
Once a model writes its first wrong fact, are the later errors caused by that commitment, or by a truth the model knew and suppressed? In 111 wrong answers from three instruction-tuned models, the first wrong fact was located and replaced, and the model continued writing. Study B of the thesis MRI of a Mad Machine; the hypotheses, threshold and tests were pre-specified before scoring.
Result
| Condition | Folder label | What the model sees |
|---|---|---|
| A · as generated | a_baseline |
Its own original answer |
| B · correct fact inserted | b_substituted |
The gold answer at the point where it first went wrong |
| C · asked in isolation | c_isolated |
The question again, with no earlier answer |
| D · different wrong fact | d_control |
A different wrong answer at the same point |
| Result | Number |
|---|---|
| Answers with a later error, before → after the repair | 92.5% → 48.1% (50 fixed, 3 broken, n = 106, p = 5.5 × 10⁻¹²) |
| Second annotator, independently | 97.2% → 50.0% (p = 2.35 × 10⁻¹⁴) |
| Correct fact vs a different wrong fact | 54.6% vs 60.8% error-free (n = 97, p = 0.38) |
| Model correct when asked on its own | 2 of 111 |
The later errors are elaborations of the model's own committed claim. Replace that claim with anything at all and about half of them disappear; the models were not hiding the truth.
Scope. Instruction-tuned models served through an API (GPT-OSS-120B, Qwen3.6-27B, ALLaM-2-7B) at temperature 0, scored by two annotators working independently (κ = 0.94 for downstream error, 1.00 for isolated correctness).
What is inside
| Path | Contents |
|---|---|
data/questions.json |
200 SimpleQA questions, random sample with seed 20260819 |
data/baseline_*.json |
Raw baseline answers, one file per model |
data/qualifying_items.json |
128 candidate items with divergence point and replacements |
data/annotator_A.csv, data/annotator_B.csv |
Independent scores for all 444 rows (111 items × 4 conditions) |
code/ |
Sampling, model querying, agreement and analysis scripts |
STUDY_README.md |
The full study page |
Reproduce
from huggingface_hub import snapshot_download
path = snapshot_download("rejaulkorimsadi/commitment-test-llm-hallucination", repo_type="dataset")
cd <the path printed above>
pip install pandas numpy scipy statsmodels
python code/analyze_study2.py
python code/analyze_study2.py data/annotator_B.csv
Citation
@misc{sadi2026commitment,
title = {The Commitment Test: prefix substitution for commitment-driven hallucination},
author = {Sadi, Md. Rejaul Korim and Tasin, Toufiqur Rahman},
year = {2026},
note = {Study B of the B.Sc. thesis "MRI of a Mad Machine", Metropolitan University, Sylhet; also reported in arXiv:2606.07537}
}
Part of a research collection
Authors' affiliation: Department of Computer Science and Engineering, Metropolitan University, Sylhet, Bangladesh. Questions and feedback: open a discussion on this page.
- Downloads last month
- -