README.md

metadata

license: cc-by-4.0
pretty_name: What Arranges Features in Activation Space? (checkpoints and analysis)
tags:
  - interpretability
  - computational-mechanics
  - belief-states

What Arranges Features in Activation Space? Checkpoints and analysis outputs

Trained checkpoints and pre-computed analysis outputs for the NeurIPS 2026 paper "What Arranges Features in Activation Space? Non-Classical Predictive Geometry in Next-Token Predictors" (arXiv:2507.07432). Code: https://github.com/adamimos/what-arranges-features.

The original single-seed (seed 42) models used for the geometry panels of Fig. 2, Fig. 3 and the RSA figure are in a separate dataset, SimplexAI/quantum-representations.

Upload in progress (November 2026). The files listed below are being uploaded; until then only this card is present.

Layout

Download into the root of the code repository (see its README); paths match what the scripts expect.

Path	Contents
results/multiseed_{bloch,mess3,moon,frdn}_transformer/	5 seeds (0, 1, 2, 3, 42) x transformer
results/multiseed_{bloch,stage2,frdn}_rnn/	5 seeds x LSTM / GRU / vanilla RNN (stage2 = Mess3 and Moon)
results/qrrxor_robust_{transformer,lstm,gru,rnn}/	qRRXOR, 5 report seeds (42, 123, 456, 789, 1000)
results/20260424220339/, results/20260504050111_lstm_b64_full/	qRRXOR transformer and LSTM shown in Fig. 4
results/20260503215800_lstm_default_cpu/	qRRXOR LSTM at lr 1e-4 / batch 16 (comparison runs)
results/paper_metrics/	Fig. 2 and appendix bar metrics, order-k loss, PCA views
results/*.csv, results/causal_transplant/, results/heldout_suffix/	scores behind the paper's numbers, including the scores (not the checkpoints) of the qRRXOR hyperparameter screen (dev seeds 7, 11, 13)
shuffle_results_quantum_rrxor_hero_3seed/, figures/beyond_ntp/	cached curves read by the Fig. 4 script

Each run directory run_<i>_.../ holds a checkpoint every 100 epochs, named by tokens seen (0.pt is the untrained network), plus loss.csv, log.csv, log.json, and the run config (run_config.yaml, and hooked_model_config.json or model_config.json). Each sweep directory has the sweep_config.yaml it was launched from.

Notes

• A run counts as converged when the last val_loss_mean in loss.csv (normalized by the optimal loss) is below 1.005. All runs in the core dataset converge except some seeds of 20260503215800_lstm_default_cpu, which are included for that comparison.
• The shuffle columns of qrrxor_robust_scores.csv and qrrxor_robust_grurnn_scores*.csv come from a single no-PCA shuffle and are not the paper's numbers; the paper's shuffle results are in qrrxor_shuffle_null_analysis.csv.