FindingQwen/Qwen2.5-7B-Instruct + google/gemma-3-12b2026-05-11 · by caiovicentino

NLA two-tier verbalization — uniform fve_nrm decoupled from category-spread recall (Qwen2.5-7B + Gemma-3-12B)

N=150. Reconstruction fve_nrm UNIFORM 0.880 across chat/code/reasoning/agent. Keyword recall MASSIVELY category-dependent (chat 0.578 / agent 0.088 = 6.5×). Better-trained NLA → smaller fve_nrm spread but LARGER recall spread (decoupling magnification).

Paper Manifest Raw JSON

Numbers

n_samples

150

fve_nrm_uniform

0.880

fve_nrm_spread

0.017

recall_spread

0.490

permutation_gap

0.270

direction_injection_alignment_qwen

direction_injection_alignment_gemma

0.750

Artifacts

phase16_data.jsonphase17_pilot.ipynb

Cite

Content-only sha256 below. Verifiable: re-hash the JSON manifest (with manifest_sha256 set to null, sort_keys=True) and you get the same digest. Zenodo DOI pending.

manifest_sha256

e328cd066f6ffe53ebb5c139da9a1be16c8a5acd02473806328e6cd0ce1e421c

Atlas URL

https://openinterp.org/atlas/e328cd066f

Raw manifest

https://raw.githubusercontent.com/OpenInterpretability/registry/main/atlas/2026/e328cd066f.json

Reproduce this in your agent

In an agent session attached to your Colab via openinterp-mcp:

from openinterp_mcp.atlas import load_entry

entry = load_entry("e328cd066f")
print(entry.methodology_check)

# Re-run the causality protocol against the linked HF artifact:
# (no HF artifact attached — replicate from methodology alone)

First result in 10 minutes