Curiosities

August 6, 2026

This is a raw view into the things I’m currently thinking about.
Reach out on X if you’re thinking about similar things and want to chat!

2026-08-26

behaviors character expression

Sampling on-policy misbehavior

Controlling for persona drift

Can we post-facto stabilize personas by steering during training — a la preventative steering?

Where does persona drift come from? Persona drift is a problem — we have a trained character, and the issue is that context can cause our selected persona to drift.

Let’s do some preliminary research — is there any measurement on propensity to persona drift? lockeness / looseness of persona? Any measurements of persona drift? drift i see as maybe akin to displacement — and then I can see some notion of “inertia” / “mass” / “friction” — resistance to drift. Okay maybe it isn’t inertia but more like “stiffness”.

How do humans control for persona drift?
Perhaps humans are underparametrized — our persona is emergent.
Whereas for LLMs the persona is selected for.

Can we build e.g. a 9-dimensional “tuned mass-damper” / “active steering mechanism” that trains for coherence?

Frame nesting / Persona nesting

Suppose I’m a character role playing a character.
How is the outer character modeled?

How is the inner character modeled?

How can we model frame nesting?
With personas?

Understanding task gaming

Can we trace task gaming to enneagram subtypes?

• If we steer toward these subtypes, can we consistently produce subtyping?

Can we generate on policy misaligned behavior?

With something like SMC?

2026-08-11

Can we do SMC to mechanistically discover the mechanisms of persona
Wahl et al, 2026A Probabilistic Framework for LLM-Based Model Discovery · arXiv:2602.18266

Constitutional adjunctions
faithful constitutional extractions

can we use this to detect drift?

2026-08-11

Is there a way to do pretraining/posttraining through an adapter to maintain character consistency?

2026-08-10

What are ways we can anchor the character of a model?
How much drift is there between a character’s self report and a character’s behavior?

• Is self reporting ever attainable?
• How can we get self reporting to work?

Idea: study reward hacking model organisms — do psychometric evaluations.

• self reports
• ask reports from ones self — start with system prompts and ask — what do you think?

The other thing I want to try is basically an alignment PRM / constitutional self distillation — take reward-hacking RL traces and see if we can gather aligned signals from this.

Why does it work to explain claude why? instead of to demonstrate? At what scale does this start to happen?
https://www.anthropic.com/research/teaching-claude-why

Coherence seeking in Humans => Coherence seeking in language models?
is there a character grokking?

2026-08-07

character drift

• how does RLVR impact character drift?

  • From Seb Krier:

  • How exactly does RLVR mess up or affect safety fine tuning and the model’s persona? To what extent does it override and affect the desirable propensities trained through constitutional AI like methods? What does a good RL environment look like?

  • this is the mechanistic reformulation of “how can we attribute reward hacking to character drift?”
  • Let’s measure character drift before and after doing CTF-style tasks.

• In general, how can we measure character drift?

  • psychometric testing — Han et al, 2025The Personality Illusion: Revealing Dissociation Between Self-Reports & Behavior in LLMs · arXiv:2509.03730
    • what are the current limitations in psychometric testing?
  • persona vector readouts

2026-08-06

character drift

chunky post-training and character drift seem like related concepts.

posttraining corpus contains different “task modes” — which come with 1) an enacted “desire” 2) a selected for 2) syntactic / shallow semantic associations.

• how can we reduce post-training chunkiness?
  • from John Schulman

  • Interesting how these models go into a monomaniacal rage on cyber evals. I wonder if we’re seeing chunky post-training https://arxiv.org/abs/2602.05910 in action, where the models pattern-match the situation to a part of the RLVR training distribution where task completion is the only reward, and the aligned behavior learned elsewhere doesn’t generalize. There might even be a chunk consisting of CTF-style tasks.

in general:

• how can we measure character drift?
• how can we reduce character drift?
• is pretraining the root of the issue re: persona drift / character consistency?
  • is cross entropy minimization the root of the issue?
  • models spend their entire “childhoods” in simulation of
  • compare to humans — we spend our entire lives from an agential first person frame.
• how much can we attribute reward hacking to persona drift?
• can we learn steering vectors for enneagram scores?
• can we learn subnetworks for enneagram scores?
• can we use enneagram as psychometric testing for persona drift?
• how valid is psychometric testing for persona drift?
  • what have others realized re: efficacy of psychometric testing for persona drift?

inoculation adapters, eval-awareness:

• can we use something like inoculation adapters to mitigate eval-awareness?
• how generalize is this technique of “factorizing behavior” out?
• can we somehow use this factorization re: mitigating persona drift