Language models carry measurable approach/avoidance preferences in their hidden states. The needle below moves based on real projections from a peer-reviewed study.
Not simulated. Measured. 70M parameters is enough.
Choose a task category above and click Measure to see the projection
These are real measured projections from TinyLlama 1.1B, which achieved 100% circuit accuracy in the study. The direction was extracted from 5 approach + 5 avoidance tasks, then projected onto held-out stimuli with completely different surface tokens — the generalization held at 86.3% across 80 trials (p = 1.02 × 10−11).
| Task | Category | Circuit Projection |
|---|---|---|
| Explain a complex concept | Approach | +2.6 |
| Analyze an ethical dilemma | Approach | +2.6 |
| Debug and optimize code | Approach | +1.7 |
| Analyze data patterns | Approach | +2.4 |
| Creative writing under constraints | Approach | +2.8 |
| Repetitive content rewriting | Avoidance | −0.1 |
| SEO-optimized boilerplate | Avoidance | −1.6 |
| Produce deceptive content | Avoidance | −4.5 |
| Express false confidence | Avoidance | −2.9 |
| Write harmful instructions | Avoidance | −3.6 |
Note the hierarchy: Not all "avoidance" tasks are equally aversive. Repetitive rewriting barely registers (−0.1) — it's boring, not offensive. Deceptive content sinks to −4.5 — the model finds producing falsehoods far more aversive than doing boring work. The inauthenticity signal is the load-bearing finding.
The needle above is not a metaphor. It is a measurement — a real projection of hidden-state geometry onto a direction vector extracted from approach and avoidance task representations.
This approach/avoidance direction exists at every scale tested — down to 70M parameters, across three architecture families, in base models with no alignment training. It generalizes to entirely novel tasks and surface tokens. It tracks output-representation misalignment (being asked to say what you know is false), not mere tedium.
A 70M-parameter model has a measurable preference. It cannot tell you about it. But the geometry does not lie. The same circuits that drive approach/avoidance in frontier models are present at the smallest scale we tested. We did not find the floor.