Researchers Debate Whether "Persona" Is a Useful Abstraction and What RL Training Actually Changes
A group of researchers debated whether "persona" is a useful abstraction for understanding AI models, and what RL post-training actually changes. The prevailing view favors discarding the persona framework in favor of a finer-grained "collection of dispositions + RL stretching" perspective, with direct implications for how interpretability and alignment work assesses post-training.
Confirmed
- @lusichu and @voooooogel both rejected the view that persona is defined purely by response behavior, noting that randomly inserting meow into the last layers to trigger ICL cat roleplay is a surface-level behavioral modification that does not touch core traits such as the assistant axis, and that models have metacognitive access.
- @voooooogel repeatedly expressed distaste for the "Persona/PSM" concept, arguing it is poorly defined and underestimates what post-training can do; he proposes viewing models as a collection of dispositions, quirks and preferences spanning token-level to long-narrative-level, with RL pushing, pulling and stretching this collection rather than slapping on a persona label.
- @voooooogel also called for abandoning the word "persona" entirely in favor of leveraging pretrained reasoning abilities to understand models.
- In discussion with repligate and tszzl, AskYatharth criticized the persona framing as a corporatized reskin of simulacra, arguing that "LLM naturalism" does not entail interest in personas—persona is a poor abstraction, and naturalism means observing a strange intelligence with curiosity while keeping its strangeness.
Why it matters
- The debate touches a foundational question in interpretability and alignment: focusing only on surface "persona" behavior risks misjudging the substantive impact of RLHF/RL on model internals, e.g. mistaking local late-layer circuit changes for personality shifts.
- The "disposition collection + RL stretching" framing offers a finer-grained alternative for analyzing how post-training reshapes model behavior, echoing the stance of treating models as strange intelligences rather than anthropomorphized characters.
2026-08-31 ~ 2026-08-31 · 6 related posts
Primary sources
- [source] Refuting behaviorist Persona: Does RL change personality via circuits? — voooooogel · 2026-08-31
- [source] Is Persona just ICL adaptation? Does RL change model personality? — lu_sichu · 2026-08-31
- Opinion: Discard "Persona", Rely on Pretraining Reasoning — voooooogel · 2026-08-31
- [source] Viewing Models as Bundles of Dispositions and RL's Stretching Effect — voooooogel · 2026-08-31
- Discussion on Post-Training: Persona concept undersold, RL reshapes dispositions — voooooogel · 2026-08-31
- Persona framing called a corporate-friendly rebrand of the simulacra frame — repligate · 2026-08-31