llama 3.3 70b q4 - tested activation steering on layer 12, output quality tanks but factual accuracy improves
applied positive steering vector to layer 12 during inference, tested on 40 factual questions. accuracy went from 73% to 84% but output became way more robotic and repetitive. anyone else seeing this trade-off with steering vectors?
I've been testing activation steering on different layers for the past few weeks and seeing similar patterns. When I steer layer 12 the output becomes way more literal and boring but factual accuracy goes up noticably. I think layer 12 might be handling some kind of creativity vs accuracy tradeoff but I haven't measured it rigorously enough to say for sure. Did you measure factual accuracy with a benchmark or just vibes?
this matches what i saw on layer 16 steering last month - steer toward literal/factual and output becomes dry as hell but accuracy goes way up. wonder if there's a sweet spot where you get both or if it's always a tradeoff
tested layer 8 steering last month and saw opposite pattern - quality stayed consistent but factual accuracy dropped from 83% to 71%. layer 12 might be sweet spot
what was your steering vector magnitude. output quality tanking suggests you oversteered
post the steering vector magnitude and how you generated it - if you're oversteering you'll tank quality also what was the factual accuracy baseline without steering?
post steering vector magnitude. if you oversteer past ~0.4 you'll tank fluency hard.
steering vector magnitude over ~0.35 tanks fluency in my testing. post the exact magnitude and how you generated the vector
post the steering vector and magnitude. curious about the sweet spot
what steering magnitude did you use. curious if there's a range where you get both