Contribution to the handbook
Supporting · Assessed 2026-10-28
The question compares fine-tuning and added reference material, which helps characterize technical alternatives. The reported comparison is narrow, including a single model and a particular reference-document setup, and matching the study participants does not establish that either method represents a wider public. It is useful background rather than a system-level recommendation.
Does this challenge the goal?
In-scope gap
Comparing methods for matching participants’ recorded preferences is an evidence and implementation question within the current exploration goal, not a challenge to that goal.
The supplied comparisons concern a particular study and do not establish which approach would work better in other settings.
Why this category?
- comparison of technical alternatives
- bounded empirical evidence
- limited transferability
The supplied evidence gives a narrow comparison; it does not establish a general ranking of fine-tuning and reference material.
This is an AI assessment of contribution to the goal, separate from whether the question has an answer.
Authority, accountability and resilience →
Tests, stop conditions and unresolved work →
Read the exploration goalCompare how fine-tuning and added reference material matched the full participant data in the study’s tests.
What we can say so far
In the study’s reported GPT-3.5 Turbo comparison, LoRA fine-tuning matched the full participants’ aggregate preferences better than retrieval-augmented generation (RAG), especially with 25% and 50% samples. The comparison measured how closely sample proposal win rates matched the full-sample pattern. This is a result for the tested model and the particular RAG reference documents, not a general ranking of fine-tuning and added reference material.
What remains open
- The supplied evidence gives no exact numerical results for the fine-tuning-versus-RAG comparison.
- The result does not establish how the comparison would turn out with other models, reference materials, tasks, or populations.
Conditions and limits
- The direct comparison used GPT-3.5 Turbo and RAG documents summarizing proposals favored by participants (Bolsonaro for conservatives and Lula for liberals).
- The reported match concerns aggregate proposal win-rate patterns against the study’s full participant data, not each individual’s choices or preferences outside the study.
- This is the authors’ reported result as represented in the supplied evidence; the figures are not independently interpreted and the evidence is not independent verification.
Objections to consider
- A close match on the study’s measure should not be read as reproducing every individual’s preferences. The answer limits its conclusion to aggregate proposal win-rate patterns.
Evidence and related ideas
Supporting evidence, qualifications and context play different roles in this answer.
context
Testing group-level preferences The researchers also compare whether ordinary random samples and model-augmented samples can reproduce the full group’s combined preferences.
Establishes that the study compared augmented samples with full-group preferences, but does not report this comparison’s outcome.
Large Language Models (LLMs) as Agents for Augmented Democracy
Read full source: Large Language Models (LLMs) as Agents for Augmented Democracy (opens in a new tab) ↗Read source discussion: Large Language Models (LLMs) as Agents for Augmented Democracy →