PDF(1564 KB)
价值感知与信息整合的差异：人类与当前大语言模型在爱荷华赌博任务中的对比*
张逸飞, 侯夔南, 周禧彬, 张然, 蒲渝, 黄艾玥, 何清华
心理科学 ›› 2026, Vol. 49 ›› Issue (4) : 822-833.
PDF(1564 KB)
PDF(1564 KB)
价值感知与信息整合的差异：人类与当前大语言模型在爱荷华赌博任务中的对比*
Value Sensitivity and Outcome Integration: A Comparative Study of Humans and Current-Generation Large Language Models on the Iowa Gambling Task
大语言模型的决策潜力日益凸显，理解其系列决策机制并与人类进行比较，对于优化人机协同设计至关重要。研究采用价值序列探索模型解析大学生和大语言模型的爱荷华赌博任务数据，并比较了人类（包括全体样本及专家子样本）与大语言模型的决策表现和参数。研究结果表明：相较于其他候选模型，价值序列探索模型能较好地拟合数据；人类和大语言模型在逆衰减率、价值敏感度和稳定性参数上存在差异：人类在信息整合方面占优，而大语言模型在价值感知和策略稳定性方面占优。这些发现为优化人机协同系统的设计和决策提升提供实证基础。
Understanding the differences in the sequential decision-making between Large Language Models (LLMs) and humans is crucial for optimizing human-AI collaboration. This study employed the Iowa Gambling Task (IGT), combined with Value plus Sequential Exploration (VSE) cognitive modeling, to systematically compare behavioral performance between humans and LLMs. We contrasted a large human sample with three representative LLMs under baseline and Chain-of-Thought (CoT) conditions. We hypothesized that the VSE model could effectively characterize the exploration and exploitation parameters in samples including both humans and LLMs.
The study utilized a clinical version of the IGT with dynamic reward/loss schedules. Data were collected from 1763 Chinese undergraduate students via a standard computerized procedure. LLM data were collected via the official APIs of three models (DeepSeek-V3, GLM-4-Flash, GLM-Zero-Preview), simulating 48 participants per model for both baseline and CoT conditions. To prevent LLMs from leveraging pre-existing knowledge about the IGT task acquired during their training, the standard deck labels (A, B, C, D) were replaced with random three-letter strings as anonymized labels. Furthermore, the assignment of the underlying deck characteristics (i.e., their reward and punishment schedules) to the anonymized labels was counterbalanced across participants using a Latin square design. LLM interactions involved system prompts delivering instructions and user prompts requesting choices. They incorporated dialogue history and feedback on gains/losses from the previous trial. An error-correction mechanism ensured valid responses. Cognitive modeling compared the VSE model against four alternative models using Variational Bayesian Analysis for parameter estimation and model fitting with non-informative priors. A parallel analysis approach was adopted. We conducted comparisons between the LLMs and (1) the entire human sample, and (2) the top-performing (based on final net gain) 48 “expert” human participants, matched to the LLM sample size.
The behavioral results indicated that certain LLM conditions (baseline GLM-4-Flash and GLM-Zero-Preview) failed to follow IGT task instructions. They exhibited repetitive non-adaptive patterns and were thus excluded from further analysis. LLMs (Deepseek-V3 and GLM-Zero-Preview) achieved significantly higher final net gains compared to the overall human sample, but their performance showed no significant difference from that of the “expert” human sample. A comprehensive investigation involving model comparison analyses robustly identified the VSE model as the superior framework for explaining the choice sequences of both humans and LLMs across two parallel analyses. This conclusion was evidenced by better-fit indices (BIC, AIC, Free Energy) and higher predictive accuracy for choices. Parameter estimates from the VSE model revealed several key differences between humans and LLMs. Regarding value sensitivity (accuracy in perceiving reward/loss utility), LLMs showed greater sensitivity than the general human population, although comparable to that of expert humans. Conversely, human participants always showed a superior ability to integrate historical value information over time, reflected in higher inverse decay rate parameters compared to LLMs. Differences in exploration-related parameters (learning rate and gain) were also observed, indicating distinct exploration tendencies between the groups and across two parallel analyses. Finally, LLMs consistently demonstrated significantly higher decision stability (consistency) than human samples.
In conclusion, this research demonstrates the utility of combining the IGT, cognitive modeling, and a parallel analysis framework to dissect the nuances of sequential decision-making in both humans and LLMs. Although high-performing LLMs can achieve outcomes similar to those of human experts in the IGT, their underlying cognitive parameters differ substantially. Key distinctions include the LLMs’ heightened value sensitivity (similar to human experts) and greater strategic consistency, contrasted with humans’ superior integration of past outcomes, which is potentially linked to affective systems like somatic markers absent in LLMs. The VSE model provides a unified computational lens to examine these exploration-exploitation dynamics. These findings highlight the unique cognitive profile of LLMs. These findings offer insights for developing targeted training to enhance decision-making. They also inform the design of more effective human-LLM collaborative systems that capitalize on the distinct strengths of each agent, while acknowledging potential LLM limitations in less structured environments.
爱荷华赌博范式 / 大语言模型 / 价值序列探索模型 / 价值敏感度 / 逆衰减率 / 稳定性参数
Iowa gambling task / large language models / value plus sequential exploration model / value sensitivity / inverse decay / consistency
| [1] |
胡馨允, 沈悦, 戴俊毅. (2023). 系列决策任务中的策略转换: 来自爱荷华赌博任务的证据. 心理学报, 55(11), 1793.
|
| [2] |
徐四华. (2012). 网络成瘾者的行为冲动性——来自爱荷华赌博任务的证据. 心理学报, 44(11), 1523.
|
| [3] |
It is a hallmark of a good model to make accurate a priori predictions to new conditions (). This study compared 8 decision learning models with respect to their generalizability. Participants performed 2 tasks (the Iowa Gambling Task and the Soochow Gambling Task), and each model made a priori predictions by estimating the parameters for each participant from 1 task and using those same parameters to predict on the other task. Three methods were used to evaluate the models at the individual level of analysis. The first method used a post hoc fit criterion, the second method used a generalization criterion for short‐term predictions, and the third method again used a generalization criterion for long‐term predictions. The results suggest that the models with the prospect utility function can make generalizable predictions to new conditions, and different learning models are needed for making short‐versus long‐term predictions on simple gambling tasks.
|
| [4] |
Following damage to the ventromedial prefrontal cortex, humans develop a defect in real-life decision-making, which contrasts with otherwise normal intellectual functions. Currently, there is no neuropsychological probe to detect in the laboratory, and the cognitive and neural mechanisms responsible for this defect have resisted explanation. Here, using a novel task which simulates real-life decision-making in the way it factors uncertainty of premises and outcomes, as well as reward and punishment, we find that prefrontal patients, unlike controls, are oblivious to the future consequences of their actions, and seem to be guided by immediate prospects only. This finding offers, for the first time, the possibility of detecting these patients' elusive impairment in the laboratory, measuring it, and investigating its possible causes.
|
| [5] |
A decision-making instrument known as the "gambling task" was used, which has been shown to be sensitive to the decision-making impairment of patients with bilateral lesions of the ventromedial prefrontal cortex (VM). Three groups of subjects were tested, substance dependent individuals (SD) (n=41), normal controls (n=40), and VM patients (n=5). All SD met the DSM-IV criteria for dependence, with either alcohol or stimulants (metamphetamine or cocaine) as the primary substance of choice. The results revealed a significant impairment in the performance of SD relative to normal controls. A significantly high proportion of SD (61 vs. only 32.5% of normal controls) performed within the range of the VM patients, while the rest performed within the range of normal controls. General demographic factors such as age, sex, and level of education could not explain these differences in performance. As well, differences in performance were not explained by intelligence (IQ), memory, or performance on standard executive function/frontal lobe tests. Performance on the gambling task was best predicted by a combination of factors, including duration of abstinence, years of abuse, relapses and times in treatment, and the ability to hold gainful employment. The results support the hypothesis that impairment in decision-making linked to a dysfunctional VM cortex is associated with at least a sub-group of SD.
|
| [6] |
|
| [7] |
|
| [8] |
The aim of this study was to examine the impact of different neural systems on monetary decision making in frequent poker gamblers, who vary in their degree of problem gambling. Fifteen frequent poker players, ranging from non-problem to high-problem gambling, and 15 non-gambler controls were scanned using functional magnetic resonance imaging (fMRI) while performing the Iowa Gambling Task (IGT). During IGT deck selection, between-group fMRI analyses showed that frequent poker gamblers exhibited higher ventral-striatal but lower dorsolateral prefrontal and orbitofrontal activations as compared with controls. Moreover, using functional connectivity analyses, we observed higher ventral-striatal connectivity in poker players, and in regions involved in attentional/motor control (posterior cingulate), visual (occipital gyrus) and auditory (temporal gyrus) processing. In poker gamblers, scores of problem gambling severity were positively associated with ventral-striatal activations and with the connectivity between the ventral-striatum seed and the occipital fusiform gyrus and the middle temporal gyrus. Present results are consistent with findings from recent brain imaging studies showing that gambling disorder is associated with heightened motivational-reward processes during monetary decision making, which may hamper one's ability to moderate his level of monetary risk taking. © 2015 Society for the Study of Addiction.
|
| [9] |
|
| [10] |
|
| [11] |
|
| [12] |
|
| [13] |
|
| [14] |
|
| [15] |
|
| [16] |
|
| [17] |
|
| [18] |
|
| [19] |
|
| [20] |
It is argued that reasoning in the real world supports decision making and is aimed at the achievement of goals. A distinction is developed between two notions of rationality: rationality which is reasoning in such a way as to achieve one's goals--within cognitive constraints--and rationality which is reasoning by a process of logic. This dichotomy is related to the philosophical distinction between practical and theoretical reasoning. It is argued that logicality (rationality) does not provide a good basis for rationality and some psychological research on deductive reasoning is re-examined in this light. First, we review belief bias effects in syllogistic reasoning, and argue that the phenomena do not support the interpretations of irrationality that are often placed upon them. Second, we review and discuss recent studies of deontic reasoning in the Wason selection task, which demonstrate the decision making, and rational nature of reasoning in realistic contexts. The final section of the paper examines contemporary decision theory and shows how it fails, in comparable manner to formal logic, to provide an adequate model for assessing the rationality of human reasoning and decision making.
|
| [21] |
Noise--random disturbances of signals--poses a fundamental problem for information processing and affects all aspects of nervous-system function. However, the nature, amount and impact of noise in the nervous system have only recently been addressed in a quantitative manner. Experimental and computational methods have shown that multiple noise sources contribute to cellular and behavioural trial-to-trial variability. We review the sources of noise in the nervous system, from the molecular to the behavioural level, and show how noise contributes to trial-to-trial variability. We highlight how noise affects neuronal networks and the principles the nervous system applies to counter detrimental effects of noise, and briefly discuss noise's potential benefits.
|
| [22] |
Feher da Silva, C.,
A standard assumption in neuroscience is that low-effort model-free learning is automatic and continuously used, whereas more complex model-based strategies are only used when the rewards they generate are worth the additional effort. We present evidence refuting this assumption. First, we demonstrate flaws in previous reports of combined model-free and model-based reward prediction errors in the ventral striatum that probably led to spurious results. More appropriate analyses yield no evidence of model-free prediction errors in this region. Second, we find that task instructions generating more correct model-based behaviour reduce rather than increase mental effort. This is inconsistent with cost-benefit arbitration between model-based and model-free strategies. Together, our data indicate that model-free learning may not be automatic. Instead, humans can reduce mental effort by using a model-based strategy alone rather than arbitrating between multiple strategies. Our results call for re-evaluation of the assumptions in influential theories of learning and decision-making.© 2023. The Author(s), under exclusive licence to Springer Nature Limited.
|
| [23] |
|
| [24] |
|
| [25] |
The Iowa Gambling Task (IGT) is widely used to study decision-making within healthy and psychiatric populations. However, the complexity of the IGT makes it difficult to attribute variation in performance to specific cognitive processes. Several cognitive models have been proposed for the IGT in an effort to address this problem, but currently no single model shows optimal performance for both short- and long-term prediction accuracy and parameter recovery. Here, we propose the Outcome-Representation Learning (ORL) model, a novel model that provides the best compromise between competing models. We test the performance of the ORL model on 393 subjects' data collected across multiple research sites, and we show that the ORL reveals distinct patterns of decision-making in substance-using populations. Our work highlights the importance of using multiple model comparison metrics to make valid inference with cognitive models and sheds light on learning mechanisms that play a role in underweighting of rare events.© 2018 Cognitive Science Society, Inc.
|
| [26] |
|
| [27] |
In the past two decades neuroimaging research has substantiated the important role of the prefrontal cortex (PFC) in decision-making. In the current study, we use the complementary lesion based approach to deepen our knowledge concerning the specific cognitive mechanisms modulated by prefrontal activity. Specifically, we assessed the brain substrates implicated in two decision making dimensions in a sample of prefrontal cortex patients: (a) the tendency to differently weigh recent compared to past experience; and (b) the tendency to differently weigh gains compared to losses. The participants performed the Iowa Gambling Task, a complex experience-based decision-making task, which was analyzed with a formal cognitive model (the Expectancy-Valance model). The results indicated that decisions become influenced by more recent, as opposed to older, events when the damage reaches the posterior sectors of the ventromedial prefrontal cortex (VMPC). Furthermore, the degree of this recency deficit was related to the size of the lesion. These results suggest that the posterior area of the prefrontal cortex directly modulates the capacity to use time-delayed information. In contrast, we did not find similar modulation for the sensitivity to gains versus losses.(c) 2010 Elsevier B.V. All rights reserved.
|
| [28] |
|
| [29] |
|
| [30] |
|
| [31] |
|
| [32] |
|
| [33] |
|
| [34] |
|
| [35] |
|
| [36] |
|
| [37] |
|
| [38] |
Scientific discoveries often hinge on synthesizing decades of research, a task that potentially outstrips human information processing capacities. Large language models (LLMs) offer a solution. LLMs trained on the vast scientific literature could potentially integrate noisy yet interrelated findings to forecast novel results better than human experts. Here, to evaluate this possibility, we created BrainBench, a forward-looking benchmark for predicting neuroscience results. We find that LLMs surpass experts in predicting experimental outcomes. BrainGPT, an LLM we tuned on the neuroscience literature, performed better yet. Like human experts, when LLMs indicated high confidence in their predictions, their responses were more likely to be correct, which presages a future where LLMs assist humans in making discoveries. Our approach is not neuroscience specific and is transferable to other knowledge-intensive endeavours.© 2024. The Author(s).
|
| [39] |
|
| [40] |
|
| [41] |
|
| [42] |
|
| [43] |
|
| [44] |
In everyday life, risky decision-making relies on multiple cognitive processes including sensitivity to reinforcers, exploration, learning, and forgetting. Neuroimaging evidence suggests that the dorsolateral prefrontal cortex (DLPFC) is involved in exploration and risky decision-making, but the nature of its computations and its causal role remain uncertain. We provide evidence for the role of the DLPFC in value-independent, directed exploration on the Iowa Gambling Task (IGT) and we describe a new computational model to account for the competition of directed exploration and exploitation in guiding decisions. Forty-two healthy human participants were included in a right DLPFC, left DLPFC or sham stimulation groups using continuous theta-burst stimulation (cTBS). Immediately after cTBS, the IGT was completed. Computational modelling was used to account for exploration and exploitation with different combinations with value-based and sensitivity to reinforcers for each group. Applying cTBS to the left and right DLPFC selectively decreased directed exploration on the IGT compared to sham stimulation. Model-based analyses further indicated that the right (but not the left) DLPFC stimulation increased sensitivity to reinforcers, leading to avoidance of risky choices and promoting advantageous choices during the task. Although these findings are based on small sample sizes per group, they nevertheless elucidate the causal role of the right DLPFC in governing the exploration-exploitation tradeoff during decision-making in uncertain and ambiguous contexts.Copyright © 2021 IBRO. Published by Elsevier Ltd. All rights reserved.
|
| [45] |
|
| [46] |
From Piaget to the present, traditional and dual-process theories have predicted improvement in reasoning from childhood to adulthood, and improvement has been observed. However, developmental reversals-that reasoning biases emerge with development -have also been observed in a growing list of paradigms. We explain how fuzzy-trace theory predicts both improvement and developmental reversals in reasoning and decision making. Drawing on research on logical and quantitative reasoning, as well as on risky decision making in the laboratory and in life, we illustrate how the same small set of theoretical principles apply to typical neurodevelopment, encompassing childhood, adolescence, and adulthood, and to neurological conditions such as autism and Alzheimer's disease. For example, framing effects-that risk preferences shift when the same decisions are phrases in terms of gains versus losses-emerge in early adolescence as gist-based intuition develops. In autistic individuals, who rely less on gist-based intuition and more on verbatim-based analysis, framing biases are attenuated (i.e., they outperform typically developing control subjects). In adults, simple manipulations based on fuzzy-trace theory can make framing effects appear and disappear depending on whether gist-based intuition or verbatim-based analysis is induced. These theoretical principles are summarized and integrated in a new mathematical model that specifies how dual modes of reasoning combine to produce predictable variability in performance. In particular, we show how the most popular and extensively studied model of decision making-prospect theory-can be derived from fuzzy-trace theory by combining analytical (verbatim-based) and intuitive (gist-based) processes.
|
| [47] |
|
| [48] |
|
| [49] |
|
| [50] |
|
| [51] |
|
| [52] |
|
| [53] |
|
| [54] |
|
| [55] |
|
| [56] |
|
| [57] |
|
| [58] |
|
| [59] |
Models of human behavior in the Iowa Gambling Task (IGT) have played a pivotal role in accounting for behavioral differences during decision-making. One critical difference between models that have been used to account for behavior in the IGT is the inclusion or exclusion of the assumption that participants tend to persevere, or stay with the same option over consecutive trials. Models that allow for this assumption include win-stay-lose-shift (WSLS) models and reinforcement learning (RL) models that include a decay learning rule where expected values for each option decay as they are chosen less often. One shortcoming of RL models that have included decay rules is that the tendency to persevere by sticking with the same option has been conflated with the tendency to select the option with the highest expected value because a single term is used to represent both of these tendencies. In the current work we isolate the tendencies to perseverate and to select the option with the highest expected value by including them as separate terms in a Value-Plus-Perseveration (VPP) RL model. Overall the VPP model provides a better fit to data from a large group of participants than models that include a single term to account for both perseveration and the representation of expected value. Simulations of each model show that the VPP model's simulated choices most closely resemble the decision-making behavior of human subjects. In addition, we also find that parameter estimates of loss aversion are more strongly correlated with performance when perseverative tendencies and expected value representations are decomposed as separate terms within the model. The results suggest that the tendency to persevere and the tendency to select the option that leads to the best net payoff are central components of decision-making behavior in the IGT. Future work should use this model to better examine decision-making behavior.
|
| [60] |
|
感谢智谱MaaS开放平台的模型支持。
/
| 〈 |
|
〉 |
