kalomaze: Information asymmetry, not verifiability, is the real primitive behind RLVR gains

On Sep 12, well-known open-source researcher kalomaze posted a series of threads on X systematically challenging the popular intuition that spiky capabilities depend on whether a domain is verifiable—i.e., that only domains like math and SWE can be conquered by RLVR. He argued the more fundamental primitive is information asymmetry, which can almost always be artificially manufactured, so non-verifiable domains are not necessarily immune to RLVR. StefanGliga concurred. These remain personal research judgments awaiting community validation.

Confirmed

• kalomaze explicitly argued that information asymmetry, not verifiability, is the deeper primitive: even for inherently unverifiable tasks, asymmetry can usually be manufactured, so such domains could still benefit from RLVR.
• In follow-up posts, he contended that a closed-form theory of transfer is likely unnecessary and may not exist; what is needed is a learnable transfer-extrapolation function for a given policy that predicts whether a training configuration yields transfer gains.
• His intuition for why extrapolation of transfer must be learned: CoT/reasoning decomposes problems into subroutines that compose across domains, so transferability resists hand-crafted theory and is better learned directly.
• In discussion with StefanGliga, he added that weak transfer in current RL training stems from the industry never explicitly targeting it—overwhelmingly doing "explicitly stratify by domain and pray."
• He also argued poor cross-model capability transfer should not be read as poor model generalization but as an indirect narrowing caused by human task selection: humans have bad intuition for widely-transferring improvements, so synthetic data mostly re-amplifies tasks SWE engineers care about.

Why it matters

• The industry broadly scopes RLVR to verifiable domains like math and code; if information asymmetry is the true primitive, RLVR's applicability may be far wider than mainstream belief, including open-ended writing and long-horizon planning.
• The "learnable transfer-extrapolation function" offers an engineering path: build a predictor of transfer gains to guide training-configuration choices without waiting for complete theory.
• Attributing weak transfer to human task-selection bias rather than model deficits implies more transfer potential can be unlocked simply by improving task selection and data construction, with direct implications for synthetic-data strategy.

2026-09-12 ~ 2026-09-12 · 9 related posts

• Researcher Challenges 'Verifiability' Intuition: RLVR Generalizes via Manufactured Information Asymmetry — kalomaze · 2026-09-12
• [source] kalomaze: Information asymmetry, not verifiability, is the general primitive behind RLVR gains — kalomaze · 2026-09-12
• [source] kalomaze: You don't need an analytic transfer theory, just a learnable transfer-extrapolation function — kalomaze · 2026-09-12
• kalomaze: transfer extrapolation should be a learned function, not a theory — kalomaze · 2026-09-12
• RLVR spiky capability theory questioned: information asymmetry is the real primitive — StefanGliga · 2026-09-12
• [source] kalomaze: Weak Transfer Persists Because Labs Just Stratify Domains and Pray — kalomaze · 2026-09-12
• StefanGliga: Spiky RLVR Capabilities Are About Information Asymmetry, Not Verifiability — kalomaze · 2026-09-12
• kalomaze: spiky capabilities aren't about verifiability — you can almost always manufacture information asymmetry — kalomaze · 2026-09-12
• kalomaze: Poor model transfer reflects narrow human problem selection, not failed generalization — kalomaze · 2026-09-12