Steering aligns large language models (LLMs) by injecting a bias into selected activations at inference time, offering a far cheaper alternative to weight-update methods such as supervised fine-tuning or reinforcement learning. However, most existing training-free steering methods are input-independent: a single direction is fitted once and shared across all inputs. This is fundamentally limiting as different inputs occupy different regions of the activation space and admit different optimal steering directions toward the same target concept, much as the gradient with respect to a fixed loss varies from input to input. We close this gap with IDEEA (Input-Dependent stEEring via Activation cluster matching), a training-free framework for input-dependent steering. IDEEA clusters the positive and negative activation supports per attention head, and solves an optimal-matching problem to construct a set of cluster-conditional directions, all about the target concept. At inference time, it picks from this pool of directions and uses the one that best matches the input's own activation for steering. IDEEA aligns the model toward the target concept while preserving the input's original representation, evidence that activations encoding a concept occupy several distinct sub-regions of the representation space rather than a single one. IDEEA improves the truth $\times$ info rate in TruthfulQA by an average of 9.9% (up to 23.5%) over the best input-independent baseline.
Large language models (LLMs) are commonly evaluated under the assumption that their observable behavior is primarily determined by model weights, training data, alignment procedures, and user prompts. This view is incomplete. Modern inference pipelines may systematically modify the probability distribution produced by a model immediately before token selection, creating an additional layer of control between frozen weights and observed text. While controlled generation (e.g., PPLM, GeDi, DExperts, FUDGE) and text-watermarking systems (e.g., SynthID-Text) demonstrate the technical maturity of decoding- and logit-level interventions, the governance, security, and economic implications of an undisclosed inference policy remain comparatively underexplored. This paper examines the emergence of inference-time framing bias: the systematic modification of generated language toward political, ideological, institutional, or commercial frames via interventions applied after model inference but before token sampling. We formalize the operational reality Model != Deployed System and introduce three concepts: (1) the Inference Attribution Problem, characterizing why observed behavioral bias cannot generally be causally attributed to model weights alone under limited observability; (2) Probability Placement, defining a hypothetical advertising primitive in which commercial influence is implemented through systematic shifts in generation probabilities rather than explicit product insertions; and (3) Inference Policy Transparency, a governance principle for making deployment-layer interventions auditable. We examine these concepts in relation to Article 5 of the EU AI Act, the EU Digital Services Act, and FTC doctrines.
Moritz Zoellner, Anastasios Manganaris, Ahmed H. Qureshi +1cs.RO cs.LG
A central goal of robot learning is to enable robots to execute rich instructions specified at runtime. Large-scale language-conditioned policies have made substantial progress toward this goal, yet still struggle with temporal structure and safety constraints. Linear Temporal Logic (LTL) provides a powerful language to express complex, non-Markovian instructions. However, guiding learned manipulation policies toward LTL satisfaction remains challenging because modern policies generate short-horizon action chunks and replan in closed loop, while almost all LTL specifications are evaluated over long-horizon trajectories. In this paper, we introduce hint$^2$, a method for guiding short-horizon policies toward satisfying complex LTL specifications at inference time using hierarchical world models. Our key idea is to derive two separate guidance objectives using each world model's abstraction level. A high-level model predicts future action-induced transitions in task-relevant atomic propositions to guide progress through the LTL automaton, while a low-level dynamics model predicts immediate state evolution for accurate local safety guidance. Our results show that hint$^2$ overcomes the limitations of current LTL-guided diffusion methods, outperforms existing inference-time steering methods in CALVIN, and successfully completes instructions with complex liveness and safety constraints more elegantly than language-conditioned alternatives. Finally, we demonstrate that hint$^2$ can handle complex instructions on a real UR5e manipulator.
Multilingual large language models exhibit substantial performance differences across languages, while existing adaptation methods often require parameter updates and considerable multilingual training data. We propose an inference-time multilingual steering method that uses pretrained sparse autoencoders to identify and strengthen target-language-related features. Using multilingual parallel sentences, we compare SAE activations across languages and select a small number of layer-specific features associated with each target language. These features are decoded into steering signals and injected into the model's hidden states without additional training. Experiments with Gemma-3-12B-it show average accuracy improvements of 10.9 percentage points on XCOPA, 5.3 points on XNLI, and 1.9 points on MGSM.
Although Large Language Models (LLMs) demonstrate remarkable multilingual fluency, their internal knowledge representations remain disproportionately biased toward high-resource languages. This leads to cross-lingual factual inconsistency, where they shift their empirical answer distributions based solely on the prompt language. We investigate whether these biases can be mitigated at inference time, forcing an English-prompted model to answer as if it were queried in target languages (German, Spanish, Bulgarian), and evaluate four intervention strategies: zero-shot contextual steering (persona prompting), internal representation manipulation via Contrastive Activation Addition (CAA), and lightweight weight modification via Direct Preference Optimization (DPO) trained on benchmark-derived factual data as well as conceptual generalization data. To assess alignment, we curate a multilingual factual dataset alongside a novel generalization benchmark comprising culturally rooted queries to determine whether factual interventions transfer to broader target-centric preferences. Experiments on Gemma 3 12B Instruct reveal persona prompting to be the strongest overall intervention, balancing efficacy, safety, and out-of-domain generalization. While CAA yields sharp inconsistency benchmark shifts, it is configuration-sensitive and risks knowledge degradation. DPO-based adapters offer permanent, yet narrower and less transferable gains. These findings suggest that cross-lingual inconsistency is at least partly a selection problem, and that simple contextual interventions may outperform more invasive methods for robust, transferable alignment.
Open-weight video diffusion models can generate photorealistic unsafe content, from violence to misinformation, yet existing defenses either require expensive safety fine-tuning that degrades general capability, or apply external filters that are trivially bypassed by adversarial prompts. We present REINS (REpresentation-space INference-time Safety steering), a training-free method that aligns video diffusion models at inference time by steering their internal representations toward safe generation. Our key finding is that safety-relevant structure is linearly encoded in the hidden-state activations of video diffusion transformers, and a single direction, discovered via Supervised PCA on binary safety labels, suffices to separate safe from unsafe generation trajectories. At inference, adding this direction to hidden states at an intermediate transformer layer redirects generation from harmful content to semantically related safe alternatives, with no weight updates, no concept enumeration, and negligible computational overhead. Through mechanistic analysis, we reveal that while safety information accumulates monotonically with transformer depth, steering effectiveness peaks at intermediate layers (~50% depth), exposing a fundamental tradeoff between information availability and downstream propagation capacity. We evaluate REINS across 9 video diffusion models, multiple parameter scales (1.3B-5B), and both text-to-video and image-to-video generation, to our knowledge, the broadest safety evaluation suite in the video generation literature.
Yilin Wu, Zilin Si, Zeynep Temel +2cs.RO cs.AI cs.LG
Inference-time steering adapts pre-trained generative robot policies during deployment by verifying candidate actions before execution. While prior methods typically perform this verification only with visual observations, vision alone is often insufficient for contact-rich manipulation, where success depends on both global task progress and subtle local interactions such as contact force. We introduce ViTaL, a visuo-tactile inference-time steering framework that formulates multimodal guidance as a bi-level optimization problem. At the high level, visual sampling-and-verification performs long-horizon mode selection, deciding what behavior the robot should execute. At the low level, tactile-guided diffusion editing refines the selected action sequence over a shorter horizon to satisfy local contact requirements. To support outcome-based steering, ViTaL learns a visuo-tactile latent world model and employs semantically aligned visual and tactile verifiers, including a novel text-conditioned tactile reward that scores predicted tactile futures directly in latent space. Across three real-world contact-rich manipulation tasks, ViTaL improves overall success by 51% over the base policy, outperforms unimodal steering by at least 33%, and exceeds naive multimodal fusion by at least 20%. Website: https://yilin-wu98.github.io/vital_website.