Concept Steering in ML

• Concept steering is the inference-time control of internal model activations by intervening on hidden representations to bias output semantics.
• Canonical mechanisms include activation addition, subspace projection, and low-rank edits that enable targeted interventions in LLMs, diffusion models, and vision transformers.
• Evaluations reveal trade-offs between improved concept expression and potential declines in fluency or sample quality, underscoring challenges of reliable, identifiable intervention.

Concept steering, in contemporary machine learning, denotes the inference-time control of a model by identifying internal representations associated with a semantic concept and intervening on those representations so that generation is biased toward, away from, or between target behaviors without retraining the base model (Beaglehole et al., 6 Feb 2025). In current work this usually means modifying hidden states, residual streams, or learned low-rank subspaces in LLMs, diffusion transformers, or vision models, although some papers also study concept-level prompt editing as a steering mechanism (Konovalova et al., 10 Apr 2026). The term steering also has an older and technically distinct meaning in quantum information, where it denotes remote preparation of ensembles and can be reformulated as a separability problem via steering maps (Moroder et al., 2014).

1. Concept steering as an inference-time control problem

A basic formulation treats steering as a transformation of hidden representations during the forward pass. In LLMs this is often written as an additive perturbation at layer $l$, $h_l'(x)=h_l(x)+\delta_l$, or, in a compositional form, $$\mathbf{h}'=\mathbf{h}+\mathbf{V}\boldsymbol{\alpha}$$, where $\mathbf{V}$ is a dictionary of concept directions and $\boldsymbol{\alpha}$ are task-specific coefficients (Han et al., 7 Feb 2026). The central promise is a middle ground between prompting and fine-tuning: interventions remain lightweight, localized, and interpretable, yet can change behavior more directly than natural-language instructions (Li et al., 2 Feb 2026).

This literature also distinguishes concept detection from model steering. Detection asks whether a concept is present in a hidden state; steering asks whether an intervention on that state causes the model’s outputs to express the concept (Wu et al., 28 Jan 2025). The distinction is consequential because a method can be a strong detector yet a weak causal lever. A recurring result is that detection performance and steering performance do not coincide, particularly for SAE-based methods and feature-selection variants (Wu et al., 28 Jan 2025).

A broader implication is that “concept steering” is not one method but a family of intervention regimes. Some methods target a single static direction, some a low-dimensional subspace, some a context-dependent vector field, and some an explicitly causal subset of components such as attention heads (Sankaranarayanan et al., 17 Feb 2026). This suggests that the field is better understood as a study of controllable internal representations than as a single algorithmic template.

2. Canonical intervention mechanisms

The canonical mechanism is activation addition: learn a direction associated with a concept and add it to an internal representation during inference. For many LLM methods this is the default rank-1 intervention, and for downstream adaptation it can be extended to a compositional recipe over reusable basis vectors, $$\mathbf{v}_{\text{combined}}=\sum_{i=1}^{k}\alpha_i \mathbf{v}_i$$ (Han et al., 7 Feb 2026). In diffusion transformers, the same principle appears as manipulation of hidden intermediates, with interventions on pooled text embeddings or text-token activations such as $\tilde a=a+\alpha v$ and $Y_{\text{steered}}=Y_t+\alpha v_{\text{steering}}$ (Konovalova et al., 10 Apr 2026).

A second family uses low-rank or learned representation edits rather than a single additive vector. RePS, for example, trains three parameterizations—rank-1 steering vectors, LoReFT, and LoRA—under a bidirectional preference objective that jointly optimizes steering and suppression (Wu et al., 27 May 2025). GCM likewise combines localization with sparse steering over selected attention heads, using mean steering, difference-in-means steering, or ReFT after identifying mediators of a long-form concept (Sankaranarayanan et al., 17 Feb 2026).

A third family replaces vector translation with subspace projection. “Conceptors for Semantic Steering” models a concept as a soft projection matrix $C=R(R+\alpha^{-2}I)^{-1}$, learned from pooled activations across both poles of a bipolar concept, and applies it through rules such as $z'=(1-\beta)z+\beta Cz$ or $h_l'(x)=h_l(x)+\delta_l$0 (Triantafyllopoulos et al., 6 May 2026). This preserves a multidimensional concept subspace rather than collapsing it to a single axis.

A fourth family argues that Euclidean vector addition is geometrically mismatched to softmax representations. “The Information Geometry of Softmax: Probing and Steering” defines dual steering by moving in dual coordinates, $h_l'(x)=h_l(x)+\delta_l$1, and proves that the corresponding KL projection optimally changes the target concept while minimizing off-target distribution change under a concept-factorization assumption (Park et al., 17 Feb 2026). This is a sharper formulation of the long-standing problem of off-target drift.

3. Learning concept representations

Most steering pipelines begin by learning a concept representation from contrastive data. The simplest construction is difference-in-means, which computes the mean positive representation minus the mean negative representation and then uses the resulting vector both for detection and for activation addition (Wu et al., 28 Jan 2025). Linear probes and logistic regression provide another standard route: the learned weight vector becomes the steering direction (Zhao et al., 21 May 2025).

More recent work generalizes beyond single directions. “Toward universal steering and monitoring of AI models” trains layer-wise predictors and uses Recursive Feature Machines with the Average Gradient Outer Product, $h_l'(x)=h_l(x)+\delta_l$2, so that the top eigenvectors form concept directions or subspaces (Beaglehole et al., 6 Feb 2025). This supports graded labels, multiclass tasks, and multi-concept linear combinations across layers.

Several papers modify the data representation rather than the intervention rule. “Denoising Concept Vectors with Sparse Autoencoders” first maps hidden states into a sparse, overcomplete latent space, selects top-$h_l'(x)=h_l(x)+\delta_l$3 latents that best separate positive and negative examples, reconstructs a filtered hidden state, and only then learns the concept vector; the paper reports improved steering success rates over vanilla linear probing and difference-in-means on several behavioral concepts (Zhao et al., 21 May 2025). “Identifiable Steering via Sparse Autoencoding of Multi-Concept Shifts” moves from hidden states to embedding differences, $h_l'(x)=h_l(x)+\delta_l$4, and proves that Sparse Shift Autoencoders recover latent concept shifts up to permutation and rescaling under support-variability and sparsity assumptions, which is directly relevant when single-concept supervision is unavailable (Joshi et al., 14 Feb 2025).

Steer2Adapt reframes the problem again: rather than learn a new vector for every task, it assumes a frozen semantic prior subspace $h_l'(x)=h_l(x)+\delta_l$5 and uses Bayesian Optimization in the low-dimensional coefficient space to discover the best mixture of reusable basis vectors from only a few examples (Han et al., 7 Feb 2026).

Family	Core construction	Representative papers
Static vectors	Mean differences or probe weights added to hidden states	(Wu et al., 28 Jan 2025)
Supervised subspaces	AGOP / RFM eigenvectors, layer-wise concept directions	(Beaglehole et al., 6 Feb 2025)
Sparse latent methods	SAE-denoised representations or identifiable shift autoencoders	(Zhao et al., 21 May 2025, Joshi et al., 14 Feb 2025)
Compositional steering	Low-dimensional coefficient search over a concept dictionary	(Han et al., 7 Feb 2026)

4. Geometry, localization, and reliability

A central development after the first wave of steering vectors is the claim that steering failure is often geometric rather than merely statistical. “Steering Vector Fields” argues that a static steering vector assumes a constant concept-improving direction across representation space, whereas the locally effective direction can vary with context; it therefore defines a context-dependent field $h_l'(x)=h_l(x)+\delta_l$6, refreshed during long-form generation and composed across attributes through a softmin boundary construction (Li et al., 2 Feb 2026). This directly targets unreliability, anti-steerability, long-form drift, and multi-attribute interference.

“Predicting Where Steering Vectors Succeed” operationalizes a related idea through the Linear Accessibility Profile. Its main quantity, $h_l'(x)=h_l(x)+\delta_l$7, applies the model’s own unembedding matrix to intermediate residual states and measures whether a concept is linearly accessible in output-aligned form. Across 24 controlled binary concept families on five models, peak $h_l'(x)=h_l(x)+\delta_l$8 predicts steering effectiveness at $h_l'(x)=h_l(x)+\delta_l$9 to $$\mathbf{h}'=\mathbf{h}+\mathbf{V}\boldsymbol{\alpha}$$0 and layer selection at $$\mathbf{h}'=\mathbf{h}+\mathbf{V}\boldsymbol{\alpha}$$1 to $$\mathbf{h}'=\mathbf{h}+\mathbf{V}\boldsymbol{\alpha}$$2 (Billa, 16 Apr 2026). This is a strong rebuttal to the common heuristic of always steering at the middle layer.

A related search-theoretic perspective appears in “When Is Rank-1 Steering Cheap?”. That paper argues that many concepts are not unsteerable; rather, a useful rank-1 intervention often exists but is expensive to find under finite budget. Prompt-boundary directional alignment predicts where effective interventions occur, and restricting Bayesian optimization to the top 15 layers ranked by this alignment reduces the trials needed to recover $$\mathbf{h}'=\mathbf{h}+\mathbf{V}\boldsymbol{\alpha}$$3 of best-found utility by $$\mathbf{h}'=\mathbf{h}+\mathbf{V}\boldsymbol{\alpha}$$4 on average across three model families (Robertson et al., 9 May 2026). The same paper introduces concept granularity, $$\mathbf{h}'=\mathbf{h}+\mathbf{V}\boldsymbol{\alpha}$$5, as a measure of directional heterogeneity across contexts; higher granularity correlates with slower convergence and lower best-found utility.

Other work treats unreliability as a structural property of the representation itself. The Cylindrical Representation Hypothesis relaxes orthogonality assumptions from the Linear Representation Hypothesis and proposes a sample-specific axis–orthogonal geometry in which the axis provides the stable semantic push while a surrounding normal plane controls steering sensitivity; the paper’s theorem states that the sensitive sector in this plane is not reliably predictable from the observable difference vector and steering vector alone (Gao et al., 3 May 2026). This suggests that some steering variance is intrinsic rather than a defect of search.

Localization methods address the same problem from an interventionist angle. An attention-guided framework uses attention to the concept prefix for automatic token selection, attention-based soft labels for heterogeneity, and a concept enrichment score for layer choice; across a 512-concept benchmark it nearly doubles the number of successfully steered concepts and reaches about $$\mathbf{h}'=\mathbf{h}+\mathbf{V}\boldsymbol{\alpha}$$6 steering success on Llama-3.1-8B (Davarmanesh et al., 30 Jan 2026). GCM goes further by ranking attention heads through activation patching or attribution patching and then steering only the strongest mediators of a binary long-form concept such as refusal, sycophancy, or verse style (Sankaranarayanan et al., 17 Feb 2026).

5. Applications across modalities and domains

In LLMs, concept steering has been applied to style, persona, truthfulness, harmfulness, hallucination, multilingual output, and code behavior. “Toward universal steering and monitoring of AI models” reports that learned concept directions can steer into Mandarin, German, Spanish, Javascript, poetic style, Shakespearean style, harmful content, PII-like outputs, honesty/dishonesty, and graded sentiment, while the same layer-wise representations can be aggregated into monitors for hallucinations, toxic content, harmful instructions, and untruthfulness (Beaglehole et al., 6 Feb 2025). This suggests that steering and monitoring are two uses of the same learned internal geometry.

For code generation, “Security-by-Design for LLM-Based Code Generation” extracts a security concept vector from contrastive secure-versus-insecure code pairs and applies the activation update $$\mathbf{h}'=\mathbf{h}+\mathbf{V}\boldsymbol{\alpha}$$7 during decoding (Wendlinger et al., 11 Mar 2026). The paper argues that CodeLLMs often internally encode security-relevant concepts even while generating insecure code, and reports superior performance to prior methods across CodeGuard+ and CWEval, with hybrid combinations often producing the best trade-offs.

In diffusion models, SHIFT extends concept steering to DiT and flow-transformer image generation. It learns steering vectors from contrastive prompt pairs and injects them into pooled text embeddings or internal text-token activations to suppress, shift, or add concepts such as nudity, styles like Van Gogh, and objects such as hats or glasses (Konovalova et al., 10 Apr 2026). On I2P nudity suppression for FLUX.1[schnell], the paper reports detected nudity instances dropping from 612 for the baseline to 122 with SHIFT (250) and 97 with SHIFT (500), while CLIP remains nearly unchanged (Konovalova et al., 10 Apr 2026).

In vision transformers, ViSAE uses sparse autoencoders, a 64K-image probing suite, a 16,050-concept vocabulary, top-down concept reading, and bottom-up circuit tracing to identify concepts and intervene on them (Li et al., 4 Jun 2026). On WaterBirds, concept editing improves worst-group accuracy from 50.3% for a linear classifier on CLIP-ViT-B-32 CLS tokens to 98.5% after removing spurious background concepts, outperforming existing methods by 23.8% (Li et al., 4 Jun 2026).

A different branch of work steers by editing the prompt concepts rather than model activations. ConceptX identifies semantically rich concepts in the input prompt, estimates concept importance through Shapley-style coalition scoring using semantic similarity of generated outputs, and then removes or replaces the highest-attribution concept (Amara et al., 12 May 2025). On sentiment steering it reports a shift of 0.252 versus 0.131 for random edits, and on Salad-Bench it reduces attack success rate from 0.463 to 0.242 (Amara et al., 12 May 2025). This suggests that “concept steering” can also refer to transparent input-side interventions when the steering goal is prompt-auditable control.

6. Evaluation, misconceptions, and open technical issues

A common misconception is that accurate concept detectors are automatically good steering mechanisms. AxBench was designed partly to test that assumption and finds that it is false: prompting outperforms all evaluated steering methods, finetuning is next, and among representation-based methods only ReFT-r1 is strongly competitive, while SAEs are not competitive on either steering or detection (Wu et al., 28 Jan 2025). The same benchmark shows that difference-in-means, linear probes, and ReFT-r1 are the strongest representation-based detectors, but good detection still does not imply effective causal control (Wu et al., 28 Jan 2025).

This does not mean representation steering is static. RePS shows that a better training objective can materially narrow the gap with prompting: on Gemma models from 2B to 27B, RePS outperforms existing steering methods trained with a language modeling objective, improves suppression on larger Gemma-3 variants, and remains resilient to prompt-based jailbreaking attacks that defeat prompting (Wu et al., 27 May 2025). A plausible implication is that part of the historical weakness of steering methods came from objective mismatch rather than from an intrinsic limit of activation interventions.

Trade-offs remain fundamental. Stronger steering often improves concept expression while harming instruction-following, fluency, or sample quality; AxBench reports a steering-factor trade-off in which concept score rises as instruction-following drops (Wu et al., 28 Jan 2025). SHIFT explicitly reports that stronger steering improves erasure but worsens FID or image quality, even if the degradation is described as minor relative to the erasure gain (Konovalova et al., 10 Apr 2026). SDCV likewise notes dependence on pretrained SAEs and the fact that SAE features are imperfect proxies for causal concepts (Zhao et al., 21 May 2025).

Another unresolved issue is identifiability. SSAE argues that unsupervised steering becomes reliable only when latent shift directions are identifiable up to permutation and rescaling, not merely up to an arbitrary invertible transform (Joshi et al., 14 Feb 2025). CRH, SVF, and the geometry-guided search literature collectively suggest that some failures arise from rotating local directions, high granularity, or latent sector uncertainty rather than from poor engineering alone (Gao et al., 3 May 2026).

Finally, the field is explicitly dual-use. The same methods that suppress harmfulness can also amplify it; the same internal concept directions that monitor hallucination or toxicity can be used to induce hallucinations, harmful content, or dishonest answers (Beaglehole et al., 6 Feb 2025). For that reason, concept steering is now studied not only as a control technique but also as a framework for mechanistic auditing, capability analysis, and safety evaluation.