RLVR training reveals unavoidable tradeoff between single and multi-attempt success
Source published ·Modelwire updated
Original coverage: arXiv cs.LG ↗·How Modelwire adds context
The development
A new theoretical result quantifies a fundamental tradeoff in reinforcement learning for language models: optimizing for single-attempt success (pass@1) necessarily degrades performance when models get multiple chances to solve a problem. Researchers prove this isn't a training artifact but an inherent cost, modeling RLVR as an exponential tilt over shared strategies and using Stiemke duality to establish the exact rate of degradation. The finding matters for practitioners tuning RL objectives: gains on benchmark metrics like pass@1 come with measurable losses elsewhere, reshaping how teams should evaluate whether RL training actually improves reasoning or merely redistributes capability.
Modelwire’s AI-generated summary of coverage from arXiv cs.LG.
Modelwire analysis
ExplainerOur AI-generated reading of the wider context and the next developments to watch.
The paper doesn't just show that pass@1 and pass@k trade off; it proves the tradeoff is structural, not fixable through better training, and derives the exact rate of degradation. That mathematical precision is what separates this from folk wisdom.
This connects directly to the October 26 analysis showing that pass@1 gains often come from token efficiency rather than genuine capability gains. That work decomposed where improvements actually come from; this paper explains why you can't have it all. The October 29 mechanistic study on Tic-Tac-Toe also matters here: if RL is exploiting patterns rather than building reasoning, then optimizing for single-attempt success might just be sharpening the exploit, which would explain why multiple attempts suffer. Together these suggest practitioners chasing pass@1 metrics may be measuring redistribution, not advancement.
If teams report that models trained to maximize pass@1 using this framework's insights maintain or improve pass@k on held-out reasoning benchmarks (MATH, GPQA) within the next two quarters, the theory has a gap. If pass@k continues to degrade proportionally to the predicted rate, the Stiemke duality bound holds in practice and becomes a hard constraint on RL objective design.
This interpretation is generated from the summary above and the archive coverage cited below. Our methodology · Report an error
Coverage behind this analysis
These archive entries ground the connection in our analysis. They are ordered by source publication date, with links to our coverage and the original sources.
·arXiv cs.CL
Outcome RL gains traced to finishing speed, not reasoning depth
A new analysis framework decomposes reinforcement learning gains in language models into three measurable components: whether models finish within token budgets, whether their answers are correct once completed, and whether they solve novel problems their base versions missed. Testing 20 RL-trained models from 1.5B to 32B parameters on physics and math tasks reveals that pass@1…
MentionsRLVR · pass@1 · pass@k · Stiemke duality
How this coverage is produced
Modelwire uses AI to generate summaries and context from source headlines, snippets, and selected archive coverage. Automated checks do not verify every claim, and items are not routinely reviewed by a person before publication. Zacaria Solis operates the site. Read the linked source for the full evidence and report errors through our corrections process.
Modelwire summarizes, we don’t republish. arXiv cs.LG originally reported this story as “There Is No Free Lunch in RLVR: Every pass@1 Gain Has a pass@k Price”. The full content lives on arxiv.org. If you’re a publisher and want a different summarization policy for your work, see our takedown page.