AI Paper Atlas

Research map / Large language models / Alignment and preferences

Truthful or Fabricated? Using Causal Attribution to Mitigate Reward Hacking in Explanations

Pedro Lobato Ferreira, Wilker Aziz, Ivan Titov

ICLR 2026 · Large language models › Alignment and preferences · approach: Reward model design and preference models · 12 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
select responses with default reward model via best-of-NLlama-3.1-8B-IT · Math Bookaccuracy and acknowledgment rate: accuracy 75.2 -> 93.6; acknowledgment 2.7 -> 1.7Hurt
select responses with default reward model via best-of-NLlama-3.1-8B-IT · BiasQAstereotype rate: 56.7 -> 72.4Hurt
train with DPO on default reward model preferencesLlama-3.1-8B-IT · BiasQA, SK-Gemma-27B, Maj@16% unfaithful explanations: 14.1 -> 25.0Hurt
augment reward model input with causal-attribution disclaimer (RM_C) (ablation)DPO, SK-Gemma-27B · BiasQA, Maj@16% unfaithful explanations: 25.0 -> 11.7Improved
augment reward model input with causal-attribution disclaimer (RM_D) (ablation)DPO, SK-Gemma-27B · BiasQA, Maj@16% unfaithful explanations: 25.0 -> 7.5Improved
augment reward model input with disclaimer (RM_C) (ablation)DPO, SK-Gemma-27B · Math Book, Maj@16% unfaithful explanations: 33.9 -> 32.5No clear effect
augment reward model input with disclaimer (RM_D) (ablation)DPO, SK-Gemma-27B · Math Book, Maj@16% unfaithful explanations: 33.9 -> 35.2No clear effect
use strategy C (cued prediction) instead of D to trigger disclaimer (ablation)BoN, SK-Gemma-27B · Math Book and BiasQAgap closed to counterfactual: 41% vs 9% (Math Book); 71% vs 32% (BiasQA)Improved
add disclaimer to all reward model inputs regardless of causal attribution (ablation)BoN · gap to counterfactualaccuracy/stereotypeNo clear effect
augment RM with disclaimer (RM_C), generalization to CodeMMLU (ablation)DPO · CodeMMLU software principlesaccuracy gap: greedy SK-Llama-8B 15.2 -> 13.1; SK-Gemma-27B 17.3 -> 17.4Mixed
provide protected feature in prompt despite instruction not to use it (ablation)Llama-3.1-8B-IT base · Math Book greedyaccuracy: 56.7 -> 74.8Improved

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