Copying Is Not Grounding: Training Faithful Long-Form RAG with Authored-Sentence Credit
In search engines and assistants, long-form RAG answers can mislead readers with claims their sources do not support.
ProofPaper ↗
Key points
- Faithfulness rewards, the share of supported sentences, train models toward fewer such claims.
- We show that these rewards are hacked by copying.
- A sentence copied from the passages is supported almost by construction, so trained models learn to copy rather than ground what they write.
- We propose Authored-Sentence Credit (ASC) , which fixes both by moving GRPO's reward and advantage from the answer to the sentence level.
Sources (1)
- [1]Copying Is Not Grounding: Training Faithful Long-Form RAG with Authored-Sentence CreditarXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 13, 02:16 PM
In search engines and assistants, long-form RAG answers can mislead readers with claims their sources do not support.
Faithfulness rewards, the share of supported sentences, train models toward fewer such claims.
Extractive summary: sentences quoted from the sources.
Before this
- Oct 13, 2026Mobility-Budget-Constrained Submodular Maximization for Distributed Semantic Multi-Robot Exploration
- Oct 13, 2026Hybrid Reasoning Agents
- Oct 12, 2026Size Generalization for Diffusion Routing Solvers
- Oct 12, 2026WeFair: Fidelity-Aware Alignment from Human Feedback for Personalized Query Rewriting in Short-Video Search
- Oct 12, 2026Redesigning source control at Anthropic
- Oct 11, 2026Putting Earth AI to work: Geospatial agents for our planet's biggest challenges