviewer_calibration.py
tests/test_audit_phase7_layer_local_composition.py
tests/test_audit_phase8_reviewer_v2_diagnostics.py
tests/test_audit_phase8b_metric_attribution.py
tests/test_commutativity.py
tests/test_commutativity_data.py
tests/test_composition.py
tests/test_composition_data.py
tests/test_config.py
tests/test_construct_data.py
tests/test_construct_validation.py
tests/test_finalize_artifact_manifest.py
tests/test_layer_local_composition.py
tests/test_manipulation_calibration.py
tests/test_merge_phase4d_s5_operator_shards.py
tests/test_merge_phase4e_s5_operator_shards.py
tests/test_merge_phase4f_s5_aspect_ratio_shards.py
tests/test_metric_attribution.py
tests/test_operator_causal.py
tests/test_operator_data.py
tests/test_operator_fit.py
tests/test_phase1_intervention.py
tests/test_phase4b_config.py
tests/test_phase4c_config.py
tests/test_phase4d_config.py
tests/test_phase4e_config.py
tests/test_phase4f_config.py
tests/test_phase4f_s5_aspect_ratio.py
tests/test_phase6_config.py
tests/test_prepare_arxiv_submission.py
tests/test_reporting.py
tests/test_reviewer_calibration.py
tests/test_reviewer_v2_diagnostics.py
tests/test_s5_operator_diagnostics.py
tests/test_s5_operator_scaling.py
tests/test_s5_stage_analysis.py
tests/test_s5_world.py
tests/test_synthetic_analysis.py
tests/test_synthetic_stage_analysis.py
tests/test_synthetic_world.py
verb-operator-algebra-research-plan.md
(235 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.
Text integral sub licența deschisă a publicației · 570 cuvinte


