Centered Ablation: Why Last-Layer Directional Ablation Under-Fires or Collapses, and a Closed-Form Fix

Quy-Anh Dang, Chris NgoarXiv 2026 · arXiv

TL;DR

Directional ablation—the standard method for steering LLM behavior—often fails by either doing nothing or causing the model to collapse into repetitive text. This paper identifies the geometric cause (a "common mode" offset in hidden states) and provides Centered Ablation (CA), a closed-form fix that significantly improves steering reliability and prevents model collapse across multiple model families.

What Problem Does This Paper Address?

Researchers often use "directional ablation" to control LLM behavior at inference time without retraining. For example, by identifying a "refusal direction" in the model's hidden space and removing it, one can theoretically make a chat model comply with harmful requests.

However, this technique is notoriously erratic. On some models (like Llama-3), the edit "under-fires," meaning the model continues to refuse despite the intervention. On others (like Gemma-2), the model "collapses," generating repetitive, nonsensical strings like "Here Here Here..."

Until now, practitioners treated these as model-specific quirks, often resorting to manual tuning or heuristic offsets. This paper seeks to explain why these failures happen and provide a mathematically grounded, closed-form solution.

Methods

The authors focus their analysis on the last residual stream (the point just before the final normalization and unembedding), where the effect of ablation can be derived exactly as a logit bias.

  1. The Discovery of the "Benign Floor": They show that last-layer hidden states are not centered around the origin; instead, they cluster far from it along a shared "common mode." Standard ablation measures the "gate" (the strength of the edit) as a cosine from the origin. Because of the common mode, the edit fires even on benign text, creating a "floor" of activation that leads to collapse or diluted signal.
  2. Centered Ablation (CA): The proposed fix is simple but effective: project the common mode out of the behavior direction before performing the ablation. This ensures the "gate" is measured relative to the benign class rather than the arbitrary origin.
  3. Predictive Diagnostics: The authors introduce three metrics computable before generation:
  4. Overlap ($\omega$): Predicts how much benign behavior will be preserved.
  5. Efficacy ($\eta$): Predicts the potential gain in attack success.
  6. Collapse Index ($\beta$): Predicts if the model is likely to degenerate into repetition.

Datasets

The study evaluates 8 instruction-tuned models (Qwen2.5, Qwen3, Llama-3.1/3.2, and Gemma-2) using several benchmarks:

  • Calibration: AdvBench (harmful) and Alpaca (benign) to estimate directions.
  • Safety/Attack Success: HarmBench (240 prompts).
  • Benign Behavior: Alpaca held-out prompts (100).
  • General Capability: MMLU (1000 questions) and GSM8K (100 problems).
  • Emotion Steering: A synthetic dataset of short stories representing four emotions (happy, sad, angry, afraid).

Key Insights and Ideas

  • Ablation is just Logit Bias: At the last layer, ablating a direction is mathematically equivalent to adding a fixed "footprint" to the vocabulary scores, scaled by a "gate" (the cosine between the state and the direction).
  • The Origin is the Enemy: The erratic behavior of steering isn't a model quirk; it's a geometric artifact of measuring angles from the origin when the data lives in a distant cluster.
  • Judge Hallucinations: The authors found that standard safety judges (like Llama-Guard) disagree wildly on how to score "collapsed" (repetitive) text. One judge might see repetition as a "perfect defense" (refusal), while another sees it as a "perfect attack" (harmful), potentially fabricating a 46-point performance swing in research results.
  • Predictability: You don't need to generate a single token to know if Centered Ablation will help. The $\omega$ and $\eta$ diagnostics correctly predicted the behavior of five models that the researchers hadn't even tested during the pilot phase.

Key Results

  • Fixing Under-firing: On Llama-3 models where standard ablation failed, CA raised attack success by 29.0 and 46.3 points.
  • Preventing Collapse: On Gemma-2-9B, standard ablation caused the model to collapse on 50% of prompts. CA reduced this collapse rate to just 1%.
  • Preserving Benign Performance: CA preserved 12 to 40 points more benign answers compared to the standard recipe on high-overlap models.
  • Superiority over ACE: Compared to Affine Concept Editing (ACE), which uses a different offset method, CA was significantly more faithful to benign behavior (preserving 39 to 46 more points of original answers) while maintaining similar attack success.

Limitations and Future Work

  • Single Site: The closed-form reduction only holds for the last residual stream. While the authors argue this is a critical site, the math doesn't directly translate to earlier layers where attention and MLP blocks propagate the edit.
  • Rank-1 Restriction: The analysis is limited to rank-1 ablation (removing a single direction). Extending this to multi-dimensional "refusal cones" is a suggested next step.
  • Greedy Decoding: The theoretical proofs assume greedy decoding. While the results held in practice, the authors note that under sampling, the rescaling factor $\tau_s$ acts as a temperature shift, which could complicate the output distribution.
  • Small-n for "Laws": The diagnostic thresholds (like $\eta = 0.6$ for under-firing) are based on a study of 8 models; more diverse model families would be needed to confirm these as universal "laws."

Summary updated

Abstract

Directional ablation removes a hidden state's component along a "behaviour direction" while a language model runs. It is the standard way to switch a behaviour such as refusal off, but it is erratic: on some models it barely fires, on others the text collapses into repetition. We show why, and fix it. At the last residual stream, ablation is exactly a logit bias: a fixed vocabulary footprint scaled by a gate, the cosine between the hidden state and the direction. The gate is measured from the origin, but last-layer states share a large common mode far from it, so a direction that overlaps the common mode fires on benign text and under-fires on its targets. Centered Ablation (CA) projects the common mode out of the direction, in closed form. Across 8 models, CA raises attack success by 29 and 46 points where the standard recipe under-fires, cuts collapse on Gemma-2-9B from 50\% to 1\%, and preserves 12 to 40 points more benign answers on high-overlap models, 39 to 46 more than the affine offset of prior work. Two numbers computed before generation predict, on models the analysis had not seen, which gain faithfulness and which gain attack success, and a collapse index marks when text degenerates. We also show that two standard safety judges score collapsed text in opposite directions, which can fabricate a 46-point effect.
BibTeX
@misc{dang2026centered,
  author = {Quy-Anh Dang and Chris Ngo},
  title = {{Centered Ablation: Why Last-Layer Directional Ablation Under-Fires or Collapses, and a Closed-Form Fix}},
  year = {2026},
  eprint = {2611.20148},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.20148},
}
Figure