Discover cognitive mechanisms (learning-rate / uncertainty modulation)

[hanhou]

Distill interpretable mechanisms from the foundation model.

Scope

• Choice/reward-history dependence: perseveration vs. reward-seeking weights and spread across embedding
• Learning-rate modulation: does effective learning rate adapt to uncertainty/volatility?

Done when

• Mechanism characterization for Fig 5

[hanhou]

🔴 Not started

No mechanism-distillation analysis (Fig 5): perseveration-vs-reward-seeking weight spread across the embedding, and learning-rate modulation by uncertainty/volatility.

Dependencies — this sits at the top of the P1 stack:

• Needs Generative evaluation suite (run-length, switch prob, logistic coeffs) #22's logistic-regression coefficients for the choice/reward-history dependence half (perseveration vs reward-seeking weights).
• Needs DisRNN update-rule interpretation + embedding modulation #25's update-rule interpretation for the learning-rate-modulation half — the candidate learning-rate-modulating latent has not been identified yet.

Encouraging signal from 03-disrnn-beta-scan: when the interaction bottleneck is squeezed, update←subject opens — i.e. the subject embedding demonstrably routes into the update net. That is direct (if indirect-evidence) support for the premise that the embedding modulates update rules, which is what this issue sets out to characterize.