Advancements in Language Model Reasoning, Interpretability, and Safety

Recent advancements in large language models focus on enhancing reasoning, interpretability, and safety. One approach explores four methods for building reasoning models, from inference-time scaling to reinforcement learning and supervised fine-tuning, noting that targeted fine-tuning can yield impressive results even on a limited budget. Another work investigates feature interpretability in crosscoder models, identifying and mitigating the polysemanticity of exclusive features through shared feature strategies. Finally, a new defense mechanism against AI jailbreaks demonstrates robustness against attacks while addressing overrefusal rates and compute overhead, with ongoing efforts to adapt rapidly to novel threats and improve AI safety.