Core Concepts
Large language models' behaviors are evaluated using cognitive psychology experiments in CogBench.
Abstract
1. Abstract:
Large language models (LLMs) have advanced AI.
CogBench introduces a benchmark with behavioral metrics.
Model size and RLHF improve performance.
Open-source models are less risky.
Prompt-engineering techniques impact behaviors.
2. Introduction:
LLMs' significance and challenges in understanding behavior.
Cognitive psychology experiments offer insights.
CogBench fills the gap in evaluating LLMs holistically.
3. Methods:
CogBench evaluates LLM behaviors using cognitive experiments.
Ten metrics from seven experiments provide insights.
Prompting techniques influence behaviors.
4. The Cognitive Phenotype of LLMs:
Performance metrics and behavioral metrics compared.
Model-basedness and risk-taking behaviors analyzed.
Open-source models exhibit less risk-taking behavior.
5. Hypothesis-driven Experiments:
RLHF enhances human-likeness.
Number of parameters influences performance.
Model-basedness affected by parameters.
RLHF enhances meta-cognition.
Open-source models exhibit more risk-taking behavior.
6. Impact of Prompt-Engineering:
CoT and SB techniques enhance probabilistic reasoning and model-basedness.
CoT effective for probabilistic reasoning, SB for model-basedness.
7. Discussion:
CogBench offers a unique benchmark for LLM evaluation.
Findings on RLHF, parameters, and prompt-engineering techniques.
Challenges in proprietary model transparency and future directions.
Stats
CogBench introduces a benchmark with behavioral metrics.
Model size and RLHF improve performance.
Open-source models are less risky.
Prompt-engineering techniques impact behaviors.
Quotes
"RLHF enhances the human-likeness of LLMs."
"Number of parameters influences performance."
"Open-source models exhibit less risk-taking behavior."