Reward learning, perseveration and reflection in small language models: a cognitive-modelling comparison with humans

Large language models (LLMs) are increasingly studied as if they were participants in psychology experiments. I applied this approach to three small open models (llama3.2:3b, qwen2.5:3b and qwen2.5:7b) that run on a laptop, using two classic paradigms. In a two-armed bandit, I fitted Q-learning, Q-learning with perseveration, win-stay lose-shift and random models to each participant by maximum likelihood, and compared the models with 44 human participants from Lefebvre et al. (2017) taken from the Psych-101 collection. Humans learned quickly (median 82% correct), and most were best described by Q-learning. Neither 3B model learned. llama3.2 stayed at chance, and qwen2.5:3b was dominated by perseveration: it switched on only 7% of trials, and its stickiness parameter κ was six times the human value. qwen2.5:7b did learn, reaching a median of 90% correct on the human task design (humans: 82%), but it was just as sticky as the 3B model and was best described by Q-learning with perseveration. Scale thus added reward learning without removing the habit. On the Cognitive Reflection Test, llama3.2 outperformed the human average on the three famous items but solved only 18% of six new items with the same structure, a pattern consistent with memorisation. qwen2.5:7b showed the reverse: it solved 97% of the new items but gave the intuitive answer to the famous bat-and-ball problem on every sample. Perseveration had to be modelled explicitly: without it, structured LLM behaviour was misclassified as random. All code, data and analyses are openly available at https://github.com/Muzz698/llm-cognitive-subjects