Activation Steering

Activation steering has become a popular way to control Large Language Model (LLM) behavior without fine-tuning. Since the technique is plug-and-play, users share datasets and precomputed vectors to steer model activations. However, we show that a \emph{stealth data poisoning attack} silently compromises this pipeline. By substituting $4{-}6\%$ of tokens in the steering dataset, an attacker can silently align the resulting vector with an anti-refusal direction. This jailbreaks the target model while preserving the intended steering effect on benign prompts. Under this threat model, a malicious actor can distribute an apparently safe bundle containing texts, vectors, and weights, alongside an equivalence certificate that the end-user can verify. We test the attack on two open-weight model families and eight model-attribute combinations, observing that poisoned vectors reach an absolute attack success rate (ASR) of $20{-}55\%$, $+19\%$ to $+51\%$ over a clean reference. Finally, we find that a refusal-direction orthogonalization defense can recover ${\approx}82\%$ of the ASR gap without harming benign behavior.

Function vectors (FVs) are task representations elicited during in-context learning that can be used to steer Large Language Models (LLMs). However, design choices in their formulation remain underexplored. In this work, we study the impact of varying FV definitions for instructions along two degrees of freedom: attention head selection and steering. For head selection, using gradient-based attributions with Layer-wise Relevance Propagation (LRP) substantially improves efficiency as well as accuracy. For FV steering, applying it in a distributed manner yields a higher accuracy compared to simple aggregation. Our code is publicly available.

Text-to-video (T2V) models trained on large-scale web data can generate undesired content, motivating interventions that reduce harmful outputs without sacrificing visual quality. Activation steering offers an attractive mechanistic alternative to finetuning and prompt filtering, but existing T2V steering methods remain limited, typically applying coarse, non-anticipative interventions that can lead to oversteering and content degradation. To close this gap, we propose Latent Activation Linear-Quadratic Regulator (LA-LQR), a reduced-order optimal control framework for minimally invasive T2V steering. LA-LQR formulates T2V inference as a dynamical system and computes closed-loop feedback interventions that steer activations toward desired feature setpoints while penalizing unnecessary perturbations. To make optimal control feasible for high-dimensional video activations, we project activations onto a low-dimensional, task-relevant subspace derived from contrastive prompt pairs, estimate local linear dynamics in this latent space, and solve a latent LQR problem to obtain timestep- and layer-specific steering signals. We provide theoretical bounds relating latent setpoint tracking to raw activation-space feature control, and empirically validate the fidelity of the reduced latent dynamics. On concept steering and video safety benchmarks, LA-LQR reduces unsafe generations relative to baselines, while preserving prompt fidelity and visual quality.

Large vision language models (LVLMs) have made rapid advancements and are deployed across various applications, yet hallucinations remain a major challenge. Activation steering is appealing due to its minimal training overhead and controllability at inference time. However, we found that during autoregressive decoding, visual conditioning affects token prediction sparsely and locally across decoding steps, and many existing methods that average image-versus-no-image differences over the entire sequence dilute these critical signals, yielding low signal-to-noise ratio steering directions. Additionally, many existing methods apply a fixed steering strength, which misallocates the intervention budget, over-perturbs non-critical tokens, and can cause instability. To address these limitations, we propose Token-Level Visual-Sensitivity Steering (TLVS) for hallucination mitigation. Our approach first extracts token-level steering vectors and refines them, and then applies fine-grained, visual-sensitivity-adaptive steering only where it matters. This lightweight, plug-and-play mechanism requires only minimal training for calibration and can be applied across diverse vision-language models. It modulates the steering strength at each decoding step, selectively suppressing hallucination-prone spans while preserving evidence-grounded content. We evaluate TLVS on several benchmarks, including POPE, AMBER, CHAIR (COCO), MMHal, and HallusionBench, demonstrating consistent improvements over previous steering methods.

We introduce HERO'S JOURNEY, a benchmark for rule induction in goal-directed episodic tasks, where agents must infer hidden rules from demonstrations and act on them through multi-step execution. HERO'S JOURNEY covers eight tasks across attribute and procedural induction families, each with four structural rule forms, controllable lexical grounding, and identifiability conditions. Evaluating state-of-the-art LLMs, we find that models show evidence of rule induction, but the ability is limited and uneven across tasks. Meanwhile, process execution adds an execution bottleneck for models, whereas surface semantics has minimal effect. Induction-specific steering methods improve performance on attribute tasks but show no reliable gains on procedural tasks, suggesting the gap in procedural induction remains an open challenge.

Speech-aware large language models often generalize poorly to out-of-domain settings. We propose SALSA (Speech-Aware LLM Adaptation via Learned Steering Activations), a lightweight adaptation method that learns layer-wise steering vectors. Unlike commonly used steering approaches that rely on contrastive activation differences, SALSA directly optimizes steering vectors using a supervised objective. Across children's speech, multilingual speech, and Mandarin-English code-switching benchmarks, SALSA substantially improves performance over zero-shot inference and speech in-context learning baselines, achieving up to 46.8% relative improvements over zero-shot. Analysis further demonstrates that steering the encoder, particularly the later layers, is more effective than steering the LLM backbone. These findings suggest that steering improves downstream ASR performance by adapting higher-level acoustic and phonetic representations to better align with the pretrained language model representation space, rather than by modifying the decoder itself.

Transformer-based architectures have significantly advanced the generation of complex symbolic sequences, yet a significant gap remains in achieving fine-grained, interpretable control over discrete signal attributes. This paper investigates the mechanistic interpretability of the Multitrack Music Transformer (MMT) and proposes a framework for deterministic attribute modulation without retraining to bridge this gap via inference-time activation steering. Utilizing the Difference-in-Means (DiffMean) methodology, we isolate latent directions for signal attributes, specifically Pitch and Duration, within the residual stream. We validate the Linear Representation Hypothesis in this domain, achieving high correlation between steering magnitude and attribute shift. To address the inherent feature entanglement in multi-attribute steering, we introduce a Dual Steering framework utilizing Gram-Schmidt Orthogonalization. Experimental results demonstrate that this geometric decoupling reduces conceptual interference and signal degradation compared to naive vector addition, enabling independent deterministic control even against strong autoregressive conditioning.

Activation-based control steers large language models (LLMs) by intervening on their internal representations during inference, and has emerged as an effective paradigm for controlling behaviors such as persona and style. However, existing methods often rely on fixed steering directions or task-specific intervention modules, making them difficult to adapt to fine-grained concepts and compositional constraints. We propose UniSteer, a text-guided activation flow matching model that learns a conditional distribution over residual-stream activations from natural-language conditions. Instead of fitting a separate intervention for each target behavior, UniSteer learns a universal conditional velocity field in activation space. At inference time, UniSteer performs flow inversion by partially transporting a source activation toward a latent state and regenerating it under a target textual condition before injecting it back into the frozen LLM. The same conditional model supports activation-space classification by selecting the textual label with the lowest reconstruction energy. Experiments on three target LLMs show that UniSteer provides a unified interface across behavioral control, truthfulness steering, fine-grained concept steering, multi-constraint instruction following, and activation-space classification.

Causal interventions in language model representations have largely targeted discrete features, like grammatical number. However, language models must also make use of features that are graded. We introduce a method for causal intervention on continuous variables: given activation vectors paired with a graded target variable, we localize a low-dimensional direction for that variable and use this direction to edit a vectors toward counterfactual target values. We apply this method to a continuous feature that is well-studied in psycholinguistics, namely verb bias (which reflects which syntactic structures tend to follow a given verb). We show that verb bias is causally represented in steering vectors extracted from large language models: counterfactual edits to verb bias systematically shift downstream structural preferences. Verb bias has also previously been linked to in-context learning; in further analyses, we find that steering vectors encode error signals that could drive the error-driven update behavior seen in in-context learning but that these aspects of the steering vectors are not causally used in downstream production. Overall, these results show causal interventions can be applied to continuous variables, though connecting continuous variables to in-context learning remains a challenge.

Steering language model generation toward desired textual properties is essential for practical deployment, and inference-time methods are particularly appealing because they enable controllable generation without retraining. Recent work has also highlighted diffusion language models as an emerging generation paradigm with distinct decoding properties. However, most existing steering approaches either rely on auxiliary models or are designed for autoregressive next-token decoding, making them difficult to apply to diffusion language models DLMs, which generate text through iterative denoising of partially masked sequences. Therefore, we propose DLM-SWAI, a simple training-free steering method that biases the token distribution at each denoising step using pre-computed token-level style scores. Experiments on style and safety control tasks show that DLM-SWAI effectively steers diffusion language models while preserving generation quality and requiring minimal computational overhead. Ablations further reveal a controllable trade-off between steering strength and fluency, and our analysis links class-wise steerability to the strength of token-level attribute cues.

Safety detection models require examples of HHH (Helpful, Harmless, Honest)-violating outputs for robust generalization, however such examples are scarce. Activation Steering (AS) has emerged as a data-efficient method for generating target-concept-aligned responses. We investigate whether AS can generate high-quality training datasets for downstream classifiers, a question that remains untested. We present a two-fold study with intrinsic and extrinsic evaluation across $4$ concepts $\times\,2$ models $\times\,4$ steering methods. Intrinsically, beyond the field-standard rubric of steering success (concept alignment) and coherence, we introduce sample- and set-level diversity as a quality axis previously absent from the literature, and find that increasing steering strength reduces response diversity. Extrinsically, we replace HHH-violating examples in the available training data with steered generations and fine-tune detection classifiers. AS-generated data results in a better classifier than the prompting-generated data on $3$ of $4$ concepts. However, only $41$ of $136$ AS configurations outperform prompting, indicating that downstream utility lies in a narrow regime that jointly satisfies success, coherence, and diversity. The harmonic mean of these three axes correlates with downstream AUROC more consistently across concepts than success and coherence alone, providing a practical heuristic target for practitioners tuning AS hyperparameters. Together, our results highlight the potential of AS in synthetic data generation for improving safety detection and identify diversity as a critical, previously overlooked axis for tuning AS.

Object hallucination remains a primary obstacle to the reliable deployment of Multimodal Large Language Models (MLLMs). Current inference-time mitigation methods mainly assume hallucinations stem from visual neglect, steering models to enhance visual reliance. In contrast, our systematic interventions on multiple MLLMs show that pushing toward more visual reliance may exacerbate hallucinations on some models, while less may mitigate hallucinations. This result suggests that attributing hallucinations solely to visual insufficiency is underdetermined. We argue that the image, as a context, simultaneously competes with the model's parametric knowledge and the textual context. For this, we propose a training-free framework, Context-Preference Activation Steering (CAS). It extracts two semantically distinct Context Preference Vectors (CPVs) via two small sets of designed conflict samples and applies them via single-pass signed residual injection at mid-early MLP layers during inference to control information reliance. Experiments show that CAS substantially mitigates object hallucinations without increasing decoding latency and preserves native text-generation quality.

Alignment faking (AF) refers to a model strategically complying with a training objective to avoid behavioural modification while preserving its deployment preferences. Understanding when and why AF arises matters as models grow better at distinguishing training from deployment. Prior work finds AF fragile, prompt-sensitive, and model-dependent, leaving its underlying drivers unclear. We study AF in a controlled, minimal setup that isolates its core components, and observe it across a wider range of models than previously reported, including small-scale models. We identify three separable drivers -- values, goal guarding, and sycophancy -- and show via targeted prompt ablations and activation steering that each independently modulates AF behaviour. Our results indicate AF is more widespread than previously reported and that its occurrence is predictable from situational cues and measurable model tendencies such as baseline sycophancy and stated values. The decomposition suggests concrete directions for detecting and mitigating AF in future models.

Large language models are increasingly used as behavioral simulators, but it remains unclear when their outputs reflect human-like cognitive mechanisms rather than prompt-sensitive surface patterns. We study this question through the realization effect, a well-characterized finding in behavioral economics in which risk-taking differs systematically after paper versus realized gains and losses. We evaluate LLM behavior at three levels: prompt-only behavioral sensitivity, linear readout of internal representations, and causal control via activation steering. Prompt-only results show systematic condition sensitivity, but the directional pattern does not reproduce human realization-effect predictions. Gemma's residual stream contains a linearly decodable realization-status signal at layer 18 that generalizes to held-out prompts. Steering along this direction does not, however, reliably shift downstream risk choices, a null result that holds across positive scales and in a negative sign-symmetry run. Behavioral sensitivity, latent readout, and causal control are three distinct properties that do not automatically co-occur, and successful latent readout is insufficient evidence that a model behaviorally relies on a representation during downstream decision-making.

Achieving both efficient pushability (propulsion transmission) and proximally concentrated bending for steerability is challenging for magnetically actuated soft catheters: higher axial/bending stiffness improves force transmission but reduces steerability, whereas lower stiffness enables large, proximally concentrated bending yet increases kinking/buckling risk under compressive push loads. To address this trade-off, we propose a stiffness-optimized multi-segment magnetically actuated catheter (SO-MAC) that integrates a decoupled steering-advancement mechanism with a gradient-stiffness architecture. The SO-MAC concentrates bending about a stable proximal pivot during advancement while the distal section passively self-straightens to transmit propulsion, aided by the optimized stiffness distribution and elastic recovery of the spring backbone against friction-induced kinking/buckling. Over $0{-}180^{\circ}$ combined steering and advancement, the pivot remained stable and the distal tip advanced near-straight toward the target direction. A 1.5 mm-diameter SO-MAC achieved up to $180^{\circ}$ steering with a 3 mm bending radius at its 10 mm tip, with an average shape error of $1.39 \pm 0.56$ mm and a steering-pivot error of $0.35 \pm 0.10$ mm. Visual feedback control in a bronchial phantom further confirmed robust navigation through highly curved, bifurcating paths.

Steering a language model - intervening on its internal activations to change downstream behaviour - has recently expanded beyond linear interpolation to nonlinear methods such as angular and kernelized steering, which define intervention transformations without learning an explicit geometry over paths in activation space. Freshly introduced geometry-aware manifold methods do learn such a geometry, but require labelled class centroids together with prescribed cyclic or sequential structure. These assumptions restrict where manifold steering can be applied, since existing constructions require labelled centroids and compatible boundary conditions. We recast manifold steering more broadly as \textbf{Riemannian geodesic computation} on activation space, recovering linear and labelled-spline steering as geodesics under particular choices of metric. A principled metric within this framework is the output-space Hellinger distance pulled back to activations; we approximate this with a learned encoder trained on output distances over a small concept-token schema - no per-prompt labels, no topology prior, and no per-task curve fitting. Empirically, the method reliably drives the model onto the target class across all tasks in a standard four-task language-model arithmetic benchmark, while following more behaviourally natural trajectories than baselines on smaller output spaces. We thereby provide a unified Riemannian framework for manifold steering together with a schema-supervised, label-free instantiation that operates without labelled centroids or prescribed boundary conditions.

Video world models learn representations of physical dynamics, but controlling their physical expectations at inference time remains an open problem. Recent interpretability work identified a Physics Emergence Zone (PEZ), a group of middle transformer layers in VideoMAE where physical plausibility is represented separately from other visual features. However, it remained unclear whether this structure could be used to directly control the model's physics reasoning. We present physics steering, a training-free method that uses the weight vector of a linear probe at a PEZ layer as a Concept Activation Vector (CAV) and injects it into hidden states during inference. This shifts the model's physical expectations without changing any model weights. On the IntPhys benchmark, this intervention reliably shifts the model's plausibility judgment in either direction, depending on the steering sign. The effect appears only when the intervention is applied within the Physics Emergence Zone, suggesting that the relevant physics representation is localized there. We further find that physics is encoded separately from motion direction, and that different intuitive physics principles occupy distinct directions within this representation space. Together, these results show that physical reasoning in VideoMAE is not only readable, but also directly steerable.

Large language models (LLMs) are increasingly used across diverse linguistic and cultural contexts, yet their cultural knowledge remains uneven across regions and languages. We present the DFKI-MLT system for SemEval-2026 Task 7 on cultural awareness, where we apply activation steering to multilingual LLMs using language vectors extracted from parallel FLORES data. Our method performs inference-time adaptation by adding language-specific steering vectors to the residual stream at a selected transformer layer, without any parameter updates. We participated in both the short-answer (SAQ) and multiple-choice (MCQ) tracks; however, only our MCQ submission received an official score. In the official MCQ track, we achieved 86.96% accuracy, ranking 7th out of 17 teams. To better understand system behavior, we conduct post-hoc analyses on the shared-task MCQ and SAQ settings. These analyses show that activation steering yields modest and heterogeneous improvements on cultural reasoning: gains are strongly layer-sensitive, vary substantially across language-region pairs, with some configurations even degrading performance, and interact with prompt formulation, comparing generic and culturally conditioned prompts. Our findings suggest that prompt design and activation steering should be jointly optimized for culturally aware multilingual inference.

Latent steering exploits internal representations of Large Language Models (LLMs) to guide generation, yet interventions on dense states can entangle distinct semantic features. In this paper, we investigate attention query activations as a high-fidelity site for precise control, hypothesizing that manipulating the attention mechanism itself offers sharper steerability than general state interventions. We introduce Prototype-Based Sparse Steering, a framework that applies Sparse Autoencoders (SAEs) specifically to query activations, to decompose them into interpretable features, then apply gradient-based optimization during inference to align the sparse representation with class prototypes of target behaviors. To validate this architectural insight, we first analyze the mechanism in Textualized Gridworld, a controlled environment for verifiable planning constraints. We demonstrate that optimizing sparse query features enables effective navigation of rigid planning requirements (i.e., safe vs. short paths), confirming the method's ability to satisfy objective rules. We then demonstrate the framework's versatility by training SAEs on a high-dimensional educational domain, where the framework steers the cognitive complexity of feedback (i.e., Bloom's Taxonomy). Our experiments establish that sparse query representations provide the necessary disentanglement for unified, interpretable control over both logical planning and stylistic nuance.

Large language models frequently produce errors in reasoning tasks despite possessing the underlying knowledge required for correct reasoning. One possible approach to improve reasoning consistency is through activation steering. However, existing activation steering approaches apply fixed, pre-computed correction vectors, ignoring where the model currently sits along its generation trajectory; the result is indiscriminate perturbation that disrupts already-correct steps as freely as erroneous ones. We propose Manifold-Guided Attention Steering (MAGS), a trajectory-aware inference-time intervention grounded in a geometric observation: the output activations of specific attention heads diverge from a low-dimensional correctness manifold at the point of error, and this deviation compounds through subsequent steps. For each identified attention head, we learn a low-dimensional subspace from contrastive pairs of correct and incorrect traces that capture the directions along which error behavior deviates from correct behavior. During inference, we monitor each head's proximity to this manifold and apply a targeted projection correction when deviation exceeds a learned threshold, steering the attention output back toward the correct subspace before the error propagates. MAGS consistently outperforms both unsteered baselines and static steering approaches across benchmarks spanning mathematical reasoning (MATH-500, GSM8K), code generation (HumanEval, MBPP), and molecular generation (SMILES), suggesting that correctness manifolds are a general feature of LLM attention geometry.

Feedforward steering control is a key component of hierarchical control architectures for autonomous racing. The goal is to reduce steering corrections from the feedback controllers by predicting the vehicle's inverse lateral dynamics. This paper presents a systematic benchmark of two learning-based and two empirical (analytical) feedforward steering controllers. We introduce a new \acf{ehd} formulation based on a polynomial surface fit that captures velocity-dependent nonlinear steering behavior with minimal parametrization. We test the feedforward controllers in a high-fidelity simulation framework based on the real-world Abu Dhabi Autonomous Racing League competition, using a high-fidelity double-track vehicle dynamics simulator. Open-loop evaluation shows that the learning-based controllers achieve the lowest prediction errors; however, closed-loop testing reveals that this improved accuracy does not translate into superior path tracking performance or lap times, even after iterative fine-tuning. In contrast, the proposed EHD approach achieves the best overall closed-loop robustness and lap time, highlighting the necessity of evaluating feedforward strategies within the complete trajectory planning and control software stack. Our code is available at https://github.com/TUMRT/steering_ff_control.

Generative verifiers have emerged as a promising paradigm for step-wise verification, but their verification behavior is often poorly calibrated: they may be under-critical and miss erroneous steps, or over-critical and reject correct reasoning. We refer to this tendency to be overly lenient or overly critical as verifier strictness. In this work, we study whether verifier strictness can be controlled through hidden-state intervention. We uncover a verification-specific hidden-state signal: in step-wise verification, a verifier's tendency to accept or reject a solution step is encoded near the boundary of the corresponding verification paragraph. Exploiting this signal, we show that hidden-state steering can directly modulate verifier strictness without fine-tuning. However, uniform steering induces a trade-off between error detection and correctness certification. To address this, we propose VerifySteer, which exploits latent correctness signals for sample-level routing and selectively intervenes on paragraph boundaries. Experiments on ProcessBench and Hard2Verify show that VerifySteer outperforms prompt optimization and activation steering baselines, and is competitive with self-consistency while requiring 4-7x less inference compute. VerifySteer is also complementary to verification fine-tuning, providing further gains on top of fine-tuned verifiers. The code is available at https://github.com/YefanZhou/VerifySteer.

Recent work has demonstrated the potential of contrastive steering for jailbreaking Large Language Models (LLMs). However, existing methods rely on limited and inherently biased contrastive prompts and require laborious manual tuning of steering strength, limiting their robustness and effectiveness. In this paper, we leverage the idea of model extraction to guide the learned steering vectors to approximate the ideal one and propose tuning the steering strength adaptively based on contrastive activations' statistics. Experiments demonstrate that our method notably improves the effectiveness and robustness of probe-based steering, without any extra contrastive prompts or laborious manual tuning. Being an attack paper, this paper focuses on revealing the breakdown of fortified LLMs, raising the average harmfulness score from 6% to 70%. Our code is available at https://github.com/fhdnskfbeuv/adaptiveSteering.

Different age-related regulations have been proposed to protect minors from harmful content and interactions online. Automated age estimation is central to enforcing such regulations, and vision-language models (VLMs) achieve state-of-the-art performance on this task. However, we find that the zero-shot nature of VLM-based age estimation produces an unexpected side effect we call the identity shortcut: Instead of estimating age from visual features, VLMs tend to identify the depicted person and infer their age from memorized knowledge. This phenomenon leads to substantially incorrect predictions when non-celebrities are misidentified as celebrities. It also produces deceptively high robustness to noise and adversarial perturbations on celebrity images, which dominate popular benchmarks. To mitigate this, we propose an activation steering method that suppresses the shortcut by intervening on the hidden states of the VLM. This method improves age estimation accuracy for both memorized and unseen identities, reducing mean absolute error by up to 25% across popular benchmarks.

Activation steering methods modify intermediate representations of language models to control output behavior, but universally assume the activation space is Euclidean. We show this assumption fails drastically: the local geometry induced by the model's own output behavior -- the Fisher information metric of the softmax layer, pulled back through the Jacobian of subsequent layers -- deviates from the Euclidean metric by over 97% in relative spectral norm on GPT-2, with an effective dimensionality of only 2--17% of the ambient space. From this pullback Fisher metric, we derive a closed-form steering equation that identifies the minimum-distortion direction for any target concept, yielding a closed-form optimal direction at each point that can be applied iteratively without manifold fitting or data-driven geometry estimation. We call the resulting framework FishBack. The metric admits a layer-wise recursive decomposition, which reveals that existing methods -- CAA, ActAdd, ITI, and others -- each implicitly adopt a particular approximate metric, and that their performance gaps are quantitatively predicted by a single spectral diagnostic: the ratio of their implicit metric's cost to the Fisher-optimal cost. On GPT-2, iterative pullback steering consistently outperforms all Euclidean baselines across three verb-morphology concepts and four layers, with off-target KL reductions of $1.3\times$--$2.5\times$ relative to Euclidean gradient ascent and $1.5\times$ relative to CAA at matched concept probability.

Vision-Language-Action (VLA) models leverage powerful perceptual priors from web-scale Vision-Language Model (VLM) pre-training, yet they remain surprisingly brittle in practice, frequently failing at simple robotic tasks. To mitigate this, we propose Contrastive Conceptor Activation Steering (COAST). COAST builds on the notion of a "conceptor", a linear operator that soft-projects data into the principal components of a target distribution. COAST uses conceptors to identify success-critical subspaces for a target robotic task from a few examples of success and failure rollouts. At inference time, it steers VLA latents into these identified success subspaces to improve task outcomes. Across three architecturally distinct neural policies (flow-matching VLA, autoregressive VLA, and Diffusion Policy), COAST improves absolute mean simulation and real-robot task success rate by over 20 and 40% respectively. The activation subspace geometry reveals that failure modes share substantial structure across tasks while success representations remain largely task-specific. When tasks share similar failure modes, this structure enables previously fitted conceptors to improve performance on new tasks without refitting. Ultimately, our results suggest that current VLAs retain substantial task-relevant knowledge in their latent representations, and that the action expert's decoding bottleneck could be mitigated by steering its residual stream toward task-relevant subspaces. COAST provides a lightweight, training-free path to unlocking these latent capabilities by steering the model towards its own "success" distributions.

Multimodal large language models (MLLMs) may memorize sensitive cross-modal information during pretraining, making machine unlearning (MU) crucial. Existing methods typically evaluate unlearning effectiveness based on output deviations, while overlooking the generation quality after unlearning. This can easily lead to hallucinated or rigid responses, thereby affecting the usability and safety of the unlearned model. To address this issue, we propose ASRU, a controllable multimodal unlearning framework that incorporates generation quality as a core evaluation objective. ASRU first induces initial refusal behavior through activation redirection, and then optimizes fine-grained refusal boundaries using a customized reward function, thereby achieving a better trade-off between target knowledge unlearning and model utility. Experiments on Qwen3-VL show that ASRU significantly improves unlearning effectiveness (+24.6%) on average and generation quality (5.8X) on average while effectively preserving model utility, using only a small amount of retained supervision data.

While large language models (LLMs) are trained to align with human values, their generations may still violate safety constraints. A growing line of work addresses this problem by modifying the model's sampling policy at decoding time using a safety reward. However, existing decoding-time steering methods often intervene unnecessarily, modifying generations that would have been safe under the base model. Such unnecessary interventions are undesirable, as they can distort key properties of the base model such as helpfulness, fluency, style, and coherence. We propose a new test-time steering method designed to reduce such unnecessary interventions while improving the safety of unsafe responses. Our approach filters tokens using a value-based safety criterion and provides an explicit bound on the probability of false interventions. A single threshold hyperparameter controls this bound, allowing practitioners to trade off higher rates of unnecessary intervention for better output safety. Across multiple datasets and experiments, we show that our value-filtered decoding method outperforms existing baselines, achieving better trade-offs between safety, helpfulness, and similarity to the base model.

Large Language Models memorize vast amounts of training data, raising concerns regarding privacy, copyright infringement, and safety. Machine unlearning seeks to remove the influence of a targeted forget set while preserving model performance, ideally approximating a model retrained from scratch without the forget set. Existing approaches aim to achieve this by updating model parameters via gradient-based methods. However, these updates are computationally expensive, lead to irreversible weight changes, and degrade when the model is quantized for deployment. A recent alternative to changing model weights is activation engineering, where activations are changed during inference to steer model behavior. Despite circumventing weight editing, naive activation steering introduces its own failure modes, as a single global steering vector applies the same intervention to every input, leading to unintended changes in model behavior. We introduce Inference-Time Unlearning via Gated Activation Redirection (GUARD-IT), a training- and gradient-free method that unlearns via input-dependent activation steering at inference time. The resulting intervention is applied as a norm-preserving rotation in the residual stream, leaving model weights untouched. Experiments on TOFU and MUSE show that GUARD-IT matches or exceeds 12 gradient-based baselines across three model scales, while being the only method to simultaneously preserve utility, suppress memorization, and avoid catastrophic collapse across all settings. GUARD-IT further supports continual unlearning without retraining, and remains effective under quantization, a scenario in which parameter-editing methods degrade.

Activation steering controls language model behavior by adding directions to internal representations at inference time, but standard residual-stream steering can fail in stateful dialogue. We identify KV-cache contamination as a key failure mode: steered token states are stored and repeatedly reused, turning a local perturbation into cumulative coherence degradation. To address this challenge, we propose Gated Cropped Attention-Delta steering (GCAD), which extracts steering signals from system-prompt contributions to self-attention and applies them with token-level gating. Across persona-steering experiments, GCAD preserves trait control while substantially improving long-horizon coherence. On the main multi-turn benchmark, GCAD improves average coherence drift from -18.6 to -1.9 and raises turn-10 trait expression from 78.0 to 93.1. These results suggest that activation steering becomes more reliable when interventions follow the prompt-mediated pathways that models already use for behavioral control.

Activation steering offers a lightweight way to control LLMs without retraining, but its effectiveness varies sharply across concepts. Prior work often reads this variability as evidence that many concepts are not captured by a single steering direction. We argue instead that much of it reflects search difficulty: a useful rank-1 intervention often exists, but finding it can be expensive. We formalize rank-1 steering as a budget-constrained optimization over intervention layer and coefficient. Across concepts and model families, prompt-boundary directional alignment predicts where effective interventions occur, enabling geometry-guided search that reaches high utility with substantially fewer evaluations, reducing the trials needed to recover 95% of best-found utility by 39.8% on average across three model families. To explain why some concepts remain expensive even under better search, we introduce concept granularity, a measure of directional heterogeneity across contrastive contexts. Granularity distinguishes concepts whose difference vectors share a stable global direction from those where prompts agree locally within each input but the utility-maximizing direction rotates systematically across inputs. Higher granularity is associated with slower convergence and lower best-found performance (Pearson $r{=}0.44$ with trials-to-95%, $r{=}{-}0.46$ with best-found utility, both $p<0.001$). We present GRACE, a Granularity- and Representation-Aware Concept Engineering framework that uses activation geometry to diagnose the dominant source of steering difficulty, select the appropriate remedy, and allocate optimization effort efficiently. Our results shift the frame from "when does rank-1 fail?" to "when is rank-1 cheap and stable?", turning activation geometry from a descriptive tool into an actionable prior for LLM control.

Browser agents are increasingly deployed in long-horizon tasks, which require executing extended action chains to accomplish user goals. However, this prolonged execution process provides attackers with more opportunities to inject malicious instructions. Existing prompt injection attacks against browser agents expose two key gaps: (1) low effectiveness, as attacks optimized for toy baselines fail to achieve end-to-end goals in real-world scenarios with complex environments and longer steps; (2) weak stealthiness, since most attacks pit the attack goal against the user goal, causing a significant drop in system usability under attack. To address these gaps, we propose WebTrap, a mid-task hijacking injection attack. It employs multi-step instruction fusion steering to seamlessly combine both goals, enabling the agent to resume the original user task after executing the attack goal. Furthermore, we design a context-grounded generation method to align the injected content with the task environment and system instructions, maximizing the hijacking success rate. Extensive experiments on two browser agent tasks, based on extended WASP and InjecAgent environments, demonstrate that our method achieves a high attack success rate while preserving the usability of the original system. We find that WebTrap exploits the agent's navigation vulnerabilities, binding the two goals so tightly that standard defense mechanisms cannot restore the system to normal operation. These findings reveal a critical vulnerability in agent systems during long-horizon tasks that they can be stealthily hijacked.

As foundation models grow in capability, the ability to efficiently and reliably control their behavior becomes critical. Fine-tuning these models can be costly, and while prompting can be practical for controllability, it remains fragile due to models' high sensitivity to exact prompt wording and structure. This brittleness has driven interest in activation steering techniques that offer more stable and predictable control over model behavior. However, existing activation steering methods require per-concept optimization, which makes them ill-suited to deployment scenarios where the concept set is large, evolving, or only specified at request time: each new concept incurs at least minutes of optimization on the target model. We propose HyperTransport, a hypernetwork framework that amortizes this cost by mapping embeddings from a pretrained encoder (CLIP in our instantiation) directly to intervention parameters, trained end-to-end using an optimal transport loss. Once trained, HyperTransport produces each new intervention in a single hypernetwork forward pass, 3600-7000x faster than per-concept fitting. On concepts unseen during training, it matches the strongest per-concept baselines at inducing the target concept. By decoupling concept representation from intervention prediction, HyperTransport combines three capabilities that no existing approach offers as a set: amortized steering for open-ended concept sets, continuous interpretable strength control, and cross-modal conditioning where reference images can directly steer text-based generation. We validate HyperTransport on DMD2 and Nitro-1-PixArt across 167 held-out test concepts via CLIP-based metrics, a VLM-as-a-judge evaluation, and a user study. In pairwise comparisons, both human and VLM judges prefer HyperTransport over prompting ~2x as often.

Activation steering controls LLM behaviour towards target behaviour by intervening in internal representations, yet it often degrades reasoning and retrieval performance. We argue that a primary cause of this trade-off is attention rerouting: steering vectors alter query-key matching, shifting attention away from contextually important tokens toward less informative ones. To address this, we propose Steering via Key-Orthogonal Projections (SKOP), a steering method that constrains harmful attention rerouting without eliminating steering efficacy. SKOP achieves this by preserving attention patterns on a small set of focus tokens the model relies on for reasoning and retrieval, while allowing redistribution among less critical tail tokens. Across multiple steering benchmarks, we show that SKOP achieves the best joint steering-utility trade-off, reducing utility degradation by 5-7x while retaining over 95% of vanilla steering efficacy. Our results further suggest that, in long-context retrieval settings where vanilla steering approaches are ineffective, SKOP can maintain robust performance by avoiding attention rerouting.

Steering large language models (LLMs) is usually done by either instruction prompting or activation steering. Prompting often gives strong control, but caches guidance tokens at every layer and can clutter long interactions; activation steering is compact but typically weaker and does not support large structured reminders. We introduce memory inception (MI), a training-free method that steers in latent attention space by inserting text-derived key-value (KV) banks only at selected layers. Rather than materializing reminder content throughout the prompt cache, MI treats steering as selective KV allocation, injecting latent slots only where the model routes to them. On matched personality-steering tasks, MI gives the best overall control--drift trade-off, remaining competitive with prompting while consistently outperforming CAA. On updateable guidance, MI supports mid-conversation behavior shifts without rewriting the visible transcript, achieving the highest post-shift alignment on Qwen3. On structured reasoning, MI outperforms visible prompting on HARDMath and PHYSICS (10/12 subject$\times$mode cells), serving as proxies for structured reasoning in verifiable domains, while cutting content-matched KV storage by up to 118$\times$. These results position MI as a powerful steering method when guidance is persistent, structured, or expensive to keep in the visible transcript.

Recently, steering vectors (SVs) have emerged as an effective and lightweight approach to steer behaviors of large language models (LLMs), among which fine-tuned SVs are more effective than optimization-free ones. However, current approaches to fine-tuned SVs suffer from two limitations. First, they require careful selection of steering factors on a per-SV basis to balance steering effectiveness and generation quality at inference time. Second, they operate as full-sequence SVs (FSSVs), which can sacrifice generation quality regardless of factor selection due to excessive intervention on the model generation process. To address the first limitation, we propose joint training of steering factors and directions, such that post-hoc factor selection is no longer required. Using neural network scaling theory, we find that moderately large initialization sizes and learning rates for steering factors are essential for stability and efficiency of joint training. To tackle the second limitation, we draw inspiration from representation fine-tuning and introduce Prompt-only SV (PrOSV), an SV that intervenes only on a few prompt tokens. Our empirical results show that PrOSV outperforms traditional FSSVs on AxBench when using our joint training scheme. We also find that PrOSV achieves a better tradeoff between general model utility and adversarial robustness than FSSV.

When language model agents tackle complex software engineering tasks, they often degrade over long trajectories, which we define as agent drift. We focus on two recurring failure modes overthinking and overacting, i.e., where the agent repeatedly reasons over information it already has, and where it issues tool calls without integrating recent observations or acquiring new evidence. In this paper, we introduce TACT (Think-Act Calibration via activation Steering), to detect and mitigate agent drift in the residual stream before it surfaces as a behavioral failure. In specific, we label trajectory steps as overthinking, overacting, or calibrated, and find that their hidden states can separate linearly along two drift axes, pointing from calibrated behavior toward each failure mode (AUC $\approx$ 0.9). To mitigate agent drift, we project each step's activation onto these axes at test time and pull drifted ones back toward the calibrated region. Experiments show that TACT outperforms unsteered baselines across SWE-bench Verified, Terminal-Bench 2.0, and CLAW-Eval, lifting average resolve rate by $+5.8$ pp on Qwen3.5-27B and $+4.8$ pp on Gemma-4-26B-A4B-it while cutting steps-to-resolve by up to $26\%$. These gains frame agent drift as a steerable direction in the residual stream, and position TACT as a viable handle for reliable long-horizon agents.

Activation steering has emerged as a promising alternative for controlling language-model behavior at inference time by modifying intermediate representations while keeping model parameters frozen. However, large-scale evaluations such as AxBench show that existing steering methods are often outperformed by simple in-context prompting and generalize poorly to unseen concepts. We hypothesize that these limitations arise from unvalidated simplifying assumptions shared across prior methods, which typically restrict steering interventions to fixed, single-step, position-invariant transforms. We propose FLAS (Flow-based Activation Steering), which learns a general, concept-conditioned velocity field $v_t(h,t,c)$ that transports unsteered activations to steered ones without relying on these assumptions. On AxBench, FLAS is the first learned method to consistently outperform prompting, reaching held-out harmonic means of $1.015$ on Gemma-2-2B-IT and $1.113$ on Gemma-2-9B-IT without per-concept tuning. Analysis of the learned flow shows curved, multi-step, token-varying trajectories, which suggests that previous hypotheses on activation space geometry might be incomplete.

Activation-based steering provides control of LLM behavior at inference time, but the dominant paradigm reduces each concept to a single direction whose geometry is left largely unexamined. Rather than selecting a single steering direction, we use conceptors: soft projection matrices estimated from activations pooled across both poles of a bipolar concept, which preserve the concept's full multidimensional subspace. A geometric analysis shows the bipolar subspace strictly subsumes the single-vector baseline. We further show that the conceptor quota provides a parameter-free layer-selection diagnostic, predicting concept separability with Pearson correlations up to r=0.96 across three instruction-tuned models and three semantic dimensions. Beyond selection, conceptors admit a closed-form Boolean algebra (AND, OR, NOT): we evaluate conceptor compositionality on thematically related sub-concepts. Across a systematic five-axis design-space evaluation, conceptors match or outperform additive baselines at layers where concept subspaces are multi-dimensional while producing substantially fewer degenerate outputs. Conceptor steering is a geometrically principled, compositional, and practically safer alternative to single-direction steering from a limited number of contrastive pairs.

Large language models can be steered at inference time through prompting or activation interventions, but activation steering methods often underperform compared to prompt-based approaches. We propose a framework that formulates prompt steering as a form of activation steering and investigates whether distilling successful prompt steering behavior into simpler, interpretable models can close this gap. Our analysis reveals that popular activation steering methods are not faithful to the mechanics of prompt steering, which applies strong interventions on some tokens while barely affecting others. Based on these insights, we introduce Prompt Steering Replacement (PSR) models that estimate token-specific steering coefficients from the activations themselves and are trained to imitate prompt-based interventions. Experiments on three steering benchmarks across multiple language models show that PSR models outperform existing activation steering methods, especially when controlling for high-coherence completions, and also compare favorably to prompting on AxBench and persona steering.

Steering is a widely used technique for controlling large language models, yet its effects are often unstable and hard to predict. Existing theoretical accounts are largely based on the Linear Representation Hypothesis (LRH). While LRH assumes that concepts can be orthogonalized for lossless control, this idealized mapping fails in real representations and cannot account for the observed unpredictability of steering. By relaxing LRH's orthogonality assumption while preserving linear representations, we show that overlapping concept contributions naturally yield a sample-specific axis-orthogonal structure. We formalize this as the Cylindrical Representation Hypothesis (CRH). In CRH, a central axis captures the main difference between concept absence and presence and drives concept generation. A surrounding normal plane controls steering sensitivity by determining how easily the axis can activate the target concept. Within this plane, only specific sensitive sectors strongly facilitate concept activation, while other sectors can suppress or delay it. While the surrounding normal plane can be reliably identified from difference vectors, the sensitive sector cannot, introducing intrinsic uncertainty at the sector level. This uncertainty provides a principled explanation for why steering outcomes often fluctuate even when using well-aligned directions. Our experiments verify the existence of the cylindrical structure and demonstrate that CRH provides a valid and practical way to interpret model steering behavior in real settings: https://github.com/mbzuai-nlp/CRH.

Activation steering is a method for controlling Large Language Model (LLM) behavior by intervening in its internal representations to increase the alignment with a specific target feature direction. However, standard interventions, such as vector addition, often cause ``collateral damage", defined as unintended changes in the alignment of activations along other non-target feature directions. This damage occurs because standard methods implicitly assume the isotropy of non-target features. In this work, we provide a mathematical formalization of collateral damage and introduce a principled framework that models steering as a constrained optimization problem. Our method finds a new activation that minimizes the expected squared collateral change weighted by the empirical second-moment matrix of activations. This weighting encodes the nonuniform cost of the perturbation in different feature directions, in contrast to isotropic approaches that penalize changes uniformly in all feature directions. By accounting for the empirical second-moment of activations, our approach achieves more precise control while reducing the degradation of model performance on unrelated tasks.

Linear activation steering is a powerful approach for eliciting the capabilities of large language models and specializing their behavior using limited labeled data. While effective, existing methods often apply a fixed steering strength to all tokens, resulting in inconsistent steering quality across diverse input prompts. In this work, we introduce Contextual Linear Activation Steering (CLAS), a method that dynamically adapts linear activation steering to context-dependent steering strengths. Across eleven steering benchmarks and four model families, it consistently outperforms standard linear activation steering and matches or exceeds the performance of ReFT and LoRA in settings with limited labeled data. We therefore propose CLAS as a scalable, interpretable, and accurate method for specializing and steering large language models.

Thinking LLMs produce reasoning traces before answering. Prior activation steering work mainly targets on shaping these traces. It remains less understood how answer tokens actually read and integrate the reasoning to produce reliable outcomes. Focusing on quantitative reasoning, we analyze the answer-to-reasoning attention and observe a benign self-reading pattern aligned with correctness, characterized by a forward drift of the reading focus along the reasoning trace and a persistent concentration on key semantic anchors, whereas incorrect solutions exhibit diffuse and irregular attention pattern. We interpret this as internal certainty during answer decoding, where the model commits to a viable solution branch and integrates key evidence. Following this, we propose a training-free steering method driven by Self-Reading Quality (SRQ) scores combining geometric metrics for process control with semantic metrics for content monitoring. SRQ selects data to build steering vectors that guide inference toward benign self-reading and away from uncertain and disorganized reading. Experiments show that our method yields consistent accuracy gains.

Inference-time LLM alignment methods, particularly activation steering, offer an alternative to fine-tuning by directly modifying activations during generation. Existing methods, however, often rely on non-anticipative interventions that ignore how perturbations propagate through transformer layers and lack online error feedback, resulting in suboptimal, open-loop control. To address this, we show empirically that, despite the nonlinear structure of transformer blocks, layer-wise dynamics across multiple LLM architectures and scales are well-approximated by locally-linear models. Exploiting this property, we model LLM inference as a linear time-varying dynamical system and adapt the classical linear quadratic regulator to compute feedback controllers using layer-wise Jacobians, steering activations toward desired semantic setpoints in closed-loop with minimal computational overhead and no offline training. We also derive theoretical bounds on setpoint tracking error, enabling formal guarantees on steering performance. Using a novel adaptive semantic feature setpoint signal, our method yields robust, fine-grained behavior control across models, scales, and tasks, including state-of-the-art modulation of toxicity, truthfulness, refusal, and arbitrary concepts, surpassing baseline steering methods. Our code is available at: https://github.com/trustworthyrobotics/lqr-activation-steering

Steering vectors work for some concepts and layers but fail for others, and practitioners have no way to predict which setting applies before running an intervention. We introduce the Linear Accessibility Profile (LAP), a per-layer diagnostic that repurposes the logit lens as a predictor of steering vector effectiveness. The key measure, $A_{\mathrm{lin}}$, applies the model's unembedding matrix to intermediate hidden states, requiring no training. Across 24 controlled binary concept families on five models (Pythia-2.8B to Llama-8B), peak $A_{\mathrm{lin}}$ predicts steering effectiveness at $ρ= +0.86$ to $+0.91$ and layer selection at $ρ= +0.63$ to $+0.92$. A three-regime framework explains when difference-of-means steering works, when nonlinear methods are needed, and when no method can work. An entity-steering demo confirms the prediction end-to-end: steering at the LAP-recommended layer redirects completions on Gemma-2-2B and OLMo-2-1B-Instruct, while the middle layer (the standard heuristic) has no effect on either model.

Applying steering vectors to large language models (LLMs) is an efficient and effective model alignment technique, but we lack an interpretable explanation for how it works--specifically, what internal mechanisms steering vectors affect and how this results in different model outputs. To investigate the causal mechanisms underlying the effectiveness of steering vectors, we conduct a comprehensive case study on refusal. We propose a multi-token activation patching framework and discover that different steering methodologies leverage functionally interchangeable circuits when applied at the same layer. These circuits reveal that steering vectors primarily interact with the attention mechanism through the OV circuit while largely ignoring the QK circuit. Freezing all attention scores during steering drops performance by only 8.83% across three model families. A mathematical decomposition of the steered OV circuit further reveals semantically interpretable concepts, even in cases where the steering vector itself does not. Leveraging the activation patching results, we show that steering vectors can be sparsified by up to 85-96% while retaining most performance, and that different steering methodologies agree on a subset of important dimensions.

Alignment in LLMs is more brittle than commonly assumed: misalignment can be induced by adversarial prompts, benign fine-tuning, emergent misalignment, and goal misgeneralization. Recent evidence suggests that some misalignment behaviors are encoded as linear structure in activation space, making it tractable via activation steering, which could be used as a lightweight runtime defense. We implement three methods: Steer-With-Fixed-Coefficient (SwFC), which applies uniform additive steering, and two novel projection-aware methods, Steer-to-Target-Projection (StTP) and Steer-to-Mirror-Projection (StMP), that use a logistic regression decision boundary to selectively intervene only on tokens whose activations fall below the threshold. We evaluate these methods on two threat models, dishonesty and dismissiveness, using malicious system prompts as a controlled proxy for misalignment. We conduct our experiments on two architectures (Llama-3.3-70B-Instruct and Qwen3.6-27B). All methods substantially recover alignment. StTP and StMP preserve general capabilities (MMLU, MT-Bench, AlpacaEval) better than uniform steering. Finally, we show that our honesty steering generalizes to out-of-distribution scenarios: a single honesty direction extracted from the aligned model significantly raises scores on the MASK benchmark, suppresses deception in multi-agent settings (Among Us), doubles the hidden-behavior discovery rate on AuditBench, and restores honesty in an emergently misaligned model

Activation-based steering enables inference-time personalization of large language models, but its effects in educational applications are not well understood. We study activation-based persona vectors representing seven character traits in short-answer generation and automated scoring on the ASAP-SAS benchmark, across three language models spanning dense and mixture-of-experts architectures. Persona steering lowers answer quality overall, with much larger effects on open-ended English Language Arts (ELA) prompts than on factual science prompts. Interpretive and argumentative tasks are particularly sensitive, showing up to 11$\times$ larger degradation. On the scoring side, we observe predictable valence-aligned calibration shifts: evil'' and impolite'' scorers grade more harshly, while good'' and optimistic'' scorers grade more leniently. ELA tasks are 2.5-3$\times$ more susceptible to scorer personalization than science tasks, and the mixture-of-experts model shows roughly 6$\times$ larger calibration shifts than the dense models. To our knowledge, this is the first study to systematically examine the effects of activation-steered persona traits in educational generation and scoring. Our findings highlight the need for task- and architecture-aware calibration when deploying personalized models in educational settings.

Language models are commonly fine-tuned for safety alignment to refuse harmful prompts. One approach fine-tunes them to generate categorical refusal tokens that distinguish different refusal types before responding. In this work, we leverage a version of Llama 3 8B fine-tuned with these categorical refusal tokens to enable inference-time control over fine-grained refusal behavior, improving both safety and reliability. We show that refusal token fine-tuning induces separable, category-aligned directions in the residual stream, which we extract and use to construct categorical steering vectors with a lightweight probe that determines whether to steer toward or away from refusal during inference. In addition, we introduce a learned low-rank combination that mixes these category directions in a whitened, orthonormal steering basis, resulting in a single controllable intervention under activation-space anisotropy, and show that this intervention is transferable across same-architecture model variants without additional training. Across benchmarks, both categorical steering vectors and the low-rank combination consistently reduce over-refusals on benign prompts while increasing refusal rates on harmful prompts, highlighting their utility for multi-category refusal control.

Jailbreak attacks remain a persistent threat to large language model safety. We propose Context-Conditioned Delta Steering (CC-Delta), an SAE-based defense that identifies jailbreak-relevant sparse features by comparing token-level representations of the same harmful request with and without jailbreak context. Using paired harmful/jailbreak prompts, CC-Delta selects features via statistical testing and applies inference-time mean-shift steering in SAE latent space. Across four aligned instruction-tuned models and twelve jailbreak attacks, CC-Delta achieves comparable or better safety-utility tradeoffs than baseline defenses operating in dense latent space. In particular, our method clearly outperforms dense mean-shift steering on all four models, and particularly against out-of-distribution attacks, showing that steering in sparse SAE feature space offers advantages over steering in dense activation space for jailbreak mitigation. Our results suggest off-the-shelf SAEs trained for interpretability can be repurposed as practical jailbreak defenses without task-specific training.

Steering diffusion models toward conditions unseen during training typically requires either retraining with conditional inputs or per-step gradient computations, both of which incur substantial computational overhead. We present Noise-Aligned RFM Steering (NA-RFM), a general recipe for efficiently steering diffusion models without gradient guidance during inference, enabling fast controllable generation. The method combines two offline-computed signals: noise alignment, a high-noise correction from PCA statistics of the target examples and the full data, and Recursive Feature Machine (RFM) activation steering, which learns a target-discriminative direction from labeled forward-process activations. During sampling, noise alignment provides coarse control at high noise, while the RFM direction is reused over intermediate/late timesteps through lightweight activation edits. Experiments on CIFAR-10, ImageNet, CelebA, and fine-grained bird species show improved target accuracy over gradient-based post-hoc guidance baselines, improved FID on the class-guidance benchmarks, and substantial inference speedups. Code: https://github.com/isotrivial/na-rfm.

Large language models can recover mid-generation from task-misaligned activation steering, producing explicit verbal restarts (e.g., ``wait, that's not right'') and continuing on-topic even while the steering perturbation remains active. We term this Endogenous Steering Resistance (ESR). Using sparse autoencoder (SAE) latents to steer model activations, we find that Llama-3.3-70B exhibits explicit ESR at 3.8%, with smaller models from the Llama-3 and Gemma-2 families showing the explicit form less frequently. Two controls dissociate ESR into a detection event and a sustained-resistance component that conditioning on recent on-topic tokens does not fully explain. We identify 26 SAE latents through contrastive on-topic/off-topic search; zero-ablating them reduces the multi-attempt rate by 25%, with random-latent and held-out-prompt controls supporting specificity. ESR can also be deliberately enhanced through both meta-prompting and fine-tuning on synthetic self-correction examples. ESR has dual implications for safety: it could harden models against adversarial activation-space manipulation, but may equally interfere with beneficial steering-based interventions, since the model has no way to distinguish the two. Code is available at https://github.com/agencyenterprise/endogenous-steering-resistance.

Modern deployments require LLMs to enforce safety policies at scale, yet many controls rely on inference-time interventions that add recurring compute cost and serving complexity. Activation steering is widely used, but it requires runtime hooks and scales cost with the number of generations; conditional variants improve selectivity by gating when steering is applied but still retain an inference-time control path. We ask whether selective refusal can be moved entirely offline: can a mechanistic understanding of category-specific refusal be distilled into a circuit-restricted weight update that deploys as a standard checkpoint? We propose C-Δθ Circuit Restricted Weight Arithmetic}, which (i) localizes refusal-causal computation as a sparse circuit using EAP-IG and (ii) computes a constrained weight update ΔθC supported only on that circuit (typically <5% of parameters). Applying ΔθC yields a drop-in edited checkpoint with no inference-time hooks, shifting cost from per request intervention to a one-time offline update. We evaluate category-targeted selectivity and capability retention on refusal and utility benchmarks.

A popular approach to post-training control of large language models (LLMs) is the steering of intermediate latent representations. Namely, identify a well-chosen direction depending on the task at hand and perturbs representations along this direction at inference time. While many propositions exist to pick this direction, considerably less is understood about how to choose the magnitude of the move, whereas its importance is clear: too little and the intended behavior does not emerge, too much and the model's performance degrades beyond repair. In this work, we propose the first theoretical analysis of steering strength. We characterize its effect on next token probability, presence of a concept, and cross-entropy, deriving precise qualitative laws governing these quantities. Our analysis reveals surprising behaviors, including non-monotonic effects of steering strength. We validate our theoretical predictions empirically on eleven language models, ranging from a small GPT architecture to modern models.

Large Audio-Language Models (LALMs) excel in Audio QA but often suffer from hallucinations ungrounded in the audio. To our knowledge, we are the first to propose applying vector steering to the audio domain to mitigate this. Unlike text-based steering, our silence-anchored contrastive approach steers the model away from hallucinations by contrasting active audio against a silent baseline. Probing internal states reveals a strong correlation between specific layer representations and output correctness. Leveraging this, we introduce Layer-Weighted Vector Steering (LWVS), a training-free intervention that increases steering strength at influential layers. On the Audio Hallucination QA dataset, LWVS significantly outperforms baselines, boosting Recall on the Gemma model by 15.6% (53.4% to 69.0%). Crucially, MMAU benchmark tests confirm LWVS preserves and even enhances general audio understanding, achieving an 8% relative accuracy increase on the Qwen model (54.8% to 59.2%).

As Vision Language Models (VLMs) are deployed across safety-critical applications, understanding and controlling their behavioral patterns has become increasingly important. Existing behavioral control methods face significant limitations: system prompting approaches could easily be overridden by user instructions, while applying activation-based steering vectors requires invasive runtime access to model internals, precluding deployment with API-based services and closed-source models. Finding steering methods that transfer across multiple VLMs is still an open area of research. To this end, we introduce universal visual input based steering for output redirection (VISOR++), to achieve behavioral control through optimized visual inputs alone. We demonstrate that a single VISOR++ image can be generated for an ensemble of VLMs to emulate each of their steering vectors. By crafting universal visual inputs that induce target activation patterns, VISOR++ eliminates the need for runtime model access while remaining deployment-agnostic. This means that when an underlying model supports multimodal capability, model behaviors can be steered by inserting an image input replacing runtime steering vector based interventions. We first demonstrate the effectiveness of the VISOR++ images on open-access models such as LLaVA-1.5-7B and IDEFICS2-8B along three alignment directions: refusal, sycophancy and survival instinct. Both the model-specific steering images and the jointly optimized images achieve performance parity closely following that of steering vectors for both positive and negative steering tasks. We also show the promise of VISOR++ images in achieving directional behavioral shifts for unseen models including both open-access and closed-access ones. Furthermore, VISOR++ images are able to preserve 99.9% performance on 14,000 unrelated MMLU evaluation tasks.

Vision Language Models (VLMs) are increasingly being used in a broad range of applications, bringing their security and behavioral control to the forefront. While existing approaches for behavioral control or output redirection, like system prompting in VLMs, are easily detectable and often ineffective, activation-based steering vectors require invasive runtime access to model internals--incompatible with API-based services and closed-source deployments. We introduce VISOR (Visual Input-based Steering for Output Redirection), a novel method that achieves sophisticated behavioral control through optimized visual inputs alone. By crafting universal steering images that induce target activation patterns, VISOR enables practical deployment across all VLM serving modalities while remaining imperceptible compared to explicit textual instructions. We validate VISOR on LLaVA-1.5-7B across three critical alignment tasks: refusal, sycophancy and survival instinct. A single 150KB steering image matches steering vector performance within 1-2% for positive behavioral shifts while dramatically exceeding it for negative steering--achieving up to 25% shifts from baseline compared to steering vectors' modest changes. Unlike system prompting (3-4% shifts), VISOR provides robust bidirectional control while maintaining 99.9% performance on 14,000 unrelated MMLU tasks. Beyond eliminating runtime overhead and model access requirements, VISOR exposes a critical security vulnerability: adversaries can achieve sophisticated behavioral manipulation through visual channels alone, bypassing text-based defenses. Our work fundamentally re-imagines multimodal model control and highlights the urgent need for defenses against visual steering attacks.

Traditional topic models are effective at uncovering latent themes in large text collections. However, due to their reliance on bag-of-words representations, they struggle to capture semantically abstract features. While some neural variants use richer representations, they are similarly constrained by expressing topics as word lists, which limits their ability to articulate complex topics. We introduce Mechanistic Topic Models (MTMs), a class of topic models that operate on interpretable features learned by sparse autoencoders (SAEs). By defining topics over this semantically rich space, MTMs can reveal deeper conceptual themes with expressive feature descriptions. Moreover, uniquely among topic models, MTMs enable controllable text generation using topic steering vectors. To properly evaluate MTM topics against word list approaches, we propose \textit{topic judge}, an LLM-based pairwise comparison evaluation framework. Across eight datasets, MTMs match or exceed traditional and neural baselines on coherence metrics, are consistently preferred by topic judge, and enable effective LLM steering.

Inference-time steering aims to alter a large language model's (LLM's) responses without changing its parameters, but a central challenge is identifying the internal modules that most strongly govern the target behavior. Existing approaches often rely on simplistic cues or ad hoc heuristics, leading to suboptimal or unintended effects. We introduce REAL, a framework for identifying behavior-relevant modules (attention heads or layers) in Transformer models. For each module, REAL trains a vector-quantized autoencoder (VQ-AE) on its hidden activations and uses a shared, learnable codebook to partition the latent space into behavior-relevant and behavior-irrelevant subspaces. REAL quantifies a module's behavioral relevance by how well its VQ-AE encodings discriminate behavior-aligned from behavior-violating responses via a binary classification metric; this score guides both module selection and steering strength. We evaluate REAL across eight LLMs from the Llama and Qwen families and nine datasets spanning truthfulness enhancement, open-domain QA under knowledge conflicts, and general alignment tasks. REAL enables more effective inference-time interventions, achieving an average relative improvement of 20% (up to 81.5%) over the ITI method on truthfulness steering. In addition, the modules selected by REAL exhibit strong zero-shot generalization in cross-domain truthfulness-steering scenarios.