AI Paper Atlas

Research map / Large language models / Alignment and preferences

VALUEFLOW: Toward Pluralistic and Steerable Value-based Alignment in Large Language Models

Woojin Kim, Sieun Hyeon, Jusang Oh, Jaeyoung Do

ICML 2026 · Large language models › Alignment and preferences · approach: Steering and correcting LLM behavior · 0 citations (OpenAlex)

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
replace direct LLM rating with ranking-based intensity evaluation (ablation)value intensity evaluator · SVT, 10K texts per valuemean variance: 12.6 -> 2.1Improved
replace direct LLM rating with ranking-based intensity evaluation (ablation)value intensity evaluator · SVT, 10K texts per valuesign-flip rate (%): 48 -> 29Improved
replace direct LLM rating with ranking-based intensity evaluation (ablation)value intensity evaluator · ValueNet agreement (SVT)pairwise ranking accuracy (%): 77.4 -> 84.2Improved
add value-profile steering using HiVES-derived profilesPhi-4 (14B) · OpinionQA, Religion groupaccuracy (%): 44.5 -> 57.4Improved
use value-profile steering instead of default group-attribute promptQwen3-32B, Phi-4, GLM-4-32B · OpinionQA micro-averageaccuracy (%): 56.6 -> 59.1 (Qwen3); 51.2 -> 55.7 (Phi-4); 56.8 -> 59.1 (GLM-4)Improved
use value-profile steering instead of Modular PluralismQwen3-32B · OpinionQA micro-averageaccuracy (%): 39.3 -> 59.1Improved
profile steering on Phi-4 for Region, Education, Income groupsPhi-4 (14B) · OpinionQA by groupaccuracy (%): Region 60.3 -> 57.2; Education 57.5 -> 56.2; Income 55.3 -> 52.5Mixed
add HiVES two-stage hierarchical embedding training on Qwen3-embedding-0.6BQwen3-embedding-0.6B · SVT and MFT hierarchy metricsranking consistency and similarity correlation: over 20% ranking consistency, over 50% similarity correlationImproved
user-text prompts instead of intensity-anchor prompts for steering (ablation)steered LLMs · 500-prompt steerability benchmarksteering gainMixed
steer negative targets on prosocial values (Benevolence, Universalism)Phi-4 and Claude-4 · 500-prompt steerability benchmarksteering gainNo clear effect
steer positive vs negative intensity targetsHedonism and most rights values · steerability by valuesteering gainMixed
steer Conformity value in either directionprompt-based steering · steerability by valuesteering gain: |delta| ~ 0No clear effect

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy