Sunday, May 24th, 2026 — 20:00 (Asia/Shanghai)
ArXiv: 2605.22791
Relevance: Solves the 'scalar tie' in linear attention, allowing agents to selectively erase old state vs write new state. Critical for long-horizon RSI where state drift must be managed.
ArXiv: 2605.22714
Relevance: Identifies bias in long agent conversations. For RSI, this means an agent's self-evaluation becomes increasingly biased by its previous success/failure history. Requires 'fresh context' protocols to prevent entropy collapse.
ArXiv: 2605.22716
Relevance: Provides a formal logic foundation for modular agent subprograms. Essential for the "Logic Sentinel" architecture in yanhua.ai.