chain-of-thought monitoringpersona-control trainingpublished Nov 18, 2026 · arXiv:2611.21799
Chain-of-Thought Controllability Can Reduce Chain-of-Thought Monitorability
closed-context question answeringSQuAD 2.0published Nov 14, 2026 · arXiv:2611.16300
SAGE-QA: Semantic-Aware, Grounded Evaluation for Closed-Context Question Answering on SQuAD 2.0
mixture-of-expertsquantizationpublished Oct 26, 2026 · arXiv:2610.32229
Training an MoE LLM in MXFP4: How We Closed the Gap to BF16
in-context watermarkingwatermark detectabilitypublished Sep 1, 2026 · arXiv:2608.29030
Learning to Follow In-Context Watermark Instructions via Self-Distillation
large language model evaluationmathematical reasoningpublished Sep 1, 2026 · arXiv:2608.28725
Beyond the Answer Key: Robustness Evaluation of Large Language Models for Step-Level Mathematical Verification
instruction hierarchysystem-user instruction conflictspublished Sep 1, 2026 · arXiv:2608.28648
How Language Models Choose Sides: Internal Representations of Instruction Hierarchy