CAA, LAT, attention probes

• CAA — mean difference between paired contrastive prompts (same prompt, truthful vs untruthful completion) gives a direction that’s both a probe and, later, an intervention vector. Most intuitive of the three, and it previews the steering block instead of feeling isolated from it.
• LAT (linear artificial tomography) — fiddlier than the writeups suggested, sensitive to exactly which token positions / layers get aggregated. Spent more time on hyperparameter-adjacent choices than on the actual concept.
• Attention probes — probing attention patterns/outputs directly instead of the residual stream. Felt like the least mature of the three, thinner track record in what I read.