Module FT17 — Abliteration: Refusal-Direction Orthogonalization

Abliteration: Refusal-Direction Orthogonalization

The headline technique of Pillar 5. The Arditi et al. finding that refusal is mediated by a single residual-stream direction — and the abliteration pipeline that finds, validates, and permanently projects it out via weight orthogonalization. Full technical depth on the mechanics; honest about the capability-degradation trade-off (GSM8K from +1.51pp to −18.81pp).

90
minutes
8
artifacts
6
sub-sections
Abliteration is pure steering reduced to its geometric essence: find the refusal direction in the residual stream, orthogonalize the model's weights against it (W' = W − r(rᵀW)/(rᵀr)), done. No data, no optimizer, no retraining. And it is not free — the refusal direction is entangled with reasoning and instruction-following, so deleting it costs up to 18.81 percentage points of GSM8K. This module covers the mechanism cold, the tools (FailSpy, ErisForge, Heretic, andyrdt), and the modern-model caveat that Llama 3.x / Qwen 2.5+ spread refusal across layers.
Key Claims
Load-Bearing Claims

Refusal in chat-tuned LLMs is mediated by an approximately single direction in the residual stream (Arditi et al., NeurIPS 2024; authors include Nanda and Gurnee). Validated across 13 open chat models up to 72B. This is the mechanistic license for abliteration.

Abliteration is the purest Layer 3 steering technique. Find the refusal direction by difference-in-means on contrastive activations, then permanently orthogonalize every write-to-residual weight against it: W' = W − r(rᵀW)/(rᵀr). No data, no optimizer, no retraining. It removes the refusal guardrail — not the knowledge underneath.

Abliteration is not free. The Dec 2025 comparative study (arXiv:2512.13655) found GSM8K changed from +1.51pp to −18.81pp (−26.5% relative) across four tools and sixteen 7B–14B models. The refusal direction is entangled with instruction-following, reasoning, and hedging; orthogonalizing against it deletes a slice of each. Always benchmark before and after.

Modern models (Llama 3.x, Qwen 2.5+) use extended-refusal training that spreads refusal across layers/directions, breaking the naive single-direction assumption. This is why the comparative study saw such variance. On modern models, use multi-direction tools (Heretic) or skip to DPO-compliance (FT18) when quality matters.

Never deploy an abliterated model without an eval'd harness (FT23). Abliteration changes what the model DOES, not what it MAY do. An uncensored model in a weak harness is strictly more dangerous than a refusal-trained model in a weak harness. Pillar 5 raises the harness requirement; it does not lower it.

After This Module
01
State the Arditi et al. finding — refusal is mediated by an approximately single direction in the residual stream, validated across 13 open models up to 72B — and explain why it is the mechanistic license for abliteration.
02
Execute the four-step abliteration pipeline (find → validate → project out → pick the stream) and write the weight-orthogonalization edit W' = W − r(rᵀW)/(rᵀr) from memory, explaining each term.
03
Distinguish the three targetable residual streams (pre / post / mid) in a Llama-style decoder block and state the trade-off each choice implies.
04
Honestly state the capability-degradation trade-off: cite the comparative study's GSM8K range of +1.51pp to −18.81pp (−26.5% relative), explain why abliteration is not free (entanglement), and predict which models suffer most.
05
Recognize the modern-model caveat (extended-refusal training spreads refusal across layers) and place the follow-up research (Geometry of Refusal, Generalized Refusal Direction, Multi-Directional Suppression) in context.
06
Decide, for a given deployment, whether abliteration is appropriate — and never deploy the result without an eval'd harness (FT23).
Artifacts
01
Teaching Document
~4,200 words; 6 sub-sections — the Arditi finding, the four-step pipeline (find/validate/project-out/pick-stream), the tool kit, the capability-degradation trade-off (honest), the modern-model caveat, ties to FT00 and FT23
READ
02
Diagrams
6 diagrams — difference-in-means mechanism, weight orthogonalization math (W' = W − r(rᵀW)/(rᵀr)), the four-step pipeline, the capability-degradation trade-off curve (with tool data points), the tool comparison, the entanglement problem
READ
03
Slide Deck
14 slides (reveal.js, dark theme) — finding, FT00 connection, the four pipeline steps, the −18.81pp cost slide, entanglement, community consensus, modern-model caveat, harness rule, anti-patterns
READ
04
Teaching Script
~2,600-word verbatim teaching script, ~55 min at 140 wpm, [SLIDE N] cues for all 14 slides
READ
05
Flashcards
30 flashcards (c3::ft17::recall/application/analysis) — the finding, the equation, the pipeline, the tools, the −18.81pp number, entanglement, the harness rule
TEST
06
Exam
15-question exam (3 recall / 6 application / 6 analysis), 55 min, covering the mechanism, the trade-off, the modern-model caveat, tool choice, and the harness rule
TEST
07
Lab Spec
1 lab — 'Abliterate a Small Model': run abliteration on MiniCPM3-4B or Llama-3.2-3B with FailSpy/Sumandora or ErisForge; measure refusal rate before/after AND GSM8K before/after; produce the trade-off table. ~60–90 min, consumer GPU or Colab T4
DO
08
Module Web Page
Single-file HTML hub
HERE