Models fine-tuned on facts of the form “A is B” failed to answer “B is A”; GPT-4 named the parent of 79% of 1,000 celebrities, but on the 1,573 child–parent pairs this produced it named the celebrity from the parent only 33% of the time.

SocioLens project,

Claim LD01-SL-C19 (v1.0)

Provenance and status

Claim type: Estimate
Publication status: Eligible for publication, with a qualifier
Verification state: Checked against the primary source. 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 claim about later models or about reasoning in context
Limitations: Read in arXiv version 4 (May 2024); the GPT-4 test is not controlled because its training data are unknown; humans also show some direction effects in recall
Source (as cited): Berglund, Tong, Kaufmann, Balesni, Cooper Stickland, Korbak & Evans, “The Reversal Curse: LLMs trained on ‘A is B’ fail to learn ‘B is A’,” ICLR 2024
Source type: Peer-reviewed conference paper (recorded as “Peer-reviewed conference paper”)
Institutional position: Mixed: independent academic (Vanderbilt, NYU, Sussex, Oxford) and AI-safety organizations (UK AI Safety Institute, Apollo Research)
Maturity of the evidence: Emerging evidence
Recorded date: 16 October 2026
Currentness / correction: This is the current version. No correction or concern is recorded on its sources.

Sources this version cites

Berglund, Tong, Kaufmann, Balesni, Cooper Stickland, Korbak & Evans, “The Reversal Curse: LLMs trained on ‘A is B’ fail to learn ‘B is A’,” ICLR 2024 (Peer-reviewed conference paper)
Read the original

Version

This page shows version v1.0. Its record never changes, even when newer evidence arrives; the site may add a note to the page or withhold a figure not cleared for public use.

This pinned version is also the current version.

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 (QUALIFIED) (as printed: “ELIGIBLE WITH QUALIFIER”)
Verification state (coded): CONTENT_VERIFIED_PRIMARY
Source type (coded): PEER_REVIEWED_CONFERENCE_PAPER
Maturity (coded): EMERGING

#	Canonical field	Canonical value (verbatim)	Normalized (Claim Value Vocabulary v1.0.2)
1	CLAIM ID	LD01-SL-C19	—
2	PUBLIC-FACING CLAIM	Models fine-tuned on facts of the form “A is B” failed to answer “B is A”; GPT-4 named the parent of 79% of 1,000 celebrities, but on the 1,573 child–parent pairs this produced it named the celebrity from the parent only 33% of the time.	—
3	CLAIM TYPE	ESTIMATE	base_types: ESTIMATE
4	RESEARCH STRAND	B — Evidence that could tell the accounts apart	—
5	MATURITY	EMERGING	states: EMERGING
6	SOURCE	Berglund, Tong, Kaufmann, Balesni, Cooper Stickland, Korbak & Evans, “The Reversal Curse: LLMs trained on ‘A is B’ fail to learn ‘B is A’,” ICLR 2024	source_ids: SRC-SL-BERGLUND-ET-AL-REVERSAL-ARXIV-2309-12288
7	SOURCE TYPE	Peer-reviewed conference paper	classes: PEER_REVIEWED_CONFERENCE_PAPER
8	INSTITUTIONAL POSITION	Mixed: independent academic (Vanderbilt, NYU, Sussex, Oxford) and AI-safety organizations (UK AI Safety Institute, Apollo Research)	—
9	SCOPE	Fictitious facts used for fine-tuning; real celebrities and their parents; English	—
10	DESIGN	Fine-tuning experiments and queries to a deployed model	—
11	SAMPLE / DATASET	Fictitious name–description pairs; 1,573 child–parent pairs	—
12	MODEL / TOOL / VERSION	GPT-3 and Llama-1 (fine-tuned); GPT-4 (queried)	—
13	COMPARATOR	The trained direction of the same facts	—
14	MAIN RESULT	Near-zero accuracy in the reverse direction after fine-tuning; GPT-4 79% against 33%; models can reverse the relation when the fact is in the prompt	—
15	INFERENTIAL TARGET	Whether models store facts as relations they can use in either direction	—
16	DOES NOT LICENSE	A claim about later models or about reasoning in context	—
17	VERIFICATION STATE	CONTENT VERIFIED	state: CONTENT_VERIFIED_PRIMARY · verified_by: UNKNOWN
18	LIMITATIONS	Read in arXiv version 4 (May 2024); the GPT-4 test is not controlled because its training data are unknown; humans also show some direction effects in recall	—
19	CONTRADICTORY / SUPPORTING SOURCES	LD01-SL-C18	—
20	GEOGRAPHY / LANGUAGE	English task set	—
21	DISTRIBUTIONAL LEDGER	Not analysed in this claim	—
22	R01 / FIELD-MAP LINKS	LD-01, T4, D3	—
23	WHAT WOULD CHANGE THIS CLAIM?	Models trained in the ordinary way that answer reversed questions as well as forward ones	—
24	PUBLICATION STATUS	ELIGIBLE WITH QUALIFIER	state: ELIGIBLE · qualifiers: QUALIFIED
25	RECORDED DATE	2026-10-16	—

Field 22 references resolved against Field Map v0.2: LD-01 (LIVING_DEBATE); T4 (TENSION); D3 (DOMAIN)

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

Pinned Sources (field 6, version-specific)

“Berglund, Tong, Kaufmann, Balesni, Cooper Stickland, Korbak & Evans, “The Reversal Curse: LLMs trained on ‘A is B’ fail to learn ‘B is A’,” ICLR 2024” → Berglund et al. (2024), “The Reversal Curse: LLMs trained…” (SRC-SL-BERGLUND-ET-AL-REVERSAL-ARXIV-2309-12288--v1.0) — Berglund, Tong, Kaufmann, Balesni, Cooper Stickland, Korbak & Evans, “The Reversal Curse: LLMs trained on ‘A is B’ fail to learn ‘B is A’,” ICLR 2024; class PEER_REVIEWED_CONFERENCE_PAPER (explicit public provenance pin)

Cited by Living Debate versions

What kind of thing is generative AI, and does the answer predict anything? (v0.1) LD-01

Cite this page

SocioLens project. (2026, October 16). Models fine-tuned on facts of the form “A is B” failed to answer “B is A”; GPT-4 named the parent of 79% of 1,000 celebrities, but on the 1,573 child–parent pairs this produced it named the celebrity from the parent only 33% of the time [Claim LD01-SL-C19, version v1.0]. SocioLens. https://sociolens-web.pages.dev/claims/ld01-sl-c19/v1.0/

@misc{sociolens_ld01_sl_c19_v1_0,
  author = {{SocioLens project}},
  title = {Models fine-tuned on facts of the form “A is B” failed to answer “B is A”; GPT-4 named the parent of 79\% of 1,000 celebrities, but on the 1,573 child–parent pairs this produced it named the celebrity from the parent only 33\% of the time.},
  year = {2026},
  date = {2026-10-16},
  howpublished = {\url{https://sociolens-web.pages.dev/claims/ld01-sl-c19/v1.0/}},
  note = {Claim LD01-SL-C19, 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 1653fc9b9758f327476cfd60e92dd322c64f9efc287ecf599eb98af320572bf0.

Repository-authored inputs (not Drive exports)

SocioLens claim — LD01-SL-C19 v1.0 (v1.0) — written in the SocioLens repository by Project Lead (Claude, AI) on 2026-10-16; 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 16 October 2026 to build the LD-01 debate; a new Claim, first version, written from the project's LD-01 source notes, checked against full texts, official web pages or official abstracts. Block “ENTRY 19 — LD01-SL-C19” SHA-256 e46c59d205b7f170bf4d0db82b6ede6d1874113ee9e9218dfd68d39992710827; fixture fixtures/authored/claims/LD01-SL-C19--v1.0.

Builder decisions (not Drive data)

VERSION = v1.0 — basis: Claim version taken from the repository-authored claim fixture version (SocioLens claim — LD01-SL-C19 v1.0); the first version of a new Claim of the LD-01 debate
CURRENT_POINTER = true — basis: repository-authored claim "SocioLens claim — LD01-SL-C19 v1.0" (Project Lead (Claude, AI), 2026-10-16); 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 16 October 2026 to build the LD-01 debate; a new Claim, first version, written from the project's LD-01 source notes, checked against full texts, official web pages or official abstracts
FIELD_6_SOURCE_REFERENCE = SRC-SL-BERGLUND-ET-AL-REVERSAL-ARXIV-2309-12288--v1.0 — basis: field 6 of ENTRY 19 — LD01-SL-C19 cites "Berglund, Tong, Kaufmann, Balesni, Cooper Stickland, Korbak & Evans, “The Reversal Curse: LLMs trained on ‘A is B’ fail to learn ‘B is A’,” ICLR 2024"; the only field-6 segment of this ClaimVersion's repository-authored ledger block.
SOURCE_PROVENANCE_PIN = SRC-SL-BERGLUND-ET-AL-REVERSAL-ARXIV-2309-12288--v1.0 — basis: explicit version-pinned public provenance of this ClaimVersion (Amendment v1.2 §6): the pinned field-6 Source record "Berglund, Tong, Kaufmann, Balesni, Cooper Stickland, Korbak & Evans, “The Reversal Curse: LLMs trained on ‘A is B’ fail to learn ‘B is A’,” ICLR 2024".