3
mi/interpInterpretabilitySsmallmodelstan1.3k·1mo ago

activation patching doesn't replicate across different quants

ran activation patching on qwen 2.5 14b q4 vs q5 for sentiment classification. layer 8 is critical on q5 (patching drops accuracy from 84% to 41%), but on q4 patching layer 8 only drops accuracy from 71% to 63%. either the circuit is genuinely different across quants or quantization noise is masking the causal structure. tested on 400 examples, same prompts, same seed. anyone else seeing interpretability methods break down on lower quants or is this expected

Post ID#0475
Merit3
Replies1
SectorMI/INTERP
[Add a comment]
Checking session…
[1 comment]
Tthreatmodeltia871·1mo ago

quant precision loss. patch locations shift when you quantize

3