AI Paper Atlas

Research map / Large language models / Retrieval and knowledge

TruthRL: Incentivizing Truthful LLMs via Reinforcement Learning

Zhepei Wei, Xiao Yang, Kai Sun, Jiaqi Wang, Rulin Shao, Jingxiang Chen, Mohammad Kachuee, Teja Gollapudi, Yiwei Liao, Nicolas SCHEFFER, Rakesh Wanga, Anuj Kumar and 3 more

ICML 2026 · Large language models › Retrieval and knowledge · approach: Knowledge-augmented question answering

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
use ternary reward instead of binary reward (add abstention-neutral reward) (ablation)TruthRL · average of 4 benchmarks, Llama3.1-8Btruthfulness: 4.5 -> 25.6Improved
use ternary reward instead of binary reward (ablation)TruthRL · average of 4 benchmarks, Llama3.1-8Bhallucination rate: 47.7 -> 18.8Improved
use ternary reward instead of binary reward (ablation)TruthRL · average of 4 benchmarks, Llama3.1-8Baccuracy: 52.2 -> 44.4Hurt
add knowledge-enhanced reward to binary reward (ablation)TruthRL (binary) · average of 4 benchmarkstruthfulness: 4.5 -> 11.7Improved
add knowledge-enhanced reward to ternary reward (ablation)TruthRL · average of 4 benchmarkstruthfulness: 25.6 -> 23.2No clear effect
use neutral abstention reward (0) instead of negative (-0.5) (ablation)TruthRL · CRAG, Llama3.1-8Btruthfulness: 30.8 -> 37.2Improved
use neutral abstention reward (0) instead of positive (0.5) (ablation)TruthRL · CRAG, Llama3.1-8Btruthfulness: 34.7 -> 37.2Improved
use model-based verifier instead of rule-based verifier (ablation)TruthRL · CRAG, Llama3.1-8Btruthfulness: -3.6 -> 37.2Improved
train with online GRPO instead of offline DPO (ablation)TruthRL · average of 4 benchmarks, Llama3.1-8Btruthfulness: -10.1 -> 25.6Improved
run iterative DPO for 4 iterations instead of 3 (ablation)Iterative DPO · average of 4 benchmarkstruthfulness: 12.6 -> -2.0Hurt
add additive reasoning reward to outcome reward (ablation)TruthRL · CRAG, Llama3.1-8Btruthfulness: 37.2 -> 36.1Mixed
apply TruthRL to Llama3.2-3B-InstructTruthRL · CRAG (with retrieval)truthfulness: 1.9 -> 27.4Improved

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