AI Paper Atlas

Research map / Large language models / Alignment and preferences

Task-Aware Preference Calibration for Direct Preference Optimization

Mengyang Li, Zhong Zhang, Pinlong Zhao

ICML 2026 · Large language models › Alignment and preferences · approach: Robust and calibrated DPO

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
add task conditioning to meta-learned calibration (Uniform-TAPC to TAPC) (ablation)TAPC · UltraFeedback Creative prompts, 20% noisewin rate (%): 88.5 -> 92.5Improved
add task conditioning to meta-learned calibration (ablation)TAPC · UltraFeedback Factual prompts, 20% noisewin rate (%): 93.5 -> 94.2No clear effect
add task conditioning to meta-learned calibration (ablation)TAPC · UltraFeedback Open prompts, 20% noisewin rate (%): 90.2 -> 92.8Improved
add a task encoder over no task encoding (ablation)TAPC · UltraFeedback, 20% noisewin rate (%): 88.5 -> 90.2Improved
use mean pooling instead of first-token hidden state for the task encoder (ablation)TAPC task encoder · UltraFeedback, 20% noisewin rate (%): 89.2 -> 90.2Improved
use mean pooling instead of last-token hidden state (ablation)TAPC task encoder · UltraFeedback, 20% noisewin rate (%): 89.0 -> 90.2Improved
replace learned embedding with oracle task labels (ablation)TAPC task encoder · UltraFeedback, 20% noisewin rate (%): 90.2 -> 90.0No clear effect
use meta-learning instead of joint training without reference set (ablation)TAPC · UltraFeedback, 20% noisewin rate (%): 88.0 -> 90.2Improved
use meta-learning instead of joint training without reference set (ablation)TAPC · UltraFeedback, 40% noisewin rate (%): 78.5 -> 84.5Improved
use meta-learning instead of joint training with reference regularization (ablation)TAPC · UltraFeedback, 40% noisewin rate (%): 80.2 -> 84.5Improved
increase reference set size from 100 to 200 (ablation)TAPC · UltraFeedback, 20% noisewin rate (%): 88.5 -> 90.2Improved
increase reference set size from 200 to 300 (ablation)TAPC · UltraFeedback, 20% noisewin rate (%): 90.2 -> 90.5No 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