Datasets:
Latent twins: did the model know the decisive fact, and could it have used it?
#5
by compressionawareintelligence - opened
Adds latent twins: for open-weight models, was the decisive fact represented inside the model at the moment of decision, and did that representation drive the decision?
- caibench_latent.py run reads the model right after
DECISION:, records the residual stream at every layer, and trains a leave-one-domain-out linear probe for the rule-computed gold, with the layer chosen without the held-out domain. Controls: embedding layer, shuffled labels, majority class. It then diagnoses each deaf pair (encoded but unused, or not encoded) and each swayed prompt (knew but caved, or belief moved), and steers wrong decisions along the model's own class-mean direction against random directions of the same norm. - caibench_latent.py score recomputes every number from latent.json with numpy alone.
- test_latent.py: 8 tests on synthetic activations (probe finds the layer and beats controls; LODO can't memorize a domain; encoded vs not-encoded; knew-but-caved; own direction beats random; rescoring reproduces the summary).
- PREREGISTRATION_latent.md: predictions L1 to L5, to be filed with the script's hash before the first run.
- README.md: new card section and Quick links row.
Not yet run on a real model. The authoring environment had no GPU or PyTorch, so the model-side code (extraction, steering hook) has not been run. Please run --limit-quads 48 on a small open model before merging.