RLVR Explained: How AI Learned to Prove Things It Wasn't Taught

RLVR explained in one sentence: reinforcement learning from verifiable rewards is a training method where a model gets rewarded only when an automatic checker confirms its answer is correct, instead of when a human rates it as good. That distinction sounds small. It produced the single strangest result in AI this year. In May 2026, an OpenAI reasoning model disproved the Erdős unit-distance conjecture, an open problem from 1946, by borrowing tools from algebraic number theory that have nothing obvious to do with geometry. Nine mathematicians reviewed it. Tim Gowers said that if a human had written the paper, he would have recommended acceptance.

The same model produced the correct solution in only about 50% of repeated attempts.

Key Takeaways

• RLVR rewards a model based on an automatic verifier (does the math check out, does the code pass the tests) rather than a human preference rating. It was named as a method in Lambert et al.’s Tulu 3 paper, November 2024.
• The reward is binary and cheap to compute, which means you can run it millions of times. That’s why it works on math and code and doesn’t work on essay quality.
• The Erdős result (OpenAI, May 20, 2026) is the clearest payoff so far: a genuine counterexample to a 1946 conjecture, reviewed by nine mathematicians. Sébastien Bubeck reported the model got it right in about 50% of trials.
• The failure that came before matters: seven months earlier, a claim that GPT-5 had solved 10 open Erdős problems was retracted after the “solutions” turned out to exist already in the literature.
• For kids, RLVR is a lesson about which kinds of answers can be checked automatically and which need a person, which is a genuinely useful distinction for homework.

What RLVR actually is, and how it differs from RLHF

Start with what came before. Reinforcement learning from human feedback, or RLHF, trains a model by having people rate responses. A human sees two answers, says which is better, and a reward model learns to predict human preference. That’s how models learned to be helpful, polite, and readable. We covered it in how AI gets its values.

RLHF has a ceiling. Human preference is expensive, slow, and noisy. Worse, humans are bad at spotting subtle errors in long math proofs, so a model optimized on human ratings learns to produce convincing math rather than correct math.

RLVR swaps the judge. Instead of a person, use a program that can verify the answer. The method was named in Lambert et al., “Tulu 3: Pushing Frontiers in Open Language Model Post-Training”, released November 2024 by the Allen Institute for AI, where RLVR appears alongside supervised fine-tuning and DPO as one of three post-training approaches.

The mechanism is simple enough to explain to a twelve-year-old:

1. Give the model a problem with a known, checkable answer.
2. Let it generate a solution.
3. Run the answer through a checker. Math answer matches? Code passes the unit tests? Formal proof compiles in Lean? Reward = 1. Otherwise reward = 0.
4. Update the model to make rewarded behavior more likely.
5. Repeat a few million times.

The crucial property is that step 3 costs almost nothing and involves no human. You can run it at enormous scale, and the reward is not a matter of taste. A proof either checks or it doesn’t.

What that produced: the Erdős result and its asterisks

Paul Erdős posed the unit-distance conjecture in 1946. In May 2026, OpenAI announced that a general-purpose reasoning model, not a specialized math system, had disproved it by constructing a counterexample.

The detail that makes mathematicians pay attention is how. According to Science News’ June 8, 2026 account, the model built its counterexample using tools from algebra and number theory, fields that Melanie Matchett Wood of Harvard noted seem unrelated to the discrete geometry problem at hand. Transferring machinery across distant subfields is the kind of move that reads as insight rather than pattern matching.

Verification was serious. Nine mathematicians reviewed the solution in a companion paper. Quanta Magazine reported on August 3, 2026 that Tim Gowers said of it: “if a human had written the paper… I would have recommended acceptance.” Noga Alon, Melanie Wood, and Thomas Bloom also endorsed the work. Bloom’s framing: “AI is helping us to more fully explore the cathedral of mathematics we have built over the centuries.”

Now the asterisks, and there are three.

Reliability. OpenAI researcher Sébastien Bubeck reported that when the prompt was run multiple times, the model “produced the correct solution in 50 percent of those trials.” A coin flip. Our piece on what a 50% success rate means for how kids should treat AI answers unpacks the implications.

Transparency. Science News noted that OpenAI would not reveal how much time the model spent, and that the data supporting reliability claims had not been released or peer-reviewed.

The prior retraction. Seven months before, an OpenAI executive claimed GPT-5 had found solutions to 10 previously unsolved Erdős problems. It hadn’t. The solutions already existed in the mathematical literature; the model had rediscovered them. The post was retracted after criticism from Yann LeCun and Demis Hassabis, as TechCrunch documented on May 20, 2026. That episode is the reason careful coverage of the May 2026 result says “for real this time.”

Bloom raised the deeper worry: “A big problem is AI is being used a lot by people who aren’t mathematicians… no human has read it.” And separately: “Who’s going to be able to check this?” about hundreds-of-pages proofs. Verifiable rewards solve the training problem and create a reading problem.

Why verifiability is the whole ballgame

Here’s the sentence that makes RLVR click: you can only train this way on tasks where a machine can tell right from wrong.

That’s a sharp line, and it explains the uneven progress of the last two years. Competition math improved enormously: multiple AI systems were graded at a perfect score at IMO 2026 in Shanghai. Code generation improved enormously. Essay quality, historical judgment, and emotional attunement improved much less, because there’s no checker for them.

The line also explains why formal proof assistants like Lean matter so much right now. Lean turns a mathematical proof into code that either compiles or doesn’t, which converts an unverifiable task into a verifiable one. Harmonic’s Aristotle tool does proof certification along similar lines. Our explainer on how proof assistants turn math into code covers the mechanism, and how mathematicians verify an AI proof covers the workflow.

One more honest note. Quanta’s reporting emphasizes that the current practice is iterative prompting and human collaboration, not fully automated verification. One methodology described: ask a model for a solution, feed the output to a fresh instance for checking, repeat until something workable emerges. Humans validate, simplify, and generalize. The models generate candidates, not finished proofs.

How to Teach Your Kid About Verifiable Rewards

The idea to install is “who or what decides if you got it right?” That question is useful far beyond AI.

Ages 5–8: The two-judge game

Play two rounds of a guessing game. Round one: your kid draws a picture and you say whether it’s good. Round two: your kid guesses a number you wrote on paper, and the paper says yes or no. Ask which round was fairer and why. Name it: “The second one had a checker. Computers learn fastest on games with a checker, which is why they’re better at math than at drawing something beautiful.” Five minutes, and they have the core distinction.

Ages 9–12: Sort the homework

Take your kid’s actual homework from this week and sort each item into two piles: “a machine could check this” (arithmetic, spelling, multiple choice, code that runs) and “a person has to judge this” (was the essay persuasive, was the drawing expressive, was the explanation clear). Then explain that AI got dramatically better at the first pile over the last two years, partly because of RLVR, and only modestly better at the second. Ask them to predict which of their subjects will be most changed by AI in five years.

Ages 13+: Build a verifier

Have your teen write a checker. It can be as simple as a five-line Python function that takes an answer and returns True or False, or a spreadsheet formula, or even a written rubric with unambiguous pass/fail criteria. Then have them try to write a verifier for something that resists it, like “is this paragraph well argued.” The failure is the point. Finish by having them read the Quanta piece and identify one place where a human, not a program, did the verifying.

The question to ask: “Could a computer tell whether your answer is right without asking a person? If not, why not?”

Reward type by task: what can be trained this way and what can’t

Reward type	Who or what judges	Example task	Trains well?	Why
Verifiable (exact match)	A program comparing to a known answer	Competition math, arithmetic	Yes	Binary, instant, free to compute
Verifiable (execution)	A test suite running the code	Writing a function that passes tests	Yes	Same, with the compiler as judge
Verifiable (formal proof)	A proof assistant like Lean	Proving a theorem formally	Yes, and growing	Compiles or doesn’t; no opinion involved
Verifiable (constraint)	A rule checker	”Answer in exactly 3 bullet points”	Yes	Precise instruction following is checkable
Human preference (RLHF)	People rating pairs of answers	Being helpful, polite, readable	Partly	Expensive, slow, and blind to subtle errors
Unverifiable	Nobody agrees on the standard	”Is this essay moving?” “Is this art good?”	No	No automatic checker exists

The last row is not a temporary gap. It’s a category difference. A kid who understands that row understands why AI writing help feels smoother than it feels right.

What to do at home

Ask “what’s the checker?” before trusting an AI answer

For a math problem, the checker might be plugging the answer back in. For a factual claim, it might be finding a primary source. For an essay, there may be no checker at all, which means the judgment stays with your kid. Making this explicit builds the habit that matters most.

Use the 50% number as the anchor

The Erdős result worked about half the time. That’s the most useful statistic a family can carry about frontier AI: even the celebrated results are unreliable per attempt. Ask the same question twice and compare the answers. If they differ, you’ve learned something.

Reward the checking, not the answer

If your kid uses AI on homework, praise the moment they catch an error, not the moment they produce a finished answer. This is RLVR applied to parenting, and it’s the version of the concept that actually changes behavior.

Keep the retraction story handy

The GPT-5 Erdős claim that turned out to be literature rediscovery is a better teaching story than the success. It shows that a confident announcement from a major lab can be wrong, that experts caught it, and that the correction happened publicly.

What not to do

Don’t tell your kid that AI “solved” a famous math problem, full stop. It disproved one specific conjecture, once, reproducibly about half the time, verified by nine humans, in a field where the prior claim from the same company had been retracted. All of that is more interesting than the simplified version, and all of it is true.

What to Watch For Over the Next 3 Months

• Week 4: Your kid can sort five tasks into “machine-checkable” and “needs a person” and explain one of the sorts.
• Month 2 red flags: They treat an AI answer on an unverifiable question with the same confidence as one on arithmetic. Or they’ve stopped checking because the model “gets math right now.”
• Month 3 self-check: Ask them what a verifier is and to describe one they could build for a subject they’re studying. A working answer, even a crude one, means they’ve got it.

Frequently Asked Questions

What’s the difference between RLVR and RLHF?

RLHF uses human preference ratings as the reward signal; RLVR uses an automatic checker. RLHF made models pleasant and helpful. RLVR made them much better at tasks with right answers, because the reward is binary, instant, and costs nothing to compute at scale.

Did an AI really disprove a math conjecture from 1946?

Yes, with qualifications. OpenAI announced on May 20, 2026 that a general reasoning model produced a counterexample to the Erdős unit-distance conjecture, reviewed by nine mathematicians. Tim Gowers said he would have recommended acceptance had a human written it. The model reproduced the result in roughly 50% of trials.

Why can’t this method improve AI writing the same way?

Because there’s no automatic checker for whether a paragraph is good. RLVR needs a program that can return right or wrong without a human. Grammar can be checked. Persuasiveness can’t. That’s why the gains have been so uneven across subjects.

Does this mean my kid doesn’t need to learn proofs?

The opposite. Someone has to read and check what these systems produce, and Thomas Bloom’s stated worry is that AI results are increasingly used by people who aren’t mathematicians and can’t verify them. The ability to follow and evaluate an argument became more valuable, not less.

What’s Lean and why does it keep coming up?

Lean is a proof assistant: software where a mathematical proof is written as code that either compiles or fails. It converts proof-checking from a human judgment into a machine check, which makes formal mathematics compatible with verifiable-reward training and gives humans a reliable way to audit AI-produced proofs.

---

About the author

Ricky Flores is the founder of HiWave Makers and an electrical engineer with 15+ years of experience building consumer technology at Apple, Samsung, and Texas Instruments. He writes about how kids learn to build, think, and create in a tech-saturated world. Read more at hiwavemakers.com.

---

Sources

1. Lambert, N., Morrison, J., Pyatkin, V., et al. (2024). “Tulu 3: Pushing Frontiers in Open Language Model Post-Training.” arXiv:2411.15124. https://arxiv.org/abs/2411.15124
2. Science News. (2026, June 8). “AI guardrails and the Erdős math problem.” https://www.sciencenews.org/article/ai-guardrails-erdos-math-problem
3. Quanta Magazine. (2026, August 3). “Why the Legendary Erdős Problems Are Falling to AI.” https://www.quantamagazine.org/why-the-legendary-erdos-problems-are-falling-to-ai-20260803/
4. TechCrunch. (2026, May 20). “OpenAI claims it solved an 80-year-old math problem, for real this time.” https://techcrunch.com/2026/05/20/openai-claims-it-solved-an-80-year-old-math-problem-for-real-this-time/
5. Snell, C., Lee, J., Xu, K., & Kumar, A. (2024). “Scaling LLM Test-Time Compute Optimally can be More Effective than Scaling Model Parameters.” arXiv:2408.03314. https://arxiv.org/abs/2408.03314
6. International Mathematical Olympiad. (2026). “IMO 2026, Shanghai.” https://www.imo-official.org/editions/2026/
7. “From Solvers to Research: AI in Mathematical Discovery.” arXiv:2607.07779. https://arxiv.org/pdf/2607.07779