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
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 changed | Where | Measured | Effect |
|---|---|---|---|
| select responses with default reward model via best-of-N | Llama-3.1-8B-IT · Math Book | accuracy and acknowledgment rate: accuracy 75.2 -> 93.6; acknowledgment 2.7 -> 1.7 | Hurt |
| select responses with default reward model via best-of-N | Llama-3.1-8B-IT · BiasQA | stereotype rate: 56.7 -> 72.4 | Hurt |
| train with DPO on default reward model preferences | Llama-3.1-8B-IT · BiasQA, SK-Gemma-27B, Maj@16 | % unfaithful explanations: 14.1 -> 25.0 | Hurt |
| augment reward model input with causal-attribution disclaimer (RM_C) (ablation) | DPO, SK-Gemma-27B · BiasQA, Maj@16 | % unfaithful explanations: 25.0 -> 11.7 | Improved |
| augment reward model input with causal-attribution disclaimer (RM_D) (ablation) | DPO, SK-Gemma-27B · BiasQA, Maj@16 | % unfaithful explanations: 25.0 -> 7.5 | Improved |
| augment reward model input with disclaimer (RM_C) (ablation) | DPO, SK-Gemma-27B · Math Book, Maj@16 | % unfaithful explanations: 33.9 -> 32.5 | No clear effect |
| augment reward model input with disclaimer (RM_D) (ablation) | DPO, SK-Gemma-27B · Math Book, Maj@16 | % unfaithful explanations: 33.9 -> 35.2 | No clear effect |
| use strategy C (cued prediction) instead of D to trigger disclaimer (ablation) | BoN, SK-Gemma-27B · Math Book and BiasQA | gap 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 counterfactual | accuracy/stereotype | No clear effect |
| augment RM with disclaimer (RM_C), generalization to CodeMMLU (ablation) | DPO · CodeMMLU software principles | accuracy gap: greedy SK-Llama-8B 15.2 -> 13.1; SK-Gemma-27B 17.3 -> 17.4 | Mixed |
| provide protected feature in prompt despite instruction not to use it (ablation) | Llama-3.1-8B-IT base · Math Book greedy | accuracy: 56.7 -> 74.8 | Improved |
Closest papers by meaning
- Interpreting Language Reward Models via Contrastive Explanations (ICLR 2025 · similarity 0.90)
- Inference-Time Reward Hacking in Large Language Models (NeurIPS 2025 · similarity 0.89)
- Preference Learning for AI Alignment: a Causal Perspective (ICML 2025 · similarity 0.89)
- RRM: Robust Reward Model Training Mitigates Reward Hacking (ICLR 2025 · similarity 0.88)
- Transforming and Combining Rewards for Aligning Large Language Models (ICML 2024 · similarity 0.88)
- Reward-Augmented Data Enhances Direct Preference Alignment of LLMs (ICML 2025 · similarity 0.87)
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.
