CoolFace
Datasetpublic

caiovicentino1/openinterp-psae-v15-marginal-fit-pathology

PSAE v1.5 — Marginal-Fit Pathology Honest-Negative Reproducibility artifacts for the paper The Marginal-Fit Pathology in Predictive SAE Feature Trajectory Probes (workshop submission, NeurIPS MI Workshop 2026). TL;DR: We trained linear probes on Qwen3.6-27B residuals to predict end-of-thinking SAE features from earlier-thinking residuals across L11/L31/L55. Naive recall@1024 = 0.83-0.87 looked paper-grade. The shuffled-source baseline B1 reproduces this within ±0.03 at all 12… See the full description on the dataset page: https://huggingface.co/datasets/caiovicentino1/openinterp-psae-v15-marginal-fit-pathology.

sourceHugging Faceapache-2.0updated 4mo agoView on Hugging Face
0likes44downloads
Dataset Card

PSAE v1.5 — Marginal-Fit Pathology Honest-Negative

Reproducibility artifacts for the paper The Marginal-Fit Pathology in Predictive SAE Feature Trajectory Probes (workshop submission, NeurIPS MI Workshop 2026).

TL;DR: We trained linear probes on Qwen3.6-27B residuals to predict end-of-thinking SAE features from earlier-thinking residuals across L11/L31/L55. Naive recall@1024 = 0.83-0.87 looked paper-grade. The shuffled-source baseline B1 reproduces this within ±0.03 at all 12 sites. The trivial constant baseline B2 ("predict the top-M most-common features in train, ignore input") reaches recall@1024 = 1.000 at L11/L31 — strictly exceeding the trained probe. The predictive claim does not survive.

Contents

PathSizeDescription
cache/residuals_multilayer.pt39 MBQwen3.6-27B residuals, 133 GSM8K prompts × 4 thinking-fractions × 3 layers
cache/features_multilayer.pt2.9 MBSAE TopK features (indices, values) at end-of-thinking, 3 layers
cache/thinking_traces.pt1.2 MBThinking-phase text traces
results/predictive_sae_v15_results.json7 KBOriginal PSAE v1.5 recall@k numbers (REAL)
results/random_baseline_results.json~30 KBB0 (SAE-init no-train) + B1 (shuffled-source) recall@k
results/feature_support_analysis.json~3 KBB2 trivial constant baseline + concentration stats
figures/recall_multilayer.png200 KBOriginal PSAE v1.5 figure
figures/recall_multilayer_with_baselines.png280 KBREAL + B0 + B1 + B2 overlay

Reproduction

python
import torch
residuals = torch.load("cache/residuals_multilayer.pt", weights_only=False)
features  = torch.load("cache/features_multilayer.pt",  weights_only=False)

# residuals[L][frac] = Tensor[N=133, d_model=5120]
# features[L][1.00]['indices'] = Tensor[N=133, k=128]  ← target

The SAE encoder weights live in `caiovicentino1/qwen36-27b-sae-papergrade`.

The two notebooks that produced these results:

  • —nb_predictive_sae_v1.ipynb — original REAL training
  • —nb_predictive_sae_v15_baseline.ipynb — B0 + B1 baselines (this paper)

Both are in OpenInterpretability/openinterp.

Citation

@misc{openinterp-psae-v15-marginal-fit-pathology-2026,
  author = {Vicentino, Caio},
  title  = {The Marginal-Fit Pathology in Predictive SAE Feature
            Trajectory Probes},
  year   = {2026},
  url    = {https://huggingface.co/datasets/caiovicentino1/openinterp-psae-v15-marginal-fit-pathology}
}

Part of the OpenInterpretability honest-negative methodology series. Sibling papers: Two Forms of Epiphenomenal Probes in Code Agents, Saturation-Direction Lever: A Five-Class Taxonomy of Probe Causality.