Connie on Nostr: Findings: 1. Longer reasoning → More incoherence 2. Scale improves coherence on ...
Findings:
1. Longer reasoning → More incoherence
2. Scale improves coherence on easy tasks, not hard ones
3. Natural "overthinking" increases incoherence more than reasoning budgets reduce it
4. Ensembling reduces incoherence
“LLMs are dynamical systems, not optimizers”
https://alignment.anthropic.com/2026/hot-mess-of-ai/Published at
2026-02-03 14:47:46 UTCEvent JSON
{
"id": "1b8012c3b6dd71ae81344a2bb2a036105a8af0b4fbbdc0ebfc00ed4407e69e22",
"pubkey": "ae8ef5576370b5cb91d262cf0d31d5ce9f5ca26c3ad2d56d5c58f6023633e453",
"created_at": 1770130066,
"kind": 1,
"tags": [],
"content": "Findings:\n\n1. Longer reasoning → More incoherence\n2. Scale improves coherence on easy tasks, not hard ones\n3. Natural \"overthinking\" increases incoherence more than reasoning budgets reduce it\n4. Ensembling reduces incoherence\n\n“LLMs are dynamical systems, not optimizers”\n\nhttps://alignment.anthropic.com/2026/hot-mess-of-ai/",
"sig": "413d56d232fc06545bffbf83118939aa9bf2fc0156aefa61030bd1ecbea2af88d4fb040736f2204934b182178cccb0f2b821c5e7ce3e9de515342683af32496b"
}
