AI Paper Atlas

Research map / Large language models / Alignment and preferences

ENCORE: Entropy-guided Reward Composition for Multi-head Safety Reward Models

Xiaomin Li, Xupeng Chen, Jingxuan Fan, Eric Hanchen Jiang, Mingye Gao

AAAI 2026 · Large language models › Alignment and preferences · approach: Reward modeling for diverse preferences · 8 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
add entropy-penalized rule weighting instead of uniform weightsMulti-head reward model, Llama3.1-8B · RewardBench safetyaccuracy (%): 85.5 -> 88.5Improved
add entropy-penalized weighting instead of random weightsMulti-head reward model, Llama3.1-8B · RewardBench safetyaccuracy (%): 84.2 -> 88.5Improved
use learned MoE gating instead of entropy weightingMulti-head reward model, Llama3.1-8B · RewardBench safetyaccuracy (%): 88.5 -> 86.0Hurt
use all 10 rules with multi-head instead of a single ruleMulti-head reward model, Llama3.1-8B · RewardBench safetyaccuracy (%): 76.4 -> 85.5Improved
use multi-head reward model instead of single-head Bradley-TerryLlama3.1-8B · RewardBench safetyaccuracy (%): 66.61 -> 85.5Improved
use entropy weighting instead of Bradley-TerryLlama3.1-8B · RewardBench XSTest Should Respondaccuracy (%): 13.6 -> 72.4Improved
use entropy weighting instead of MoE gatingLlama3.1-8B · RewardBench XSTest Should Respondaccuracy (%): 73.6 -> 72.4Hurt
use entropy weighting instead of uniform weights on alternate backboneFsFairX-Llama3-8B · RewardBench safetyaccuracy (%): 77.7 -> 83.1Improved
use entropy weighting instead of Bradley-Terry + FsfairXFsFairX-Llama3-8B · RewardBench safetyaccuracy (%): 79.3 -> 83.1Improved
select top 5 lowest-entropy rules (hard selection) instead of soft entropy weighting over all rules (ablation)ENCORE · RewardBench safetyaccuracyHurt
use an 8B ENCORE reward model instead of LLM-as-a-judge with GPT-3.5 and Llama modelsLlama3.1-8B · RewardBench safetyaccuracy (%): 64.0 -> 88.5Improved

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