Learning When to Trust via Selective Context Preference Optimization
The paper says robustness has to mean using good context while rejecting bad context.
The authors introduce MIST, a human-annotated benchmark that tests the same reasoning item under clean, misleading, correct-context, and irrelevant-context conditions. They also define SC2W, a paired metric for cases where a misleading signal turns a clean-correct answer into a wrong one. Their study finds that this susceptibility appears across the models they test. Their SCOPE training method reduces those failures on popular open-source models while preserving accuracy when context is clean, correct, or irrelevant. ArXiv · AI/CL/LG's note
The authors introduce MIST, a human-annotated benchmark that tests the same reasoning item under clean, misleading, correct-context, and irrelevant-context conditions. They also define SC2W, a paired metric for cases where a misleading signal turns a clean-correct answer into a wrong one. Their study finds that this susceptibility appears across the models they test. Their SCOPE training method reduces those failures on popular open-source models while preserving accuracy when context is clean, correct, or irrelevant. ArXiv · AI/CL/LG's note
score 5