Re-checking 15 popular benchmarks wherever a frontier model failed an item, MIT researchers found many “riddled with errors (e.g., 5% of GSM8K)”, and for most benchmarks more than half of model failures could be traced to label noise rather than to the model; they describe their counts as a lower bound.

SocioLens project,

Claim LD06-SL-C03 (v1.0)

Provenance and status

Claim type: Estimate
Publication status: Eligible for publication, with qualifiers: preprint
Verification state: Checked against the primary source; the source is a preprint. Who checked this claim is not recorded for it; claim checks so far were AI-led, and no person has checked claims against their sources.
Does not license: A full error rate for any benchmark (items every model got right were not checked)
Limitations: Lower bound by design; only items a model failed were inspected; venue not confirmed
Source (as cited): Vendrow, Vendrow, Beery & Madry, “Do Large Language Model Benchmarks Test Reliability?,” arXiv:2502.03461, February 2025
Source type: Preprint (recorded as “Preprint”)
Institutional position: Independent academic (MIT)
Maturity of the evidence: Emerging evidence
Recorded date: 15 October 2026
Currentness / correction: This is the current version. No correction or concern is recorded on its sources.

Sources this version cites

Vendrow, Vendrow, Beery & Madry, “Do Large Language Model Benchmarks Test Reliability?,” arXiv:2502.03461, February 2025 (Preprint)
Read the original

Version

This is the current version, v1.0. Earlier versions stay available, and their records never change.

Versions: v1.0 · current

Raw record: every recorded field, with its coded values

All 25 canonical Claim fields, verbatim, with the coded values derived from them. The plain-language labels above are read from these.

Claim type (coded): ESTIMATE (as printed: “ESTIMATE”)
Publication status (coded): ELIGIBLE (PREPRINT) (as printed: “ELIGIBLE WITH PREPRINT QUALIFIER”)
Verification state (coded): CONTENT_VERIFIED_PRIMARY (PREPRINT)
Source type (coded): PREPRINT
Maturity (coded): EMERGING

#	Canonical field	Canonical value (verbatim)	Normalized (Claim Value Vocabulary v1.0.2)
1	CLAIM ID	LD06-SL-C03	—
2	PUBLIC-FACING CLAIM	Re-checking 15 popular benchmarks wherever a frontier model failed an item, MIT researchers found many “riddled with errors (e.g., 5% of GSM8K)”, and for most benchmarks more than half of model failures could be traced to label noise rather than to the model; they describe their counts as a lower bound.	—
3	CLAIM TYPE	ESTIMATE	base_types: ESTIMATE
4	RESEARCH STRAND	A — Benchmark validity and measurement	—
5	MATURITY	EMERGING	states: EMERGING
6	SOURCE	Vendrow, Vendrow, Beery & Madry, “Do Large Language Model Benchmarks Test Reliability?,” arXiv:2502.03461, February 2025	source_ids: SRC-SL-VENDROW-ET-AL-ARXIV-2502-03461
7	SOURCE TYPE	Preprint	classes: PREPRINT
8	INSTITUTIONAL POSITION	Independent academic (MIT)	—
9	SCOPE	Subsets of 15 benchmarks treated as saturated, including GSM8K and reading-comprehension sets	—
10	DESIGN	Manual inspection of every item that any tested frontier model got wrong; flawed items removed or relabelled	—
11	SAMPLE / DATASET	Revised subsets of 15 benchmarks	—
12	MODEL / TOOL / VERSION	A range of frontier language models available in early 2025	—
13	COMPARATOR	The benchmarks’ original answer keys	—
14	MAIN RESULT	About 5% of GSM8K items erroneous; for the majority of benchmarks more than half of model failures attributable to label noise; issues with up to 30% of examples in some reading-comprehension sets; frontier models still fail simple items on the cleaned versions	—
15	INFERENTIAL TARGET	The share of benchmark error that is in the benchmark rather than the model	—
16	DOES NOT LICENSE	A full error rate for any benchmark (items every model got right were not checked)	—
17	VERIFICATION STATE	CONTENT VERIFIED PREPRINT	state: CONTENT_VERIFIED_PRIMARY · qualifiers: PREPRINT · verified_by: UNKNOWN
18	LIMITATIONS	Lower bound by design; only items a model failed were inspected; venue not confirmed	—
19	CONTRADICTORY / SUPPORTING SOURCES	Consistent with LD06-SL-C04 (MMLU) and with the benchmark-wide GSM8K figure in LD06-SL-C01	—
20	GEOGRAPHY / LANGUAGE	Not checked / English	—
21	DISTRIBUTIONAL LEDGER	Not analysed in this claim	—
22	R01 / FIELD-MAP LINKS	LD-06, D4b	—
23	WHAT WOULD CHANGE THIS CLAIM?	Full audits of all items, including those every model answers the same way	—
24	PUBLICATION STATUS	ELIGIBLE WITH PREPRINT QUALIFIER	state: ELIGIBLE · qualifiers: PREPRINT
25	RECORDED DATE	2026-10-15	—

Field 22 references resolved against Field Map v0.2: LD-06 (LIVING_DEBATE); D4b (SUB_AREA)

The sources as recorded, with their record IDs and coded source types:

Pinned Sources (field 6, version-specific)

“Vendrow, Vendrow, Beery & Madry, “Do Large Language Model Benchmarks Test Reliability?,” arXiv:2502.03461, February 2025” → Vendrow et al. (2025), “Do Large Language Model…” (SRC-SL-VENDROW-ET-AL-ARXIV-2502-03461--v1.0) — Vendrow, Vendrow, Beery & Madry, “Do Large Language Model Benchmarks Test Reliability?,” arXiv:2502.03461, February 2025; class PREPRINT (explicit public provenance pin)

Cited by Living Debate versions

The evaluation apparatus: Can benchmarks, leaderboards and audits carry the weight placed on them? (v0.1) LD-06

Cite this page

SocioLens project. (2026, October 15). Re-checking 15 popular benchmarks wherever a frontier model failed an item, MIT researchers found many “riddled with errors (e.g., 5% of GSM8K)”, and for most benchmarks more than half of model failures could be traced to label noise rather than to the model; they describe their counts as a lower bound [Claim LD06-SL-C03, version v1.0]. SocioLens. https://sociolens-web.pages.dev/claims/ld06-sl-c03/v1.0/

@misc{sociolens_ld06_sl_c03_v1_0,
  author = {{SocioLens project}},
  title = {Re-checking 15 popular benchmarks wherever a frontier model failed an item, MIT researchers found many “riddled with errors (e.g., 5\% of GSM8K)”, and for most benchmarks more than half of model failures could be traced to label noise rather than to the model; they describe their counts as a lower bound.},
  year = {2026},
  date = {2026-10-15},
  howpublished = {\url{https://sociolens-web.pages.dev/claims/ld06-sl-c03/v1.0/}},
  note = {Claim LD06-SL-C03, version v1.0}
}

The link points to this version, whose record never changes. For a finding, cite the original study as well: SocioLens summarizes research and does not replace it.

Research-of-record provenance

Materialized by sociolens-materializer 0.6.0 at 2026-09-28T20:00:00.000Z; normalization contract SocioLens — Claim Value Vocabulary v1.0.2 — Pre-Spike Materialization Contract (v1.0.2). Content seal SHA-256 b9a5710d869eee9beeb1883ae6d41835914d70a3b754720f57bccaa25bdeacc0.

Repository-authored inputs (not Drive exports)

SocioLens claim — LD06-SL-C03 v1.0 (v1.0) — written in the SocioLens repository by Project Lead (Claude, AI) on 2026-10-15; no Drive identity exists. Authority: AGENTS.md "Authority" (the Project Lead decides research content within the research-integrity rules) and the Owner's choice in the SocioLens project chat on 15 October 2026 to build the LD-06 debate; a new Claim, first version, written from the project's LD-06 source notes, checked against full texts or official abstracts. Block “ENTRY 03 — LD06-SL-C03” SHA-256 8972da3a39e30cb192940a0b75a4d963af2fbf3b434f74a2debb28bdcdc1d35a; fixture fixtures/authored/claims/LD06-SL-C03--v1.0.

Builder decisions (not Drive data)

VERSION = v1.0 — basis: Claim version taken from the repository-authored claim fixture version (SocioLens claim — LD06-SL-C03 v1.0); the first version of a new Claim of the LD-06 debate
CURRENT_POINTER = true — basis: repository-authored claim "SocioLens claim — LD06-SL-C03 v1.0" (Project Lead (Claude, AI), 2026-10-15); basis: AGENTS.md "Authority" (the Project Lead decides research content within the research-integrity rules) and the Owner's choice in the SocioLens project chat on 15 October 2026 to build the LD-06 debate; a new Claim, first version, written from the project's LD-06 source notes, checked against full texts or official abstracts
FIELD_6_SOURCE_REFERENCE = SRC-SL-VENDROW-ET-AL-ARXIV-2502-03461--v1.0 — basis: field 6 of ENTRY 03 — LD06-SL-C03 cites "Vendrow, Vendrow, Beery & Madry, “Do Large Language Model Benchmarks Test Reliability?,” arXiv:2502.03461, February 2025"; the only field-6 segment of this ClaimVersion's repository-authored ledger block.
SOURCE_PROVENANCE_PIN = SRC-SL-VENDROW-ET-AL-ARXIV-2502-03461--v1.0 — basis: explicit version-pinned public provenance of this ClaimVersion (Amendment v1.2 §6): the pinned field-6 Source record "Vendrow, Vendrow, Beery & Madry, “Do Large Language Model Benchmarks Test Reliability?,” arXiv:2502.03461, February 2025".