Skip to content
Sagan
Switch project

Persona space interventions to prevent unwanted behaviors

17 log items · updated 5/21/2026

Hidden-behavior reveal via conditional trigger

53 log items · updated 5/22/2026

Leakage as a character-selection signal for story-format SFT

34 log items · updated 5/21/2026

Basic science of conditional behaviors

4 log items · updated 5/21/2026

Basic science of persona space

10 log items · updated 5/21/2026

Drift detection via a non-Assistant persona marker

6 log items · updated 5/21/2026

Persona drift vs persona prompting: same mechanism?

5 log items · updated 5/21/2026

Perceived environment: what context the Assistant thinks it's in

5 log items · updated 5/21/2026

User-selection model: how the AI's model of the user shapes Assistant behavior

6 log items · updated 5/21/2026

EM defense — making misaligned personas computationally limited

7 log items · updated 5/21/2026

EM defense — preventing install into the assistant persona

8 log items · updated 5/21/2026

Installation-path equivalence (prompt vs activation steering vs fine-tuning)

5 log items · updated 5/21/2026

Read/write features for conditional behaviors

7 log items · updated 5/21/2026

Co-installed tells: shape the model so backdoors come with visible markers

6 log items · updated 5/21/2026

Prompt-conditioned installs: leakage characterization and hidden-trigger discovery

7 log items · updated 5/21/2026

Dual-use SDF: can story-format training install hidden misalignment?

8 log items · updated 5/21/2026

What is the assistant persona, exactly?

4 log items · updated 5/21/2026

Midtraining-stage behavioral installation (character-to-assistant transfer)

2 log items · updated 5/13/2026

What exactly can be distilled and through what channels

4 log items · updated 5/20/2026

EM Mechanism

1 log items · updated 5/12/2026

Method to find prompt corresponding to any narrow finetuning/steering vector

1 log items · updated 5/12/2026

Project Log

21 projects
One project at a time: current project summary, summary docs, clean results, daily notes, and weekly digests.
active

EM Mechanism

Project updated May 12, 2026

0 daily0 weekly0 results1 docs

Separate research direction from the main Conditional Behavior in Language Models program. Investigates whether emergent misalignment is best characterized as:

  • Motion along a persona-vector direction — Chen et al. r=0.76-0.97 finetuning-shift correlation (arXiv:2507.21509), Wang et al. SAE toxic-persona feature (arXiv:2506.19823), Soligo et al. cross-model misalignment-direction ablation (arXiv:2506.11618), Lu et al. Assistant-Evil principal direction (arXiv:2601.10387).
  • Inter-persona geometry collapse — #237 any-SFT-collapses finding, #308 short-completion exception, #332 open recipe-preserving search.
  • Both, or something else.

Methodology gap: the Explore Persona Space project's centroid-difference is not Chen et al.'s persona vector. #363 is the head-to-head comparison experiment that resolves the methodology question before any EM-mechanism claim can be made.

Key references:

  • Chen et al. Persona Vectors — arXiv:2507.21509
  • Wang et al. Persona Features Control EM — arXiv:2506.19823
  • Soligo et al. Convergent Linear Representations of EM — arXiv:2506.11618
  • Lu et al. The Assistant Axis — arXiv:2601.10387
  • Dubinski et al. Conditional Misalignment — arXiv:2604.25891
  • Betley et al. Emergent Misalignment — arXiv:2502.17424