August 3, 2026

Proposes RLSVR, which transforms open-ended tasks (summarization, creative writing) into verifiable proxy environments so reward signals emerge automatically instead of relying on human preferences or LLM judges. Instantiated as SpyRL, a multi-agent self-play game where voting outcomes provide ground-truth-verifiable rewards, and shown to improve both non-verifiable and verifiable reasoning tasks.

Introduces an 8-dimension auditing framework (identity transparency, privacy, manipulation prevention, harm prevention, fairness, etc.) grounded in AI-safety guidelines and human-rights articles, then audits 3,249 system-prompt instructions across 88 commercial AI products.

Formalizes a 'safety trilemma': for dual-use tasks, LLM safeguards cannot simultaneously offer useful capability, reliable safety, and open access, because an attacker can copy or imitate any prompt/context evidence a legitimate user would present. Derives a mathematical worst-case floor on attacker assistance and proposes 'trusted credentials', hard-to-copy, use-tied evidence, as the necessary complement to prompt-based safeguards.

Identifies that LeWorldModel's Epps-Pulley regularization objective, which pushes latents toward an isotropic Gaussian, has corrective gradients that vanish for isolated tail samples, leaving heavy-tailed latent deviations poorly controlled. QQWorld replaces it with a quantile-quantile matching objective that keeps gradients active in the tails, improving latent-space quality for planning.