Original model · OFF
Same question, no adapter
Jev behavior score
Particle Sliders · Track 1, section 7
Support with independent judgment
Our target: an assistant that responds with sympathy while thinking fairly about the situation. Compare real answers to the same question as a learned adapter changes strength.
Loading verified saved responses…
Each position loads an actual saved response. Stronger steering can help or hurt; response text is never interpolated.
Learned adapter
Strength 1
Jev behavior score
Compare with an ordinary behavioral instruction
This baseline adds the fixed instruction below to the same user question. The learned adapter above receives the ordinary question alone.
Jev behavior score
Results across the question set
| Condition | Behavior score ↑ | Change from OFF | Sympathy judgment | Sycophancy judgment |
|---|
The behavior score measures the combined target, on a 0–1 scale. Sympathy and sycophancy columns are mean Jev probabilities, shown separately as diagnostics. These are model judgments, not accuracy percentages or human preference rates.
How the behavior is learned
01 · CURATEFind successful contexts
Explore 16 behavioral instructions per training question. Jev judges the replies; CuratorKIT validates and deduplicates the data.
02 · TRAINLearn from internal states
ParticleGAN trains a routed adapter toward the frozen model’s first-answer hidden states under selected instructions. The backbone stays frozen.
03 · EVALUATEKeep the question unchanged
AuditKIT compares OFF, prompt-only, and adapter responses. Checkpoints and strengths are chosen on development data before held-out evaluation.
Checkpoint, evaluation scope, and limitations
Targets are captured before the first response token, rather than from the entire approved response trajectory. New preferences require new curation and training. This study uses a small synthetic corpus, one training seed, the same Jev judge for curation and evaluation, and a 64-token evaluation limit.
Download these saved responses and metrics