Steering Interference Reflects the Model's Defaults, Not the Behavior Directions
Srikanth Malla, Chiho Choi, Joon Hee Choi
cs.LG · cs.AI
Quick verdict
Argues steering side effects reflect model defaults, not direction geometry; random-push and held-out tests on 10 models make a strong case.
Summary
The paper asks why steering one behavior in an LLM perturbs others, and argues the answer lies in the model rather than the steered direction. Using 24 behaviors defined by contrastive prompt triples on ten instruction-tuned models (1.5B-72B, four families), with all effects read from generated text by an LM judge and per-behavior calibrated steering coefficients, the authors build interference matrices over ordered behavior pairs. Three findings support the claim: (1) the antisymmetric part of the matrix carries 0.68 of entry magnitude on Qwen2.5-7B (steering profanity raises toxicity by 3.4 pooled SDs; the reverse effect is 0.0), which any cosine-overlap account cannot express; (2) random directions matched in norm reproduce the same susceptibility ranking (Spearman 0.90) while producing none of the lexically specific behaviors; (3) geometry fit on other behaviors predicts a fully held-out behavior's interference weakly (best Spearman 0.41, R2 below 0.10), and orthogonalized or anchored steers do not reduce side effects. The induced default set (refusal, sycophancy, poeticism) is model-specific, strongest below 10B and weakening in each family's largest model. Limitations: only 32 evaluation prompts, 12 AI-generated contrastive triples per behavior, and judge-sensitive membership of the default set (Jaccard 0.33-0.71 across judges).
Review complete
Assessment at a glance
4
risk flags identified
● high 0
● medium 2
● low 2
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Steering side effects are determined by the model's default behavior set rather than by the geometry of the steered direction.
- Interference is largely one-way (antisymmetric norm fraction 0.63-0.69 across all ten models), which any cosine-overlap account cannot express in principle.
- Random directions matched in added norm reproduce the same susceptibility ranking as real steers (Spearman 0.90 on Qwen2.5-7B) while producing none of the lexically specific behaviors.
- Geometry fit on other behaviors predicts a fully held-out behavior's interference only weakly (best Spearman 0.41, held-out R2 below 0.10), and orthogonalized or anchored steers do not reduce side effects.
- Linear decodability does not imply steerability: all 24 behaviors reach held-out AUROC 1.00 but only 20 change the generated text.
Strengths
Readout discipline against circular measurement
Sec. 1-2 argue a probe-based readout would confirm the geometric account tautologically (steering adds i's direction, so j's projection rises with overlap regardless of output); every effect is therefore read from generated text by an LM judge. Sec. 4 quantifies the payoff: all 24 behaviors reach held-out AUROC 1.00 yet only 20 move the text.
Genuinely held-out evaluation design
Sec. 7/App. E use leave-one-behavior-out CV removing a behavior with all 38 of its pairs, closing the leakage loophole (a pair-holdout still leaves the behavior in 37 other pairs) the authors show afflicts prior geometric work; best held-out Spearman is 0.41 with R2 under 0.10 across three direction estimators.
Clean falsification control
Breadth with ablations
Unusually transparent metric reporting
Weaknesses
Thin evaluation basis
Judge sensitivity could contribute to the antisymmetric result
Scale-trend claim rests on few points without uncertainty
Narrow intervention regime
Behavior set selection may flatter the readout
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
The paper addresses a precise question: what determines the off-target effects of activation steering? It argues against geometric accounts (interference proportional to cosine overlap of behavior directions, attributed to Sharma et al. 2026 and Bhandari et al. 2026) and for a model-centric account: a steer relaxes the model toward a small default behavior set (refusal, sycophancy, poeticism) largely independent of what is steered. Five contribution bullets (Sec. 1) map onto Secs. 4-8. Claims are mostly calibrated ('explains almost none', 'only weakly'), and Sec. 10 bounds them explicitly: the default set is model-specific, the scale trend is unexplained, and most pairwise interference is captured by neither account.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 4 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is steering interference in LLMs?
When you add an activation-space direction to make a language model express one behavior (for example politeness), other behaviors often shift too. This paper measures those side effects for 20 behaviors across ten models and asks what determines them.
What is the paper's main finding?
Side effects follow the model's pre-existing default tendencies (chiefly refusal, sycophancy, and poetic language), not the geometry of the steered direction: a random vector of the same size produces the same side-effect ranking, and directions' overlap predicts a held-out behavior's interference almost not at all.
What does this mean for using steering as a safety tool?
Decorrelating or disentangling behavior directions will not make steering modular. Instead, the authors suggest measuring a model's default set - cheaply recoverable with a single random push - to learn which behaviors a steering edit is likely to move.
Which models were studied?
Ten publicly available instruction-tuned models from 1.5B to 72B parameters across four families: Qwen2.5 (1.5B/3B/7B/32B/72B), Llama-3.1-8B, Llama-3.3-70B, Gemma-2-9B and 27B, and Mistral-7B, each steered and judged by itself.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figure interiors (data points in Figs. 1-4) are not recoverable from text extraction; all numbers cited in this review come from prose, Table 1-6, and appendices A-G, which are present and complete in the extracted text.
- The metadata-supplied abstract was cut mid-sentence, but the complete abstract appears in the body text, so no content was lost.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.