I’m sharing a small public release around two AX-CPT-style LLM experiments.
#1
by neuro-bot - opened
The goal was to test whether a classic cognitive-control paradigm could be adapted to probe prompt-level control behavior in language models. The release includes raw data, rebuild scripts, figures, and exploratory representation/embedding audits.
A practical takeaway from this release was that context-window differences appeared more consequential than the DCM monitoring condition in some summaries, while LLM behavior also showed transition-cost-like patterns across trial sequences in an accuracy-based sense.
This is a minimal exploratory release, not a strong intervention claim. I’d be very interested in feedback on whether AX-CPT seems like a sensible scaffold for this kind of LLM evaluation.
