PaperlayerView exact revision on arXiv

Source-linked AI summary

Aligning Language Models to User Opinions

EunJeong Hwang, Bodhisattwa Prasad Majumder, Niket Tandon

arXiv:2305.14929v1cs.CL

TL;DR

Aligning LLMs to individual users remains challenging because demographics and ideology do not reliably predict their opinions. The paper combines these persona attributes with users’ opinions, including relevant past opinions, to predict survey responses. It reports accuracy gains of up to 7 points on certain topics and improved predictions from relevant past opinions.

  • Problem

    How to align LLMs with individual users rather than demographic or ideological groups remains open, since people sharing demographic traits can still have different opinions.

    Source: §1 · p. 1

  • Method

    The study analyzes OpinionQA and evaluates GPT-3 using combinations of users’ demographics, ideology, and past opinions as persona information.

    Source: §1 · p. 2

  • Results

    Incorporating user opinions, demographics, and ideology yields up to 7 points in QA accuracy gains on certain topics, while relevant past opinions improve predictions.

    Source: §1 · p. 2

  • Takeaways & Limitations

    The paper concludes that aligning to users is best done by modeling demographics, ideologies, and the most relevant past opinions.

    Source: §6 · p. 8

Abstract

from arXiv · show

An important aspect of developing LLMs that interact with humans is to align models' behavior to their users. It is possible to prompt an LLM into behaving as a certain persona, especially a user group or ideological persona the model captured during its pertaining stage. But, how to best align an LLM with a specific user and not a demographic or ideological group remains an open question. Mining public opinion surveys (by Pew Research), we find that the opinions of a user and their demographics and ideologies are not mutual predictors. We use this insight to align LLMs by modeling both user opinions as well as user demographics and ideology, achieving up to 7 points accuracy gains in predicting public opinions from survey questions across a broad set of topics. In addition to the typical approach of prompting LLMs with demographics and ideology, we discover that utilizing the most relevant past opinions from individual users enables the model to predict user opinions more accurately.

1 Introduction

Aligning LLMs to individual users is difficult because group personas do not capture the variation in users’ opinions, even among people with the same demographics. The paper studies how demographics, ideology, and past opinions can be combined to predict individual responses.

1 Introduction

  • Users can share all demographic traits yet differ substantially in opinion, so group-based alignment may not represent an individual user.The introduction frames the research question as whether alignment should use demographics, fine-grained opinions, or both.Source: §1 · p. 1
  • User personas based on normalized groups such as religion or political inclination leave individual and long-tail opinions less addressed.The paper notes that users’ opinions can change over time and vary with context.Source: §1 · p. 1
  • The paper models individual opinions using demographic traits, ideological inclinations, and past opinions.It presents these inputs as complementary persona information for predicting user responses.Source: §1 · p. 2
  • Combining user opinions, demographics, and ideology yields up to 7 points in QA accuracy gains on certain topics; relevant past opinions further improve predictions.The study analyzes OpinionQA and evaluates GPT-3 with different combinations of these inputs.Source: §1 · p. 2

2 Related work

Prior personalization work often focused on normalized groups or domain-specific systems, while effective alignment to individual users remained an open question. This paper positions opinion-based personalization as a way to address that gap.

2 Related work

  • Earlier systems for individual users mainly came from recommender systems and relied on domain-specific annotated data or latent user information.The passage gives previously written reviews as an example of user information that can be sparse.Source: §2 · p. 2
  • Whether LLMs can align effectively with individual personas, and how demographics compare with past opinions, remained open questions.The authors connect this gap to limited long-tail representation in pretraining and a lack of large-scale individual-opinion datasets until recently.Source: §2 · p. 2
  • Related studies connect ideology with consumer behavior and decision strategies; this paper examines ideology alongside other factors in LLM prediction of user opinions.The cited studies include Crockett and Wallendorf (2004) and Chan and Palmeira (2021).Source: §2 · p. 2
  • The paper builds on retrieval-based QA and user-opinion memory work, but focuses on tuning LLMs to individual users rather than user groups.It distinguishes this focus from prior LLM work on group-level personalization.Source: §2 · p. 3

3 What makes a persona?

The paper examines how demographics, ideology, and opinions contribute to a user’s persona, then analyzes agreement patterns in OpinionQA. Users with shared demographics or ideology do not always share opinions, and agreement varies by topic.

3 What makes a persona?

  • OpinionQA covers 15 topics, averaging about 100 questions and 5,340 users per topic.The study analyzes components of user persona using this public-survey dataset.Source: §3 · p. 3

3.1 Demographics

  • The dataset records eight demographic attributes, including region, sex, age, education, race, citizenship, marital status, and income.The paper notes that relying on demographic information alone risks stereotyping and is not enough to model an individual.Source: §3.1 · p. 3

3.2 Ideology

  • Ideology is represented through each subject’s political affiliation and inclinations toward political ideologies such as conservative or liberal.Source: §3.2 · p. 3

3.3 Opinions

  • OpinionQA captures opinions through subjective public-survey questions intended to reflect what distinguishes one respondent from another.The paper treats these survey responses as users’ implicit opinions.Source: §3.3 · p. 3

3.4 Deriving insights from public surveys

  • Among users with the same demographics, implicit-opinion agreement scores cluster around 0.5, indicating demographics alone are insufficient for personalization.The authors calculate agreement with Cohen’s kappa, which ranges from −1 to 1.Source: §3.4 · p. 3
  • Agreement varies by topic: Family & Relationships and Guns show relatively higher agreement, while Race and America in 2050 show lower agreement.Lower agreement indicates greater variability in opinions across users.Source: §3.4 · p. 3
  • Among pairs with similar opinions, around 80% have different ideologies, while similar ideology is associated with a greater likelihood of similar opinions.The comparison uses pairs who answered more than 10 common questions.Source: §3.4 · p. 4

4 Aligning LLMs with persona

The paper evaluates prompting an LLM to predict individual survey answers from persona information, including demographics, ideology, and past opinions. It compares persona variants and tests whether question-relevant past opinions improve prediction.

4.1 Setup

  • The setup samples 100 users per topic, uses 20% of each user’s implicit questions as persona information, and tests personalization on the rest.The task predicts survey multiple-choice answers with text-davinci-003.Source: §4.1 · p. 5
  • The evaluated variants range from no persona and ideology to combinations of demographics, ideology, and past opinions.The comparisons include ideology plus opinions and the combination of demographic, ideological, and opinion information.Source: §4.1 · p. 5
  • Accuracy is measured both by matching the original answer choice and by collapsing answer choices into broader classes.The collapsed measure groups choices such as “Very likely” and “Somewhat likely” together as “Likely.”Source: §4.1 · p. 5

4.2 LLM as a person with ideologies, demographic, and opinions

  • Past opinions are selected by embedding similarity to the question, and the approach compares all available opinions with top-k selections.The top-k settings are k = 3, 5, or 8; Figure 3 shows a prompt using demographics, ideology, and embedding-based past opinions.Source: Figure 3 · p. 5, §4.2 · p. 6

5 Results and Analysis

The individual-user experiments find that demographic and ideology information combined with past opinions improves prediction, while selecting relevant opinions affects exact-match results more than collapsed accuracy. Group-level tests show that the LLM predicts majority answers moderately well and that adding ideology offers little help.

5.1 LLM for an individual

  • Adding demographics and ideology with up to 16 past opinions substantially improves overall exact-match and collapsed accuracy over demographic information alone.Demographics and ideology also outperform using no persona information.Source: §5.1 · p. 6
  • Top-k opinions improve exact-match accuracy by up to 3 points, but do not significantly increase collapsed accuracy.Top-3 opinions perform similarly to top-8 and to demographic, ideology, and 16 random past opinions.Source: §5.1 · pp. 6–7
  • Topic-wise, demographics plus implicit opinions score higher on Biomedical-food and Guns, but performance decreases slightly on Automation.The authors note that implicit opinions may include viewpoints misaligned with a user’s demographics or ideology, potentially confusing predictions.Source: §5.1 · p. 7, Table 4 · p. 8
  • An irrelevant past opinion can confuse the model: an answer based on demographics becomes incorrect after it sees a mismatched opinion.The example involves an opinion stating “does not describe me well” that also appears among the question’s answer choices.Source: §5.1 · p. 7, Figure 4 · p. 7

5.2 LLM with majority answer choices

  • Without ideology information, majority-answer prediction reaches 0.549 exact-match accuracy and 0.659 collapsed accuracy.The experiment treats the most common answer within an ideological group as the gold answer and prompts the model as a group representative.Source: §5.2 · p. 7

5.3 LLM as a person with an ideology

  • With ideology information, group-majority prediction reaches 0.566 exact-match accuracy and 0.667 collapsed accuracy; the added ideology is not particularly helpful.The authors relate this pattern to individual opinions not necessarily aligning with their group’s majority opinion.Source: §5.3 · p. 8, Table 5 · p. 8

6 Discussion and Conclusion

The conclusion presents demographic traits, ideology, and relevant past opinions as ingredients for aligning models with individual users. It also notes echo-chamber risks and proposes ways to support reflection and future personalization.

Ethical concerns

  • Personalized answers may amplify biased or unethical views and reinforce undesirable or polarized views, potentially causing harm.Source: §6 · p. 8
  • Showing demographic or ideology-group answers alongside personalized answers is proposed as a mitigation for echo-chamber risk.Explanations grounded in a user's past opinions may also help users introspect on those opinions.Source: §6 · p. 8

Extensions

  • Future work could store user opinions in a growing memory and increasingly rely on them, backing off to demographic or group-level opinions as needed.Source: §6 · p. 8

Conclusion

  • The paper concludes that aligning LLMs to users is best done by modeling demographics, ideologies, and the most relevant past opinions.Large-scale OpinionQA experiments show approximately 7% absolute QA accuracy over strong demography-based baselines.Source: §6 · p. 8

A Prompt

The appendix provides prompts used in individual-user and group-level tests, including prompts for models using demographics, ideology, and opinions.

A Prompt

  • The appendix provides prompts for models incorporating user demographics, ideology, and opinions in individual-user tests.Source: A Prompt · p. 10
  • The passage identifies prompts used in models incorporating user demographics, ideology, and opinions, which were employed for individual-user-level tests shown in Figures 5 and 6.Source: A Prompt · p. 10
  • Figures 7 and 8 present the prompts utilized for group-level experiments, distinct from the individual-user level tests associated with Figures 5 and 6.Source: A Prompt · p. 10

B Similar ideologies and different opinions

This appendix includes prompts illustrating opinion and demographic inputs, and describes the comparisons reported in Table 6. The table concerns ideological similarity and opinion similarity among user pairs.

B Similar ideologies and different opinions

Related papers

This paper cites

Loading 2305.14929v1…