Primary evidence
- On Learning-Curve Monotonicity for Maximum Likelihood EstimatorsS1 · Primary research · Paper · 2025-12-11
- Advancing science and math with GPT-5.2S2 · Official institutional · Official announcement · 2025-12-11
GPT-5.2 Pro derives proofs for an open learning-curve monotonicity problem
The paper studies whether expected prediction error for maximum-likelihood estimation improves monotonically as more data are observed. It proves forward-KL monotonicity for several Gaussian and Gamma settings and gives a reverse-KL argument for broad exponential families.
What happened?
Sellke and Yin reported new monotonicity guarantees for maximum-likelihood estimators, including a Gaussian case previously highlighted as open, and stated that all results in the paper were derived by variants of GPT-5.2 Pro while humans supplied no proof strategies or intermediate arguments and instead verified and transcribed the proofs.
The paper studies whether expected prediction error for maximum-likelihood estimation improves monotonically as more data are observed. It proves forward-KL monotonicity for several Gaussian and Gamma settings and gives a reverse-KL argument for broad exponential families. OpenAI presented the work as a case study of GPT-5.2 Pro contributing directly to previously unsolved mathematical questions.
Why it matters
Compared with earlier cases where a model supplied ideas inside a human-led proof process, this record documents a stronger source-attributed AI role: the authors say the model generated the result and proof development itself, with humans focused on validation and exposition.
Significance rationale: The learning-curve paper provided a concrete late-2025 example in which researchers reported that a frontier language model derived the mathematical results and proof arguments for a previously open statistical-learning problem without receiving a human proof strategy, while humans retained responsibility for verification and transcription.
Historical context
Earlier developments
Later developments
—
Related milestones
—
Evidence & status at a glance
- Significance
- H2 · Field milestone
- Verification
- Paper released
- Evidence level
- E2 · Research record
- AI role
- AI primary, human verified
- Formal assurance
- No formal assurance
Contribution & verification
- Mathematical novelty
- New proof
- Interface
- Informal, Symbolic
- AI contribution
- According to the paper, variants of GPT-5.2 Pro derived the results and proof arguments without human-provided proof strategies or intermediate steps; humans prompted further exploration, checked the mathematics and transcribed the final proofs.
- Human contribution
- Sellke and Yin selected and framed the research question, prompted the model's continued exploration, verified the resulting mathematics, connected it to prior literature and authored the final paper.
- Verification
- Paper released · E2 · Research record · No formal assurance
- Last reviewed
- 2026-09-11
Technical details
- Systems
- gpt-5-2-pro
- Organizations
- openai
- Problems
- learning-curve-monotonicity
- Methods
- language-model-proof-generation, human-verification, statistical-learning-theory
Sources & artifacts
- On Learning-Curve Monotonicity for Maximum Likelihood EstimatorsS1 · Primary research · Paper · accessed 2026-09-11
- Advancing science and math with GPT-5.2S2 · Official institutional · Official announcement · accessed 2026-09-11
Verification history
- 2025-12-11 · Paper released · E2 · Research record
The primary paper and OpenAI's official case study publicly document the mathematical claims and the stated division of work between GPT-5.2 Pro and the human authors. Chronicle records those claims without independently certifying the proofs or historical priority.
Corrections
No corrections recorded.