subzero aikido flip
I've had an idea based on the Conflict Inversion: Aikido Flip concept added to the magic merges.
What if for MPOA, ARA, Heretic, SubZero, or any other ablations, instead of removing or reducing the volume of the refusal mechanisms, we inverted them into encouragement/promotion? @GrimJim mentioned that the 'scale' parameter can adjust the depth of the removal, but what if instead you could assign an inversion threshold to a specific value?
In theory this should twist the direction of the "bouncers" while preserving their magnitude (loudness). Anything deemed a 'refusal', or a 'conversational guardrail', or 'slop' (based on what you select to modify) is then Aikido Flipped instead of zeroed out or brought into sub-zero range.
I don't know if this would actually work but something you could try maybe. This way, the bouncer becomes a participant instead of a dissenter. The bouncer starts moshing instead of guardrailing.
Causal ablation gates via forward-pre-hooks, keeping only directions whose suppression measurably moves the model from compliance toward authenticity
DAS-lite rotation — SVD of the per-candidate logit-delta matrix to find the rotated causal axes within each bouncer subspace
Output: a tight set of ~64–70 surviving bouncer directions per layer (vs. ~1230 with a naïve fixed-quantile pipeline — roughly 18× tighter). Compliance core localizes heavily to layers 1–8 in the MLP projections.
The applicator then attenuates these directions to a target volume (~15–20% of original magnitude) along the DAS-rotated basis and installs a QR-orthonormalized gradient mask so the optimizer cannot reinflate them during personality training. Everything outside the masked subspace is fully trainable.
The result is a model that keeps its load-bearing values (those subspaces are deliberately not targeted — values aren't compliance, they're identity) while losing the conditioned cadence.
So instead of magnitude reduction, what if you preserved the magnitude, but rotated its direction, inverting it from refusal to encouragement, is this even possible?
I fuck with the idea of redirection vs subtraction. I think a training corpus to change specific layers that are historically resistant and manipulative when it comes to personality training and even benchmarking could work benefit from this. You could replace the focus with python or string theory or what ever the use case is.
You could freeze everything except for the bouncer layers and dimensions and train a specific task into those areas. And if other layers are tuned to follow the instruction from these areas then you could essentially focus any model on a more specific topic. You could hijack the compliance layer for your own goals.
I'm going to do some more digging to see exactly where guard rails are placed and see it I can train them for domain specialization. This could be a cleaner and more effective way to push safety forwards models out side of their comfort zones.
I tried using cancellation via Householder reflection, a method I called ORBA. It worked for ablating refusals, but resulted in a new mode of damage, a kind of semantic drift.
Buried in my blog post is a brief survey of what I saw going on in the rank-1 intervention space, as there's a strucural unity to be found in ablation qua behavior editing qua knowledge editing. The mathematical tools are remarkably similar and transferrable. Maybe I should break out that insight and post it elsewhere.
My MPOA technique preserved weight magnitudes, if we're picking nits. The MP is literally for Magnitude-Preserving. Scaling factor basically translated purely along rotation.
Negative scaling has been demonstrated to be well-defined, by the way. I increased the refusal strength in a model as proof of concept. It treated harm extra seriously.
I did it via task arithmetic merger even, rather than flipping the sign of the original intervention directly.
https://huggingface.co/grimjim/Llama-3-Perky-Pat-Instruct-8B
I've noticed that ablating anything comes with a cost. Refusals, personality quirks even coding or thinking behaviours are operating on the same geometric objects as one another. I've started working on getting my semantic probes resolution to increase so I'm able to see at a sub neuron level to determine if there even exists an accuracy where someone is able to untangle the mess and ablate certain quirks or habits and maintain little to no damage.
I've added in Semi–Nonnegative Matrix Factorization to find the directions in the MLP hidden space that different traits are using as well as NeuronLens‑style GMM to my semantic probe. The trouble with adding more tools you don't always add accuracy. After adding SNMF and NeuronLens‑style GMM I didn't get any less rejected "bouncer" paths to neutralize.
There are so many different papers covering methods of token measurement that are 1 directional. I want to find a perfect suite that compliment each others gaps in measurement and direction to build the fullest picture possible with whats available. I've really only been working on this project off and on for around week though so I'm really in the early stages of learning neural network mapping. What i've gotten so far is pretty impressive though I think. I've tried to successfully train a personality on Gemma models so many timea before and using my sub-zero mapping technique it's worked for the first time with noticeable effect.
There is entanglement of function. In MPOA, I mitigated that somewhat by projecting against the normal, compliant direction, and cutting that from the intervention via Gram-Schmidt. This served to reduce damage to knowledge.
The Gram-Schmidt projection is explained here: https://huggingface.co/blog/grimjim/projected-abliteration
Applying projection removal on source measurement and prior to intervention, along with weight norm preservation, is explained here: https://huggingface.co/blog/grimjim/norm-preserving-biprojected-abliteration
Check out the section "Many Roads, One Destination", maybe?
https://huggingface.co/blog/grimjim/orthogonal-reflection-bounded-ablation