Steering

Ill-posed questions, including ambiguous, underspecified, or contradictory queries, may admit no valid answer or multiple plausible answers, posing a challenge for large language models (LLMs). Existing approaches largely analyze ill-posedness through model outputs and often focus on specific subclasses. We investigate whether diverse sources of ill-posedness can be represented within a unified topology of LLM internal states and whether this structure can be used to steer response behavior. We model the contextual hidden states of prompt tokens at each transformer layer as a point cloud and characterize its geometry using finite zero-dimensional persistent homology. Each layer is summarized by three compact descriptors: mean finite lifetime, normalized lifetime entropy, and largest-lifetime concentration. Concatenating these descriptors across layers yields a topology representation of the question. We further introduce topology-conditioned activation steering, which retrieves topologically similar examples and constructs query-specific activation interventions that encourage source-aware clarification or abstention. Across three open-weight LLMs, topology features consistently outperform prompt-based and pooled-hidden-state baselines for ill-posedness classification, improving average accuracy from $67.4\%$ to $78.9\%$ on AmbigQA, from $79.9\%$ to $88.5\%$ on SituatedQA, and from $57.6\%$ to $69.6\%$ on CLAMBER 9-way classification. Topology-conditioned steering increases the average total acceptable response rate from $61.4\%$ to $70.6\%$ and grounded acceptable responses from $11.9\%$ to $16.4\%$. These results show that persistent homology provides both an interpretable representation of ill-posedness and an effective mechanism for targeted response steering.

Visual text editing aims to precisely modify text in images and videos while preserving stylistic consistency and visual realism. Despite significant advances in the image domain, video text editing remains largely unexplored: it is a localized task demanding stroke-level precision within small text regions, which compounds the challenges of cross-frame accuracy, temporal coherence, and stylistic fidelity. We introduce SteerVTE, a unified framework that \underline{\textbf{steer}}s a frozen video diffusion model to perform precise \underline{\textbf{V}}ideo \underline{\textbf{T}}ext \underline{\textbf{E}}diting through style and glyph control. Built on a frozen diffusion transformer, SteerVTE attaches a lightweight text context adapter with two complementary modules: a style encoder capturing the original text's visual attributes, and dual-granularity glyph encoders encoding the target text at both the line and character levels. To overcome the inherently weak text rendering priors of video foundation models, we further propose a glyph-aware spatial-focal loss and a three-stage progressive training curriculum that scales from image to video data. To support large-scale training, we also develop an automatic synthesis pipeline and construct SteerVTE-1M, a dataset of one million triplets spanning diverse scenes, fonts, and stylistic effects. Extensive experiments demonstrate that SteerVTE substantially outperforms existing video editing baselines across text accuracy, style consistency, and temporal coherence.

Steerability is a defining capability of generalist robot policies, yet remains largely absent in dexterous-hand systems for lack of large-scale, language-aligned, and action-accurate demonstration data. To address this bottleneck, we present a full-stack system that scales dexterous VLA pre-training from egocentric human videos and enables data-efficient real-robot post-training. It integrates EgoSmith, a data pipeline that curates in-the-wild egocentric videos into 9.6K hours of high-quality pre-training data with 9x higher throughput and better accuracy than prior SOTA; a unified robot stack for teleoperation and human-in-the-loop correction; and EgoSteer, a world-model-enhanced VLA trained on optimized infrastructure. Human-data pre-training equips EgoSteer with language-guided manipulation priors, which are grounded through robot post-training and improved by DAgger refinement. Empirically, EgoSteer robustly executes free-form instructions across 40+ diverse tasks, demonstrating failure recovery, dexterity, and generalization. The pre-trained model also few-shot adapts to complex long-horizon tasks, including box folding, on two embodiments with 75+% success. We open-source the system, data, and model at https://egosteer.github.io/.

Language models are widely used in assistant settings, where controlling behavioral attributes is often essential. Activation steering modifies hidden-state representations at inference time, providing a lightweight, training-free mechanism that can be toggled at runtime. Existing methods, however, have focused primarily on steering a single attribute at a time. When multiple attributes must be controlled simultaneously, naive summation of per-attribute steering vectors suffers from norm imbalance and directional cancellation, while classifier-based approaches require retraining whenever the attribute set changes. We introduce ORBIT (Orthogonal Rotation-Based Intervention Technique), a training-free extension of rotation-based steering to the multi-attribute setting. Our method constructs a joint subspace from per-attribute steering planes via singular value decomposition and applies a single norm-preserving rotation within that subspace toward a combined target direction. Adaptive per-token gating identifies which attributes need correction at each position, and an optional additive boost strengthens attributes with weak initial projection. We also introduce TraitFactory, a new multi-attribute benchmark that focuses on behavioral tendencies rather than surface-level style. We evaluate ORBIT on TraitFactory and ToneBank across three models (Llama-3.2-3B, Qwen-2.5-7B, Llama-3.1-8B) while steering multiple attributes simultaneously, showing that it achieves stronger and more balanced multi-attribute steering than existing training-free baselines while better preserving output coherence.

Coding tasks are typically complicated and require multiple capabilities, ranging from high-level planning to low-level implementation. While coding agents are optimized for the joint capabilities, individual capabilities such as high-level planning may have different optima and remain a major bottleneck. To address this challenge, we train a separate critic model that is specialized in high-level planning to steer the coding agent in inference. We construct SFT and DPO data to train the critic model to identify errors made by the coding agent and provide correct and clear high-level guidance without generating concrete actions. Experiments show that our fine-tuned 4B and 8B critic models significantly improve the performance of 6 larger coding agents (e.g., improving the resolved rates of GLM-4.7-Flash-30B-A3B and GPT-OSS-120B by 16.0% and 14.4% on SWE-Bench Verified). The critic model also reduces the total inference costs for some coding agents by solving tasks in fewer steps (e.g., reducing the per-example inference cost for GPT-OSS-20B from $0.07 to $0.03). Code: https://github.com/shubhamrgandhi/critic-training

Multi-task robot manipulation policies are challenging to learn from demonstration because traditionally a single network must select among qualitatively different action modes from a multimodal demonstration distribution, conditioned on language and visual context. A wrong mode selection means executing the wrong task or an action infeasible in the scene. Tokenizing continuous actions into a learned discrete codebook separates these modes at the representation level, offering structural advantages for multi-task learning. We propose VQActFlow, a multi-task manipulation policy that tokenizes action chunks and generates code sequences via Variational Flow Matching. VQActFlow maintains an explicit preference over action modes throughout generation. Inference-time guidance acts on this preference to steer mode commitment. We instantiate this with classifier-free guidance over language conditioning, which steers the policy toward the instructed action mode, and a learned codebook critic that supplies a complementary feasibility signal. We evaluate VQActFlow on three platforms: the LIBERO simulation benchmarks, a Unitree G1 humanoid performing whole-body pick-and-place, and an ALOHA-style bimanual platform performing contact-rich tasks. Across these benchmarks, VQActFlow outperforms both continuous and discrete baselines.

Large language models can store both outdated facts and newer superseding facts in their parameters, but standard prompting may still elicit the outdated answer. We formalize this problem as Parametric Temporal Conflict (PTC) and introduce Temporal Attractor Steering (TAS), a three-stage test-time intervention that detects likely conflicts, identifies a conflict-critical layer, and steers hidden states toward newer-fact representations without retraining or external retrieval. We construct an 8,746-record verified benchmark across five Wikidata relations and evaluate four open-weight language models from three families: Qwen-2.5-1.5B/7B, Mistral-7B-v0.3, and Llama-3.1-8B. Single-layer activation patching achieves answer-flip rates of 0.72-0.85 across all models. End-to-end TAS resolves 29-57% of PTC cases while preserving 85-99% accuracy on non-conflict queries, outperforming a matched ITI baseline on three of four models. These results show that outdated parametric knowledge can be selectively overridden at inference time.

Vision-language models (VLMs) often generate fluent but visually unsupported descriptions, especially by mentioning objects absent from the image. We propose QK Product Steering, a data-free, training-free, and zero-inference-cost weight edit for reducing object hallucination. The method directly edits the per-head query-key product, the operator that produces pre-softmax attention logits, by suppressing a small number of dominant singular modes in selected middle layers. The edited product is then mapped back to the query weights through a closed-form query-only update while keeping shared key weights fixed, making the edit compatible with grouped-query attention. We further decompose the QK product into symmetric and antisymmetric components to distinguish mutual content-similarity patterns from directional attention patterns. Across three GQA-based VLMs, QK Product Steering achieves an average relative CHAIR$_s$ reduction of $4.0\%$, while matched random-mode controls show negligible change. Interpretability ablations show that the hallucination signal is specific to dominant QK modes and is primarily localized to the symmetric mutual-attention channel. Overall, QK Product Steering offers a simple alternative to decoding-time mitigation, requiring no additional data, fine-tuning, or inference-time overhead while largely preserving general multimodal capability.

Activation steering controls model behavior by modifying intermediate hidden states at inference time without retraining. Existing methods handle only single-direction injection; when multiple semantic directions are superposed without constraints, the model collapses. We show that this collapse decomposes into two independently acting sources: distributional deviation, where additive perturbations accumulate in norm across layers and drive activations outside the training distribution, and directional interference, where non-orthogonal semantic vectors mutually dampen when superposed. These two sources define the design constraints that any training-free multi-directional intervention must address. As one instantiation of these principles, we propose GEMS, a training-free method that maps each source to a corresponding geometric constraint: norm-preserving weighted superposition and targeted attention-pathway injection for distributional deviation, and real-time orthogonalization for directional interference. On GSM8K, injecting three concurrent non-mathematical directions preserves accuracy at 98% (baseline 92%), while unconstrained addition collapses to 4%; on Wikitext-2, the same injection incurs only 2.2% PPL increase. Component ablation isolates the causal role of each constraint, and layer-level probes confirm that orthogonalized signals survive the FFN pathway and reach the output distribution with semantic specificity. Qualitative steering effects transfer across architectures from 3B to 31B.

Recovering complete 3D representations of objects from few casual image captures remains a significant challenge. Recent 3D generative models, particularly those based on Flow-Matching (FM), can synthesize high-quality textured assets; however, they often suffer from ''synthetic bias'' where learned priors override observational evidence, alongside a lack of alignment with the observed instance. Conversely, optimization-based methods like 3D Gaussian Splatting (3DGS) provide high fidelity on visible surfaces but fail to reason about unobserved geometry. In this paper, we present FlowObject, a framework that reformulates sparse-view 3D reconstruction as a training-free, guided inverse problem. Our approach applies a dual-space guidance strategy to steer the Ordinary Differential Equation (ODE) trajectory of a flow-matching model, enabling the completion of unseen regions through learned generative priors while enforcing strict consistency with real-world observations. By integrating a 3DGS refinement stage, FlowObject further bridges the gap between ''synthetic-looking'' generative outputs and photorealistic reconstructions. Comprehensive benchmarks on synthetic and real-world datasets demonstrate that current state-of-the-art methods often struggle to achieve geometric completeness and observational consistency simultaneously, especially under severe occlusions. In contrast, our method significantly outperforms state-of-the-art generative models and optimization-based frameworks in both geometric completeness and view-dependent appearance fidelity.

Protein language models (PLMs) have emerged as powerful tools for controllable biomolecular design, yet their post-training adaptation typically relies on costly wet-lab validation or curated preference datasets. To overcome this supervision bottleneck, we introduce unsupervised reward optimization of PLMs, a comprehensive framework for steerable protein generation without ground-truth labels. Our key insight is that task-agnostic rewards, which combine intrinsic model uncertainty with extrinsic semantic consistency informed by protein representation models, exhibit strong correlation with controllability measures across base models and temperature regimes. Building upon this discovery, we propose two offline algorithms: Soft Reward Optimization (SRO) and Binarized Reward Optimization (BRO), which effectively maximize the classical RLHF objective induced by these proxy rewards. Extensive experiments on compositional out-of-distribution prompts demonstrate that both methods significantly outperform competitive baselines (DPO, KTO), while approaching oracle performance across multiple sampling temperatures, model scales and protein families. Moreover, PLMs fine-tuned with unsupervised rewards can achieve consistently higher coverage compared to their base model in pass@k evaluations. By enabling self-improvement of PLMs through their own generated experience, our framework provides a scalable pathway toward controllable biomolecular design in settings where labeled preferences or experimental feedback are scarce or unavailable.

Robots deployed in the real world should learn from their experience and improve over time. This requires a mechanism of practicing and learning from feedback. In this paper, we propose VERITAS, a generator-verifier framework for generalist robot policies for inference-time policy steering and self-improvement. We use a pre-trained generalist robot policy as a generator'' and pair it with a gradient-free visual verifier'' that evaluates actions at inference time. This framework enables inference-time steering that improves policy performance without additional training. We demonstrate that inference-time verification consistently outperforms vanilla generalists without training on additional demonstration data. Additionally, we demonstrate that the verified rollouts provide effective supervision for offline policy improvement: policies fine-tuned on verified self-generated trajectories achieve consistent performance gains. Notably, we find that post-training with verified rollouts achieves comparable efficiency to expert demonstrations, while requiring no human interventions. Our results highlight inference-time verification as a practical and scalable mechanism for improving robotic policies during deployment.

Steerable convolutional neural networks (Steerable-CNNs) guarantee SE(3)-equivariance by parameterizing kernels as linear combinations of steerable basis functions, but their deterministic nature precludes uncertainty quantification - limiting their use in settings where confidence estimates are essential. We propose a Bayesian Steerable-CNN that places posterior distributions over the basis coefficients, yielding stochastic kernels while preserving equivariance exactly. The loss function of the model is obtained via variational inference and minimized by Bayes-by-Backpropagation. The framework admits a decomposition of predictive uncertainty into epistemic and aleatoric components. Empirically, the model attains competitive classification accuracy alongside an expected calibration error of 0.0263 and outperforms its deterministic counterpart by up to 6.17% under distributional shift induced by additive Gaussian noise. Furthermore, we leverage the model's uncertainty estimates to enhance its performance significantly, achieving a notable gain - approximately 4% higher accuracy across 84% of the test dataset. A statistically significant negative correlation between epistemic uncertainty and prediction error confirms that the learned posterior variance is semantically meaningful. The framework unifies Bayesian uncertainty quantification with the inductive bias of equivariant CNNs.

Generalist policies can learn a wide range of skills from diverse robot datasets. In order to solve or improve on challenging new tasks, we need a way to infer and invoke the appropriate actions from the policy's rich behavioral prior, especially when directly commanding the policy fails. We focus on flow matching generalists and propose Flow Reversal Steering (FRS): a method that takes suboptimal but ``reasonable'' actions, finds their latent noises by passing them through the flow policy in reverse, and maps them to nearby generalist action modes. We evaluate FRS across many simulated and real-world manipulation settings. First, FRS can turn coarse semantic guidance from humans or vision-language models (VLMs) into corresponding good robot actions, improving zero-shot control. These gains can be distilled with behavioral cloning by training an auxiliary policy to output noises that the generalist maps to good actions -- showing up to 95% absolute task success rate boosts in under a minute of training. Finally, FRS enables policy improvement by bootstrapping reinforcement learning with semantic knowledge, improving on several tasks that standard RL fails to improve on.

The standard basis of transformer hidden states is a training-free, architecture-general feature basis for detecting concepts and, in language models, steering them; with no learned dictionary. Individual dimensions act as binary registers read one at a time: their signs (+/-1) encode content, their magnitudes strength. A feature is just a subset of dimensions with a consistent sign pattern, read by counting sign agreements. We validate this Bag of Dims (BoD) framework across seven models spanning language, vision, and audio; reading dimensions one at a time loses nothing, as a full-capacity MLP adds zero AUC over per-dim reading. The same per-dimension signs appear in every modality, so they reflect transformer training itself, not the language objective. Sign alone carries predictive content: setting all magnitudes to unity preserves 60-93% top-5 next-token accuracy through the LM head. From a single-token cache (one forward pass per token, no labels) we detect 175 categories at AUC 0.97-0.99 by counting sign agreements, and from random seeds alone discovery scales to 1500 features per model. A trained probe adds only +0.018 AUC and converges to axis-aligned weights: the rotation dictionaries learn buys little. Signs are causally operative: they survive the attention projections, and flipping a concept's sign pattern in the live forward pass suppresses it. Reading and steering are separate roles in the same basis: a concept's reader dimensions are not its writer dimensions. The writer target is just as cheap, the sign of the summed unembedding rows over a few seeds, no training. Injected through the attention output pathway under closed-loop control, it steers concepts into fluent text on four language models (62-92% of twelve concepts). The signs were in the standard basis all along; the open problem is no longer finding the right rotation but cataloging what each dimension encodes.

Vision-Language-Action (VLA) models provide a natural language interface to robot control, but the mapping from language to behavior is often brittle and unintuitive: semantically similar instructions can induce drastically different behaviors, while some capabilities may not be elicitable through prompting alone. As a result, both human instructions and zero-shot language models can fail to reliably steer VLAs toward successful task execution. In this work, we propose a framework that interactively searches for language sequences that improve closed-loop VLA task performance, distills these sequences into a test-time language feedback policy (LFP), and learns an improvement head that predicts when language steering will improve performance. We conformalize this improvement head to prevent harmful steering interventions, where the LFP decreases task performance relative to the original instruction on out-of-distribution scenarios. Crucially, our approach operates on arbitrary frozen pre-trained VLAs, requiring neither access to the original training distribution nor fine-tuning of the underlying model. On seen environments, our conformalized LFP improves base VLA performance by 24.7% in simulation and 65.0% in hardware. On visual and semantic perturbations, our conformalized LFP has strong harmlessness guarantees, and produces recovery behaviors not observed with open-loop prompting.

Activation steering offers a lightweight approach to control language models' behavior at inference time, but whether it succeeds or fails heavily depends on the prompt, concept, model, and steering configuration. Finding the regime and boundaries of successful steering typically requires expensive grid searches and post-hoc evaluation of full autoregressive rollouts. In this work, we investigate whether steerability can be predicted from the model's internal states at the beginning of the generation process, e.g., after generating the first few tokens, and how to leverage such a predictor to improve steering success rate. To this end, we first introduce ASTEER, a testbed including 1.4M steered generations, spanning 150 concepts with each steering success/failure labeled. Leveraging this testbed, we analyze the model's early decoding dynamics by extracting features that compare hidden states before and after steering across layers and initial decoding steps. These features help us understand how steering's effects propagate along layers and token positions, which provide key information for steerability prediction. We then train a Gradient Boosting Decision Trees (GBDT) classifier on these features to predict whether an intervention will under-steer, succeed, or over-steer without requiring full rollout. Our predictor achieves around 0.7 macro-F1 score on unseen concepts, demonstrating that early hidden states encode substantial, structured information about eventual steering efficacy. We further leverage this steerability predictor as guidance for steering strength searching, achieving near-optimal performance with a small fraction of decoding cost.

Real-world robots need to adapt their behavior beyond the envelope of their pre-trained policy. Policy finetuning or retraining are options, but they risk catastrophic forgetting, degrading the pretrained policy's base performance. To combat this, we introduce CLAE: Closed-Loop Affine Activation Editing, an inference-time framework for steering the behavior of a frozen policy by editing intermediate activations while keeping the base policy weights and downstream action head untouched. CLAE approaches behavior steering as a closed-loop problem whose outputs edit policy activations that adapt online to the robot state, environment, target behavior, and multi-robot context. It trains a sparse autoencoder over frozen-policy activations, selects behavior-relevant latent features via post-hoc probing, and learns a lightweight RL-based steering policy that applies state-dependent affine edits to selected latents during inference. We validate CLAE on a frozen multi-quadrotor navigation policy trained to perform a single task: navigating robots to a set of goal locations while avoiding obstacles. Through extensive simulations and physical tests, we show that while navigating to their goal positions, CLAE can 1. steer individual robot behavior by controlling each robot's velocity profile; 2. coordinate multirobot behavior by preserving a desired formation; and 3. produce entirely new behavior wherein robots are required to reduce their exposure to surveillance cameras in the environment. Our project website is available at https://clae-robotics.github.io/.

Large Audio-Language Models (LALMs) excel at audio understanding but expose little about where in an audio signal they attend. We introduce instruction-based vector steering, which constructs a steering vector by contrasting activations from differently instructed prompts while keeping the audio fixed. Through a systematic probe of LALM attention, we find that - unlike standard prompting or audio-based steering - this intervention significantly redistributes the temporal attention allocated to audio tokens, concentrating it on acoustically relevant regions. We then show that this attention shift is behaviorally meaningful: in a controlled three-event setting, reading out the temporal position of maximal steering-induced attention change recovers the location of a queried sound event without any training, attaining 60.87% and 68.72% overlap with ground-truth intervals on Qwen2-Audio and Audio Flamingo 3, far above direct prompting (31.84%, 46.75%) and random baselines (27.74%). Our results characterize a mechanistic property of instruction-based steering in LALMs and provide a training-free probe for the latent temporal structure these models encode.

Large Language Models (LLMs) are deployed across cultural contexts but often reflect homogenized values inherited from training data. Evaluations of cultural alignment typically rely on direct prompting with survey-style questions, which frequently elicit neutral or safety-aligned responses and fail to capture underlying model preferences. We propose a framework for probing and steering latent cultural representations in LLMs along the two Inglehart--Welzel axes of the World Values Survey (WVS). By translating social value questions into scenario-based behavioral dilemmas, we extract token-level probabilities to measure implicit values and apply activation steering, optionally combined with country-conditioned prompting, to shift model behavior without retraining. Across three open-source LLMs and four target cultures, we find substantial variation in steerability and identify latent entanglement, where interventions along one cultural dimension induce shifts along another. This coupling mirrors correlations in human WVS data and persists across activation, prompt, and hybrid steering. It constrains axis-independent alignment, though general task performance is largely preserved.

Deployed large reasoning models (LRMs) often behave unexpectedly. Test-time steering controls LRM outputs by intervening on their hidden representations, but it can degrade output quality. We argue that prior steering work implicitly relies on internal features that detect behavior in already generated text. We show that these detection features are poor predictors of future behavioral outcomes, and thus not the natural intervention target. Instead, we train activation probes to predict future behavior likelihoods from intermediate reasoning steps. These probes predict the most likely behavior with 64%-91% accuracy, revealing a separate type of internal prediction features. Building on these prediction features, we introduce a text-level steering method, Future Probe Controlled Generation. FPCG samples multiple candidate sentences and chooses the best one according to a probe predicting the future behavior likelihood. This enables steering with almost no output quality degradation. FPCG also enables steering in several evaluations where activation steering fails. These results show that distinguishing detection and prediction features enables a more nuanced approach to controlling LRM behaviors.

Federated learning (FL) is often subject to aggregation variance if clients do not consistently participate in training rounds. While reusing stale model updates from inactive clients is a common technique to reduce this variance, we find that with skewed client participation, the resulting update staleness can become severe enough to destabilize training. To remedy this, we propose FedSteer, a novel method that constructs a gradient subspace from a cache of recent client gradients to serve as a low-dimensional representation of the current optimization landscape. FedSteer projects an active client's true gradient onto this subspace to find a set of optimal coordinates. For an inactive client, FedSteer reuses these coordinates with the now-evolved subspace drifted by other active clients. This process effectively "steers" outdated gradients toward the current global objective. This is complemented by a selective caching strategy that identifies a representative client subset to form the subspace, reducing server memory. Experiments demonstrate that FedSteer significantly outperforms baselines, preventing performance collapse in challenging scenarios while delivering accuracy gains of over 7% in others.

Autonomous Racing has seen remarkable progress through deep Reinforcement Learning (RL), primarily for four-wheeled vehicles. However, motorbikes introduce substantially greater complexity due to the need to manage balance and lean angle, in addition to more reactive steering and throttle control, and a smaller weight. In this work, we present a framework for training an autonomous agent to race a superbike in VRider SBK, a physics-accurate Unity-based motorbike simulator. Our approach integrates Soft Actor-Critic (SAC) with Self-Paced curriculum Deep reinforcement Learning (SPDL), which dynamically generates progressively more challenging tasks based on the agent's performance, without requiring manual curriculum design. The agent's state space comprises proprioceptive features extended with lean-angle history, along with global track features via course points. The reward signal is shaped to encourage progress along the track while penalizing instability-inducing behaviors specific to two-wheeled dynamics. Preliminary experimental results demonstrate that SPDL outperforms SAC alone in training efficiency, lap time, and driving stability across multiple tracks and motorbike models, establishing a first baseline for RL-based autonomous motorbike racing.

Activation steering provides a lightweight inference-time mechanism for controlling large language models (LLMs) by modifying their internal activation vectors toward desired behaviors. Most existing methods compute a fixed steering direction in the original activation space, typically from pairs of contrastive examples using mean differences, linear probes, or arbitrary separability criteria. While effective to a certain extent, these methods treat behavioral control as a global, linear, additive offset: the same direction is applied across inputs, and behaviors are linearly separable. This can be restrictive when behavioral features vary nonlinearly across the activation space or lie on curved and anisotropic manifolds, where the optimal intervention may be input-dependent. To address this limitation, we propose INNSteer, a nonlinear activation steering framework based on invertible latent transformations. Rather than searching for a better steering vector in the original representation space, INNSteer learns a lightweight invertible neural network $φ$ that maps an LLM's activations into a latent space where behavioral classes are more amenable to linear control. At inference time, activations are mapped through $φ$, steered in the latent space, and mapped back through the exact inverse transformation $φ^{-1}$. This makes a simple latent-space translation become a nonlinear, input-dependent intervention in the original activation space. Across experiment settings on multiple LLM families, scales, behavioral traits, and safety benchmarks, INNSteer consistently improves model control over linear, transport-based, and nonlinear steering baselines while largely preserving generation fluency.

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.

Linear activation steering has gained popularity as a simple and empirically effective way to control language model behavior. More recently, spherical steering paradigms have been proposed to address limitations of additive interventions, often motivated by the assumption that hidden-state norm does not carry concept-relevant information. In this work, we revisit this assumption through a controlled empirical study designed to disentangle the roles of angular and radial components. We show that steering methods differ mainly in how they couple two geometric effects: changing a token's angular alignment with a concept direction and changing its hidden-state norm. Across seven language models, we find that concepts are represented primarily in angular structure, supporting the motivation for spherical methods, but that norm remains important for the stability and downstream effects of steering. Our results explain why interventions with similar concept-level effects can behave differently, and suggest that activation steering should be parameterized by interpretable angular and radial components of the intervention, rather than by a single additive coefficient that entangles these two effects.

Recent advances in interpretability suggest that large language models (LLMs) implicitly encode signals in their generated text that enable self-recognition of their outputs. We demonstrate that this capability is reliable, even in low-entropy scenarios, and that it can be amplified through targeted intervention. By steering the internal residual stream during generation with a random sparse vector, we create a detectable fingerprint that enables attribution of a given text to a specific LLM. This signal is recoverable from the activations of an LLM used as a detector, achieving over 98% accuracy across multiple detection settings while preserving the quality of generated text. As AI-generated content proliferates, this approach offers a practical alternative to traditional detectors by leveraging the model's natural representation structure for attribution rather than embedding a signal externally. Our contributions include: (i) establishing reliable self-recognition capabilities in LLMs, (ii) a simple steering mechanism enabling multi-LLM identification with no quality degradation, (iii) demonstrating that activation spaces contain exploitable structure for encoding signals without semantic interference.

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.

We propose GLASS, a framework for composable acoustic style control in zero-shot autoregressive text-to-speech (TTS) that learns controls from post-generation rewards rather than style labels. In zero-shot TTS, a speaker prompt often entangles speaker identity with prosodic attributes such as speaking rate and pitch, making it difficult to change style without changing the prompt itself. GLASS instead treats each acoustic attribute as a reward-defined control direction. For each control axis, GLASS freezes the TTS backbone and trains one lightweight LoRA adapter with Group Relative Policy Optimization (GRPO), using speech-token length and mean F0 as style rewards and WER as an intelligibility anchor. Because each control is represented as a LoRA weight update, independently trained adapters can be swapped, interpolated, and composed through linear LoRA arithmetic without retraining the backbone. Experiments on speaking rate and pitch control show targeted style shifts while preserving naturalness, speaker similarity, and intelligibility, and demonstrate smooth interpolation and multi-axis composition across independently trained adapters.

Large language models (LLMs) now generate substantial production code, often for tasks with multiple valid algorithmic solutions. Incidental prompt cues, meaning contextual words or metadata outside the task specification, can steer which algorithm the model selects, even when all outputs pass the same tests. Prompt sensitivity is well studied as a tool to improve output quality. Here, output policy means algorithm choice under fixed correctness. We define algorithm steering as cue-induced shifts in algorithm-family distributions and run 46,535 controlled experiments across 11 tasks, 19 cue types (18 channels plus a memoization semantic-vs-surface ablation that preserves meaning while changing typography and punctuation), and 15 model configurations. We find large, systematic shifts in algorithm-family distributions (up to 100 pp), largely consistent with cue semantics, including in applied tasks such as rate limiting. Direct algorithm naming is the most reliable mitigation we tested. Accidental context therefore creates an "invisible lottery" over performance, security, and maintainability.

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.

Vision-Language-Action (VLA) models trained on large-scale data have made remarkable progress, but they remain vulnerable to distribution shifts at deployment time. Recent VLA models suggest that prompts can serve as an efficient interface for steering policy behavior, but existing prompt-based steering typically relies on external guidance. This raises a natural question: can test-time training (TTT) for VLA be achieved by optimizing a prompt, so that the steering interface itself can be learned and adapted from interaction? We address this question with TTT-VLA, a test-time training framework based on Latent Prompt Optimization (LPO). During training, the latent prompt is learned with an additional proxy task, providing an extra learned conditioning signal for policy learning. At test time, TTT is performed by collecting interaction data from the current environment and optimizing only the latent prompt on those data using the proxy task's self-supervised signal, without modifying the policy itself. Experiments on SimplerEnv demonstrate that the proposed method consistently improves task success rates in both single- and multi-embodiment settings. Further analysis shows that the gains arise primarily from correcting a small number of critical decisions rather than globally altering policy behavior. These results suggest that LPO provides an effective and practical pathway for deployment-time improvement of foundation manipulation policies.

Prompting steers large language models (LLMs) and vision-language models (VLMs) without weight updates, but it remains unclear how instruction changes reshape internal representations to produce behavior. We introduce a nested geometric decomposition framework that treats prompting as a transformation of the representational geometry of the content following the prompt. For each prompt pair, we align representations of the same stimuli under two prompts using increasingly expressive stimulus-invariant maps: translation, rigid transformation with uniform scaling, sequential axis scaling, affine transformation, and nonlinear transformation. We then causally test each map by replacing a single layer's prompt-A hidden state for held-out stimuli with its mapped counterpart and measuring recovery of prompt-B representational geometry and behavior. Across three LLMs, three VLMs, and six text or image datasets spanning style, emotion, scene content, and number, prompts consistently reshape representations toward the instructed task structure. Cross-validated variance decomposition shows that much prompt-induced activation change is captured by shape-preserving maps, especially translation and rigid transformation with uniform scaling, while tier profiles reveal model- and task-specific routing strategies across layers. Crucially, although translation and rigid tiers already improve behavioral agreement, affine transformation is the first tier to nearly recover target-prompt task geometry and yields corresponding behavioral gains. This suggests that cross-dimensional linear mixing is a key mechanism by which prompts reorganize representations toward instructed task structure. Our framework decomposes prompt-induced representational change into interpretable geometric components and reveals how models route task-relevant structure to produce prompt-driven behavior.

We argue that the AI community is now ready to move beyond benchmarking and consolidate scattered efforts in model analysis into a systematic discipline, a direction we term Model Science. Complex AI models now serve billions of users, yet our understanding of how they work lags far behind our ability to deploy them. Decades of benchmark-driven research have delivered remarkable progress: extensive leaderboards, a wide range of performance metrics, tracking capability gains across diverse tasks; yet this success has also revealed the limits of benchmarks as they tell us whether models perform but not why they succeed or fail, they miss critical failure modes, such as hallucinations or shortcuts. Precedents from established sciences point the way forward: cognitive science shows that understanding complex systems requires complementary levels of analysis; neuroscience demonstrates that deep study of single cases reveals what population studies miss; medicine teaches that specialised training must develop alongside research practice; and agriculture models how shared infrastructure and principles enable cumulative progress. These lessons inform three foundations for Model Science. First, we propose to consolidate research around four functional perspectives: Verify, Explore, Steer, and Refine that address complementary questions about model behaviour. Second, we discuss the required infrastructure for cumulative knowledge: catalogues of datasets, models and findings. Third, we highlight the need for deep analysis of individual model instances, not just model families, because single cases can reveal what population studies miss.

We propose Parameterized Diffusion Policy (PDP), a framework for learning diffusion policies conditioned on low-dimensional, continuous parameters embedded in a learned behavior manifold. By constructing this manifold so that distances between latent representations reflect the semantic similarity between physical trajectories, we transform diffusion from a mechanism for stochastic diversity into a precise and optimizable tool for behavior steering. Our approach enables smooth interpolation between known strategies and efficient adaptation to novel constraints without updating policy weights. We demonstrate that PDP significantly improves adaptation performance on complex multimodal benchmarks in both simulated and real-robot experiments compared to standard diffusion policies, particularly in scenarios requiring the synthesis of novel behaviors.

Vision-Language-Action (VLA) models can be steered at test time by intervening on semantically meaningful internal directions, but existing methods use a fixed steering coefficient, effectively operating in open loop. This is poorly suited to embodied control, where task state and concept error evolve over time, often causing overcorrection, oscillation, and reduced task success, especially for temporal behaviors such as speed and smoothness. We propose CTRL-STEER, a closed-loop framework that replaces static intervention strength with adaptive, time-varying control signals. The key idea is to decouple representation from regulation: rather than assuming temporal concepts are directly controlled by individual neurons, we steer along motion-aligned residual directions while a feedback controller adjusts intervention magnitude online. We instantiate this framework with both PID and reinforcement learning based controllers. Experiments with a fine-tuned OpenVLA policy on four LIBERO task suites show that CTRL-STEER achieves more stable concept regulation and a better steering-task success trade-off than fixed-coefficient baselines, without modifying or retraining the base model.

Reasoning-based robotic policies using large language and vision-language models achieve strong semantic planning capabilities but mostly suffer from a high inference latency that limits practical real-time deployment. In this work, we observe that robotic reasoning workloads contain substantial temporal redundancy, where consecutive observations frequently produce identical actions and subgoals. Based on this insight, we present REIS, a human cognition inspired robotic decision-making framework that minimizes unnecessary reasoning while preserving semantic adaptability. REIS combines lightweight scene gating, KV-steered affordance routing, and deliberative reasoning to accelerate robotic control under embodied constraints. Experiments on ALFRED, and real-world robotic tasks demonstrate that REIS significantly suppresses reasoning overhead while maintaining competitive task performance.

Hateful meme detection remains a formidable challenge for vision-language models, as existing benchmarks are structurally observational - confounding rhetorical hate mechanisms with target community features and preventing causal evaluation of model vulnerabilities. To address this, we introduce FBHM, a systematically curated benchmark of Functionality Based Hateful Memes constructed along two orthogonal axes: 25 distinct rhetorical functionalities and 10 target communities (5,000 memes total). Benchmarking state-of-the-art VLMs reveals a severe generalization gap: models highly accurate on standard datasets catastrophically drop to near-random performance on FBHM, proving they exploit dataset-specific heuristics rather than robust multimodal reasoning. To efficiently close this gap, we propose LSV (learnable steering vectors), an ultra-low data regime strategy that applies a causal intervention objective on as few as 500 steering samples (50 unique base memes), boosting FBHM performance by ~30 Macro-F1 points while outperforming in-context learning and PEFT without degrading source-domain performance.

Sparse Autoencoders (SAEs) have been seen as a promising avenue for exploring the internals of Large Language Models (LLMs) and for steering model output generation. When AxBench - a model steering benchmark - was introduced in Wu et al. (2025), SAEs did not seem to live up to their original hype due to poor steering performance relative to a set of simple baselines. This work serves as a partial rebuttal for Sparse Autoencoders and suggests that the results of Wu et al. (2025) did not do them full justice. We find that Sparse Autoencoders can, in fact, perform close to on par with the reference LoRA performance on the AxBench benchmark, when features are selected and labelled with our supervised pipeline. We also find that our pipeline selects features that are surprisingly causal of their identified labels when using only its interpretability-based components. Lastly, we present evidence that high sparsity (low l0) may not be crucial for successful steering based on interpretability, which is in contrast to the earlier findings in Wang et al. (2025).

Controllable music editing is to modify high-level attributes while strictly preserving rhythmic and melodic structures. However, this task is challenged by a semantic-structural entanglement: steering methods often degrade structure to achieve editing performance, while structural adaptors suppress semantic responsiveness. We propose AnchorSteer, a framework that disentangles this tension by coupling structural anchoring with self-discovered semantic steering. The proposed approach probes internal representations to extract interpretable, label-free concept vectors via a self-supervised reconstruction objective, isolating attributes without curated data. During editing, these portable, plug-and-play concept vectors are injected into diffusion hidden manifolds while a structural adaptor enforces consistency. Variants for unconditioned and conditioned injections are provided to balance robustness and semantic strength. Experiments on ZoME-Bench and subjective tests show that the proposed framework outperforms both steering-only and anchoring-only baselines, enabling significant semantic transformations with high-fidelity structural preservation.

The shortage of legally compliant data for face recognition training has sparked growing interest in using synthetic data as an alternative. While recent diffusion-based methods enable the generation of photorealistic face images with strong identity adherence and data diversity, their downstream recognition performance still exhibits a significant synthetic-real gap. This paper identifies visual tendency as a previously underexplored limitation, whereby synthetic data exhibit an unrealistic prevalence of visual attributes and thus deviate from the real-data distribution. Visual tendency can be attributed to the generator's conditioning on identity embeddings, through which co-occurring residual visual cues are unintentionally absorbed into learned identity semantics. To discourage the generator from exploiting such visual cues, this paper proposes SteerFace, a simple and efficient training framework that perturbs identity embeddings by steering them toward random orthogonal directions on the embedding hypersphere. The perturbation serves as an identity-preserving regularizer that penalizes the generator's reliance on non-identity components, as supported by theoretical analysis. This paper further introduces an adaptive strategy that learns perturbation strengths with both sample-wise preference and favorable overall statistics. Extensive experiments show that SteerFace effectively mitigates visual tendency, outperforms prior methods in downstream face recognition, and generalizes well across different training datasets and generation pipelines.

Multilingual large language models can generate figurative language, but whether the internal signals driving this behavior are language-specific or reusable across languages is unclear. Using activation steering as a probe, we estimate a direction for a figurative category from figurative--literal activation differences in one language and apply it during generation. Across five figurative categories, six languages, and four multilingual LLMs, these directions steer reliably within their own language, most robustly for metaphor and simile. More importantly, they transfer across languages: a direction learned in one increases the target behavior when applied to another, with German among the most receptive targets. Going further, directions assembled from other languages can match or even surpass a target language's own native direction, while removing this shared component weakens native steering. Together, these results provide direct evidence of a reusable but target-dependent cross-lingual signal for figurative generation.

Vision-Language-Action (VLA) models have recently shown strong potential for robot learning by following language instructions. However, in practice, language alone is often insufficient to precisely convey human intent. It is difficult to describe which exact object to interact with among similar candidates, where to act on the object, or how the target may change during execution. To address this limitation, we propose Gaze2Act, a novel VLA framework that leverages human gaze as a dynamic and intuitive intent signal for complex interactive manipulation. Gaze2Act first bridges the ego-exo view gap by mapping first-person gaze into the robot's perspective through cross-view semantic matching, producing both an object mask and a gaze point for coarse-to-fine target specification. These cues are then integrated into the policy through perception-level prompting and action-level conditioning, allowing the robot to attend to relevant regions and execute precise interactions under dynamic intent. In a systematic evaluation across seven task categories and 16 real-robot tasks on a Unitree G1 humanoid, Gaze2Act achieves state-of-the-art performance in both intent accuracy and task success rate. It notably outperforms baselines in object disambiguation, fine-grained interaction, and dynamic intent steering. These results demonstrate that human gaze provides a natural, low-burden, and highly expressive modality for human-in-the-loop VLA control.

Large vision-language models (LVLMs) often hallucinate objects that are not present in the input image, largely because visual grounding weakens as decoding progresses. Existing inference-time mitigation methods modify logits or hidden states throughout generation, but they suffer from three key limitations: they lack an explicit grounding objective, intervene even when the model is already well-grounded, and use fixed correction strengths that do not adapt to the severity of grounding failure. We propose BRACS (Barrier-Regulated Adaptive Closed-form Steering), a training-free steering framework that addresses these issues through barrier-regulated adaptive closed-form steering. BRACS monitors the model's own attention to measure visual grounding and applies corrections to the hidden states only when grounding deteriorates. The corrective update is computed analytically in closed form, requiring no training of auxiliary networks or model retraining. Experiments on LLaVA-1.5-7B and Qwen-VL-Chat show that BRACS consistently outperforms prior methods on hallucination benchmarks, reducing CHAIR$_s$ by 9.4 points and improving POPE F1 by 2.7 points, while matching or improving performance on four general multimodal benchmarks. BRACS also remains efficient, operating at 80% of greedy decoding throughput and achieving 1.3 times higher speed on average than the baselines.

Recent advances in large language models have enabled natural language to serve as a flexible interface for controlling complex systems, but often at the cost of large-scale multimodal training or weakened domain-specific inductive biases. In structured decision-making domains such as chess, specialized policy networks achieve strong performance but lack semantic controllability, while prompt-conditioned language models are more flexible yet typically exhibit weaker domain grounding. We propose $\textbf{UniMaia}$, a framework for prompt-conditioned policy modulation that adapts a frozen Lc0-based chess policy network using a parameter-efficient text encoder and a ControlNet-style conditioning mechanism. UniMaia enables semantic control over gameplay, including opening selection and player strength, while preserving the pretrained policy representations. We further introduce $\textbf{UniMaia-Aux}$, which incorporates auxiliary temporal conditioning and behavioral prediction objectives. To support this work, we construct a large-scale metadata-augmented Lichess dataset, develop a semi-automated prompt-generation pipeline, and introduce benchmarks spanning both prompt-conditioned and metadata-conditioned settings. UniMaia achieves state-of-the-art expected accuracy on several prompt-conditioned benchmarks and competitive top-move accuracy on general instruction-following tasks, while remaining competitive with dedicated metadata-conditioned approaches on human move prediction benchmarks. UniMaia-Aux further improves expected accuracy and behavioral modeling across several evaluation settings, with modest trade-offs in top-move accuracy. Overall, our results demonstrate that prompt-conditioned control of domain-specific policy networks is feasible without end-to-end multimodal training, while highlighting trade-offs between controllability and predictive performance.

Large reasoning models (LRMs) generate chain-of-thought (CoT) traces before producing final outputs, introducing a dynamic internal state that may complicate control mechanisms such as refusal. Unlike instruction-tuned LLMs, where refusal is mediated by a single directional subspace, refusal in large reasoning models (LRMs) additionally depends on the CoT. In DeepSeek-R1-Distill-LLaMA-8B, activation steering reverses refusal in only 39% of cases when the CoT is kept fixed, but removing the CoT entirely increases this to 70%, indicating that the CoT actively reinforces refusal. In a two-stage intervention where the model regenerates its CoT under activation steering, refusal is reversed in 94% of cases, while the resulting CoT alone retains 48% of this effect even after steering is removed. This suggests that the CoT can carry and reconstruct the compliance signal independently. These findings indicate that refusal in LRMs is jointly encoded in residual stream activations and CoT. This joint activation makes LRM more robust against activation-level interventions alone, but exposes CoT to a possible alternative surface attack.

Medical vision-language models (VLMs) often hallucinate findings when generating chest X-ray reports: they fabricate findings that are not present in the image, miss important ones, or locate them incorrectly. We mitigate this without weight updates by decoding-time residual steering on a per-token sparse autoencoder (SAE) basis: Top-$K$ SAEs on late layers, causal steering against clinical errors, then combined suppress/boost intervention at inference time. On the MIMIC-CXR test split, our inference-only method improves the quality of generated reports for three radiology VLMs (RadVLM, LLaVA-Rad, and CheXOne), with relative improvements of +5.4%, +7.2%, and +17.0% in the clinical composite metric, and statistically significant GREEN gains on all backbones. A cross-model feature alignment shows that the quality-promoting (boost) directions overlap strongly across architectures, whereas hallucination-linked (suppress) directions are model-specific. Therefore, transferable steering must treat suppression per-backbone, rather than sharing a universal suppress list. The same recipe transfers zero-shot to IU-Xray (Green $+7.7\%$ rel.) without retraining, confirming that the identified features are properties of the model, not of the training corpus. We release causal feature sets and an interactive feature dashboard: https://cxr-sparse-feature-dashboard.netlify.app/.

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.

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.

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.

We study the effect of different persona on \textbf{sycophancy}: model's agreement with users even when the user is incorrect. The standard mitigation, Contrastive Activation Addition (CAA), derives a steering direction from labelled pairs of sycophantic and honest responses. This study evaluates whether off-the-shelf persona steering vectors, originally developed for general role-playing and not trained on sycophancy data, can serve as an alternative. In two instruction-tuned models, steering toward personas characterised by doubt or scrutiny reduces sycophancy to approximately $68\%$ and $98\%$ of CAA's effect, and, unlike CAA, maintains accuracy when the user is correct. The effect is also asymmetric: steering toward agreeable personas does not produce a mirror increase in sycophancy. Geometrically, the persona vector is largely independent of the direction of sycophancy in activation space. Collectively, these findings suggest that sycophancy is better understood as a persona-level property rather than a single steerable direction. We release our code here: https://anonymous.4open.science/r/Sycophancy-Steering-9DF0/.

We present a modular framework to benchmark new and existing methods for trajectory planning and control in high-acceleration maneuvers that push autonomous driving to the limits. Our framework includes time-optimal raceline generation, online time-optimal velocity replanning, geometric path tracking controllers, and a new model-structured neural network (MS-NN) to learn the inverse dynamics for steering control. We deploy our framework on a 1:10-scale RoboRacer platform, using two circuits. Through several ablations with cautious and aggressive racelines, we study the performance of single modules and their combinations. We show that our MS-NN significantly improves tracking accuracy, decreases steering oscillations, and is physically interpretable. Moreover, online velocity replanning improves lap times by compensating for execution errors, and enables the vehicle to safely reach higher speeds and accelerations. To support future research, our code, datasets, videos and results are publicly available at https://roboracer-benchmark.github.io/planning_control_benchmark/.

We study selective refusal editing as a three-way control problem: induce non-refusal on designated edit prompts while preserving benign behavior and harmful refusals outside the edit set. We introduce Residual Paving, a routed residual editing method for frozen instruction-tuned transformers that separates route selectivity, whether to intervene, from residual-edit capacity, what edit to apply. An early-layer router predicts a scalar gate and expert mixture; when active, prompt-conditioned bottleneck residual experts apply later-layer residual updates while leaving the backbone unchanged. This decomposition supports an oracle-routing diagnostic where only the learned scalar gate is replaced with the held-out edit/keep label, leaving the residual editor and frozen backbone fixed. On the primary Gemma-3-4B-IT held-out split, learned Residual Paving reduces edit refusal from 88.6% to 4.0%, with 95.5% benign distribution preservation and 87.3% harmful distribution preservation. Same-protocol one-direction steering controls are much weaker on edit success, leaving edit refusal at 86.8% for Edit-target ActAdd and 78.9% for DIM-style refusal steering. The remaining failure is off-target harmful-keep degradation: harmful refusal remains below the frozen-base rate, 65.3% vs. 81.6%. Across six backbones, oracle routing improves the keep-side diagnostic score on every reported row, with median gain +12.9 pp, supporting the interpretation that learned route selectivity is the main observed bottleneck. Trajectory diagnostics on two backbones further suggest directed movement toward edit-target continuations rather than generic refusal suppression.

The promise of Rectified Flow rests on producing self-generated couplings whose trajectories are straight, or nearly so. In practice, trajectories generated by the base flow model can bend and intertwine, and the resulting coupling inherits this distortion. In this paper, we identify that such trajectory entanglement is often associated with regions of nonzero divergence in the learned velocity field, where local expansion or contraction distorts trajectories and steers particles away from their ideal endpoints. We then propose divergence-suppressing couplings for Rectified Flow, an offline correction that attenuate the divergent component of the learned velocity during coupling generation. The correction is paid only once per coupling pair and amortized over training, so deployment runs plain Euler at identical wall-clock cost to standard Rectified Flow. Empirically, this offline modification yields consistent improvements on 2D synthetic benchmarks and on image generation.

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.

Video reasoning segmentation requires localizing objects across video frames from natural language expressions, often involving spatial reasoning and implicit references. Recent approaches leverage frozen large vision-language models (LVLMs) by extracting attention maps and using them as spatial priors for segmentation, enabling training-free grounding. However, these attention maps are optimized for text generation rather than spatial localization, often resulting in diffuse and ambiguous grounding signals. In this work, we introduce SteerSeg, a lightweight framework that identifies attention misalignment as the key bottleneck in attention-based grounding and proposes to steer attention at its source through input-level conditioning. SteerSeg combines learnable soft prompts with reasoning-guided Chain-of-Thought (CoT) prompting. The soft prompts reshape the attention distribution to produce more spatially concentrated maps, while CoT-derived attributes resolve ambiguity among similar objects by guiding attention toward the correct instance. The resulting attention maps are converted into point prompts across keyframes to guide a segmentation model, while candidate tracklets are ranked and selected using correlation-based scoring. Our approach freezes the LVLM and segmentation model parameters and learns only a small set of soft prompts, preserving the model's pretrained reasoning capabilities while significantly improving grounding. Despite being trained only on Ref-YouTube-VOS, SteerSeg generalizes well across diverse benchmarks, significantly improving the spatial grounding capability of LVLMs. Project page: https://steerseg.github.io

While existing text-to-speech (TTS) models exhibit high expressiveness, fine-grained control over composite instructions remains challenging due to the structural mismatch between discrete textual intents and continuous acoustic realizations. Inspired by human cognitive decoupling, we introduce AgentSteerTTS, a multi-agent closed-loop framework designed for intent-faithful expressive control of composite instructions. First, in our framework, an adversarial disentanglement agent mitigates speaker-emotion leakage by learning separable identity and emotion-prosody subspaces with leakage-suppressing regularization. Next, a Dual-Stream Anchoring Controller grounds abstract intents using a large-scale acoustic prototype library: a Retrieval Agent selects expressive anchors, while a Synthesis Agent fuses them into continuous control vectors via gated attention. Finally, a Fast-Slow Feedback Agent refines output intensity through latent gradient correction and resolves semantic-acoustic mismatches using high-level perceptual critique. Experiments on a composite-instruction benchmark and public test sets show that AgentSteerTTS yields consistent and significant improvements to the baselines, demonstrating the effectiveness of the proposed method.

We introduce WriteSAE, a sparse autoencoder for the matrix updates written into recurrent language-model state. In Gated DeltaNet, Mamba-2, and RWKV-7, each token writes a matrix-shaped update to a recurrent cache; a residual-stream SAE has vector-shaped atoms and cannot replace that update directly. WriteSAE learns rank-1 matrix atoms with the same shape as the model's own write. This lets us test a direct replacement: at positions where the SAE activates an atom, we remove the model's write, insert the atom scaled by its SAE activation, and continue the forward pass. The atom gives a closer final token distribution than deleting the write on 92.4% of evaluated positions; averaged per atom, the rate is 89.8%. For Gated DeltaNet, a formula using the forget gate, read query, and output embedding predicts the resulting logit change with $R^2 = 0.98$. The same replacement test transfers to Mamba-2-370M at 88.1%. In generation, the formula chooses a write direction; writing it into three consecutive cache positions at $3\times$ the norm of the model's write makes tokens initially ranked 100--1000 by the unmodified model appear in 100% of continuations, up from 33.3%. To our knowledge this is the first cache-level steering intervention reported in a state-space or hybrid recurrent layer.