Stuck in a changing world: small language models do not adapt to volatility, unlike humans

Adapting to change is central to human learning, and failures to adapt, such as perseveration after a reversal or insensitivity to environmental volatility, are markers studied in computational psychiatry. I tested whether three small open language models (llama3.2:3b, qwen2.5:3b and qwen2.5:7b) adapt when the rewards they learn from change. In a probabilistic reversal-learning task (80/20 reward probabilities, reversal at mid-game), qwen2.5:7b learned the initially better option (median 100% correct before the reversal) but kept choosing it afterwards (25% correct in the 10 trials after the reversal; 7.5 of 10 perseverative errors, compared with 4.6 for a simulated Q-learner, p < 0.001). In a change-point bandit, I compared the models with 30 humans from Xiong et al. (2023), who played the same task at three announced hazard rates (0.1, 0.2, 0.4). Humans switched more often as volatility increased (median switch rate 23%, 29%, 46%), switched after about two thirds of noticeable drops in payout, and earned well above chance. No model adapted to the hazard rate (adaptation index ≈ 0 vs 0.27 for humans, all p < 10^-6), and all earned chance-level points. The two qwen2.5 models switched on 1–3% of trials and almost never after a drop in payout, including after drops to 1–10 points. A model with a separate bias toward one response letter showed that qwen2.5:7b's lock-in is genuine perseveration (stickiness κ ≈ 4 with or without the bias term), whereas qwen2.5:3b and llama3.2 mainly expressed a fixed letter preference. When the long history was replaced by a short per-option summary, qwen2.5:7b chose the same letter on every trial. Reward-insensitive repetition was therefore robust across prompt formats, although its form changed. Small language models are not volatility-sensitive learners, and response biases must be modelled before LLM "stickiness" is interpreted. Code, data and analyses: https://github.com/Muzz698/llm-cognitive-subjects (study 2). Follow-up to Muzamil (2026), https://doi.org/10.5281/zenodo.24023655

Authors

Publication Details

Journal
Zenodo (CERN European Organization for Nuclear Research)
Published
2026-11-21
DOI
https://doi.org/10.5281/zenodo.24420866
Primary Topic
Neural and Behavioral Psychology Studies
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

Stuck in a changing world: small language models do not adapt to volatility, unlike humans

Muzamil
Zenodo (CERN European Organization for Nuclear Research)
Neural and Behavioral Psychology Studies
preprint

Stuck in a changing world: small language models do not adapt to volatility, unlike humans

Muzamil
preprint en

Abstract

Adapting to change is central to human learning, and failures to adapt, such as perseveration after a reversal or insensitivity to environmental volatility, are markers studied in computational psychiatry. I tested whether three small open language models (llama3.2:3b, qwen2.5:3b and qwen2.5:7b) adapt when the rewards they learn from change. In a probabilistic reversal-learning task (80/20 reward probabilities, reversal at mid-game), qwen2.5:7b learned the initially better option (median 100% correct before the reversal) but kept choosing it afterwards (25% correct in the 10 trials after the reversal; 7.5 of 10 perseverative errors, compared with 4.6 for a simulated Q-learner, p < 0.001). In a change-point bandit, I compared the models with 30 humans from Xiong et al. (2023), who played the same task at three announced hazard rates (0.1, 0.2, 0.4). Humans switched more often as volatility increased (median switch rate 23%, 29%, 46%), switched after about two thirds of noticeable drops in payout, and earned well above chance. No model adapted to the hazard rate (adaptation index ≈ 0 vs 0.27 for humans, all p < 10^-6), and all earned chance-level points. The two qwen2.5 models switched on 1–3% of trials and almost never after a drop in payout, including after drops to 1–10 points. A model with a separate bias toward one response letter showed that qwen2.5:7b's lock-in is genuine perseveration (stickiness κ ≈ 4 with or without the bias term), whereas qwen2.5:3b and llama3.2 mainly expressed a fixed letter preference. When the long history was replaced by a short per-option summary, qwen2.5:7b chose the same letter on every trial. Reward-insensitive repetition was therefore robust across prompt formats, although its form changed. Small language models are not volatility-sensitive learners, and response biases must be modelled before LLM "stickiness" is interpreted. Code, data and analyses: https://github.com/Muzz698/llm-cognitive-subjects (study 2). Follow-up to Muzamil (2026), https://doi.org/10.5281/zenodo.24023655

Zenodo (CERN European Organization for Nuclear Research)
Neural and Behavioral Psychology Studies
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon