SCIT: Testing Causal Cache Carriers in Latent Chain-of-Thought Models
The paper’s causal tests point to value-cache suffixes as the main carrier for counterfactual arithmetic in specific latent-CoT checkpoints.
SCIT patches declared cache segments to test where hidden reasoning is carried. In CODI-GPT2, the authors report sufficiency and necessity evidence for a late value-suffix mechanism. A Sim-CoT-style GPT-2 reproduction shows similar sufficiency, but not enough matched-corruption evidence for necessity. The paper frames the result as checkpoint-specific, with other cells shifting to prompt-prefix or full-cache K/V routing. ArXiv · AI/CL/LG's note
SCIT patches declared cache segments to test where hidden reasoning is carried. In CODI-GPT2, the authors report sufficiency and necessity evidence for a late value-suffix mechanism. A Sim-CoT-style GPT-2 reproduction shows similar sufficiency, but not enough matched-corruption evidence for necessity. The paper frames the result as checkpoint-specific, with other cells shifting to prompt-prefix or full-cache K/V routing. ArXiv · AI/CL/LG's note
score 5