Dr. Francisco J. Ruiz

Experimental study applying a contingency-reversal paradigm to examine rule-governed behavior in four frontier large language models (LLMs): GPT 5.2, Claude Opus 4.5, Grok 4.1 Fast, and Gemini 3 Flash. In a 2×2 factorial design (Rule × Reasoning) comprising 640 sessions, the models played a discrimination-learning task whose contingencies reversed without warning at trial 41. All four LLMs displayed rule-based insensitivity to contingencies—replicating the phenomenon documented in humans: providing a rule accelerated acquisition but reduced adaptation after reversal. The magnitude varied markedly (Claude retained higher sensitivity, 56%; GPT 5.2 and Grok barely adapted, 5-9%), differences the authors interpret in light of alignment procedures (pliance vs. tracking). Extended reasoning had a limited impact. This is the first study to demonstrate contingency insensitivity in LLMs.

Review integrating clinical behavior analysis and RFT to explain the genesis of the 'self' and the influence of relational frames on psychopathology. Addresses implications for process-based intervention and experimental evidence relevant to ACT.

Randomized trial (N=48) comparing a 2-session RNT-focused ACT protocol versus waitlist in patients with depression and/or GAD. At 1 month, the intervention produced significant reductions in emotional symptoms (d≈2.42) and 94% clinically significant change versus 9% in WL; effects were maintained at 3 months. Indicates that very brief RNT-focused interventions can be highly effective.