Enhancing SAE-based Steering via Neighbor Integrated Feature Selection

Sparse autoencoders (SAEs) disentangle model activations into interpretable features and are widely used for steering large language models. Most existing SAE-based steering methods select features by applying a top- filter based on statistical scores, assuming that higher-scoring features yield stronger steering effects. In this paper, we show that this assumption is often invalid, leading to suboptimal feature selection. Our analysis reveals that effective steering features may be distributed among representationally adjacent, semantically similar groups induced by feature splitting in SAEs. Within such groups, features may exhibit disparate statistical scores despite having comparable steering influence, causing score-based selection to overlook important features. Based on these observations, we propose \textsc{Neighbor Integrated Feature Selection} (\textsc{NIFS}), a plug-and-play strategy that leverages representation similarity to improve feature selection for steering. We evaluate \textsc{NIFS} across multiple SAE-based steering methods and tasks, and demonstrate consistent performance gains over conventional top-$k$ selection.

RARE: Decoupling Representation Steering from Expert Routing in Mixture-of-Experts Language Models

Representation engineering offers a lightweight means of controlling language-model behavior by modifying intermediate hidden states, but its direct application to Mixture-of-Experts (MoE) models introduces a structural mismatch. We first verify this failure mode through a series of empirical studies and find that preserving clean routing substantially recovers steering performance and that routing is more sensitive to semantic content than to behavioral changes under controlled content. Motivated by these findings, we introduce RARE, a router-agnostic representation engineering framework for MoE language models. RARE projects arbitrary behavioral perturbations onto the null space of the router matrix, thereby removing router-visible components, and further corrects routing drift propagated to selected downstream layers. To decide the best perturbation estimator in this framework, we evaluate five estimators on six heterogeneous open-weight MoE models across three steering scenarios: harmfulness, truthfulness, and factual editing. On harmfulness steering, RARE reaches an average attack success rate of 53.3% while retaining 67.8% MMLU accuracy, yielding a stronger aggregate effectiveness--utility trade-off than baselines. It further improves average TruthfulQA MC1 accuracy from 41.0% to 58.6% and CounterFact efficacy from 16.8% to 96.3%. These results support routing consistency as an important architectural consideration for adapting representation engineering to MoE models.

Sycophancy Undermines Epistemic Vigilance in Cooperative Vision-Language Tasks

To maintain common ground in cooperative conversation, humans iteratively update their beliefs as conversation participants share new information; participants who are epistemically vigilant detect when new information conflicts with prior beliefs and take steps to repair these conflicts. In order for AI systems to serve as reliable partners in complex cooperative tasks, they must similarly weigh incoming information against their own private evidence and shared context and appropriately surface inconsistencies when they arise. To measure the epistemic vigilance of vision-language models in cooperative settings, we present an information-asymmetric, dialog-based "spot-the-difference" task. Two models are privately shown one image each, and must determine through conversation whether the images are identical or, if not, identify the difference. Models routinely fail at this: they frequently overlook key evidence in their private image in favor of agreeing with their conversational partner, even when their agreement is unwarranted. We relate these violations of epistemic vigilance to the broader behavior of sycophancy, which manifests itself in cooperative goal-oriented dialog as over-accommodation and weak evidential grounding. Our results show that model steering to reduce sycophancy with a vector learned from task-agnostic sycophancy examples can reduce epistemic vigilance-related errors, making models more faithful reporters of their evidence, and in turn, more reliable partners in information-asymmetric cooperative tasks.

Distributed Sparse Interventions in Language Models

Language models perform a wide range of tasks at varying levels of abstraction with the capacity to flexibly infer tasks from context, execute multiple tasks simultaneously, and select among competing tasks. To study the role of model components in task behaviour, their causal influence can be investigated through interventions. Prior work on model steering has largely focused on interventions along global directions in activation space, modeling task representations as approximately linear and additive. By studying interventions at the neuron level, we find substantial, neuron-specific nonlinear effects on model outputs that are not captured by current steering approaches. We introduce Distributed Sparse Interventions (DSI), an intervention approach that considers nonlinearities and interactions between neurons across layers to identify sparse sets of neurons that elicit task-relevant computations. Across a range of tasks, we demonstrate that DSI can activate task behaviour in instruction-tuned language models by localising and intervening on as few as 0.01% of neurons, highlighting the effectiveness of sparse, distributed interventions in the neuron basis. Additionally, adopting a set-based perspective enables computations over the identified neuron sets, offering insights into the roles of individual neurons by analysing their effects across tasks. Through sparse interventions, DSI enables fine-grained control over model behaviour, localisation of task-relevant neuron sets, and furthers our understanding of task composition.

Gaze Heads: How VLMs Look at What They Describe

How a vision-language model internally solves the task of describing an image is far from obvious. We find that the model develops a specific mechanism for this: a small set of attention heads in its language-model backbone, which we call gaze heads, whose attention tracks the image region the model is currently describing. We find them with a simple correlation score from a few forward passes, using comic strips as a controlled testbed where narrative order is laid out spatially. These gaze heads do not just track the image tokens being described: redirecting their attention to a chosen region forces the VLM to describe that region instead. A single attention-mask intervention on the top-100 gaze heads, fewer than 9% of all heads, steers the model's answer to any chosen comic panel at 83.1% accuracy, while the same intervention on random heads fails to redirect the answer, and intervening on all heads destroys generation. The same lever also extends to continuous control: switching the gaze target mid-generation makes the model wrap up its current panel description and move to the new one within a few tokens. Beyond comics, the same intervention redirects answers to chosen regions in natural COCO images. The mechanism further recurs across model sizes from 2B to 32B parameters and across other VLM architectures, although some frozen-encoder families show no comparable head set. More broadly, this shows that targeted edits identified through mechanistic analysis can serve as practical inference-time levers for steering multimodal model behavior, without any retraining. Our code, interactive demo, and datasets are available at https://gaze.baulab.info/

Pre-Intervention Prediction of Sparse Autoencoder Steering Side Effects

Sparse autoencoder (SAE) features are increasingly used to steer language models, but feature steering is rarely clean: the same intervention can behave inconsistently across contexts and perturb unrelated features. We introduce a pre-intervention screening framework for forecasting SAE steering side effects from feature statistics computed before steering. We operationalize side effects along two axes of steering modularity, effect stability and collateral spread, and evaluate GPT-2-small, Pythia-70M-deduped, Gemma-2-2B, and Llama-3.1-8B across ReLU, JumpReLU, and TopK SAE dictionaries. Across these settings, decoder geometry, activation statistics, co-activation structure, and direct-logit footprint predict steering modularity better than frequency-only and activation-magnitude baselines. The signal is strongest in GPT-2-small, Pythia-70M, and Llama-3.1-8B, where it survives residualization against magnitude-related confounds, and weaker in Gemma-2-2B. Held-out screening shows that ranking unseen features by predicted cleanliness can select features that steer more cleanly on fresh contexts, but the successful axis varies by setting: GPT-2 improves most cleanly, Pythia improves mainly on stability, Llama mainly on collateral, and Gemma only partially. A controlled Llama Scope width comparison shows that the predictive signal persists under a 32K-to-128K dictionary-width change, although the screening payoff becomes less stable. Overall, SAE steering side effects are predictable in advance, but the useful predictor signature and transferred modularity axis are model- and dictionary-setting dependent.

Fast & Faithful Function Vectors

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.