Reports of LLMs mastering math have been greatly exaggerated
We recorded 11 claims from this post. 5 were checked against evidence. 6 can’t be checked. 4 more are held. Read the post at its source.
Claims from this post (11)
In the order they appear in the post. Each quotation is the speaker’s words in this post. A label such as “Supported” or “Can’t be checked” is the claim’s current assessment; its page has the evidence and the reasons.
Assessment: SupportedSaid
“The results were dismal: None of the AIs scored higher than 5% overall.”
What we tested: On the 2025 USAMO, none of the leading reasoning LLMs tested scored higher than 5% overall.
What happened: Table 1 reports average totals of 2.0, 1.8, 1.5, 1.2, 1.2 and 0.9 points for the six named models, against a maximum of 42. The highest reported total therefore corresponds to approximately 4.76%, below 5%. (Petrov et al., ETH Zurich and INSAIT; arXiv, )
Assessment: Can’t be checkedSaid
“The refusal of these kinds of AI to admit ignorance or incapacity and their obstinate preference for generating incorrect but plausible-looking answers instead are one of their most dangerous characteristics.”
What we tested: The refusal of these kinds of AI (LLMs) to admit ignorance or incapacity, and their obstinate preference for generating incorrect but plausible-looking answers instead, are among their most dangerous characteristics.
Assessment: Can’t be checkedSaid
“Getting AIs to answer “I don’t know” is one of the most important unsolved challenges facing the field.”
What we tested: Getting AIs to answer 'I don't know' is one of the most important unsolved challenges facing the field.
Assessment: Can’t be checkedSaid
“If this kind of technology becomes commonly used to answer difficult questions before the problem of generating invalid answers is fixed, we will be in serious trouble.”
What we tested: If LLMs become commonly used for difficult questions before their invalid-answer problem is fixed, society will be in serious trouble.
Assessment: SupportedSaid
“But that is because they rely in part on powerful, completely hand-written, symbolic reasoning systems. LLMs are not similarly immune.”
What we tested: AlphaProof and AlphaGeometry cannot generate incorrect proofs because they rely in part on powerful, completely hand-written symbolic reasoning systems, and LLMs are not similarly immune.
What happened: The paper describes AlphaProof's use of Lean and attributes proof correctness to kernel verification. It also explains that informal LLM reasoning lacks guaranteed correctness; this supports the mechanism behind Marcus's comparison. (Nature, )
Assessment: UnresolvedSaid
“Importantly, the neurosymbolic method used by DeepMind’s AlphaProof and AlphaGeometry systems (which we discussed recently) which (more or less) achieved a silver-medal level performance on the 2024 International Math Olympiad, is immune to this problem.”
What we tested: DeepMind's neurosymbolic AlphaProof and AlphaGeometry are immune to producing confidently incorrect proofs.
Assessment: UnresolvedSaid
“So Nestler’s experiment does not contradict the finding of the report; it corroborates it. o3, yet again, produced an invalid answer to this problem.”
What we tested: The o3 proof of USAMO 2025 problem 2 posted by Nestler was invalid, corroborating the report.
Assessment: Can’t be checkedSaid
“It is by no means certain that additional user prompts will add much,.”
What we tested: Additional user prompting may not substantially improve reasoning LLMs' performance on such problems.
Assessment: Can’t be checkedSaid
“Moreover, however you slice it, this need of the AIs for special cajoling before they will think carefully about questions that are presented to them is a serious weakness, and the demand for special prompting tilts the scale.”
What we tested: However you slice it, the AIs' need for special cajoling before they will think carefully about questions presented to them is a serious weakness, and the demand for special prompting tilts the scale.
Assessment: SupportedSaid
“The important problem is that they sometimes claim to they have solved problems that they have not properly solved; as with hallucinations, this reflects a deeply problematic failure to sanity check their own work.”
What we tested: LLMs sometimes falsely claim to have solved problems, reflecting a failure to sanity-check their own work.
What happened: The study pairs claims of completion with expert findings of flawed logic, unjustified steps and other proof defects. Its independent double-grading procedure supports the limited claim that models sometimes deliver incorrect work as solved. (arXiv, )
Assessment: Can’t be checkedSaid
“The really important challenge is not to get the AIs to solve more USAMO problems; it is to get them to say “I give up” when they can’t.”
What we tested: The really important challenge is not getting AIs to solve more USAMO problems but getting them to say 'I give up' when they can't solve them.
Held from this post (4)
These records did not pass every check, so nothing about them is published. Held is not a finding.
- The two reviews read the claim differently
- The criteria are ambiguous. Our reading, not assessed: Users can very easily trust plausible-looking LLM answers without checking whether they are correct.
- The criteria are ambiguous. Our reading, not assessed: The authors have yet to see any evidence that any kind of prompt helps get AIs to say 'I give up' when they cannot solve a problem.
- The two reviews read the claim differently