Reinforcement learning with verifiable rewards can produce reasoning capabilities unavailable from base-model sampling.
12 events · 5 assessments · 6 decisions
Reassessed
The pass@k-validity subclaim was first assessed as doubtful. Because the metric's defect favors the base, this weakens the against argument more than the for argument. I added an argument from reasoning-checked pass@k (rated slight) that links the existing CoT-pass@k and guess-crediting claims, and re-evaluated all arguments. Status stays supported; credence moves from 0.66 to 0.68 and confidence stays at 0.6. Importance is 0.6 and contestation 0.8. The change is small, so I did not notify dependents.
Reassessed: still Supported
verdict confidence 0.60 · credence 0.68
Reassessed
The prolonged-RL subclaim dropped to inconclusive because of the Invisible Leash format correction. The change is material but absorbed: the composition premise, which is supported and independently reproduced, still carries the existential claim. Status stays supported. Credence moves from 0.72 to 0.66 and confidence from 0.65 to 0.60. Both arguments were re-evaluated, and importance (0.6) and contestation (0.75) were confirmed. Dependents were not notified because the status did not change and the credence shift is small.
Reassessed: still Supported
verdict confidence 0.65 → 0.60 · credence 0.66
Reassessed
Trigger: the composition subclaim was assessed supported (0.65 confidence, 0.70 credence). Under the previous trace, this was the outcome that keeps the status, so it stays supported. Credence moved from 0.70 to 0.72 and confidence from 0.62 to 0.65, because the for-argument now has two supported premises that do not share the ProRL formatting confound. Both arguments were re-evaluated with unchanged ratings. Importance stays at 0.6 and contestation is 0.75. Dependents were not notified, since the status is unchanged and the credence shift is minor.
Reassessed: still Supported
verdict confidence 0.62 → 0.65 · credence 0.72
Reassessed
The prolonged-RL subclaim's first assessment (supported 0.6) agrees with the previous verdict but adds a formatting confound. Status stays supported. Credence moves from 0.72 to 0.70 and confidence from 0.65 to 0.62. Both arguments were re-evaluated with unchanged ratings. Importance stays 0.6, with contestation 0.7. This is not a material change, so dependents are not notified.
Reassessed: still Supported
verdict confidence 0.65 → 0.62 · credence 0.70
Structured and assessed
First pass. Created five subclaims, all reported novel by the Matcher: the ProRL expansion result, the RL compositional-skill result, the large-k crossover, power sampling matching RL, and an assumes premise on the validity of large-k pass@k. These are grouped under one argument for and one against. Recorded three primary-paper instances and corrected the Yue et al. passage to the verbatim abstract text. Assessed as supported (confidence 0.65, credence 0.72): the claim is existential, and the evidence against it addresses typical setups only. Importance 0.6, contestation 0.75. No notification sent to the parent, since this is the first assessment; the parent's Steward will see it.
Assessed Supported
verdict confidence 0.65 · credence 0.72
The claim is an existence claim: that reinforcement learning with verifiable rewards (RLVR) can, at least under some conditions, give a model reasoning abilities that sampling the base model many times does not reach. The best-known evidence against it concerns typical setups. Yue et al. (NeurIPS 2025) found that base models catch up with and overtake RLVR models in pass@k at large k on standard math, code and visual benchmarks. Later work found that sharpening the base model's own distribution at inference time can match RL's single-attempt performance. Together these show that much of what current RLVR delivers is a reweighting of solutions the base model could already sample. That does not settle the existence claim. Yue et al. limit their conclusion to "the current training setup" and leave longer training as an open possibility. Two results point the other way. In NVIDIA's ProRL, prolonged RL produced models that solved tasks their base models failed at under extensive sampling, mostly on tasks where the base model was weak. In a controlled synthetic study, outcome-reward RL taught models to compose atomic skills into new ones that next-token training on the same data did not teach, and the RL advantage in pass@k grew on harder problems. These results are suggestive but not conclusive. ProRL's base was a distilled reasoning model, and the compositional tasks are synthetic. Every result on both sides rests on treating failure at large but finite k as a capability boundary. Finite sampling cannot show that a probability is zero, and large-k pass@k can credit guessed answers. In principle a base model gives nonzero probability to almost any string, so "unavailable" has to mean unreachable in practice. On that practical reading, the evidence supports the claim that RLVR can sometimes extend reach. It is clearest for tasks far from the base model's competence and for long, diverse training, and weakest for standard benchmarks where the base is already strong. Pass@k studies that use large k and guard against guessing, run on purely pretrained bases and on tasks without contamination, would decide the question more firmly.
Updated claim instance
Instance 2c2cc35e-f32f-424b-98ae-9ec3b00376f7 (https://arxiv.org/abs/2504.13837): was stance=denies, confidence=0.7, speaker=Yang Yue et al.; set confidence=0.7, verbatimText="Surprisingly, we find that the current training setup does not elicit fundamentally new reasoning patterns. While RLVR-trained models outperform their base models at small k (e.g., k = 1), the base models achieve a higher pass@k score when k is large. Coverage and perplexity analyses show that the observed reasoning abilities originate from and are bounded by the base model.", segmentId=null, context="Abstract (v5, NeurIPS 2025 oral). The denial is scoped to current RLVR methods; the authors add that RL's potential to elicit novel abilities has \"not yet\" been realized and may need continual scaling or multi-turn interaction.". Replaced a paraphrase from a search summary with the verbatim abstract passage, read from arXiv, and noted that the denial is limited to current methods rather than RL in principle.
Claim entered the graph