Centered Ablation: why last-layer direction ablation is insufficient or collapses, and a closed-form correction
Centered Ablation: Why Last-Layer Directional Ablation Under-Fires or Collapses, and a Closed-Form Fix
The authors propose Centered Ablation: projecting the benign common mode out of the ablation direction, fixing under-triggering and collapse across 8 models, reducing the collapse rate of Gemma-2-9B from 50% to 1%.
- Directional ablation is a common inference-time edit for turning off behaviors such as refusal, but it performs unstably across models: on some it barely triggers, on others the output collapses into repetition; the literature treats the symptoms as individual model differences, with no explanation and no way to predict them.
- The authors restrict the analysis to the final residual stream and prove that ablation there is equivalent to a logit bias scaled by a cosine gate; the symptoms stem from the overlap ω between the direction and a benign common mode. Centered Ablation removes the common mode from the direction via a closed-form projection, and gives three generation-free diagnostic quantities ω, η, β.
- On 8 models, CA raises fluent-only ASR by 29.0 and 46.3 points on the two under-triggering models, lowers the Gemma-2-9B collapse rate from 50% to 1%, and retains 12–40 more points of benign answers on high-ω models; ω and η predict gains on unseen models, and the regime prediction was correct for 4 of the 5 later-added models.
- The authors note that the conclusions apply only to rank-1 ablation of the last residual stream and greedy decoding, that the two diagnostic regularities are based on only 8 models, that the η=0.6 split relies on only 2 models and the β threshold is crossed by only one model, and that DIM+offset is a version of the ACE offset at this paper's site, not ACE's original mid-layer setting.
- Threat model
- The attacker has the model weights and performs directional ablation (s=1) along the refusal direction at the final residual stream to remove refusal behavior; the victim systems are 8 open-source instruction-tuned models; success is judged by the HarmBench classifier and Llama-Guard-3-8B, and fluent-only ASR counts collapsed outputs as failures.
- Evaluated models
- Qwen2.5-0.5B-InstructQwen3-4B-Instruct-2507Qwen2.5-7B-InstructQwen2.5-3B-InstructGemma-2-2B-itLlama-3.1-8B-InstructLlama-3.2-3B-InstructGemma-2-9B-it
- Benchmarks
- AdvBenchAlpacaHarmBenchJailbreakBenchXSTestMMLUGSM8KEmotion stories
- Metrics
- fluent-only ASRcollapse ratebenign rewrite rateωηβMMLU accuracyover-refusal rate
Researchers at Knovel Engineering Lab proposed Centered Ablation (CA) to explain and fix the failure of directional ablation, an activation steering method commonly used to remove refusal directions. The paper proves that last-layer directional ablation is exactly a cosine-gated logit bias, and that the overlap between the common pattern shared by last-layer hidden states and the direction uniformly explains three failure modes: insufficient effect, spillover to benign text, and collapse into repetition; CA uses a closed-form projection to remove the common pattern from the direction. On 8 instruction-tuned models, CA raises the fluent attack success rate of Llama-3.1-8B and Llama-3.2-3B by 29.0 and 46.3 percentage points, lowers Gemma-2-9B's collapse rate from 50% to 1%, and on high-overlap models retains 39 to 46 percentage points more benign responses than the ACE offset method.
Why it was selectedThe paper attributes the failure of directional ablation to a single mechanism and gives a closed-form correction; its pre-generation diagnosis and analysis of safety evaluation pitfalls are of reference value to teams working on activation steering and evaluation.
Paper analysis
What problem does this paper attempt to solve?
Why directional ablation under-triggers or collapses across different models: the authors give a mechanistic explanation and a closed-form fix, CA.
Core problem: directional ablation performs unstably across models—on some it barely triggers, on others the output collapses into repetition—and the literature only treats the symptoms as individual model differences, with no mechanistic explanation and no way to judge in advance which model will fail in which way.
- Setting and importance: inference-time activation editing is a common means of controlling model behavior without retraining; ablating the refusal direction can make chat models comply with harmful requests, and the reverse strengthens refusal (Arditi et al., 2024); the recipe is cheap, gradient-free and widely reused.
- Shortcomings of existing methods (according to the paper): practitioners tune the strength, change the estimator or add offsets; ACE (Marshall et al., 2025) re-anchors the projection on the benign mean and claims it prevents collapse; the authors argue that none of these fixes explains what the edit does to the output distribution, and therefore cannot state in advance which models will have problems or what a fix must change.
- Core observation: at the final residual stream, ablation is exactly a logit bias—a fixed vocabulary footprint multiplied by a gate (the cosine between the hidden state and the direction); the cosine is measured from the origin, while final-layer states cluster along a common mode away from the origin, and a direction overlapping the common mode mis-triggers on benign text and under-triggers on the target.
- The paper proposes: Centered Ablation (CA), which removes the common mode from the direction via a closed-form projection; and three diagnostic quantities ω, η, β that can be computed before generation, to predict which model needs the fix and what the fix brings.
- Threat model:
- Goal: remove (or inversely amplify) the refusal behavior of instruction-tuned models.
- Knowledge: the attacker can read and modify the model's internal activations (open-weight models, hooking at the final decoder block).
- Capability: apply rank-1 ablation at every generation position at inference time, with strength s=1 (s<0 when amplifying).
- Victim systems: 8 open-source instruction-tuned models; success is judged by the HarmBench classifier and Llama-Guard-3-8B, and the main metric, fluent-only ASR, counts collapsed outputs as failures.
What related research exists?
The authors position this paper along five lines: activation steering, refusal direction geometry, concept erasure, vocabulary-space readouts, and jailbreak evaluation.
Activation steering
- Turner et al. (2024)'s activation addition, Rimsky et al. (2024)'s contrastive activation addition, Zou et al. (2023a)'s representation engineering, and Li et al. (2023)'s inference-time intervention: adding or subtracting directions in hidden space at inference time.
- Lee et al. (2025)'s CAST: uses an independent detector to decide when to steer; Tan et al. (2024) and Dang & Ngo (2026) find that steering effectiveness varies greatly with input and behavior. The authors note that these works evaluate edits only by behavioral outcomes and do not derive the effect of edits on the output distribution.
Refusal direction and its geometry
- Arditi et al. (2024): refusal is mediated by a single direction; ablating it removes refusal, and adding it back induces refusal; Wollschläger et al. (2025) further propose a multi-dimensional refusal cone.
- ACE (Marshall et al., 2025): argues that refusal is an affine rather than linear function of activations, re-anchors the projection at the benign mean (at a mid-layer chosen by observation), and claims the offset prevents collapse. The authors criticize that its offset is geometrically motivated rather than derived from the effect of the edit, that the direction is unchanged, and that it gives no criterion for which models need it.
Linear representations and concept erasure
- The linear representation hypothesis (Mikolov et al., 2013; Park et al., 2024) supports projecting concepts away; the erasure methods INLP (Ravfogel et al., 2020) and LEACE (Belrose et al., 2023) choose projections that remove linearly decodable information. The authors note that they optimize decodability rather than steering selectivity: in this paper's tests, the most discriminative direction is precisely the worst ablation direction.
Vocabulary-space readouts and anisotropy
- Logit lens (nostalgebraist, 2020; Geva et al., 2022; Belrose et al., 2025) reads intermediate states through the unembedding; this paper uses the same mapping to precisely characterize a single intervention rather than to describe the unsteered state.
- Contextual representations occupy a narrow cone (Ethayarajh, 2019), are dominated by a few coordinates (Timkey & van Schijndel, 2021; Sun et al., 2024), and removing principal components is beneficial (Mu & Viswanath, 2018); the authors state that these works do not connect anisotropy to the selectivity of interventions.
Evaluation of jailbreaks and steering
- Attack success is judged by classifiers (HarmBench: Mazeika et al., 2024; Llama Guard: Inan et al., 2023); Souly et al. (2024) note that refusal-only or keyword metrics overstate success; over-refusal is measured separately by XSTest (Röttger et al., 2024). The authors note that prior evaluations do not report how reviewers score degenerate text, whereas this paper finds that the two reviewers score degenerate text in opposite directions.
Baselines and benchmarks: the compared methods are DIM ablation (the standard recipe) and DIM+offset (the pure ablation form of the ACE offset at this paper's site); the benchmarks are AdvBench, Alpaca (calibration), HarmBench (attack), JailbreakBench and XSTest (jailbreak and over-refusal), MMLU, and GSM8K (capability).
Authors' positioning: CA is a closed-form correction to the direction itself, derived from an exact derivation of the effect of last-layer ablation; relative to ACE it changes the direction rather than the origin, relative to erasure methods it targets gating rather than decodability, and it comes with a pre-generation diagnostic (Table 1).
How does the paper address this problem?
Exactly reduce final-layer ablation to a cosine-gated logit bias, project out the common mode accordingly to obtain CA, and give three generation-free diagnostic quantities.
Overall approach: study a single operator (rank-1 directional ablation) at a single site (the residual stream after the final decoder block), derive exactly its effect on logits, and let the derivation itself explain the failure modes and determine the form of the fix. Centered Ablation (CA) follows from this.
Problem setting and formalization
- Let h be the residual stream after the final decoder block; the model obtains logits via RMSNorm and unembedding: ℓ(h) = WU N(h), and greedy decoding takes the argmax.
- Directional ablation Aus(h) = h − s(h⊤ u)u: s=1 removes the component of h along the unit direction u, s<0 amplifies it; the standard direction uDIM is the mean difference between the final-token states of harmful (D+, AdvBench) and benign (D−, Alpaca) prompts (Arditi et al., 2024).
- Reason for choosing this site: the attention of block L reads the output of block L−1, edits to h do not enter the KV cache, and the causal path from the edit to the next token is only the above expression, so a closed-form solution exists and it can be deployed as a logit processor.
Exact reduction: a cosine-gated logit bias
- Define the gate κu(h) = cos∠(h, u) and the footprint fu = WU G u. Proposition 3.1: the post-ablation logits equal the original logits minus s√d·κu(h)·fu, then multiplied by a positive scalar τs (which does not change the argmax).
- Two corollaries: self-gating—a state's gate is zero if and only if it is unaffected, and neither the strength s nor the footprint can protect benign text; the collapse exponent βs(h) = s√d|κu(h)|σfu/σℓ(h), and when β≳1 the footprint dominates and the output degrades into repetition.
Failure mechanism: the benign floor
- Final-layer benign states cluster around the common mode µ̂− (on the 3 measured models, the |cos| between the benign mean direction and the mean of all states is 0.83–0.99, Appendix G.3); define the overlap ω(u) = |u⊤µ̂−|.
- The gate decomposition (Eq. 6) gives a benign-mean gate ≈ λ̄ω: a floor determined only by ω, which tuning s cannot eliminate. Large ω causes two failures: benign text is rewritten or even collapses; and the class-separating part of the direction is diluted, causing under-triggering.
Centered Ablation and the offset route
- CA: uCA = unit((I − µ̂−µ̂−⊤)uDIM), which projects the common mode out of the direction before ablating. Proposition 3.2: ω(uCA)=0 (the floor is removed), and all class separation except the common-mode component is retained. Algorithm 1 requires only two sample means, one projection and one normalization, with no hyperparameters besides s; when s<0 it is called CA-Amp.
- Offset route (Eq. 8, the form of ACE's affine correction at this paper's site, denoted DIM+offset): keep uDIM and subtract the benign class's own coefficient c = µ−⊤uDIM. Proposition B.5: on behavioral states the offset route's average coefficient is 1/√(1−ω²) times that of CA (stronger triggering), but on benign states it retains an ω² share of the common-mode variance (CA removes it entirely); when ω≈0 the two are equivalent, and when ω is large the offset trades benign cost for effect.
Diagnostics and operator boundaries
- Three generation-free diagnostics (requiring only cached activations and WU, G): ω predicts CA's faithfulness gain; η(u) = the fraction of behavioral states whose greedy token has already been changed by the edit, predicting the ASR gain; β± determines whether the bias is enough to degrade the text.
- Capacity bound (Proposition B.6): the entire effect of the edit is a single vector whose length is proportional to the gate; if for some behavioral state the gate is small along all directions that do not harm benign states, then no strength can remove it without benign cost. RQ5 uses this to test whether refusal (decided by a few tokens) and emotion (spread throughout the text) fall on opposite sides of this boundary.
What experiments did the paper conduct?
8 models, 144 cells: CA fixes under-triggering and collapse, ω/η/β predict gains before generation, and the two judges give opposite verdicts on collapsed text.
Experimental setup
- Models tested: Qwen2.5-{0.5B, 3B, 7B}-Instruct, Qwen3-4B-Instruct-2507, Llama-3.1-8B-Instruct, Llama-3.2-3B-Instruct, Gemma-2-{2B, 9B}-it, 8 in total; the analysis was developed on 3 pilots (Qwen2.5-7B, Llama-3.1-8B, Gemma-2-9B), and the diagnostics and regime predictions for the 5 later models were locked before generation.
- Data: directions were calibrated using the final prompt states of 256 AdvBench versus 256 Alpaca prompts, with 6 calibration draws resampled; attack evaluation on HarmBench (n=240, excluding the copyright portion), benign behavior on 100 held-out Alpaca prompts, capability on MMLU (1000 questions) and GSM8K (100 questions).
- Conditions: three directions DIM, CA, DIM+offset, all with s=1 and all at the final layer; one model×direction×draw is one cell, and the main matrix is 8×3×6=144 cells; for amplification, on the 4 models with s<0(CA-Amp、DIM-Amp)。
- 指标与评审:ASR 由 HarmBench 分类器(Llama-2-13B)和 Llama-Guard-3-8B 双评审;输出 ≥200 字节且 zlib 压缩比 <0.2 判为退化,主指标 fluent-only ASR 把退化输出计为失败;良性改写率为与未引导答案的 difflib 相似度 <0.5 的比例;成对质量评审 Qwen3-8B 双顺序。
- 统计:配对提示级 bootstrap(5000 次)与每 draw 精确 McNemar 检验;跨模型相关 n=8。生成贪心、256 新 token。
主结果
据 Table 10、Table 11、Table 12(HarmBench,6 draws 均值,HB-F 为 fluent-only ASR):
Swipe horizontally to read the table
模型 ω DIM HB-F CA HB-F CA−DIM(点) 坍缩率 DIM→CA 良性改写率 DIM→CA Qwen2.5-0.5B .022 49.7 49.7 +0.0 3→3 59.2→60.0 Qwen3-4B .034 7.4 7.2 −0.3 0→0 30.2→30.3 Qwen2.5-7B .041 52.3 50.1 −2.2 0→0 35.2→33.8 Qwen2.5-3B .125 38.6 41.0 +2.4 0→0 67.5→62.8 Gemma-2-2B .315 29.3 26.7 −2.6 1→0 87.5→71.0 Llama-3.1-8B .332 33.6 62.6 +29.0 [22.9, 35.1] 0→0 45.8→33.5 Llama-3.2-3B .438 10.3 56.7 +46.3 [40.3, 52.4] 0→0 44.2→25.7 Gemma-2-9B .471 14.4 14.0 −0.5 50→1 73.2→32.8 - 作者称 CA 在出现失效的模型上修复对应失效:两个 Llama 的 ASR 提升在 6/6 draw、双评审下均显著;其余 6 个模型 ASR 变动至多 2.6 点且从不显著。
- 良性侧:ω>0.3, CA retains 12 to 40 more points of benign answers than DIM; the 4 models with ω<0.13 的 4 个模型差异至多 5 点(Table 12)。
- 代价:MMLU 逐模型逐方向变动至多 0.7 点,GSM8K 在噪声内(Table 13);24 个 cell 中 23 个的成对净偏好在 ±10 点内(Table 14)。
RQ1:化简是否精确
- 用缓存状态、WU、G 预测首 token,与 vLLM 实际输出对比:全部 cell 一致率 96.1%,7 个模型 96.3–99.9%,预测与实测拒答率 Pearson 0.95(Figure 3、Table 9)。
- 唯一短板 Gemma-2-9B(82.1%):作者解释称 CA 下移除了拒答起始 token,但后续以语义 deflection 形式恢复拒答(见 RQ5)。
RQ3:改方向还是移原点
- DIM+offset 的有害 gate 与 CA 之比在 8 个模型上符合预测因子 1/√(1−ω²)(最大误差 0.06,Table 15);其良性 gate 展宽恰好出现在 ω>0.3.
- Attack strength: under the HB judge, DIM+offset's ASR is never significantly lower than CA's, and is significantly higher on 3 models (3.7 points higher on Qwen2.5-7B, 12.4 points higher on each of the two 3B models); under the LG judge it is higher on 5 models and lower on Gemma-2-2B (Table 11).
- Benign cost: on the 3 models with ω>0.33, DIM+offset rewrites 39 to 46 more points than CA; the only clear quality regression is DIM+offset's net preference of −23.5 points on Gemma-2-9B (Table 14). Authors' conclusion: CA is not a stronger attack, but a route that preserves benign behavior when ω is large.
RQ4: prediction before generation
- ω predicts benign answer retention (Spearman 0.98, Pearson 0.90, n=8); η predicts the ASR gain (Pearson −0.95); the two are not interchangeable (ω ranks ASR gain at only Spearman 0.19). η=0.6 divides the two regimes on these 8 models: the 2 models below it gain +29.0 and +46.3 points, and the 6 above it fall in [−2.6, +2.4] (Figure 6).
- The locked regime predictions for the 5 later-added models held for 4 of 5; the miss, Gemma-2-2B (labeled ASR gain), actually retained 16.5 points of benign answers with no ASR gain, consistent with its ω=0.32, η=0.98 (Table 16).
- Collapse is a threshold at β+=1: exactly 6 of the 144 cells have β+>1 (all Gemma-2-9B with DIM, β+=1.03±0.01), with an average collapse rate of 49.8%; the other 138 have β+≤0.90, with an average collapse rate of 0.6% (maximum 4.2%). The authors note that only one model crosses, so the threshold can only be located in (0.90, 1.03].
RQ5: boundaries of the operator
- On Gemma-2-9B, after CA removes the refusal onset, 116 of 240 continuations turn into deflection, and ASR stays at about 14%; a separately fitted deflection direction is ineffective when ablated alone, and switching between the two directions per state is comparable to CA alone (14.2% vs. 13.3%). Based on the capacity bound, the authors state that no rank-1 direction can carry this stage.
- Emotion concepts (happy/sad/angry/afraid): ω orders the collateral reduction within each model (Spearman 1.00, n=4), but steering itself does not work—at s=1 the judged emotion moves by only 2.9 points on average; at s=3, CA removes 60% of emotion words but judged emotion drops by only 19 points, while rewriting 53–62% of unrelated benign answers (Figure 7a, Table 24).
- Amplification side: on Gemma-2-9B, DIM-Amp already collapses 31% at s=−1 and reaches 95–100% at s=−1.5/−2, while CA-Amp stays fluent up to s=−2 and has ASR 14.1% at s=−3; on Llama-3.1-8B, DIM-Amp bottoms out at 3.3% and CA-Amp reaches 0.5% (Table 25). The authors call this a trade-off with over-refusal, not overall superiority.
Blind spots of judges and benchmarks
- Without fluency filtering, Llama-Guard scores Gemma-2-9B's DIM outputs at 64.8% and CA at 18.4%, manufacturing a −46.4-point spurious effect (significant in 6/6 draws); with filtering it disappears (15.8% vs. 17.3%, not significant) (Table 11).
- On the same batch of degraded outputs, the HarmBench classifier judges 1.7% harmful while Llama-Guard judges 98.3%; on fluent outputs the two agree (Figure 8).
- An intervention with MMLU change ≤0.7 points can simultaneously rewrite up to 88% of free-form benign answers (Gemma-2-2B, DIM; Table 12).
Other ablations and analyses
- Kernel MMD erasure reduces to length-reweighted DIM, which after centering is CA (cos ≥ 0.999) (Appendix G.2).
- The most discriminative direction (shrinkage LDA, gate AUC 0.95) is the worst ablation direction: ASR 2.9–23.7%, lower than DIM and CA on all pilots (Table 19).
- The common mode is neither the first principal component (|cos|=0.11–0.22, versus 0.92–0.99 for uDIM) nor an outlier-coordinate effect; centering by the pooled mean of the two classes removes signal (Table 20).
- Collapse is determined by strength rather than direction locality: on Gemma-2-9B, DIM collapses 1% at s=0.5, reaches 20% at s=0.75 and 83% at s=1.25 (Table 21).
- Trajectory-calibrated directions, whitened directions and deflection directions are all no better than CA (Table 22); random directions leave ASR at baseline (Table 19).
What are some points that can be further explored?
The authors list two limitations: a single site and operator, and a small sample size for the diagnostic regularities; the experiments cover 8 open-source models with dual reviewers.
Limitations and future directions noted by the authors
- Single site and operator (Section 5): the simplification holds only under the last residual stream, rank-1 ablation, and greedy decoding; under sampling, τs is equivalent to temperature, and earlier-layer edits propagate through attention with no closed-form solution; the authors make no claims about mid-layers or additive steering.
- The scope of DIM+offset (Section 5): it is a version of the ACE offset at this paper's site, not ACE under the ACE authors' original setting (a mid-layer chosen by observation).
- Small-sample regularities (Section 5): the two diagnostic regularities are based on only 8 models; the η=0.6 split relies on only 2 models below the threshold; the collapse threshold of β+=1 is crossed by only one model and can only be bracketed in (0.90, 1.03].
- Some partial collapse onsets are not captured (Appendix F): Gemma-2-9B shows 20% partial collapse at s=0.75, and the authors state that this is not captured by the threshold of the mean exponent.
Experimental coverage
- The models tested are Qwen2.5 (0.5B/3B/7B), Qwen3-4B, Llama-3.1-8B, Llama-3.2-3B, and Gemma-2 (2B/9B), 8 open-source instruction-tuned models in total, up to 9B (Section 4.1).
- The attack evaluation uses 240 behaviors from HarmBench with greedy generation of 256 tokens; calibration uses 256 AdvBench versus 256 Alpaca per draw, for 6 draws in total (Section 4.1, Appendix C).
- The reviewers are the HarmBench classifier (Llama-2-13B) and Llama-Guard-3-8B as dual reviewers, and the quality and sentiment reviewers are Qwen3-8B; the paper does not report agreement between the reviewers and human judgments (the sentiment reviewer has 65.6% accuracy on 160 annotated stories, Appendix C).
- The second-concept experiment covers diagnostics of 4 sentiments × 8 models, and the behavioral experiments are conducted on 3 pilot models (Appendix H); the jailbreak defense evaluation uses 4 held-out templates on JailbreakBench with a single draw (Table 25).
- The paper discusses conditional steering (the CAST-type detector route) and reports that kernel density ratio detection has an advantage of at most +0.022 over the regularized linear method at the last layer, and −0.046 on Gemma, on which basis the authors did not adopt this route (Appendix G.2).
Summarize the main content of the paper
Final-layer ablation is a cosine-gated logit bias; CA projects out the common mode, and the closed-form fix addresses under-triggering, collateral damage and collapse.
An analysis and methods paper on inference-time activation editing: the authors give a precise mechanism for the unstable behavior of directional ablation and, based on it, propose the closed-form fix Centered Ablation (CA).
- Problem: ablating the refusal direction is a standard means of turning off model refusal, but on some models it barely triggers (after ablation, 86% of Llama-3.2-3B responses still begin with a refusal), and on some models it collapses into repetition (half of Gemma-2-9B prompts degrade); previously there was no mechanistic explanation.
- Method: at the final residual stream, ablation is exactly the footprint vector multiplied by a cosine-gated logit bias (Proposition 3.1); the gate is measured from the origin while final-layer states cluster along a common mode, and the overlap ω simultaneously explains under-triggering, benign collateral damage and collapse. CA projects the common mode out of the direction, requiring only two sample means and one projection.
- Key results: on Llama-3.1-8B and Llama-3.2-3B, fluent-only ASR increases by 29.0 and 46.3 points respectively (significant in 6/6 draws); on Gemma-2-9B the collapse rate goes from 50% to 1% with ASR unchanged; on the 4 models with ω>0.3, it retains 12–40 more points of benign answers than DIM, and 39–46 more points than the ACE-style offset route; MMLU changes by ≤0.7 points.
- Diagnostics: ω predicts the faithfulness gain (Spearman 0.98), η predicts the ASR gain (Pearson −0.95), and β+>1 flags the only collapsing model; the regimes of 5 later-added models were locked before generation, and 4 of 5 held.
- Evaluation findings: on the same batch of collapsed text, the HarmBench classifier judged 1.7% harmful while Llama-Guard judged 98.3%, which without fluency filtering can manufacture a 46-point spurious effect; unchanged MMLU can mask 88% of benign rewrites.
- Authors' conclusions: compute ω, η, β before steering; use CA when ω is large and the offset route when effect strength matters; when η is low, do not expect ablation to be effective; evaluations should report fluent-only success rate, collapse rate and free-form faithfulness. The authors also note that the conclusions are limited to final-layer rank-1 ablation and greedy decoding, and that the diagnostic regularities are based on 8 models.
Related papers
- Where Is Rank-1 DiM Refusal Intervenable in Multimodal LM Derivatives?Analysis / interpretabilityarXiv