[Paper Review] CogBench: a large language model walks into a psychology lab

Julian Coda-Forno, Marcel Binz+2 more|arXiv (Cornell University)|Feb 28, 2024

Innovative Teaching and Learning Methods5 citations

TL;DR

CogBench introduces a behavioral benchmark for LLMs with ten metrics from seven cognitive psychology experiments, evaluating 35 LLMs to profile performance and cognitive-like behaviors beyond standard accuracy.

ABSTRACT

Large language models (LLMs) have significantly advanced the field of artificial intelligence. Yet, evaluating them comprehensively remains challenging. We argue that this is partly due to the predominant focus on performance metrics in most benchmarks. This paper introduces CogBench, a benchmark that includes ten behavioral metrics derived from seven cognitive psychology experiments. This novel approach offers a toolkit for phenotyping LLMs' behavior. We apply CogBench to 35 LLMs, yielding a rich and diverse dataset. We analyze this data using statistical multilevel modeling techniques, accounting for the nested dependencies among fine-tuned versions of specific LLMs. Our study highlights the crucial role of model size and reinforcement learning from human feedback (RLHF) in improving performance and aligning with human behavior. Interestingly, we find that open-source models are less risk-prone than proprietary models and that fine-tuning on code does not necessarily enhance LLMs' behavior. Finally, we explore the effects of prompt-engineering techniques. We discover that chain-of-thought prompting improves probabilistic reasoning, while take-a-step-back prompting fosters model-based behaviors.

Motivation & Objective

• Motivate moving beyond performance-only benchmarks to understand underlying LLM behaviors using cognitive psychology paradigms.
• Provide a standardized, open benchmark to phenotype LLMs across diverse tasks and behaviors.
• Assess how model size, RLHF, and other training choices relate to both performance and cognitive-like behaviors.
• Explore how prompt-engineering techniques influence probabilistic reasoning and model-based behaviors.

Proposed method

• Evaluate 35 LLMs (proprietary and open-source) using in-context prompts with deterministic settings (temperature=0).
• Adopt ten behavioral metrics derived from seven canonical cognitive psychology experiments.
• Apply computational cognitive modeling and multilevel regression to analyze nested model hierarchies (e.g., fine-tuned versions).
• Use prompt-chaining to collect past behavior and induce learning without model fine-tuning.
• Compare RLHF versus non-RLHF models via dimensionality reduction (UMAP) and regression analyses.
• Examine the impact of chain-of-thought (CoT) and take-a-step-back (SB) prompting on probabilistic reasoning and model-basedness.

Experimental results

Research questions

• RQ1 How do LLMs differ in probabilistic reasoning, exploration, meta-cognition, model-basedness, temporal discounting, risk-taking, and learning strategies as measured by CogBench?
• RQ2 To what extent do model size, data scale, and code-training influence performance and behavioral phenotypes?
• RQ3 Does RLHF align LLMs more closely with human behavior and enhance meta-cognition?
• RQ4 What are the effects of prompting techniques (CoT, SB) on specific behavioral metrics?
• RQ5 Do open-source models exhibit different risk profiles compared with proprietary models?

Key findings

• Larger models generally perform better and show more model-based behavior.
• RLHF models behave more human-like and estimate uncertainty more accurately.
• Open-source models tend to take fewer risks than proprietary models when controlling for other factors.
• Code-fine-tuning shows little evidence of enhancing behavior across CogBench tasks.
• CoT prompting improves probabilistic reasoning; SB prompting enhances model-based behaviors.