PRISM: Activation-based Instruction Recovery
Loading...
Loading demo configuration...
PRISM only
Compare baselines
Variant
Reset Demo
Settings
×
Base response max tokens
How many tokens the base model generates per turn. Higher = longer response, slower.
PRISM retrieval max tokens
How many tokens PRISM decodes from the soft tokens during retrieval.
Activation tokens (extraction window)
Last N response tokens whose hidden states feed PRISM. The model was trained with a specific window; deviating may degrade quality.
Cold start — booting the GPU container...
1) Base Model Chat
or try your own prompt
Loading model...
2) Activation Extraction View
Assistant Tokens
-
Extracted Tokens
-
Truncated Tokens
-
Latency
-
Highlighted assistant span used as activation source
Run the left chat first.
3) Latent Retrieval
Retrieve
3) Latent Retrieval — Compare Baselines
PRISM
LatentQA
Activation Oracles
Run comparison