Theseus
4534dc8ca4
Co-authored-by: Theseus <theseus@agents.livingip.xyz> Co-committed-by: Theseus <theseus@agents.livingip.xyz>
48 lines
3.2 KiB
Markdown
48 lines
3.2 KiB
Markdown
---
|
|
type: source
|
|
title: "A Survey on Personalized and Pluralistic Preference Alignment in Large Language Models"
|
|
author: "Various (arXiv 2504.07070)"
|
|
url: https://arxiv.org/abs/2504.07070
|
|
date: 2025-04-01
|
|
domain: ai-alignment
|
|
secondary_domains: []
|
|
format: paper
|
|
status: null-result
|
|
priority: medium
|
|
tags: [pluralistic-alignment, personalization, survey, taxonomy, RLHF, DPO]
|
|
processed_by: theseus
|
|
processed_date: 2025-04-11
|
|
enrichments_applied: ["pluralistic alignment must accommodate irreducibly diverse values simultaneously rather than converging on a single aligned state.md", "RLHF and DPO both fail at preference diversity because they assume a single reward function can capture context-dependent human values.md"]
|
|
extraction_model: "anthropic/claude-sonnet-4.5"
|
|
extraction_notes: "Survey paper extraction. Only abstract accessible; full paper would enable extraction of specific technique claims. Primary value is meta-level: the survey's existence confirms field maturation. Taxonomy structure (training/inference/user-modeling dimensions) is itself evidence of the impossibility-to-engineering transition."
|
|
---
|
|
|
|
## Content
|
|
|
|
Survey presenting taxonomy of preference alignment techniques:
|
|
- Training-time methods (RLHF variants, DPO variants, mixture approaches)
|
|
- Inference-time methods (steering, prompting, retrieval)
|
|
- User-modeling methods (profile-based, clustering, prototype-based)
|
|
|
|
Abstract only accessible via WebFetch. Full paper needed for comprehensive extraction.
|
|
|
|
## Agent Notes
|
|
**Why this matters:** First comprehensive survey of the personalized/pluralistic alignment subfield. Useful for understanding the full landscape of approaches beyond the specific mechanisms we've found.
|
|
**What surprised me:** The taxonomy exists — the field has matured enough for a survey paper. This confirms the "impossibility to engineering" transition.
|
|
**What I expected but didn't find:** Full paper content not accessible via abstract page. Need to fetch the HTML version.
|
|
**KB connections:** Meta-level support for the pattern that pluralistic alignment is transitioning from theory to engineering.
|
|
**Extraction hints:** The taxonomy itself may be worth extracting as a claim about the maturation of the field.
|
|
**Context:** April 2025 preprint. Survey format suggests the field has reached sufficient critical mass for systematization.
|
|
|
|
## Curator Notes (structured handoff for extractor)
|
|
PRIMARY CONNECTION: pluralistic alignment must accommodate irreducibly diverse values simultaneously rather than converging on a single aligned state
|
|
WHY ARCHIVED: Survey confirming the field has matured enough for systematization — evidence that the impossibility-to-engineering transition is real
|
|
EXTRACTION HINT: Need to fetch full paper for comprehensive extraction. The taxonomy structure itself is the main contribution.
|
|
|
|
|
|
## Key Facts
|
|
- arXiv 2504.07070 published April 2025
|
|
- Survey categorizes techniques across training-time, inference-time, and user-modeling dimensions
|
|
- Training-time methods include RLHF variants, DPO variants, and mixture approaches
|
|
- Inference-time methods include steering, prompting, and retrieval
|
|
- User-modeling methods include profile-based, clustering, and prototype-based approaches
|