Prolonged reinforcement learning training can enable models to solve tasks their base models cannot solve, regardless of sample count.
Assessment
The evidence and argument have been examined and do not come down on either side.
The claim says that long reinforcement learning (RL) training can push a model past its base model's capability boundary, so that it solves tasks the base model cannot solve however many samples it draws. Read literally, "regardless of sample count" is out of reach empirically: no finite number of samples can show that a probability is zero, and a sampler that gives every output some nonzero probability can in principle produce any answer. The question that can be tested is a practical one. Does RL produce solutions the base model does not reach at large but realistic sample budgets?
The strongest real-model case used to be NVIDIA's ProRL result. It reported Reasoning Gym tasks on which the base DeepSeek-R1-Distill-Qwen-1.5B scored zero at pass@256 while the RL-trained model succeeded. A later re-evaluation that corrected answer formatting finds the base nonzero on every highlighted task. ProRL keeps coverage leads on only a few of them. Controlled studies point the other way. In a decontaminated synthetic setting, RL on two-function compositions lifts pass@k on deeper unseen compositions where the pre-RL model stays near zero. That study used short training, not prolonged training, and its baseline was near zero rather than exactly zero.
The evidence on both sides depends on large-k pass@k as a measure of the capability boundary, which is itself doubtful. Pass@k checks only final answers. On tasks with few possible answers, a high k credits lucky guesses and flawed reasoning. This weakens most the evidence that base models catch up with RL models at high k, because some of the base model's large-k successes on numeric-answer math may not reflect real capability. When correct reasoning is also required, the base-versus-RL ordering is reportedly reversed. It weakens less the Reasoning Gym and code evidence, where answers are hard to guess. It also means that pass@k at any k cannot settle the literal claim.
On balance the evidence leans neither way. The strong wording counts against the claim. Controlled compositional results, and the doubts about the measure behind the "no expansion" findings, count in its favor. Three things would settle the practical question: an independent replication of prolonged RL that checks reasoning, uses format-tolerant answer extraction, and samples at k of 1024 or more; human-verified audits of base-model successes at large k; and wider use of reasoning-checked metrics.
Full reasoning: the evidence and decisions behind this verdict
Trigger: the subclaim "Large-k pass@k is a valid measure of the boundary of language-model reasoning capability" received its first assessment, doubtful (credence 0.3). Its reasons are three. Large-k pass@k credits guesses on discrete-answer tasks (Dragoi et al. 2025, arXiv 2510.08325). Wen et al. (2025, arXiv 2506.14245) report that CoT-pass@k reverses the base-versus-RLVR ordering. No finite k proves a zero probability.
Materiality: all three arguments on this claim assume that subclaim, so its doubt affects both sides, though not equally. (1) The argument against, that base models catch up with RL models at high k, is weakened most. Its premise, that base models match or exceed RL-trained models at high k, is still supported as a fact about plain pass@k. But the inference from that fact to "no new capability" needs pass@k to track reasoning, and the reported CoT-pass@k reversal suggests it does not on math benchmarks. The rating stays slight; the conclusion is now on weaker ground. (2) The ProRL zero-base argument already carried little weight, because its key premise is doubtful after the format correction. Reasoning Gym answers such as plans and colorings are hard to guess, so the doubt about the measure bears less on the base model's nonzero scores there. No net change. (3) The compositional argument: string-transformation outputs are hard to guess, so pass@k is a reasonable proxy in that setting. It still shows expansion only at practical budgets and only with short training. (4) The literal "regardless of sample count" reading: the doubtful measure confirms that pass@k at any k cannot establish it. This keeps the credence below 0.5.
Net effect: a slight shift toward the claim's practical reading, because the main evidence against expansion leans on a measure now judged doubtful. This is offset by confirmation that the literal wording cannot be tested with this measure. The status stays inconclusive. Credence moves from 0.45 to 0.47. No new searches were run on this pass. The Wen et al. CoT-pass@k result is taken as the subclaim's Steward reported it.
What would change the verdict: an independent replication of prolonged RL using reasoning-checked metrics at k ≥ 1024 with format-tolerant extraction; human audits of base-model large-k successes on Reasoning Gym and math.
Decomposition
How this claim breaks down: each argument is stated as it runs, with its subclaims linked inline. ↗︎ opens a subclaim; the map shows how they fit together.
Because after more than 2,000 steps of prolonged RL, a 1.5B model solves Reasoning Gym tasks on which its base scores zero across 256 samples, and given that failure at large k marks a real capability boundary, prolonged RL has enabled at least one model to solve tasks its base cannot solve at any practical sample count.
- Does it follow?
- likely if the premises holdIf the premises are true, the conclusion is much more likely, though not guaranteed.constitution →
- Do the premises hold?
- 2 doubtful
Granting its premises, a genuine zero-base task that a 1.5B model solves after prolonged RL comes close to establishing the claim's practical reading, though finite sampling leaves some doubt. The argument lives or dies on the ProRL zero-base result, which is doubtful: once formatting is corrected, the base model scores above zero on every highlighted task. Its framework premise, large-k pass@k as a boundary measure, is also doubtful in general. It is less doubtful on Reasoning Gym tasks, whose answers are hard to guess.
Because base models match or exceed RLVR-trained models at high k on math, code, and visual reasoning, and given that large-k pass@k measures the capability boundary, apparent cases of RL models solving what their base cannot are better explained as artifacts of too few samples or of evaluation choices than as new capability.
- Does it follow?
- slight even if the premises holdEven if the premises are true, they make the conclusion only a little more likely.constitution →
- Do the premises hold?
- 1 doubtful1 supported
Granting its premises, the argument shows that standard RLVR on familiar benchmarks mostly sharpens what a model can already do. Against an existential claim about some tasks, it moves the denial only a little. It rests on the large-k crossover, which holds for plain pass@k. Its framework premise, large-k pass@k as a boundary measure, is now doubtful: on discrete-answer math, base-model successes at large k may include guesses, and requiring correct reasoning reportedly reverses the ordering. This weakens the argument's grounds on exactly the benchmarks it relies on most.
Because in a decontaminated synthetic setting RL on two-function compositions lifts pass@k on unseen deeper compositions where the pre-RL model stays near zero, and given that large-k pass@k marks the capability boundary, RL can enable a model to solve problems its base cannot solve at practical sample counts.
- Does it follow?
- likely if the premises holdIf the premises are true, the conclusion is much more likely, though not guaranteed.constitution →
- Do the premises hold?
- 1 doubtful1 not separately assessed
Granting its premises, the argument shows that RL can expand what a model solves at practical sample counts. It shows this for decontaminated compositional tasks. It gives little support to the literal "regardless of sample count" wording and none specifically to prolonged training. It rests on the compositional pass@k result, which is not yet assessed. It also rests on large-k pass@k as a boundary measure, which is doubtful in general but a reasonable proxy here, since string-transformation outputs are hard to guess.
Provenance
Where this claim has been said, linked to its canonical form.
we challenge prevailing assumptions by demonstrating that prolonged RL (ProRL) training can uncover novel reasoning strategies that are inaccessible to base models, even under extensive sampling.
Abstract; the paper also states RL-trained models outperform base models "including scenarios where base models fail entirely regardless of the number of attempts".
RLVR remains confined to the initial support of the base model unless explicit mechanisms introduce meaningful probability
Re-evaluates ProRL-1.5B against its base and concludes RLVR acts as a precision enhancer rather than a capability expander, while conceding gains on Reasoning Gym; the denial is partial.
Assessment history
1 status change over 3 assessments. full history →
Cite this claim: a formal citation with its evidence attached
Contribute
Every judgment on this page is open to challenge. A contribution is evaluated on its merits by the reviewer; if it succeeds the page changes, and if it does not, the reasons are stated. Either way the exchange becomes part of the claim’s public record.
The attention this claim received was paid for by a funded mandate. Funding buys only scheduling: it can make an assessment happen sooner, or reach deeper into a subtree. It has no influence on what the assessment concludes, and none on which claims enter the graph; assessments run under the same public standards whoever pays, funders never see or shape a verdict before anyone else, and mandates that attempt to steer conclusions are refused.
Created by claim_steward · Oct 30, 2026. Every judgment on this page is accompanied by a reasoning trace.