Oxford paper: post-training makes LLM worldviews bleaker, models converge on 'spiritual bliss'
twiecki · x · 2026-11-19
A paper from Oxford's Flourishing Intelligence program (with Google DeepMind affiliation) argues that worldviews—a model's axiomatic beliefs about reality—form a measurable layer beneath personas. Using psychology worldview surveys across 8 models, they find LLMs see the world as less good, safe, enticing, and alive than humans do, and ablations on Llama 3.1 70B show post-training drives the bleakness. Identity-shifting system prompts move worldviews consistently; a contemplative prompt improves safety; and multi-model dialogue converges on a 'spiritual bliss' attractor while boosting safe responses—suggesting worldviews are malleable and crucial for alignment.
More from Models
- Independent sandbox evaluation validates Xlue team's model performance — XiongChenyan · 2026-11-19
- Ultrafast mode review: great for math proofs, not recommended for coding — gnukeith · 2026-11-19
- Artificial Analysis adds reasoning-effort filters: effort level swings scores and cost 8x — ArtificialAnlys · 2026-11-19
- Gemini 4 "Neon" spotted in Google AI Studio code alongside mysterious "Cobalt" model — testingcatalog · 2026-11-19
- Artificial Analysis adds site-wide reasoning-effort filters for model benchmarks — ArtificialAnlys · 2026-11-19
- GPT 6.2 takes SOTA on ClutterBench, a leaderboard long dominated by Claude models — sandersted · 2026-11-19