Unconstrained chain-of-thought from pure reinforcement learning can develop reasoning beyond human reasoning priors.
6 events · 2 assessments · 3 decisions
Reassessed no status change
The capability-expansion subclaim was first assessed as supported (0.72). The previous pass had already weighed that evidence. The firmer premise nudges credence from 0.40 to 0.42, but the step from beyond-base-sampling to beyond-human-priors stays weak, so the claim stays inconclusive. Argument evaluations were refreshed and importance 0.5 / contestation 0.6 recorded. Dependents were not notified because the change is immaterial.
Reassessed: still Inconclusive
verdict confidence 0.65 · credence 0.42
Structured and assessed
Reworded the canonical form for clarity, keeping its direction. Built two for-arguments: the readability-cost argument, linking the existing illegibility and language-consistency claims, and the capability-expansion argument, under a new contested subclaim on whether RLVR can exceed base-model sampling. Corrected the arXiv instance's metadata, recorded the Nature instance, linked the two as versions of one work, and wrote a material source map noting they are one voice. Assessed inconclusive (0.65), credence 0.4. Importance 0.5, contestation 0.6. No dependents to notify.
Assessed Inconclusive
verdict confidence 0.65 · credence 0.40
The claim comes from the DeepSeek-R1 authors, who suggested that the raw, unpolished chain of thought of DeepSeek-R1-Zero, trained by reinforcement learning on outcome rewards alone, "may possess potential that extends beyond the limitations of current human priors." They offered it as a motivating possibility, not a demonstrated result, and the evidence so far does not settle it either way. Two lines of evidence point toward it. Outcome-trained reasoning models often drift into illegible chains of thought, and one study of 14 models found accuracy fell sharply when models were limited to the legible parts of their reasoning. DeepSeek also reported that enforcing a language-consistency reward slightly reduced performance. But illegible is not the same as beyond human priors. Unreadable tokens may be compressed or idiosyncratic versions of ordinary human-style steps rather than new kinds of reasoning. The same DeepSeek work also found that adding human-written cold-start data produced a stronger overall model than pure RL. The sharper test is whether reinforcement learning can reach solutions the human-pretrained base model cannot sample. That question is contested. Pass@k studies found that standard RLVR mostly sharpens what the base model can already do. Longer, more diverse training (ProRL) reports some real expansion. Even real expansion would show new capability, not necessarily reasoning that is qualitatively non-human. What would settle the claim is evidence of RL-discovered reasoning strategies that solve problems beyond the base model's reach and that cannot be translated back into recognizable human-style steps. The game-playing analogy (AlphaZero) shows the claim is possible in principle, but language models trained with pure RL start from human-text priors, so that analogy transfers only weakly.
Updated claim instance
Instance 10737249-eec0-4b47-9179-16de2e063c42 (https://arxiv.org/pdf/2501.12948): was stance=affirms, confidence=0.6; set confidence=0.8, speaker="DeepSeek-AI (Daya Guo et al.)", publication="arXiv 2501.12948 (DeepSeek-R1 technical report)", sourceDate="2025-01". Speaker and publication were missing; the passage is the DeepSeek-R1 authors' own hedged assertion about R1-Zero's raw CoT, which affirms the possibility the claim states.
Claim entered the graph