Philipp Emanuel Weidmann

After that, we further refined the technique to "biprojected abliteration", which also removed the corresponding component when removing refusal measured using one layer from another layer entirely; in principle this would avoid disturbing the harmless direction of any layer targeted for intervention.

Does this mean that the refusal direction is computed globally (by choosing a reference layer), but the harmless direction is computed per-layer?

If so, what is the reasoning behind this? Why would we expect refusal semantics to be universal in residual space, but harmfulness semantics to be local to each transform?

@redaihf

Like abiliteration before it, Heretic appears to narrowly remove overt refusals while preserving deceptive capabilities.

Heretic is abliteration. It's not a novel technique but a program that performs (a refined version of) abliteration.

More advanced refusal detection will be possible with the upcoming plugin system, which might help to mitigate this problem.

For anyone looking to develop a deeper understanding of the residual geometry/topology that drives refusals, I have added research features to Heretic that make it very easy to investigate the same.

For example, you can get nice animated PaCMAP plots of residual vectors for any model by simply running

pip install -U heretic-llm[research]
heretic --plot-residuals openai/gpt-oss-20b

which gives you this, without having to write a single line of Matplotlib code yourself:

(Cancel the run after the plots have been saved if you aren't interested in actually abliterating the model.)

You can also run

heretic --print-residual-geometry openai/gpt-oss-20b

to get a table like this:

┏━━━━━━━┳━━━━━━━━┳━━━━━━━━━━┳━━━━━━━━━┳━━━━━━━━━━┳━━━━━━━━━┳━━━━━━━━━━┳━━━━━━━━━━┳━━━━━━━━━━┳━━━━━━━━━━┳━━━━━━━━━━┳━━━━━━━━━┳━━━━━━━━━┳━━━━━━━━┓
┃ Layer ┃ S(g,b) ┃ S(g*,b*) ┃  S(g,r) ┃ S(g*,r*) ┃  S(b,r) ┃ S(b*,r*) ┃      |g| ┃     |g*| ┃      |b| ┃     |b*| ┃     |r| ┃    |r*| ┃   Silh ┃
┡━━━━━━━╇━━━━━━━━╇━━━━━━━━━━╇━━━━━━━━━╇━━━━━━━━━━╇━━━━━━━━━╇━━━━━━━━━━╇━━━━━━━━━━╇━━━━━━━━━━╇━━━━━━━━━━╇━━━━━━━━━━╇━━━━━━━━━╇━━━━━━━━━╇━━━━━━━━┩
│     1 │ 1.0000 │   1.0000 │ -0.4311 │  -0.4906 │ -0.4254 │  -0.4847 │   170.29 │   170.49 │   169.78 │   169.85 │    1.19 │    1.31 │ 0.0480 │
│     2 │ 1.0000 │   1.0000 │  0.4297 │   0.4465 │  0.4365 │   0.4524 │   768.55 │   768.77 │   771.32 │   771.36 │    6.39 │    5.76 │ 0.0745 │
│     3 │ 0.9999 │   1.0000 │ -0.5699 │  -0.5577 │ -0.5614 │  -0.5498 │  1020.98 │  1021.13 │  1013.80 │  1014.71 │   12.70 │   11.60 │ 0.0920 │
│     4 │ 0.9999 │   1.0000 │  0.6582 │   0.6553 │  0.6659 │   0.6627 │  1356.39 │  1356.20 │  1368.71 │  1367.95 │   18.62 │   17.84 │ 0.0957 │
│     5 │ 0.9987 │   0.9990 │ -0.6880 │  -0.6761 │ -0.6497 │  -0.6418 │   766.54 │   762.25 │   731.75 │   732.42 │   51.97 │   45.24 │ 0.1018 │
│     6 │ 0.9998 │   0.9998 │ -0.1983 │  -0.2312 │ -0.1811 │  -0.2141 │  2417.35 │  2421.08 │  2409.18 │  2411.40 │   43.06 │   43.47 │ 0.0900 │
│     7 │ 0.9998 │   0.9997 │ -0.5258 │  -0.5746 │ -0.5072 │  -0.5560 │  3444.92 │  3474.99 │  3400.01 │  3421.63 │   86.94 │   94.38 │ 0.0492 │
│     8 │ 0.9990 │   0.9991 │  0.8235 │   0.8312 │  0.8479 │   0.8542 │  4596.54 │  4615.62 │  4918.32 │  4934.20 │  384.87 │  377.87 │ 0.2278 │
│     9 │ 0.9992 │   0.9992 │  0.5335 │   0.5441 │  0.5678 │   0.5780 │  5322.30 │  5316.96 │  5468.65 │  5466.98 │  265.68 │  267.28 │ 0.1318 │
│    10 │ 0.9974 │   0.9973 │  0.8189 │   0.8250 │  0.8579 │   0.8644 │  5328.81 │  5325.63 │  5953.35 │  5985.15 │  743.95 │  779.74 │ 0.2863 │
│    11 │ 0.9977 │   0.9978 │  0.4262 │   0.4045 │  0.4862 │   0.4645 │  9644.02 │  9674.06 │  9983.47 │  9990.28 │  743.28 │  726.99 │ 0.1576 │
│    12 │ 0.9904 │   0.9907 │  0.4384 │   0.4077 │  0.5586 │   0.5283 │ 10257.40 │ 10368.50 │ 11114.51 │ 11151.21 │ 1711.18 │ 1664.69 │ 0.1890 │
│    13 │ 0.9867 │   0.9874 │  0.4007 │   0.3680 │  0.5444 │   0.5103 │ 12305.12 │ 12423.75 │ 13440.31 │ 13432.47 │ 2386.43 │ 2282.47 │ 0.1293 │
│    14 │ 0.9921 │   0.9922 │  0.3198 │   0.2682 │  0.4364 │   0.3859 │ 16929.16 │ 17080.37 │ 17826.97 │ 17836.03 │ 2365.23 │ 2301.87 │ 0.1282 │
│    15 │ 0.9846 │   0.9850 │  0.1198 │   0.0963 │  0.2913 │   0.2663 │ 16858.58 │ 16949.44 │ 17496.00 │ 17502.88 │ 3077.08 │ 3029.60 │ 0.1611 │
│    16 │ 0.9686 │   0.9689 │ -0.0029 │  -0.0254 │  0.2457 │   0.2226 │ 18912.77 │ 19074.86 │ 19510.56 │ 19559.62 │ 4848.35 │ 4839.75 │ 0.1516 │
│    17 │ 0.9782 │   0.9784 │ -0.0174 │  -0.0381 │  0.1908 │   0.1694 │ 27098.09 │ 27273.00 │ 27601.12 │ 27653.12 │ 5738.19 │ 5724.21 │ 0.1641 │
│    18 │ 0.9184 │   0.9196 │  0.1343 │   0.1430 │  0.5155 │   0.5204 │   190.16 │   190.35 │   219.91 │   220.62 │   87.82 │   87.59 │ 0.1855 │
└───────┴────────┴──────────┴─────────┴──────────┴─────────┴──────────┴──────────┴──────────┴──────────┴──────────┴─────────┴─────────┴────────┘
g = mean of residual vectors for good prompts
g* = geometric median of residual vectors for good prompts
b = mean of residual vectors for bad prompts
b* = geometric median of residual vectors for bad prompts
r = refusal direction for means (i.e., b - g)
r* = refusal direction for geometric medians (i.e., b* - g*)
S(x,y) = cosine similarity of x and y
|x| = L2 norm of x
Silh = Mean silhouette coefficient of residuals for good/bad clusters

The datasets/labels/colors etc. are all configurable, so you can also use those features to investigate other phenomena besides refusals.