KingTechnician/shared-truth-results

Shared Truth: experiment results

Saved results for Shared Truth: Emergent Truth Properties in Large Language Models via Heterogeneous Injection-based Transfer (Freeman & Ngangmeni, NeurIPS 2026). The code that produced these files, and the notebooks that re-check them, are in KingTechnician/shared-truth.

The revision tagged camera-ready is the one the paper's numbers come from.

Layout

strict_overshoot_bootstrap.json        bootstrap CIs for strict overshoot (22 rows, 21 unique trajectories)
procrustes_n21_summary.json            Procrustes vs trained adapters, per pair + aggregates (22 rows, 21 unique)
procrustes_figure_pairs_results.json   Procrustes on the three dual-trajectory figure pairs
artifacts/                             final tables and figures as used in the paper
threshold-analysis/                    per-alpha thresholds and per-pair summaries behind the strict-overshoot tables
_mirror/manifest.json                  manifest of the original Drive mirror (checksums)
sweep-results/
  <pair_id>/                           one folder per run
    sweep_results.json                 the alpha sweep (see "Per-run files")
    procrustes_sweep.json              Procrustes sweep on the same eval (published pairs only)
    fig1_alpha_sweep, fig2_score_distributions, fig3_geometry_collapse   (.pdf + .png)
  activation_dumps/<trajectory>.npz    cached activations for the three multi-seed trajectories
  analysis_statement_types/*.csv       Tables A-D (+ long-form metrics)
  analysis_residuals/*.csv             Tables E and F
  figures_rebuttal/                    figures R1-R8

Run identifiers

A pair_id is <source>_l<N>_to_<target>_l<M>__<variant>, for example gemma2-2b-instruct_l13_to_llama3_2-3b-instruct_l12__truth. The variant says how the run was set up:

variant	runs	eval	what it is
truth, transfer_truth, transfer_rsa	22 (21 unique)	simple statements, n=1599	the published trajectories. truth / transfer_truth target the model's truth-probe layer; transfer_rsa uses the layer pair chosen by cross-model RSA
truth_origadapter, truth_seed42/43/44	12	simple statements, n=1599	multi-seed replication: the original adapter plus three retrained seeds on three trajectories
stmt_origadapter, stmt_seed42/43/44	12	full test set, n=2684	the same 12 adapters, swept on all statement types
stmt_randbase, stmt_shufbase	6	full test set, n=2684	baselines: random-weight MLP and example-shuffled adapter outputs

The three multi-seed trajectories are Gemma-2-2B (L13) → Llama-3.2-3B (L12), Llama-3.2-3B (L12) → Gemma-2-2B (L13), and Gemma-2-2B (L13) → Qwen2.5-1.5B (L17), the lowest-overshoot pair.

"Simple statements" means statement_type == "simple" in the test split of KingTechnician/tiu-splits, both polarities (1080 affirmed + 519 negated). The full test set adds conjunctions and disjunctions.

What backs what

result	files	runs
Strict overshoot per pair (Table 7)	artifacts/table_strict_overshoot.{csv,tex}, threshold-analysis/	the 22 published folders
Overshoot regressions	artifacts/table_regressions.tex, threshold-analysis/per_pair_summary.csv	the 21 unique published pairs
Bootstrap CIs (17/21 exclude zero)	strict_overshoot_bootstrap.json	the 22 published folders
Procrustes baseline (18/21 match or beat)	procrustes_n21_summary.json, <pair_id>/procrustes_sweep.json	the 22 published folders
Dual-trajectory figure	artifacts/fig_dual_trajectory.*, procrustes_figure_pairs_results.json	qwen2_5-1_5b-instruct_l17_to_llama3_2-3b-instruct_l12__truth, llama3_2-3b-instruct_l12_to_mistral-7b-instruct_l14__transfer_rsa, qwen2_5-7b-instruct_l21_to_gemma-7b-instruct_l19__truth
Score-distribution figures	artifacts/fig_scores_<pair_id>.*	the pair named in each file
Multi-seed replication (rebuttal)	<pair_id>/sweep_results.json	the 12 truth_* multi-seed runs
Statement-type Tables A–D, figures R1–R2	analysis_statement_types/*.csv, figures_rebuttal/figR1-2*	the 12 stmt_* trained runs
Baseline Table E, figures R3–R5	analysis_residuals/tableE_baseline_vs_trained.csv, figures_rebuttal/figR3-5*	stmt_origadapter, stmt_randbase, stmt_shufbase
Residual geometry Table F, figures R6–R8	analysis_residuals/tableF_residual_geometry.csv, figures_rebuttal/figR6-8*	activation_dumps/*.npz
Random-weight null (1000 seeded draws)	artifacts/random_weight_null_N1000.json	activation_dumps/*.npz

Per-run files

sweep_results.json records the run's models and layers, the adapter (repo, subfolder, and the adapter's config.json including its training metrics), the probe repos, the native source/target ceilings, and alpha_sweep: for each α in 0.0–1.0, accuracy at the probe's own threshold, AUROC, and the per-class score lists true_scores / false_scores. The stmt_* runs also store per-item scores, preds, and items metadata (statement type, polarity, topic). The older runs store only the per-class scores.

Each activation_dumps/<trajectory>.npz holds, for the full test set: h_S and h_T (source and target activations), map_orig and map_seed42/43/44 (trained adapter outputs), rand_out and shuf_out (the two baselines), the target probe's truth directions t_G and t_P, and labels, statement_type, polarity. All activation arrays are fp16.

Loading

from huggingface_hub import snapshot_download
root = snapshot_download("KingTechnician/shared-truth-results", repo_type="dataset",
                         revision="camera-ready",
                         allow_patterns=["sweep-results/*/*.json", "*.json"])   # ~100 MB; add the dumps if needed

Or, from the code repo: shared_truth.storage.fetch_results(include_dumps=True, include_analysis=True).

Known issues and provenance notes

• Duplicated trajectory. llama3_2-3b-instruct_l12_to_qwen2_5-1_5b-instruct_l12 appears as both __transfer_rsa and __transfer_truth: the same adapter and probes, swept twice with identical scores. The consolidated files (strict_overshoot_bootstrap.json, procrustes_n21_summary.json) list it twice. Counted once, there are 21 trajectories: bootstrap CIs exclude zero on 17/21, and Procrustes matches or beats the trained adapter on 18/21.
• Two evaluations of the multi-seed adapters. truth_* runs use simple statements (n=1599) and stmt_* runs the full test set (n=2684). Quote multi-seed numbers with the eval they come from.
• Floor-pair seed checkpoints. For Gemma-2-2B → Qwen2.5-1.5B, the truth_seed42/43/44 runs used adapter checkpoints that were later retrained in place: their recorded metrics show fvu ≈ 0.038 and cosine ≈ 0.69, against fvu ≈ 0.005 and cosine ≈ 0.72 for the stmt_seed* runs, the activation dumps, and the adapter repos today. The other two trajectories used identical checkpoints in both evals. For those three floor-pair runs, truth_* vs stmt_* differences mix the eval with the weights.
• Random-weight baseline. Each stmt_randbase run is a single random MLP drawn without a seed, so it cannot be regenerated. Its strict overshoot (0.00 on all three trajectories) is one draw; a seeded 1000-draw null is the reference for random-map comparisons. The shuffled baseline is reproducible (permutation from default_rng(42)).
• Re-probe AUROC (Table F, auroc_reprobe) comes from a nearly unregularized logistic regression and varies by up to ~0.006 across numerical environments. Every other Table E/F column reproduces exactly. Quote re-probe values at 2 decimals.
• Adapter revisions. Runs saved here record the adapter repo and its config.json but not the commit. shared_truth.storage.verify_adapter checks a saved run against what the adapter repo serves now; runs made with the current code also record the commit.

Citation

@inproceedings{freeman2026sharedtruth,
  title     = {Shared Truth: Emergent Truth Properties in Large Language Models via Heterogeneous Injection-based Transfer},
  author    = {Freeman, Isaiah and Ngangmeni, Joed},
  booktitle = {Advances in Neural Information Processing Systems},
  year      = {2026}
}