pith:ACIPAPZY
Embedding Perturbation may Better Reflect Intermediate-Step Uncertainty in LLM Reasoning
Embedding perturbations on preceding tokens reveal uncertainty in LLM reasoning steps better than existing methods.
arxiv:2602.02427 v2 · 2026-02-02 · cs.LG
Add to your LaTeX paper
\usepackage{pith}
\pithnumber{ACIPAPZYL7MOMYKA4MVAYLQX3C}
Prints a linked badge after your title and injects PDF metadata. Compiles on arXiv. Learn more · Embed verified badge
Record completeness
Claims
an LLM's incorrect reasoning steps tend to contain tokens which are highly sensitive to the perturbations on the preceding token embeddings, indicating the model's uncertainty among multiple competing continuations
That the observed sensitivity to embedding perturbations directly reflects the model's internal uncertainty in reasoning steps rather than being driven by other factors such as token frequency or architecture-specific artifacts.
Embedding perturbation sensitivity provides a stronger signal for uncertainty in intermediate LLM reasoning steps than probability, sampling, or Bayesian baselines.
Formal links
Receipt and verification
| First computed | 2026-05-17T23:39:16.432927Z |
|---|---|
| Builder | pith-number-builder-2026-05-17-v1 |
| Signature | Pith Ed25519
(pith-v1-2026-05) · public key |
| Schema | pith-number/v1.0 |
Canonical hash
0090f03f385fd8e66140e32a0c2e17d8940b1c79c972160dc6fe81dac406c767
Aliases
· · · · ·Agent API
Verify this Pith Number yourself
curl -sH 'Accept: application/ld+json' https://pith.science/pith/ACIPAPZYL7MOMYKA4MVAYLQX3C \
| jq -c '.canonical_record' \
| python3 -c "import sys,json,hashlib; b=json.dumps(json.loads(sys.stdin.read()), sort_keys=True, separators=(',',':'), ensure_ascii=False).encode(); print(hashlib.sha256(b).hexdigest())"
# expect: 0090f03f385fd8e66140e32a0c2e17d8940b1c79c972160dc6fe81dac406c767
Canonical record JSON
{
"metadata": {
"abstract_canon_sha256": "85e5e3eea633bf39d658a3659a0f1928a9ba19b08e4f5c8bada3f5a6ddf99818",
"cross_cats_sorted": [],
"license": "http://creativecommons.org/licenses/by/4.0/",
"primary_cat": "cs.LG",
"submitted_at": "2026-02-02T18:27:26Z",
"title_canon_sha256": "052935d791537b07cfb9f5d46b6fb642e3c0ab6124d30f7114e9c46e5eb899e9"
},
"schema_version": "1.0",
"source": {
"id": "2602.02427",
"kind": "arxiv",
"version": 2
}
}