Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Understanding The Effect Of Temperature On Alignment With Human Opinions

Maja Pavlovic, Massimo Poesio

TL;DR

An empirical analysis of three straightforward methods for obtaining distributions suggests that sampling and log-probability approaches with simple parameter adjustments can return better aligned outputs in subjective tasks compared to direct prompting.

Abstract

With the increasing capabilities of LLMs, recent studies focus on understanding whose opinions are represented by them and how to effectively extract aligned opinion distributions. We conducted an empirical analysis of three straightforward methods for obtaining distributions and evaluated the results across a variety of metrics. Our findings suggest that sampling and log-probability approaches with simple parameter adjustments can return better aligned outputs in subjective tasks compared to direct prompting. Yet, assuming models reflect human opinions may be limiting, highlighting the need for further research on how human subjectivity affects model uncertainty.

Understanding The Effect Of Temperature On Alignment With Human Opinions

TL;DR

An empirical analysis of three straightforward methods for obtaining distributions suggests that sampling and log-probability approaches with simple parameter adjustments can return better aligned outputs in subjective tasks compared to direct prompting.

Abstract

With the increasing capabilities of LLMs, recent studies focus on understanding whose opinions are represented by them and how to effectively extract aligned opinion distributions. We conducted an empirical analysis of three straightforward methods for obtaining distributions and evaluated the results across a variety of metrics. Our findings suggest that sampling and log-probability approaches with simple parameter adjustments can return better aligned outputs in subjective tasks compared to direct prompting. Yet, assuming models reflect human opinions may be limiting, highlighting the need for further research on how human subjectivity affects model uncertainty.

Paper Structure

This paper contains 16 sections, 2 equations, 5 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Related work
  3. Methodology
  4. Methods
  5. Evaluation metrics
  6. Datasets
  7. Results
  8. Discussion
  9. Conclusion
  10. Limitations
  11. Prompts
  12. Hate-Speech Brexit:
  13. Conversational Abuse:
  14. Multi-Domain Agreement:
  15. Datasets
  16. ...and 1 more sections

Figures (5)

  • Figure 1: MD-Agree. T=2.0; ideally, if entropies align, a model should display a diagonal line from bottom left to top right; lower L1 is better
  • Figure 2: Histograms of human and GPT entropy levels for SemEval2023 datasets: directly generated GPT distributions vs. those generated with the MCE method.
  • Figure 3: Ideally, if entropies align, a model should display a diagonal line from bottom left to top right; A lower manhattan distance (L1) is better. The majority of samples fall under human agreement (0 entropy) with GPT’s prediction confidence high (0 entropy) on such samples.
  • Figure 4: Average L1-distance manhattan-distance from model distributions in relation to entropy values of human opinion distributions: The sampling approach (MCE) performs more in line with expectations across all datasets, showing that this method exhibits greater confidence on samples with full human agreement (0 entropy); while the direct method doesn't capture this confidence on easier instances as effectively.
  • Figure 5: Distribution for each run (direct, MC and LP). Human Labels have more variation in HS-Brexit and MD-Agreement datasets.