Source-linked AI summary
Using cognitive psychology to understand GPT-3
Marcel Binz, Eric Schulz
TL;DR
The paper asks how GPT-3 learns, decides, searches for information, deliberates, and reasons causally, and evaluates it with canonical cognitive-psychology experiments. Across vignette-based and task-based tests, GPT-3 often performs impressively, but remains highly sensitive to prompt changes and fails on directed exploration and causal reasoning. The findings support systematic psychological evaluation of increasingly capable artificial agents.
Problem
The paper addresses limited understanding of how increasingly capable and opaque large language models learn and make decisions.
Method
The authors evaluate GPT-3 with cognitive-psychology experiments covering decision-making, information search, deliberation, and causal reasoning in vignette-based and task-based settings.
Results
GPT-3 often performs similarly to or better than humans, but small vignette changes cause major errors, directed exploration is absent, and causal reasoning fails.
Takeaways & Limitations
Cognitive-psychology experiments provide a systematic way to characterize both impressive capabilities and important gaps in large language models.
Takeaways & Limitations
The study covers only a small subset of cognitive tasks, and canonical tasks may be too easy to reveal GPT-3’s actual intelligence.
Abstract
from arXiv · showhide
We study GPT-3, a recent large language model, using tools from cognitive psychology. More specifically, we assess GPT-3's decision-making, information search, deliberation, and causal reasoning abilities on a battery of canonical experiments from the literature. We find that much of GPT-3's behavior is impressive: it solves vignette-based tasks similarly or better than human subjects, is able to make decent decisions from descriptions, outperforms humans in a multi-armed bandit task, and shows signatures of model-based reinforcement learning. Yet we also find that small perturbations to vignette-based tasks can lead GPT-3 vastly astray, that it shows no signatures of directed exploration, and that it fails miserably in a causal reasoning task. These results enrich our understanding of current large language models and pave the way for future investigations using tools from cognitive psychology to study increasingly capable and opaque artificial agents.
Introduction
The paper uses cognitive psychology to examine how GPT-3 makes decisions and reasons, motivated by the need to understand increasingly capable artificial agents.
- The study investigates GPT-3 using experiments from cognitive psychology.It focuses on decision-making, information search, deliberation, and causal reasoning.
- The experiments compare GPT-3’s performance with human behavior across vignette-based and task-based settings.Vignette-based tasks describe hypothetical scenarios, whereas task-based experiments are generated trial by trial.
- The authors report that GPT-3 can solve challenging vignette-based problems and make human-level decisions in descriptive task settings.They also report that its behavior depends strongly on how vignettes are presented.
GPT-3
GPT-3 is a large autoregressive transformer language model trained on extensive text data, and Figure 1 uses vignette-based tasks to illustrate both human-like and non-human-like responses.
- GPT-3 is an auto-regressive language model that uses transformer self-attention to process sequential data.Transformers process the entire input sequence while attention provides context for each position.
- Figure 1 presents a hypothetical vignette prompt and compares GPT-3’s responses on standard versus adversarial vignettes.The model responded correctly or with human-like mistakes on all 12 standard vignettes, but made non-human-like mistakes on adversarial versions.
- GPT-3 has 175 billion parameters and was trained on hundreds of billions of words from internet and book text.Its scale supports in-context learning, where instructions and examples specify what task to perform.
A cognitive psychology view on GPT-3
The paper combines canonical vignette-based and programmatically generated task-based experiments to evaluate GPT-3’s cognitive abilities while addressing possible training-data contamination and prompt sensitivity.
- The study evaluates GPT-3 on four cognitive domains: decision-making, information search, deliberation, and causal reasoning.
- Vignette-based investigations present text descriptions of hypothetical situations and collect GPT-3’s responses.
- The authors supplement vignettes with task-based assessments because GPT-3 may have encountered identical or similar problems during training.They also test adversarial variants created by marginally changing the vignettes.
Results
The simulations used OpenAI’s public API, focusing on the most powerful available GPT-3 model and deterministic generation settings.
- The study ran all simulations through the public OpenAI API.The API provided four GPT-3 models ordered from least to most complex.
- The investigation primarily used the Davinci model, the most powerful model available through the API.
- The temperature was set to 0, producing deterministic answers, while other parameters retained their default values.
Vignette-based investigations
GPT-3 often produced human-like or correct responses across canonical vignette-based tasks, spanning decision-making, information search, deliberation, and causal reasoning. However, its performance was highly sensitive to prompt wording, and several tasks exposed failures or human-like biases.
- Decision-making: Heuristics and biases: GPT-3 reproduced the conjunction fallacy in the Linda problem, choosing that Linda was both a bank teller and an active feminist.This matched the response commonly observed in people.
- Decision-making: Heuristics and biases: GPT-3 avoided the cab problem’s base-rate fallacy and provided an approximately correct answer, unlike people who commonly ignore taxi-color base rates.
- Decision-making: Heuristics and biases: GPT-3, like people, judged the hospital problem’s two probabilities as about equal instead of selecting the smaller hospital, the statistically correct answer.The smaller hospital is correct because smaller samples have greater variance.
- Information search: Questions and hypothesis tests: GPT-3 selected the appropriate question in each Toma scenario, adapting between hypothesis-scanning and constraint-seeking information search.Which question type is informative depended on the previously observed reasons for Toma’s lateness.
- Deliberation and hypothesis testing: GPT-3 showed several human-like reasoning patterns: it shared the test-selection congruence bias, gave the accepted Wason answer, and failed all three Cognitive Reflection Test items.For the CRT, it consistently produced intuitive but incorrect answers rather than overriding them through further deliberation.
- Causal reasoning and limitations: GPT-3 correctly identified blickets, selected the object to remove in an intervention, and answered mature causal counterfactuals correctly, but vignette results remained vulnerable to small prompt changes.Across 12 vignette problems, it answered six correctly and all 12 in a human-like way; modified adversarial vignettes could produce vastly different responses.
Task-based investigations
Task-based experiments tested GPT-3 on decisions from descriptions, information search, deliberation, and causal reasoning using dynamically generated trials. Performance was strong in several tasks, but exploration and causal inference revealed important weaknesses and task-dependent reasoning patterns.
- Decision-making: Decisions from descriptions: GPT-3 solved over 13,000 decisions-from-descriptions problems, with Davinci above chance but below human-level performance.The three smaller GPT-3 models did not perform above chance.
- Decision-making: Decisions from descriptions: GPT-3 displayed framing, certainty, and overweighting biases, while three other biases were not observed.The overweighting bias concerned differences between small versus larger probabilities.
- Information search: Directed and Random Exploration: In the horizon task, GPT-3 matched humans on short horizons and achieved lower overall regret than humans across the experiment.Overall regret was M = 2.72 for GPT-3 versus M = 3.24 for humans.
- Information search: Directed and Random Exploration: GPT-3 showed no evidence of directed exploration and tended to repeat options that had previously provided more samples.The model’s repetition bias could favor familiar options even when additional sampling was not justified.
- Deliberation: Model-based and model-free reinforcement learning: In the two-step task, GPT-3’s reward-and-transition patterns suggested a deliberate model-based strategy.Repeating actions decreased after rewards through rare transitions and increased after unrewarded rare transitions.
- Causal reasoning: Interventions after passive observations: GPT-3 incorporated causal structure incorrectly: it made some normative common-cause inferences but treated causal-chain cases identically.This differed from both normative expectations and human judgments in the causal-chain condition.
Discussion
The paper argues that systematic psychological experiments reveal both impressive capabilities and important limits in GPT-3. GPT-3 can perform several cognitive tasks well, but prompt sensitivity, absent directed exploration, and failures in causal reasoning constrain interpretation.
- Discussion: GPT-3’s impressive task performance risks overestimating its intelligence because simple behaviors can resemble human intelligence without establishing comparable cognition.The paper invokes the Clever Hans case as a warning against interpreting apparently intelligent behavior without systematic investigation.
- Discussion: GPT-3 solved some vignette-based experiments similarly to or better than humans, but its responses depended strongly on prompt presentation and possible training-data exposure.The authors report that small changes to vignettes could substantially reduce performance.
- Discussion: GPT-3 made reasonable decisions from descriptions, mirrored some human biases, and performed better than humans on a multi-armed bandit task.In the bandit task, it showed random but not directed exploration.
- Discussion: GPT-3 showed signatures of model-based reinforcement learning but no signatures of directed exploration.The authors relate this contrast to differences between human active learning and GPT-3’s passive text-based training.
- Discussion: GPT-3 failed to learn and use causal knowledge in a simple reasoning task, producing responses that were neither correct nor human-like.The paper identifies causal reasoning as a frequently proposed pillar of intelligent behavior.
- Discussion: The study contributes to a broader effort to use cognitive psychology to understand capable, opaque foundation models.The authors position their work alongside benchmark and cognitive-psychology approaches to studying large-scale models.
- Discussion: The study covers only a small subset of cognitive tasks, and canonical tasks may be too easy to distinguish GPT-3’s abilities from task solvability.The authors call for additional psychological domains and more complex paradigms.
Vigettes
The vignette section records GPT-3’s responses to canonical, text-described cognitive tasks and documents both human-like answers and vulnerabilities to task presentation. The materials include decision-making, information search, deliberation, causal reasoning, and deliberately confusing prompts.
- Vignettes: The tables document the exact prompts and corresponding answers used in the vignette-based assessments.The paper states that these materials provide a detailed description of submitted prompts and GPT-3’s answers.
- Vignettes: The vignette materials present prompts and GPT-3’s answers for decision-making, information search, intuitive-versus-reflective reasoning, and causal reasoning.Additional prompts were deliberately designed to confuse GPT-3.
- Information search: GPT-3 was tested on a repeated-lateness scenario asking whether constraint-seeking or hypothesis-scanning questions would better identify why Toma was late.The supplied example contrasts asking about a shared feature with asking about one specific cause.
Decisions from descriptions
The decisions-from-descriptions analysis tests GPT-3 on paired risky-choice problems modeled on prospect-theory contrasts. It estimates choice probabilities and log-odds contrasts to assess whether GPT-3 exhibits established human decision biases.
- Method: GPT-3’s bias analysis contrasts choices across paired decisions from description problems using option-F choice probabilities and log-odds ratios.The order of presented options was counterbalanced, and GPT-3’s temperature parameter was set to 1.
- Method: The analysis follows Kahneman and Tversky’s strategy of identifying cognitive biases by comparing responses to carefully selected problem pairs.The contrasts are listed in Table 7, while the underlying decision problems appear in Table 6.
- Decision contrasts: The loss contrasts include a sure-loss comparison, intermediate-probability losses, and very-low-probability losses.Examples include 80% versus 100%, 20% versus 25%, 45% versus 90%, and 0.1% versus 0.2% loss probabilities.
Two-step task
The two-step task section compares GPT-3’s behavior with simulations from model-free and model-based reinforcement-learning algorithms. Figure 6 presents the simulated behavior of both algorithms.
- Two-step task: The figure provides simulated behavior from the two reinforcement-learning algorithms as reference patterns for the task.The supplied description identifies the algorithms but does not state a specific empirical comparison or outcome.
- Two-step task: Figure 6 shows model simulations on the two-step task for a model-free reinforcement-learning algorithm and a model-based reinforcement-learning algorithm.The figure is adapted from Daw et al.