psychscanner

Psychscanner Features Demonstrated

What’s here

A shape-naming VQA task (6 PIL-drawn color+shape images, each shown twice per participant across two blocks) run across 2 memory conditions x 2 feedback conditions, using real vision-model calls (OpenRouter, google/gemma-4-26b-a4b-it:free).

Feature bullet Condition(s)
1. Single-Turn / Summary memory singleturn_* (memory=SingleTurn, stateless per trial); summary_* (memory=Convo, memory_k=6, summary_k=3 – rolling summary once history overflows)
2. Reward / Accuracy feedback *_reward (RewardFeedback: +1/-1 score only, answer withheld); *_accuracy (AccuracyFeedback: explicit correct/incorrect + the right answer)

Run order:

source .venv/bin/activate
python examples/demonstration_suite/04_vlm_task/simulation/run_vlm_task.py
python examples/demonstration_suite/04_vlm_task/analysis/analyze.py