Tag
This paper extends contextual entrainment from token-level to sentence-level, showing that even counterfactual sentences in prompts increase their probability during inference. The effect decreases with model size and is driven by 2-4% of attention heads, which can be ablated without performance loss.
This paper proposes SenFlow, a method for sentence-level AI-generated text detection in hybrid documents by modeling inter-sentence dependencies using graph propagation and linear-chain CRF decoding. It also introduces the MOSAIC benchmark with 16,000 documents generated by DeepSeek-V3.2 and Kimi K2, achieving state-of-the-art performance.