WorksComputer Science
Neural Representations as New Modes of Interactivity
Abstract
Human-AI interaction today runs almost entirely through prompts and outputs, keeping AI systems opaque to general users. We introduce Neural Representation Interaction (NRI), a paradigm that treats the internal representations of neural networks as an interaction material people can inspect, steer, and edit. Situating NRI alongside graphical, tangible, and bio-computer interaction, we argue that the concepts and algorithms models encode internally offer representational affordances for richer human-AI interaction. We synthesize actionable mechanistic interpretability methods and, drawing on General Morphological Analysis, construct a design space spanning Interaction, Target, Referent, and Purpose. Mapping existing interfaces onto this space reveals that most serve expert auditors, while few support lay users. We then present speculative applications in education, personal user modeling, and hiring audits, and discuss design principles, limitations, and sociotechnical risks. NRI offers HCI researchers abstractions for designing interactions that work with AI computation rather than only through its inputs and outputs.
CCS Concepts
• Human-centered computing → Interaction paradigms; Interaction design theory, concepts and paradigms; • Computing methodologies → Learning latent representations; Artificial intelligence; Neural networks.
Introduction and Motivation
Interpretability is fundamentally a human endeavor. Throughout history, understanding phenomena has given us greater agency, comprehension, and the resilience to act on them. This effort is even more critical as we are being heralded into an era of capable generative artificial intelligence (AI) systems, which are opaque ’black-boxes’ and difficult to interpret. Initially dominated by the development of Large Language Models (LLMs), we are witnessing a new wave of foundation models beyond LLMs, such as vision [8], audio [61], computer use [129], interface generation [102], and world models [91]. As AI systems continues to grow more expressive and embedded widely, the central challenge of human-AI interaction will move from accessing to comprehending these systems. Moreover, human interaction with such powerful and ubiquitous AI may invite new modes of interactions.
Here, we aim to conceptualize a new interaction framework called Neural Representation Interaction (NRI). We define NRI as interaction that enables users to interpret and direct the neural representations of AI systems. We imagine the parameter and representation space of complex AI systems to be new media of interaction that are subject to manipulation and intervention. This is in contrast to our current paradigm in human-AI interaction (HAI), prompting, where communication is done through natural language. While intuitive, this communication mode maintains the black-box nature of LLMs and may promote automation bias [108], over-reliance [9], and anthropomorphism [1]. Moreover, this interaction does not allow users to comprehend inner workings and leverage the rich affordances offered by the representational capacities of neural networks. As such, we present NRI as an alternative communicative medium of HAI: users can modify latent representations—used as internal semantic structures by AI systems—to control responses and interaction dynamics. Our chief tool in this paradigm is called mechanistic interpretability (MI), a rapidly evolving field that boasts an ambitious vision of "reverse engineering the neural network" to characterize model computation and behavior [12, 141].
For instance, possible applications include an auditing tool for an AI resume screener, where a hiring manager can use the model’s internal structure to audit whether models are displaying systemic biases by relying on names, demographic characteristics, and geographic cues. Within the NRI framework, users could scrutinize model representations and causal mechanisms for bias and make targeted edits to remove them. Accordingly, NRI enables users to act on representational bits, or elementary units of information embedded within trained AI systems. As such, NRI is a new framework that enables new affordances, enriches user experiences, and widens the possibilities of new interaction dynamics.
This paper makes four contributions to HAI:
First, we introduce Neural Representation Interaction (NRI), a conceptual paradigm that treats the internal representations of neural networks as an interaction material that people can inspect, edit, and control, extending human-AI interaction beyond prompts and outputs.
Second, we provide an actionability-oriented synthesis of mechanistic interpretability methods and interfaces, connecting technical capabilities such as representation decoding, attribution, steering, and model editing to human goals including understanding, agency, decision support, and oversight.
Third, drawing on General Morphological Analysis, we develop a design space organized around four dimensions: Interaction, Target, Referent, and Purpose. We use this framework to characterize existing systems, identify underexplored combinations, and articulate progressive disclosure and iterative hypothesis testing as guiding principles for NRI interfaces.
Finally, we offer a set of speculative design propositions spanning education, personal model inspection, cognitive augmentation, and consequential decision auditing. These propositions surface opportunities, limitations, and socio-technical risks for designing interfaces that expose model internals.
Neural Representation Interaction as a New Framework
Interaction paradigms are often driven by the unique affordance principles of their respective technologies, a property of a system that suggests or enables certain actions, and motivated from limitations of earlier frameworks (see Table 1). For instance, graphical user interfaces (GUI) [120], pioneered by Ivan Sutherland, were driven by the limitation of human recall capabilities that hindered syntax memorization for computer control. Meanwhile, tangible user interaction (TUI) [40] aims to bridge the inconsistency of interaction on bitmapped devices with the physical world. More recently, living bits [92] were conceptualized to incorporate biological interfaces and organisms to sense and actuate for more naturalized integration with the human body. Thus, we will accordingly define the affordance principles and motivations for this new framework.
| Interaction paradigm | Objects | Affordance principles |
| Graphical user interface (GUI) [120] | Graphical objects (“painted bits”) | Leverages visual cues, spatial reasoning, and recognition abilities. |
| Tangible user interaction (TUI) [40] | Physical objects (“tangible bits”) | Leverages physical dexterity and embodied, multisensory intuitions. |
| Bio-computer interaction [92] | Microorganisms (“living bits”) | Leverages the biodegradable and continuous properties of natural systems. |
| Neural representation interaction | Latent representations (“representational bits”) | Leverages the predictive, generative, and representational capacities of AI systems. |
Table 1. Interaction paradigms, their objects, and affordance principles.
First, NRI aims to leverage the unique generative and representational capacities of AI systems. We refer to the use of these capacities to augment human capacities as the computational affordances of modern neural networks. We define generative affordances as the relation between an AI’s capacity to generate novel outputs and the user’s capacity to specify, shape, and evaluate those outputs. Correspondingly, representational affordances of an AI system are the coupling between the system’s neural representational states and the user’s capacity to perceive, interpret, and act with the system’s computation rather than through its inputs and outputs.
Thus far, the generative affordances have been the primary focus. Notably, the affordances of deep learning have spurred rapid progress across many domains. The use of these models made significant strides in solving challenging problems and addressing user needs such as cataloging and generating new protein structures [47, 66], modeling global weather forecasts [60], code generation, synthesizing literature review, and assisted writing.
Yet the use of these models in a blackboxed approach has not exhausted the representational affordances. The representational capacities of models include the ability to encode concepts and algorithmic computations not observed in outputs. In the past, interactive machine learning has aimed to move users past this input-output boundary by letting users shape a model through rapid cycles of labeling, correction, and feature weighting [4, 38]. Feature weighting perturbs the influence of inputs but does not provide us insights to the rich patterns of a model’s internal computation. An illustrative example of this is the application of AI systems to chess games, which yielded both human-identifiable [80] and machine-level concepts [109]. Other examples include new identifiable concepts in a well-played game of Go [113], new mathematical discoveries in construction of large cap sets [100], and efficient matrix multiplication algorithms [39]. In other words, neural representations can indeed yield novel insights for humans. Moreover, these models can encode concepts and algorithmic pathways that can be subject to direct manipulation and editing [42, 81, 98]. In subsection 2.1, we turn to an actionable review of the methods in mechanistic interpretability (MI) that enable such an approach by working with internal representations.
Background
To describe the available tools and methods of NRI, we conduct a review of the relevant methods and techniques of MI that also lend themselves to notions of actionable interpretability [87]. As such, we conduct a review of MI work that could both inform insights that enrich our comprehension of models and guide actions beyond interpretability, including improved model calibration, understanding of other scientific phenomena, and enhanced user experiences.
We then also survey problems and risks in the current state of HAI that NRI could stand to help address.
Actionable Methods in Mechanistic Interpretability
Probes and steering vectors. There is a large body of evidence that features within LLMs are represented as linear directions in the vector space created by its activation values [36, 85, 141]. Various concepts and behaviors, such as refusal [7], sentiment [123], truth [76], political beliefs [58], and spatial and temporal relationships [48] have been shown to be effectively mediated using such linear representations. Probes can be simple linear or logistic regression models that learn these linear directions of these concepts or behaviors. Activation values have also been projected onto linear vector directions to reveal expressions of persona traits such as evil, sycophancy or hallucination [21]. Moreover, activation steering [98, 124, 134] has been used to push activation values in representative directions to induce desired behavior in the model response.
Verbalization methods. Verbalization is a burgeoning direction that aim to decode neural activations into natural language. This represents a promising and scalable direction to elicit privileged or hidden communication directly from activations while subjecting opaque computations to open-ended query and inspection [65]. Verbalization methods first extract the latent representations of a target model and provide them to an auxiliary LLM tasked with an inspection prompt. Representations are provisioned in various ways, whether patched into the auxiliary model’s forward pass [44], concatenated into model activations [55, 89], or supplied directly in-context as token embeddings [19, 41]. Depending on the correspondence between target and auxiliary architecture, the auxiliary model may be trained on an unsupervised objective of reconstructing the activations, as in Natural Language Autoencoders [41]; finetuned on supervised question & answer pairs about the target’s context, as in LatentQA and Activation Oracles [55, 89]; or used training-free, as in Patchscopes and SelfIE [19, 44]. Currently, these techniques are integrated into interfaces primarily used by expert human auditors for safety evaluation of target model’s latent dispositions such as unverbalized evaluation awareness [41] and hidden motivations [41, 77].
Sparse autoencoders and dictionary learning. Unlike supervised probing, sparse autoencoder (SAEs) and related variants are unsupervised dictionary-learning approaches that aims to decompose features into a sparse linear combination of a higher-dimensional feature space. SAEs are trained with a reconstruction objective and sparsity penalty, under the hypothesis that a sufficiently sparse code naturally decomposes activations into monosemantic, human-interpretable concepts [36]. Transcoders [33] extend this idea from representations to computation, learning a sparse dictionary that maps a component’s input to its output; crosscoders [70], another variant, instead learn a single dictionary shared across multiple layers or models, so that each feature carries a decoder direction at every site. These methods have been used as general purpose tools to decompose and intervene on internal representations, having been trained at scale for frontier LLMs such as GPT-4 and Claude 3 Sonnet [43] [122]. SAEs have also enabled applications across multimodal processing. In auditory domains, SAE features can underpin acoustic and paralinguistic representations, used to steer models to mitigate false speech predictions, and recover human-interpretable auditory concepts [6, 88]. In computer vision, they supported suppression and debiasing of harmful and unwanted visual content [29, 51], enabled precise concept editing [57, 137], and extraction of interpretable features in medical imaging [130].
Attribution. Attributional interpretability aims to characterize which components or features within a model are responsible for a given behavior or output. Different families of attribution methods quantify the extent to which particular factors drive model behavior, including input features [90, 119], internal components [17], and training data [20]. Context attribution has enabled human verification of model citation [27] that surface context saliency has been used to expose gender bias in machine translation [105]. Internal attribution has helped enrich understanding of internal mechanisms such as planning, multi-step reasoning, and refusal [69]. Training-data attribution has allowed practitioners to audit which finetuning examples give rise to a behavior and modulate the learning signal accordingly during post-training [13, 20]. Component attribution has also further supported circuit discovery, the identification of a sparse sub-networks that faithfully recover model behaviors [28].
Comparison of MI with Explainable AI
Explainable AI (XAI) [34] methods and MI share the aim to explain and make model behaviors more transparent to users. However, unlike XAI, MI is uniquely concerned with describing how the model behaves through its internal mechanisms [12]. XAI can also include black-boxed approaches, where analysis only operates on input and output values as in LIME [97] or SHAP [73]. MI has the ambitious objective of reverse-engineering the neural network’s implemented algorithms into causal mechanisms and circuits. Moreover, post-hoc XAI methods have advocated for model-agnostic mechanisms and scalability of methods, which differ from MI which can yield different results for different DNNs. While, notion of "mechanistic" has gone lax [104], moving from definitions of causal analysis to general interaction with model internals, we adopt the definition of parsimonious and faithful explanation through characterization of internal structure and computation. Both communities have used saliency maps, counterfactuals (i.e., interchange intervention), gradient-based attribution, and autoencoders, while circuit analysis and causal scrubbing remain largely confined to MI [49].
Problems and Risks in Human-AI Interaction
Interpretability and transparency have been theorized as ways to remedy the psychological and cognitive risks that can emerge from AI interaction. These interaction risks include over-reliance or the tendency to over-trust model responses despite accuracy [9, 16], and anthropomorphization or the ascribing of human and personal qualities to AI models [1, 45, 74]. Such tendencies can exacerbate vulnerability to sycophancy [24], emotional dependence [25, 95], cognitive offloading, reduce prosocial intentionality [24], and more. By providing interfaces that expose algorithmic decision-making, users are expected to better understand and calibrate to model responses. In practice, past paradigms such as XAI—while able to affect user experience, trust, and comprehension—often negligibly affects over-reliance [9], automation bias [108, 125], task performance [107], or calibrated trust [63].
More recently, cognitive forcing functions, or an interaction design aimed to promote Systems 2 thinking and deeper engagement with explanations, can reduce over-reliance [16]. Moreover, explanations that arise when initiated by users [114, 115] or during incidences of user expectation violation [22, 59, 114] are particularly effective. Additionally, studies that use white-box-based interpretability methods, such as reporting model token probabilities could help improve the calibration gap between model and user confidence [118] and improve user’s mental schema of model behavior [56]. Interfaces that support these interactions, if properly designed, could allow users to invest cognitive effort to form hypotheses and integrate wider evidence from the model’s neural signatures. Keeping the human in the cognitive loop has been associated with improved logical discernment [30] and decision-making [83]. In addition, they have the potential to expand the possibilities that novel interaction design mechanisms that strengthen user comprehension, calibrate trust, and reduce over-reliance.
Design Space Framework
We present a design framework to conceptualize how NRI can enable users to interact with representations or parameters of neural networks as an interaction surface. At its core, this paradigm leverages the semantic and functional richness of representational bits, presenting ways to allow users to comprehend and control them to create new human-AI experiences.
Methodology for construction of design space. Given the vast solution space of NRI that MI and HAI could open up, we drew on General Morphological Analysis (GMA) [99, 142] to conceptualize a forward-looking design space for current and future methods that intersect these fields. GMA is a method for reasoning about problems that are combinatorial and resist quantification by decomposing a domain into a set of facets, each admitting a discrete range of values, such that the full space of possibilities is the Cartesian product of those facets [99, 142]. To characterize the design space, we extracted and analyzed recent surveys and scoping reviews of the MI landscape [12, 96, 106, 111] recently published from 2024-2026. We extracted all interpretability techniques and grouped techniques by corresponding families. We then, filtered based on the actionability principle by retaining if source literature demonstrates use beyond explanation and descriptive work (subsection 2.1). Finally, we conducted affinity mapping to extract the defining categories and parameters, paying careful attention to the actionability principle (subsection 2.1) and cross-consistency principle of GMA (i.e., cell products are feasibly combined) [99].
Applying GMA, we were able to decompose existing works with respect to a given morphology and then recombine ideas to craft new analytical solutions. Our framework aims to be expansive and forward-looking by allowing ideas and methods between HAI and MI to mix and cross-pollinate to generate new ideas and applications. For instance, our review of MI from the lens of actionability revealed that much work has focused on AI safety across diverse methods tailored for technical users. Our framework enables us to also ask whether we can leverage the toolset from MI to develop tools accessible for lay users. Specifically, can we combine verbalization and multi-modal interfaces to create new affordances, interactions, and enriching user experiences? Could we provide visibility to inspect model confidence for calibrated user decision-making? Given the current emphasis of actionability on AI safety and personalization (see subsection 2.2), we hope our framework inspires novel interactions and synergies between different methods, applications, and visualization mediums to broader user groups.
We describe our framework (section 3), which features 4 categories of Interaction, Target, Referent, and Purpose:
Interaction. How does the user engage with the model internals? We identified three modes of interaction that enable users to interact with model internal representations: Inspect, Steer, and Edit. Inspect allows users to read the internal representations without altering them. Edit refers to targeted parametric or representational modifications via precise edits and ablations. Steer enables users to induce or alter model behavior using changes in representations. Generally, steering refers to behavior on a continuous space and can be intervened based on magnitude.
Target. What internals or latents are accessed in the model? These internal features can be activations, components, circuits, or parameters. These are also the representational bits described earlier. We use activations to broadly describe the representations of a neural network encoded in its activation values at each layer, components as the contributions or outputs of each of its constituent architectural units. Within a transformer architecture paradigm, these may include attention heads, feedforward networks, or convolutional channels. In this sense, some models like diffusion models may not have components. Circuits are a minimal subgraph of a neural network that implements a behavior. Finally, parameters are the weights or biases of a neural network that encode learned knowledge from training.

Figure 2. A multidimensional design space framework for leveraging mechanistic interpretability to enhance human-AI interactivity.
Referent. What are the internals accessed taken to represent? Model representations have been understood to represent concepts, which are specific semantic features or factual knowledge; encode algorithms or computations that accomplish complex tasks such as indirect object identification; store knowledge or characteristics of user information akin to theory of mind construction; and are responsible for or contribute to a specific model behavior in output.
Purpose. What are the outcomes or desiderata of human-AI interaction we wish to aim toward? The broad categories identified are Mechanistic Understanding as in promoting intuition over model latents and implemented algorithms; Sensemaking that uses interpreted machine concepts or solutions to further human understanding of some phenomenon; Decision Support to provide quantifiable or structured signals to calibrate user’s confidence or decision-making; Auditing which helps debug and surface failure modes or biases; Control to modulate model behavior in such a way as to offer adaptive, personalized, and appropriate experiences to users.
Cross-dimension considerations. Across the five dimensions, we also emphasize two overarching facets worthy of consideration that do not fit as another dimension in our framework. The first is the end user at hand. For instance, a technical user may want to audit a circuit to study a model behavior, whereas a non-technical user who asks for relationship advice may want to see which parts of their personal experiences induce the model to respond in a particular way (see 4.1). Progressive disclosure design methodologies and hypothesis-feedback design theories provide principled ways to provide suitable interfaces for users by considering levels of complexity and iterative cycles of work. Secondly, we should consider second-order effects that emerge beyond static interaction. In other words, are there interaction effects and dynamics that emerge from the interplay between representational bits and user cognition? Could enabling users to programmatically interface and direct model responses lead to new understanding of AI systems? Such questions can enrich our understanding of the affordances and drawbacks to help us refine this new paradigm of interaction. Finally, applying Zwicky’s cross-consistency principle, that is ensuring all condition choices between dimensions can be feasibly recombined, we note
that steering is not compatible with weight space. Parameter have soft incompatibility issues with user information which is often measured with the model activations.
Visualizations. Different selections naturally give rise to different rendering mediums and ways to visually express certain features. For instance, spatial clusters or cluster-based visualizations are a natural medium to express multidimensional features in a way that preserves spatial relations. These can be applied to analysis with activations and the parameter space. Directed acyclic graphs (DAGs) can be used to represent attribution graphs and circuits by showing the cross-layer pathways of interconnected features [70]. Grids have been used by methods like logit lens to show important discrete and qualitative aspects at different time steps [86]. Other visualization methods like dendrograms could be used to visually describe symbolic relationships that compose across layers with increasing levels of abstraction [78].
Applying Design Space to Current Tools
We now turn our examination to previous, yet limited, work in intersection between MI and HCI to consider what directions might naturally arise in opportunities available. For each corresponding visualization, we apply our design space framework Figure 1 to showcase the general directionality most tools have been focused on.

Figure 1. Mapping and conceptualizing the design space of neural representation interaction.
Interfaces for Everyday Users
The primary motivations for AI use range from companionship [84, 101, 136, 139] and advice-seeking to domain-specific task execution [18, 50]. Much work shows lay users generally prefer signals of model confidence and context [14, 32, 138] since their aim is to gain calibrated trust [63] and control [35] over their own interactions rather than an understanding of model internals. Notably, this involves progressively disclosing [115] users to traits or characteristics that may mediate model consistency, performance, and verifiability. Current HCI applications have used dashboards to offer users a way to characterize their own interactions and form mental models of AI systems [53, 54, 127]. For instance, NEURAL
| Interface | Interaction | Target | Referent | Purpose |
| GAN Dissect [11] | Inspect, Edit | Components | Concept | Mechanistic Understanding, Control |
| GAN Paint [10] | Inspect, Edit | Components | Concept | Control |
| TalkTuner [23] | Inspect, Steer | Activations | User information | Control, Auditing |
| Neural Transparency [54] | Inspect | Activations | Behavior | Decision support |
| Multi-Turn Neural Transparency [53] | Inspect | Activations | Behavior | Decision support |
| Circuit Tracer [3] | Inspect, Edit | Circuits, Components | Computation, Behavior | Mechanistic understanding |
| Assistant Axis [72] | Edit | Activations | Behavior | Control, Auditing |
| Goodfire Ember [46] | Inspect, Steer | Activations | Concept, Behavior | Control, Auditing |
| ConceptViz [64] | Inspect, Steer | Activations | Concept | Sensemaking |
| SAE Semantic Explorer [135] | Inspect | Activations | Concept | Sensemaking |
| Concept Explorer [75] | Inspect | Activations | Concept | Sensemaking |
Table 2. Examples of existing interpretability interfaces situated in the four dimensions of the NRI design space. Only TalkTuner, Neural Transparency, Multi-Turn Neural Transparency fall under interfaces for lay users.
Transparency presents sunburst visualizations that display predicted safety-relevant trait expressions in a custom chatbot using persona vectors derived from system prompts [54]. A follow-up, Multi-Turn Neural Transparency, provides dynamic interfaces that continuously update over multi-turn conversations [53]. These novel technologies primarily emphasize inspection of model activations to probe for personality traits that may influence model responses with a primary goal of calibrating the user’s mental understanding of model behavior. Under our design framework, Neural Transparency and Multi-Turn Neural Transparency fulfill the conditions of Inspect, Activations, Behavior, Decision support. Another interactive system, TalkTuner, enables general users to view and directly manipulate the model’s internal representation of their own demographic attributes, and was found to improve transparency and user experience [23]. TalkTuner, thus, enables users to Inspect and Steer the linear directions of the activation space that encode User information in order to Control responses and Audit how behavior may be conditioned on inherent demographic biases.
Interfaces for Auditors and Safety Researchers
To date, where MI interfaces have been designed, auditors and safety researchers are the primary beneficiaries of these novel toolings [3, 46, 67, 68]. Unlike interfaces for general audiences, these tools expose model internals and provision technical users with raw weights, features, activations, and circuit-level connections. Moreover, these interfaces provide greater degrees of freedom which enables more granular exploration of novel methodologies on diverse task settings and applications. These systems are often interactionally bidirectional; as such, they allow users to perform read-and-write operations [3, 69]. Often the principle Purpose that motivates the design of the tooling is the need for rapid hypothesis-testing feedback loop and refining researcher’s raw intuition of methodologies and behavior. One example is Circuit Tracer which allows end users to Inspect by tracing Circuits in a direct graph and Edit activations in order to see the effects on the downstream computation and behavior. While this visual modality represented internal space with nodes and edges, other tools abstract away that so provide other qualitative insights integrated into a chatbot interface. For instance, Assistant Axis [72] provides counterfactual views of interactive chat interfaces with and without capped activations along a helpful, "assistant" dimension. From the NRI’s design framework, Assistant Axis enables users to Edit a low-rank Activation subspace that provides assistant-like mannerisms to Control persona drift. The counterfactual views satisfy the Auditing criterion by exposing latent biases. These tools are deployed on Neuropedia, a shared infrastructure to demo MI interfaces, and so have been accordingly designed with an emphasis on providing users raw intuition of model behavior and methodological details.
Speculative Applications for Designing with NRI
NRI invites us to imagine AI internals as a new material for HAI that users can explore, manipulate, and interact with. To demonstrate the potential of NRI, we conjecture how this interaction framework could revolutionize disparate fields such as education, critical decision making, personal reflection, and AI oversight by exploring speculative applications designing for HAI with NRI. These scenarios build on emerging methods for extracting interpretable features, intervening on activations, tracing computational circuits, and testing them through causal experiments. By providing access to model internals, these applications could enable forms of inquiry and control that is difficult or impossible in current interfaces. Collectively, they envision a future in which people have greater agency to shape how AI understands the world and themselves, and to exercise that greater agency in working with the AI systems. We consider both the benefits and potential harms that such a paradigm could result in if it came to fruition.
Mechanistic Education: Shakespeare Writing Studio
Shakespeare’s writing is celebrated for its extraordinary command of language. However, the qualities that make it remarkable also make it intimidating to study for a student new to such work. The rise of generative AI has unfortunately reduced the incentives for learning how to write in this style, since students can simply ask the models to imitate it. Deconstructing Shakespeare Studio would apply a model’s capacity for language modeling and self-explanation to this problem, letting students inspect literary topics such as the syntax, rhythm, imagery, and meaning underlying a passage’s style with explanations at different levels of sophistication and nuance. Such an application could decompose the literary concept of "Shakespearean" into combinations of sparse, multidimensional concepts that represent different semantic and syntactic notions. Rather than enabling the cognitive offloading of a student’s understanding, such a tool can enrich and deepen learning; the model’s representation of languages can be dissected into constituent interpretable components for human learning.
Shakespeare’s style has been extensively characterized in terms of identifiable methods: metrical patterning in blank verse, including feminine endings and shared lines [133]; hendiadys [132]; classical rhetorical figures such as antithesis, anaphora, chiasmus, and hyperbaton [2, 52, 79]; functional shift, in which words migrate across grammatical categories [31]; and more [15, 26, 116, 126]. A student examining a passage from Macbeth could select lines and compare the constituent components that are highly activated and track how use of these different literary devices varies across samples. Students could further explore if any new abstractions are represented in the model, aiding in their own analysis. This same approach could generalize beyond Shakespeare to any written work of any genre, author, or subject: the rhetoric of celebrated speeches, the argumentative structure of noted essayists, or the harmonic and rhythmic patterns studied in music education, letting students connect any passage to broader questions or deconstruct how style and persuasion are produced for their own personalized learning.
Under the design space framework, students could first inspect passage-specific activations, then steer models at different points in the input sequence to see how texts alter due to earlier changes. Allowing users to understand the stylistic fluency in this way, in terms of first principles, is an example of literary Sensemaking. Moreover, this example illustrates the sensemaking process and places the emphasis on decoding and reinterpreting texts with the underlying motifs and literary building blocks.
Transparent User Models: Latent Mirror
One step beyond current user modeling done by LLMs would be a Transparent User Model. Advances in AI systems and HCI techniques could lead to models that can learn your habits, preferences, and identity in personalized ways from multiple different modalities of data, such as phone and computer usage, internet of things, and AI assistant interaction. Simultaneously, advances in ML could enable us to interpret the representations models use in their hyper-personalized and aware user model. Such user models could lead to interfaces that allow users to query their own models of themselves: users could ask models what the system believes to be recurring behaviors or decision-making, grounded in traceable internal representations built from months or years of accumulated interactions [71, 140].
We imagine Latent Mirror, an application that radically extends work like TalkTuner (see § 4). Latent Mirror would allow users to inspect, attribute, and modify representations of themselves within the model’s computations. For instance, a user might begin with a concern such as, “Why does the assistant keep steering me toward low-sodium recipes and shorter walking routes?” after noticing the pattern across several unrelated conversations. Latent Mirror would retrieve relevant portions of the interaction history and identify candidate internal features associated with this recurring response pattern. The application might propose the interpretation of an inferred chronic health condition, accompanied by the excerpts that led the model to build such an implicit profile—here, a symptom mentioned once in passing and never raised again.
Correspondingly, Latent Mirror could allow the model to programmatically steer by capping the model’s activations [72] or to inhibit the model’s assumption of a user behavior. Interventions in model responses that reduce persona drifts or alter the model’s implicit profile of the user could affect downstream responses. The user could test its functional role by suppressing the responses and comparing the differences in responses prior to intervention. In terms of the design space, Latent Mirror would support inspection and even steering or persistent editing. Its primary targets would be contextual activations that encode certain identify-relevant concepts. These targets would refer to inferred user information, including preferences, goals, emotional cues, and interaction patterns, as well as the model behavior influenced by those inferences. The resulting interaction would demonstrate control, as they alter the user model that is formed. Modifications of models’ belief state could also be used to support more supportive and healthier human-AI interaction dynamics, such as extending to use cases like adaptive tutoring or accessibility.
Causal Auditor: Auditing Bias in Hiring
AI-assisted hiring systems can influence how recruiters interpret résumés and allocate attention across a candidate pool [5, 131]. A recommendation may appear reasonable because it comes with a fluent explanation, yet that explanation may not properly attribute the computation or reasoning that produced it. For instance, a model might cite limited leadership potential while actually making its decision based on institutional prestige, career gaps, or names that serve as demographic proxies. Conversely, a model detecting demographic information internally does not mean that information causally impacted the recommendation. This gap makes it difficult for recruiters or auditors to determine when an assessment deserves scrutiny.
Causal Auditor is a speculative application that would help resume reviewers and auditors investigate such factors in opaque hiring decisions, enabling the investigation of testable hypotheses about how the model used the evidence at hand to make its determination. NRI breaks this opacity, allowing users to scrutinize representations and circuits to aid in ensuring decisions are made equitably, fairly, and effectively, targeting representations responsible for bias similar to prior work [128]. For instance, a reviewer might select an outputted model justification such as “The candidate demonstrates limited leadership potential.” The interface would attribute this sequence to résumé excerpts that influenced it and surface internal features tied to management experience, institutional prestige, or employment gaps. Each of these would be backed by examples and counterexamples from an evaluation corpus [94], and counterfactuals could be tested to more rigorously establish causality. The reviewer could compare two matched résumés that differ only in a suspected proxy, such as an institutional name, and Equity Lens would evaluate where their internal pathways diverge. Swapping concepts between the two traces and observing whether the assessment shifts tests that pathway causally [82]. This would mechanistically and faithfully diagnose whether systemic biases exist in such models, ensuring transparency in previously opaque systems. These models could then be steered or these biased representations suppressed in activation to move closer to an unbiased decision, mirroring previously done interventions [141].
Within the design space, Equity Lens would enable reviewers to inspect internal representations and edit them within a controlled audit environment. Its targets would include activations tied to application evidence, components, and circuits that link sensitive conceptual referents such as gender, name, race, or biographical information to output. This supports auditing and decision support: understanding how an assessment was produced, and using that to calibrate reliance or revise decisions. The system would keep an audit record of the concern, targets, interpretations, interventions, and alternative explanations, usable for independent review. Similar approaches could extend to admissions, credit, and benefits review, anywhere outcomes may need to be properly adjudicated to avoid use of protected characteristics.
Limitations
We present NRI as a forward-looking paradigm, and its limitations are correspondingly both technical and socio-technical. There are fundamental assumptions such as model architecture, maturity of underlying methods, access provisioning, and how current interfaces present representational bits.
Architectural assumptions of the design space. Our design space, and much of the MI literature it synthesizes, is grounded in transformer language models. The Target facet in particular presumes a residual stream, discrete attention and feedforward components, and circuits as subgraphs over them. Transferring such architectural prescriptions to other architectures could not be possible. Steering, latent-prediction lenses, and factual editing have been shown to carry over to state space and recurrent models such as Mamba and RWKV, though component-level analysis must be reformulated around a compressed state rather than attention heads [93, 112]. In diffusion models, dictionary learning has enabled concept suppression and editing [29, 57], but there is no comparative analogy for a circuit. For world models [91], interpretability tooling is largely absent. As such, the cross-consistency of our facets is established for transformers but may not hold for others; some cells of the design space may not be comparable or require redefinition for non-transformer systems.
Methodological progress and reliability. The methods reviewed in subsection 2.1 are actively being developed and improved. For instance, probes and steering vectors rest on the linear representation hypothesis, yet not all features seem to be linear [37], motivating ongoing work on manifold-based steering [134]; existing steering vectors generalize inconsistently beyond the distribution they were derived from and produce unintended side effects on unrelated behavior [121]. SAEs may not recover ground-truth and human-interpretable features as concepts can split across latents as dictionary size grows [62]. Verbalization requires an auxiliary module trained or prompted to describe activations and may risk confabulating descriptions that the target’s computation does not support [44, 89]. Many open technical and reliability problems of MI constrain the feasibility of what current NRI interfaces may look like.
Sociotechnical risks of NRI.
Data provenance and personalization. Our speculative propositions, most directly Transformer User Models and Latent Mirror (subsection 5.2), raise questions of data provenance that current privacy policies may not adequately address. Representations built from accumulated interactions can encode inferences the user never stated, and LLMs already infer sensitive attributes such as location, income, and demographics from innocuous text [117]. The ability for personalized support carries a dual-use risk, since this also could enable model providers or other actors to conduct precise tailoring of messages [103] and could be used for targeted advertisement. NRI interfaces operating on user information should adhere to the rights of access, rectification, erasure, and data portability. Future work should develop privacy and ethical frameworks specific to representational bits.
Anthropomorphization. How representational bits are interfaced with may also promote the anthropomorphization which NRI aims to counter (subsection 2.2). Verbalization could convey activations in the grammar of mental states as reporting what a model "believes" or "wants" invites users to treat statistical structure as intention [1, 110]. These novel interfaces have an opportunity to be careful with their communication by describing internals in the language of algorithm and computation. Additionally, the interaction methods, such as editing, could be presented in a way that portrays a sense of determinism. Whether such framing do in fact reduce anthropomorphism is a potential empirical question.
Conclusion
We introduce NRI as a novel human-AI interaction framework. Our paper draws on affordance theories and past interaction paradigms to contextualize this new mode of interactivity. In addition to our design space, we also provided speculative examples and mapped existing work to our framework. As such, NRI aims to present a forward-looking and opportunistic vision: what are different ways we may interact with AI systems? How are our experiences and interpretations coupled with representational bits? How can we design systems that take advantage of such representational affordances? This paper is an attempt to introduce these new conceptual questions and provide HCI and HAI researchers with abstractions to think in novel modes of interactivity. Together, our study aims to augment human-AI interaction by transforming AI systems from inscrutable algorithms to transparent partners.
References
References
- [1]Gavin Abercrombie, Amanda Cercas Curry, Tanvi Dinkar, Verena Rieser, and Zeerak Talat. 2023. Mirages. On Anthropomorphism in Dialogue Systems. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, Houda Bouamor, Juan Pino, and Kalika Bali (Eds.). Association for Computational Linguistics, Singapore, 4776–4790. doi:10.18653/v1/2023.emnlp-main.290DOI
- [2]Sylvia Adamson, Lynette Hunter, Lynne Magnusson, Ann Thompson, and Katie Wales (Eds.). 2001. Reading Shakespeare’s Dramatic Language: A Guide. Arden Shakespeare, London.
- [3]Emmanuel Ameisen, Jack Lindsey, Adam Pearce, Wes Gurnee, Nicholas L. Turner, Brian Chen, Craig Citro, David Abrahams, Shan Carter, Basil Hosmer, Jonathan Marcus, Michael Sklar, Adly Templeton, Trenton Bricken, Callum McDougall, Hoagy Cunningham, Thomas Henighan, Adam Jermyn, Andy Jones, Andrew Persic, Zhenyi Qi, T. Ben Thompson, Sam Zimmerman, Kelley Rivoire, Thomas Conerly, Chris Olah, and Joshua Batson. 2025. Circuit Tracing: Revealing Computational Graphs in Language Models. Transformer Circuits Thread (2025). https://transformer-circuits.pub/2025/attribution-graphs/methods.html
- [4]Saleema Alshemi, Maya Kacimak, W. Bradley Knox, and Todd Kulesza. 2014. Power to the People: The Role of Humans in Interactive Machine Learning. AI Magazine 35, 4 (2014), 105–120. doi:10.1609/aimag.v35i4.2513
- [5]Eitan Anzenberg, Arunava Samajpati, Sivasankaran Chandrasekar, and Varun Kacholia. 2025. Evaluating the promise and pitfalls of llms in hiring decisions. arXiv preprint arXiv:2507.02087 (2025).arxiv.org/abs/2507.02087
- [6]Georgii Aparin, Tasmina Sadekova, Alexey Rukhovich, Assel Yermekova, Laida Kushnareva, Vadim Popov, Kristian Kuznetsov, and Irina Piontkovskaya. 2026. AudioSAE: Towards Understanding of Audio-Processing Models with Sparse AutoEncoders. In Proceedings of the 19th Conference of the European Chapter of the Association for Computational Linguistics (Volume 1: Long Papers). Association for Computational Linguistics, Rabat, Morocco, 3221–3254.
- [7]Andy Arditi, Oscar Obeso, Aaquiib Syed, Daniel Paleka, Nina Panickssery, Wes Gurnee, and Neel Nanda. 2024. Refusal in language models is mediated by a single direction. Advances in Neural Information Processing Systems 37 (2024), 136037–136083.arxiv.org/abs/2406.11717
- [8]Shuai Bai, Yuxuan Cai, Ruizhe Chen, Keqin Chen, Xionghui Chen, Zesen Cheng, Lianghao Deng, Wei Ding, Chang Gao, Chunjian Ge, et al. 2025. Qwen3-VL Technical Report. arXiv preprint arXiv:2511.21631 (2025). https://arxiv.org/abs/2511.21631
- [9]Gagan Bansal, Tongshuang Wu, Joyce Zhou, Raymond Fok, Besmira Nushi, Ece Kamar, Marco Tulio Ribeiro, and Daniel S. Weld. 2021. Does the Whole Exceed Its Parts? The Effect of AI Explanations on Complementary Team Performance. In Proceedings of the 2021 CHI Conference on Human Factors in Computing Systems (CHI ’21). ACM, New York, NY, USA, Article 81, 16 pages. doi:10.1145/3411764.3454717arxiv.org/abs/2006.14779
- [10]David Bau, Hendrik Strobelt, William Peebles, Jonas Wulff, Bolei Zhou, Jun-Yan Zhu, and Antonio Torralba. 2019. Semantic Photo Manipulation with a Generative Image Prior. ACM Transactions on Graphics (Proc. SIGGRAPH) 38, 4, Article 59 (2019). doi:10.1145/3306346.3323023DOI
- [11]David Bau, Jun-Yan Zhu, Hendrik Strobelt, Bolei Zhou, Joshua B. Tenenbaum, William T. Freeman, and Antonio Torralba. 2019. GAN Dissection: Visualizing and Understanding Generative Adversarial Networks. In International Conference on Learning Representations (ICLR). arXiv:1811.10597 https://openreview.net/forum?id=Hyg_X2C5FXarxiv.org/abs/1811.10597
- [12]Leonard Bereska and Efstratios Gavves. 2024. Mechanistic interpretability for AI safety—a review. arXiv preprint arXiv:2404.14082 (2024).arxiv.org/abs/2404.14082
- [13]Leon Bergen, Usha Bhalla, Sidharth Baskaran, Max Loeffler, Raphael Sarfati, Dhruvil Gala, Ryan Panwar, Santiago Aranguri, Thomas Elticus Geiger, Matthew Kowal, Siddharth Boppana, Daniel Balsam, Owen Lewis, Jack Merullo, Thomas McGrath, and Ekdeep Singh Lubana. 2026. ATM: Automated Post-Training: Using Interpretability to Characterize Data and Shape the Learning Signal. arXiv preprint arXiv:2606.12360 (2026). arXiv:2606.12360 [cs.LG]arxiv.org/abs/2606.12360
- [14]Umang Bhatt, Javier Antorán, Yunfeng Zhang, Q. Vera Liao, Prasanna Sattigeri, Riccardo Fogliato, Gabrielle Melançon, Ranganath Krishnan, Jason Stanley, Omesh Tickoo, Lama Nachman, Rumi Chunara, Madhulika Srikumar, Adrian Weller, and Alice Xiang. 2021. Uncertainty as a Form of Transparency: Measuring, Communicating, and Using Uncertainty. In Proceedings of the 2021 AAAI/ACM Conference on AI, Ethics, and Society (AIES), 401–413. doi:10.1145/3461702.3462571 Communicating uncertainty (not just point estimates) aids user decisions; verify pages against final version.DOI
- [15]Stephen Booth (Ed.). 1977. Shakespeare’s Sonnets. Yale University Press, New Haven.
- [16]Zana Buçinca, Maja Barbara Malaya, and Krzysztof Z. Gajos. 2021. To Trust or to Think: Cognitive Forcing Functions Can Reduce Overreliance on AI in AI-assisted Decision-making. Proc. ACM Hum.-Comput. Interact. 5, CSCW1, Article 188 (April 2021), 21 pages. doi:10.1145/3449287DOI
- [17]Lucius Bushnaq, Dan Braun, and Lee Sharkey. 2025. Stochastic Parameter Decomposition. arXiv:2506.20790 [cs.LG]arxiv.org/abs/2506.20790
- [18]Aaron Charterji, Tom Cunningham, David J. Deming, Zoë Hitzig, Christopher Ong, Carl Yan Shan, and Kevin Wadman. 2025. How People Use ChatGPT. Working Paper 34255. National Bureau of Economic Research. doi:10.3386/w34255 Introduces the “Asking, Doing, Expressing” taxonomy; Asking (advisory) 49% and Doing (task) 40% of messages.
- [19]Haozhe Chen, Carl Vondrick, and Chengzhi Mao. 2024. SelfIE: Self-Interpretation of Large Language Model Embeddings. In Proceedings of the 41st International Conference on Machine Learning (ICML). https://arxiv.org/abs/2403.10949
- [20]Jianhui Chen, Yuzhang Luo, and Liangming Pan. 2026. Mechanistic Data Attribution: Tracing the Training Origins of Interpretable LLM Units. arXiv:2601.21996 [cs.CL]arxiv.org/abs/2601.21996
- [21]Runjin Chen, Andy Arditi, Henry Sleight, Owain Evans, and Jack Lindsey. 2025. Reasoning over text: Monitoring and controlling character traits in language models. arXiv preprint arXiv:2507.21509 (2025).arxiv.org/abs/2507.21509
- [22]Valerie Chen, Q. Vera Liao, Jennifer Wortman Vaughan, and Gagan Bansal. 2023. Understanding the Role of Human Intuition on Reliance in Human-AI Decision-Making with Explanations. Proceedings of the ACM on Human-Computer Interaction 7, CSCW2, Article 370 (Oct. 2023), 32 pages. doi:10.1145/3610219DOI
- [23]Yida Chen, Aoyu Wu, Trevor DeP0desta, Catherine Yeh, Kenneth Li, Nicholas Castillo Marin, Oam Patel Jan Riecke, Shivam Raval, Olivia Seow, et al. 2024. Designing a dashboard for transparency and control of conversational AI. arXiv preprint arXiv:2406.07882 (2024).arxiv.org/abs/2406.07882
- [24]Myra Cheng, Cinoo Lee, Pranav Khadpe, Sunny Yu, Dyllan Han, and Dan Jurafsky. 2026. Symptomatic AI decreases prosocial intentions and promotes dependence. Science 391, 6792 (2026), eae8352.DOI
- [25]Minh Duy Chu, Patrick Gerard, Kshitij Pawar, Charles Bickham, and Kristina Lerman. 2025. Illusions of Intimacy: How Emotional Dynamics Shape Human-AI Relationships. arXiv preprint arXiv:2505.11649 (2025).arxiv.org/abs/2505.11649
- [26]Wolfgang Clemen. 1987. Shakespeare’s Soliloquies. Methuen, London.
- [27]Benjamin Cohen-Wang, Harshay Shah, Kristian Georgiev, and Aleksander Madry. 2024. ContextCite: Attributing Model Generation to Context. In Advances in Neural Information Processing Systems (NeurIPS), Vol. 37. doi:10.5220/079017-3035
- [28]Arthur Comny, Augustine Mavor-Parker, Aengus Lynch, Stefan Heimsherim, and Adria Gárriga-Alonso. 2023. Towards automated circuit discovery for mechanistic interpretability. Advances in Neural Information Processing Systems 36 (2023), 16318–16352.
- [29]Bartosz Cywiński and Kamil Deja. 2025. SAEuron: Interpretable Concept Unlearning in Diffusion Models with Sparse Autoencoders. In Proceedings of the 42nd International Conference on Machine Learning (ICML).arxiv.org/abs/2501.18052
- [30]Valdemar Darny, Pat Pataranutaporn, Yaoli Mao, and Pattie Maes. 2023. Don’t Just Tell Me, Ask Me: AI Systems that Intelligently Frame Explanations as Questions Improve Human Logical Discernment Accuracy over Causal AI Explanations. In Proceedings of the 2023 CHI Conference on Human Factors in Computing Systems (CHI ’23). Association for Computing Machinery, New York, NY, USA, Article 352, 13 pages. doi:10.1145/3544548.3580672
- [31]Philip Davis. 2007. Shakespeare Thinking. Continuum, London.
- [32]Yifan Ding, Matthew Facciani, Amrit Poudel, Ellen Joyce, Salvador Aguiñaga, Balaji Veeramani, Samirtra Bhattacharya, and Tim Weninger. 2025. Citations and Trust in LLM Generated Responses. In Proceedings of the AAAI Conference on Artificial Intelligence, Vol. 39, 23787–23795. doi:10.1609/aaai.v39i12.34550 Citations raise self-reported trust even when random; trust drops when users check them.arxiv.org/abs/2501.01303
- [33]Jacob Dunefsky, Philippe Clenski, and Neel Nanda. 2024. Transcoders find interpretable llm feature circuits. Advances in Neural Information Processing Systems 37 (2024), 24375–24410.arxiv.org/abs/2406.11944
- [34]Rudresh Dwivedi, Devam Dave, Het Naik, Smiti Singhal, Rana Omer, Pankesh Patel, Bin Qian, Zhanyu Wen, Tejal Shah, Graham Morgan, et al. 2023. Explainable AI (XAI): Core ideas, techniques, and solutions. ACM computing surveys 55, 9 (2023), 1–33.DOI
- [35]Upol Ehsan, Q. Vera Liao, Samir Passi, Mark O. Riedl, and Hal Daumé III. 2024. Explainable AI: Operationalizing Seamful Design in Explainable AI. In Proceedings of the 2024 CHI Conference on Human Factors in Computing Systems. doi:10.1145/3613904.3642755
- [36]Nelson Elhage, Tristan Hume, Catherine Olsson, Nicholas Schiefer, Tom Henighan, Shauna Kravec, Zac Hatfield-Dodds, Robert Lasenby, Dawn Drain, Carol Chen, Roger Grosse, Sam McCandlish, Jared Kaplan, Dario Amodei, Martin Wattenberg, and Christopher Olah. 2022. Toy Models of Superposition. Transformer Circuits Thread (2022). https://transformer-circuits.pub/2022/ toy-model/index.htmlarxiv.org/abs/2209.10652
- [37]Joshua Engels, Eric J. Michaud, Isaac Liao, Wes Gurnee, and Max Tegmark. 2025. Not All Language Model Features Are Linear. In International Conference on Learning Representations.arxiv.org/abs/2405.14860
- [38]Jerry Alan Fails and Dan R Olsen Jr. 2003. Interactive machine learning. In Proceedings of the 8th international conference on Intelligent user interfaces, 39–45.
- [39]Alhussein Fawzi, Matej Balog, Aja Huang, Thomas Hubert, Bernardino Romera-Paredes, Mohammadamin Barekatain, Alexander Novikov, Francisco J. R. Ruiz, Julian Schrittwieser, Grzegorz Swirszcz, David Silver, Demis Hassabis, and Pushmeet Kohli. 2022. Discovering faster matrix multiplication algorithms with reinforcement learning. Nature 610, 7930 (2022), 47–53. doi:10.1038/s41586-022-05172-4DOI
- [40]George W. Fitzmaurice, Hiroshi Ishii, and William A. S. Buxton. 1995. Bricks: Laying the Foundations for Graspable User Interfaces. In Proceedings of the SIGCHI Conference on Human Factors in Computing Systems (CHI ’05). ACM Press/Addison-Wesley, New York, NY, USA, 442–449. doi:10.1145/2294263.2293964
- [41]Kit Fraser-Taliente, Subhash Kantamneni, Euan Ong, Dan Mossing, Christina Lu, Paul C. Bogdan, Emmanuel Ameisen, James Chen, Dzmitry Kishylau, Adam Pearce, Julius Tarng, Alex Wu, Jeff Wu, Yang Zhang, Daniel M. Ziegler, Evan Hubinger, Joshua Batson, Jack Lindsey, Samuel Zimmerman, and Samuel Marks. 2026. Natural Language Autoencoders Produce Unsupervised Explanations of LLM Activations. Transformer Circuits Thread (2026). https://transformer-circuits.pub/2026/nla/index.html
- [42]Rohit Gandikota, Joanna Materzyńska, Tingrui Zhou, Antonio Torralba, and David Bau. 2024. Concept Sliders: LoRA Adaptors for Precise Control in Diffusion Models. In European Conference on Computer Vision (ECCV).arxiv.org/abs/2311.12092
- [43]Leo Gao, Tom Dupré la Tour, Henk Tillman, Gabriel Goh, Rajan Troll, Alec Radford, Jya Sutskever, Jan Leike, and Jeffrey Wu. 2024. Scaling and evaluating sparse autoencoders. arXiv preprint arXiv:2406.04093 (2024).arxiv.org/abs/2406.04093
- [44]Asma Ghandeharian, Avi Caciularu, Adam Paek, Luca Dax, and Mor Geva. 2024. Patchscopes: A Unifying Framework for Inspecting Hidden Representations of Language Models. In Proceedings of the 41st International Conference on Machine Learning (ICML). https://arxiv.org/abs/2401.06102
- [45]Sofia Gomes, João Lopes, and Elisabete Nogueira. 2025. Anthropomorphism in artificial intelligence: a game-changer for brand marketing. Future Business Journal 11 (01 2025). doi:10.1186/s43093-025-00423-yDOI
- [46]Goodfire. 2024. Goodfire Ember: Scaling Interpretability for Frontier Model Alignment. Goodfire blog. https://www.goodfire.ai/blog/announcing-goodfire-ember
- [47]Google DeepMind. 2025. AlphaFold: Five Years of Impact. Google DeepMind Blog. Accessed 2026-09-06.
- [48]Wes Gurnee and Max Tegmark. 2023. Language models represent space and time. arXiv preprint arXiv:2310.02207 (2023).arxiv.org/abs/2310.02207
- [49]Bálint Gyevnár and Atoosa Kasirzadeh. 2026. Bridging the Gap in the Responsible AI Divide. arXiv preprint arXiv:2603.14495 (2026).
- [50]Kunal Handa, Alex Tamkin, Miles McCain, Saffron Huang, Esin Durmus, Sarah Heck, Jared Mueller, Jerry Hong, Stuart Ritchie, Tim Belonax, Kevin K. Troy, Dario Amodei, Jared Kaplan, Jack Clark, and Deep Ganguli. 2025. Which Economic Tasks are Performed with AI? Evidence from Millions of Claude Conversations. arXiv:2503.04761 [cs.CY] https://arxiv.org/abs/2503.04761 Anthropocentric Economic Index; maps Claude usage onto O*NET occupational tasks (domain-specific execution).
- [51]Sonia Joseph, Praneet Suresh, Ethan Goldfarb, Lorenz Hufe, Yossi Gandelsman, Robert Graham, Danilo Bzdok, Wojciech Samek, and Blake Aaron Richards. 2025. Steering CLIP’s Vision Transformer with Sparse Autoencoders. arXiv preprint arXiv:2405.08729 (2025).
- [52]Sister Miriam Joseph. 1947. Shakespeare’s Use of Arts of Language. Columbia University Press, New York.
- [53]Sheer Karmy, Anthony Baez, and Pat Pataranutaporn. 2026. Multi-Turn Neural Transparency: Surfacing Neural Activations Improves User Calibration to LLM Behavioral Drift. arXiv preprint arXiv:2605.15455 (2026).
- [54]Sheer Karmy, Anthony Baez, and Pat Pataranutaporn. 2026. Neural transparency: Mechanistic interpretability interfaces for anticipating model behaviors for personalized ai. In Proceedings of the 31st International Conference on Intelligent User Interfaces. 868–884.
- [55]Adam Kavon, James Chua, Clément Dumas, Kit Fraser-Taliente, Subhash Kantamneni, Julian Minder, Euan Ong, Arnab Sen Sharma, Daniel Wen, Owain Evans, and Samuel Marks. 2025. Activation Oracles: Training and Evaluating LLMs as General-Purpose Activation Explainers. arXiv preprint arXiv:2512.15674 (2025). https://arxiv.org/abs/2512.15674
- [56]Eoin M. Kenny, Akshay Dharmavaram, Sang Uk Lee, Tung Phan-Minh, Shreyas Rajesh, Yungin Hu, Laura Major, Momchil S. Tomov, and Julie A. Shah. 2026. Explainable deep learning improves human mental models of self-driving cars. Nature 657, 8130 (Sept. 2026), 114–120. doi:10.1038/s41586-026-10950-5DOI
- [57]Dahye Kim and Deepti Ghadiyaram. 2025. Concept Steerers: Leveraging K-Sparse Autoencoders for Test-Time Controllable Generations. arXiv preprint arXiv:2501.19066 (2025).arxiv.org/abs/2501.19066
- [58]Junsol Kim, James Evans, and Aaron Schein. 2025. Linear representations of political perspective emerge in large language models. arXiv preprint arXiv:2503.02080 (2025).arxiv.org/abs/2503.02080
- [59]René F. Kizilcec. 2016. How Much Information? Effects of Transparency on Trust in an Algorithmic Interface. In Proceedings of the 2016 CHI Conference on Human Factors in Computing Systems (CHI ’16). Association for Computing Machinery, New York, NY, USA, 2390–2395. doi:10.1145/2858036.2858402
- [60]Remi Lam, Alvaro Sanchez-Gonzalez, Matthew Willson, Peter Wirnsberger, Meire Fortunato, Ferran Aleat, Suman Ravuri, Timo Ewalds, Zach Eaton-Rosen, Weihua Hu, Alexander Remes, Stephan Hoyer, George Holland, Oriol Vinyals, Jacklynn Stott, Alexander Pritzel, Shakir Mohamed, and Peter Battaglia. 2023. Learning Skillful Medium-Range Global Weather Forecasting. Science 382, 6677 (2023), 1416–1421. doi:10.1126/science.adi2336DOI
- [61]Siddique Latif, Moazzam Shoukat, Fahad Shamshad, Muhammad Usama, Yi Ren, Heriberto Cuayáhuitl, Wenwu Wang, Xulong Zhang, Roberto Togneri, Erik Cambria, and Björn W. Schuller. 2023. Sparks of Large Audio Models: A Survey and Outlook. arXiv preprint arXiv:2308.12792 (2023).arxiv.org/abs/2308.12792
- [62]Patrick Leask, Bart Bussmann, Michael Pearce, Joseph Bloom, Curt Tigges, Trona Al Mouayed, Lee Sharkey, and Neel Nanda. 2025. Sparse Autoencoders Do Not Find Canonical Units of Analysis. In International Conference on Learning Representations.
- [63]John D. Lee and Katrina A. See. 2004. Trust in Automation: Designing for Appropriate Reliance. Human Factors 46, 1 (2004), 50–80. doi:10.1518/hfes.46.1.50_30392pubmed.ncbi.nlm.nih.gov/15151155
- [64]Haoxuan Li, Zhen Wen, Qiji Jiang, Chenxiao Li, Yuwei Wu, Yuchen Yang, Yiyao Wang, Xiuqi Huang, Minfeng Zhu, and Wei Chen. 2025. ConceptViz: A Visual Analytics Approach for Exploring Concepts in Large Language Models. IEEE Transactions on Visualization and Computer Graphics (2025).arxiv.org/abs/2509.20376
- [65]Millicent Li, Alberto Mario Ceballos Arroyo, Giordano Rogers, Naomi Saphra, and Byron C. Wallace. 2025. Do Natural Descriptions of Model Activations Convey Privileged Information? arXiv preprint arXiv:2509.13316 (2025). https://arxiv.org/abs/2509.13316
- [66]Yuxuan Liang, Haomin Wen, Yuqi Nie, Yushan Jiang, Ming Jin, Dongjin Song, Shiru Pan, and Qingsong Wen. 2024. Foundation Models for Time Series Analysis: A Tutorial and Survey. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining (KDD). 6555–6565. doi:10.1145/3637528.3671451DOI
- [67]Tom Lieberum, Senthooran Rajamanoharan, Arthur Conmy, Lewis Smith, Nicolas Sonnerat, Vikrant Varma, Janos Kramár, Anca Dragan, Rohin Shah, and Neel Nanda. 2024. Gemma Scope: Open Sparse Autoencoders Everywhere All At Once on Gemma 2. arXiv preprint arXiv:2408.05147 (2024).arxiv.org/abs/2408.05147
- [68]Johnny Lin. 2023. Neuronpedia: Interactive Reference and Tooling for Analyzing Neural Networks. https://www.neuronpedia.org Software available from neuronpedia.org.
- [69]Jack Lindsey, Wes Gurnee, Emmanuel Ameisen, Brian Chen, Adam Pearce, Nicholas L. Turner, Craig Citro, David Abrahams, Shan Carter, Basil Hosmer, Jonathan Marcus, Michael Sklar, Adly Templeton, Trenton Bricken, Callum McDougall, Hoagy Cunningham, Thomas Henighan, Adam Jermyn, Andy Jones, Andrew Persic, Zhenyi Qi, T. Ben Thompson, Sam Zimmerman, Kelley Rivoire, Thomas Conerly, Chris Olah, and Joshua Batson. 2025. On the Biology of a Large Language Model. Transformer Circuits Thread (2025). https://transformer-circuits.pub/2025/attribution-graphs/biology.html
- [70]Jack Lindsey, Adly Templeton, Jonathan Marcus, Thomas Conerly, Joshua Batson, and Christopher Olah. 2024. Sparse Crosscoders for Cross-Layer Features and Model Diffing. Transformer Circuits Thread. https://transformer-circuits.pub/2024/crosscoders/index.html
- [71]Jiaohong Liu, Zexuan Qiu, Zhongyang Li, Quanyu Dai, Wenhao Yu, Jieming Zhu, Minda Hu, Menglin Yang, Tat-Seng Chua, and Irwin King. 2025. A Survey of Personalized Large Language Models: Progress and Future Directions. arXiv preprint arXiv:2502.11528 (2025). https://arxiv.org/abs/2502.11528
- [72]Christina Lu, Jack Gallagher, Jonathan Michala, Kyle Fish, and Jack Lindsey. 2026. The Assistant Axis: Situating and Stabilizing the Default Persona of Language Models. arXiv:2601.10387 [cs.CL] https://arxiv.org/abs/2601.10387
- [73]Scott M Lundberg and Su-In Lee. 2017. A unified approach to interpreting model predictions. Advances in neural information processing systems 30 (2017).arxiv.org/abs/1705.07874
- [74]Takuya Maeda and Anabel Quan-Haase. 2024. When Human-AI Interactions Become Parasocial: Agency and Anthropomorphism in Affective Design. In Proceedings of the 2024 ACM Conference on Fairness, Accountability, and Transparency (FAccT ’24). Association for Computing Machinery, New York, NY, USA, 1068–1077. doi:10.1145/3630106.3658956DOI
- [75]Wilson E Marcelo Jr and Danilo M Elmer. 2026. Navigating the Concept Space of Language Models. arXiv preprint arXiv:2603.23524 (2026).
- [76]Samuel Marks and Max Tegmark. 2023. The geometry of truth: Emergent linear structure in large language model representations of true/false datasets. arXiv preprint arXiv:2310.06824 (2023).arxiv.org/abs/2310.06824
- [77]Samuel Marks, Johannes Treutlein, Trenton Bricken, Jack Lindsey, Jonathan Marcus, Siddharth Mishra-Sharma, Daniel Ziegler, Emmanuel Ameisen, Joshua Batson, Tim Lehnvoxa, Samuel R. Bowman, Shan Carter, Brian Chen, Hoagy Cunningham, Carson Denison, Florian Dietz, Savik Golecha, Akbir Khan, Jan Kirchher, Jan Leike, Austin Meek, Kei Nishimura-Gaspairan, Euan Ong, Christopher Olah, Adam Pearce, Fabien Roger, Jeanne Salle, Andy Shih, Meg Tong, Drake Thomas, Kelley Rivoire, Adam Jermyn, Monte MacDiarmid, Tom Henighan, and Evan Hubinger. 2025. Auditing Language Models for Hidden Objectives. arXiv preprint arXiv:2503.10965 (2025). https://arxiv.org/abs/2503.10965
- [78]R. Thomas McCoy, Paul Soulos, Tan Linzen, and Paul Smolensky. 2026. The Emergent Symbolic Structure of Artificial Neural Networks. arXiv:2608.29530 [cs.CL] https://arxiv.org/abs/2608.29530
- [79]Russ McDonald. 2001. Shakespeare and the Arts of Language. Oxford University Press, Oxford.
- [80]Thomas McGrath, Andrei Kapishnikov, Nenad Tomašev, Adam Pearce, Martin Wattenberg, Demis Hassabis, Been Kim, Ulrich Paquet, and Vladimir Kramnik. 2022. Acquisition of chess knowledge in AlphaZero. Proceedings of the National Academy of Sciences 119, 47 (2022), e2206625119. doi:10.1073/pnas.2206625119 Preprint: arXiv:2111.09259 (2021).DOI
- [81]Kevin Meng, David Bau, Alex Andonian, and Yonatan Belinkov. 2022. Locating and Editing Factual Associations in GPT. In Advances in Neural Information Processing Systems, Vol. 35. arXiv:2202.05262 [cs.CL] https://arxiv.org/abs/2202.05262 Code and data at https://rome.baulab.info/.
- [82]Kevin Meng, David Bau, Alex Andonian, and Yonatan Belinkov. 2022. Locating and editing factual associations in gpt. Advances in neural information processing systems 35 (2022), 17359–17372.arxiv.org/abs/2202.05262
- [83]Tim Miller. 2023. Explainable AI is Dead, Long Live Explainable AI! Hypothesis-driven Decision Support using Evaluative AI. In Proceedings of the 2023 ACM Conference on Fairness, Accountability, and Transparency (FAccT ’23). Association for Computing Machinery, New York, NY, USA, 333–342. doi:10.1145/3593013.3594001DOI
- [84]Vince Miller, Mark J. Hill, and Tiago Moreira. 2026. Artificial intelligence and unconditional love: The rise of generative AI as an alternative form of mental health support. Communication and Change 2 (2026), Article 14. doi:10.1007/s44382-026-0036-1 Interviews + Reddit analysis; use blends emotional support, advice-seeking, entertainment – closer to companionship than task-specific tool use.DOI
- [85]Neel Nanda, Andrew Lee, and Martin Wattenberg. 2023. Emergent linear representations in world models of self-supervised sequence models. arXiv preprint arXiv:2309.00941 (2023).arxiv.org/abs/2309.00941
- [86]nostalgebraist. 2020. interpreting GPT: the logit lens – LessWrong – lesswrong.com. https://www.lesswrong.com/posts/AckRR8wDpd aN6v6ru/interpreting-gpt-the-logit-lens. [Accessed 12-09-2025].
- [87]Hadas Argad, Fazal Barez, Tal Haklay, Isabelle Lee, Marius Mosbach, Anja Reusch, Naomi Saphra, Byron C. Wallace, Sarah Wiegreffe, Eric Wong, Ian Tenney, and Mor Geva. 2026. Explainable AI is Dead, Fall Back. arXiv preprint arXiv:2605.11161 (2026). arXiv:2605.11161 [cs.LG]
- [88]Nathan Paek, Yongyi Zang, Qihui Yang, and Randal Leistikow. 2025. Learning Interpretable Features in Audio Latent Spaces via Sparse Autoencoders. arXiv preprint arXiv:2510.23802 (2025).arxiv.org/abs/2510.23802
- [89]Alexander Pan, Lijie Chen, and Jacob Steinhardt. 2024. LatentQA: Teaching LLMs to Decode Activations Into Natural Language. arXiv preprint arXiv:2412.08686 (2024). https://arxiv.org/abs/2412.08686
- [90]Wenbo Pan, Zhichao Liu, Xianlong Wang, Haining Yu, and Xiaohua Jia. 2026. Towards Long-Horizon Interpretability: Efficient and Faithful Multi-Token Attribution for Reasoning LLMs. In Proceedings of the 43rd International Conference on Machine Learning (ICML) (PMLR, Vol. 306). arXiv:2602.01914 [cs.LG]arxiv.org/abs/2602.01914
- [91]Jack Parker-Holder, Sholmi Fruchter, et al. 2025. Genie 3: A New Frontier for World Models. Google DeepMind. https://deepmind.google/discover/blog/genie-3-a-new-frontier-for-world-models/.
- [92]Pat Pataranutaporn, Angela Vujic, David S. Kong, Pattie Maes, and Misha Sra. 2020. Living Bits: Opportunities and Challenges for Integrating Living Microorganisms in Human-Computer Interaction. In Proceedings of the Augmented Humans International Conference (AHs ’20). ACM, New York, NY, USA, Article 30, 12 pages. doi:10.1145/3384657.3384783DOI
- [93]Gonçalo Paulo, Thomas Marshall, and Nora Belrose. 2025. Do Transformer Interpretability Methods Transfer to RNNs?. In Proceedings of the AAAI Conference on Artificial Intelligence, Vol. 39. 27565–27572.DOI
- [94]Alejandro Peña, Ignacio Serna, Ayathmi Morales, and Julian Fierrez. 2020. Bias in multimodal AI: Testbed for fair automatic recruitment. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW). IEEE, 129–137.arxiv.org/abs/2004.07173
- [95]Jason Phang, Michael Lampe, Lama Ahmad, Sandhini Agarwal, Cathy Mengying Fang, Auren R Liu, Valdemar Danry, Eunhae Lee, Samantha WT Chan, Pat Pataranutaporn, et al. 2025. Investigating affective use and emotional well-being on ChatGPT. arXiv preprint arXiv:2504.03888 (2025).
- [96]Daking Rai, Yilun Zhou, Shi Feng, Abulhair Saparov, and Ziyu Yao. 2024. A Practical Review of Mechanistic Interpretability for Transformer-Based Language Models. arXiv preprint arXiv:2407.02646 (2024). https://arxiv.org/abs/2407.02646
- [97]Marco Tulio Ribeiro, Sameer Singh, and Carlos Guestrin. 2016. Why should I trust you? Explaining the predictions of any classifier. In Proceedings of the 22nd ACM SIGKDD international conference on knowledge discovery and data mining. 1135–1144.DOI
- [98]Nina Rimsky, Nick Gabrieli, Julian Schulz, Meg Tong, Evan Hubinger, and Alexander Turner. 2024. Steering llama 2 via contrastive activation addition. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 15504–15522.arxiv.org/abs/2312.06681
- [99]Tom Ritchey. 2018. General Morphological Analysis as a Basic Scientific Modelling Method. Technological Forecasting and Social Change 126 (2018), 81–91. doi:10.1016/j.techfore.2017.05.027DOI
- [100]Bernardino Romera-Paredes, Mohammadamin Barekatain, Alexander Novikov, Matej Balog, M. Pawar Kumar, Emilien Dupont, Francisco J. R. Ruiz, Jordan S. Ellenberg, Pengming Wang, Omar Fawzi, Pushmeet Kohli, and Alhussein Fawzi. 2024. Mathematical discoveries from program search with large language models. Nature 625, 7995 (2024), 468–475. doi:10.1038/s41586-023-06924-6
- [101]Tony Rousmaniere, Y. Zhang, X. Li, and S. Shah. 2026. Large Language Models as Mental Health Resources: Patterns of Use in the United States. Practice Innovations 11, 2 (June 2026), 139–155. doi:10.1037/pri0000292
- [102]Runway. 2026. Introducing Solaris: An Interface World Model. Runway research release. Accessed 2026-09-06.
- [103]Francesco Salvi, Manoel Horta Ribeiro, Riccardo Gallotti, and Robert West. 2025. On the Conversational Persuasiveness of GPT-4. Nature Human Behaviour 9, 8 (2025), 1645–1653. doi:10.1038/s41562-025-02194-6DOI
- [104]Naomi Saphra and Sarah Wiegreffe. 2024. Mechanistic?. In Proceedings of the 7th BlackboxNLP Workshop: Analyzing and Interpreting Neural Networks for NLP. Association for Computational Linguistics, Miami, Florida, USA, 480–498. doi:10.18653/v1/2024.blackboxnlp-1.30
- [105]Gabriele Sarti, Nils Feldhus, Ludwig Sickert, and Oskar van der Wal. 2023. Inseq: An Interpretability Toolkit for Sequence Generation Models. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 3: System Demonstrations), Danushka Bollegala, Ruihong Huang, and Alan Ritter (Eds.). Association for Computational Linguistics, Toronto, Canada, 421–435. doi:10.18653/v1/2023.acl-demo.40DOI
- [106]Pranav Milind Sawant and Jakub Krejčí. 2026. Mechanistic Interpretability for Neural Networks: Circuits, Sparse Features and Symbolic Reasoning. arXiv preprint arXiv:2607.07316 (2026). https://arxiv.org/abs/2607.07316 Scoping review conducted per PRISMA-ScR.
- [107]Max Schemmer, Patrick Hemmer, Maximilian Nitsche, Niklas Kühl, and Michael Vössing. 2022. A Meta-Analysis of the Utility of Explainable Artificial Intelligence in Human-AI Decision-Making. In Proceedings of the 2022 AAAI/ACM Conference on AI, Ethics, and Society (AIES ’22). ACM, New York, NY, USA, 617–626. doi:10.1145/3514094.3534128DOI
- [108]Max Schummer, Niklas Kühl, Carina Benz, and Gerhard Satzger. 2022. On the Influence of Explainable AI on Automation Bias. In Proceedings of the 30th European Conference on Information Systems (ECIS 2022). Timisoara, Romania.
- [109]Lisa Schut, Nenad Tomasev, Thomas McGrath, Demis Hassabis, Ulrich Paquet, and Been Kim. 2025. Bridging the human-AI knowledge gap through concept discovery and transfer in AlphaZero. Proceedings of the National Academy of Sciences 122, 13 (2025), e2406675122. doi:10.1073/pnas.2406675122 Preprint: arXiv:2310.16410 (2023).
- [110]Murray Shanahan. 2024. Talking About Large Language Models. Commun. ACM 67, 2 (2024), 68–79.arxiv.org/abs/2212.03551
- [111]Lee Sharkey, Bilal Chughtai, Joshua Batson, and Lindsey, Jeff Wu, Lucius Bush, naq, Nicholas Goldowsky-Dill, Stefan Heimersheim, Alejandro Ortega, Joseph Bloom, Stella Biderman, Adria Garriga-Alonso, Arthur Conmy, Neel Nanda, Jessica Rumbelow, Martin Wattenberg, Nandi Schoots, Joseph Miller, Eric J. Michaud, Stephen Casper, Max Tegmark, William Saunders, David Bau, Eric Todd, Atticus Geiger, Mor Geva, Jesse Hoogland, Daniel Murefer, and Tom McGrath. 2025. Open Problems in Mechanistic Interpretability. arXiv preprint arXiv:2501.16496 (2025). https://arxiv.org/abs/2501.16496
- [112]Arnab Sen Sharma, David Atkinson, and David Bau. 2024. Locating and Editing Factual Associations in Mamba. arXiv preprint arXiv:2404.03646 (2024).arxiv.org/abs/2404.03646
- [113]Minkyu Shin, Jin Kim, Bas van Opheusden, and Thomas L. Griffiths. 2023. Superhuman artificial intelligence can improve human decision-making by increasing novelty. Proceedings of the National Academy of Sciences 120, 12 (2023), e2214840120. doi:10.1073/pnas.2214840120 Preprint: arXiv:2303.07462.DOI
- [114]Aaron Springer and Steve Whittaker. 2019. Progressive Disclosure: Empirically Motivated Approaches to Designing Effective Transparency. In Proceedings of the 24th International Conference on Intelligent User Interfaces (IUI ’19). Association for Computing Machinery, New York, NY, USA, 107–120. doi:10.1145/3301275.3302322DOI
- [115]Aaron Springer and Steve Whittaker. 2020. Progressive Disclosure: When, Why, and How Do Users Want Algorithmic Transparency Information? ACM Trans. Interact. Intell. Syst. 10, 4, Article 29 (Oct. 2020), 32 pages. doi:10.1145/3374218DOI
- [116]Caroline F. E. Spurgeon. 1935. Shakespeare’s Imagery and What It Tells Us. Cambridge University Press, Cambridge.
- [117]Robin Staab, Mark Wero, Mislav Balunović, and Martin Vechev. 2024. Beyond Memorization: Violating Privacy via Inference with Large Language Models. In International Conference on Learning Representations.arxiv.org/abs/2310.07298
- [118]Mark Steyvers, Heliodoro Tejeda, Aakriti Kumar, Catarina Belem, Sheer Kany, Xinyue Hu, Lukas W. Mayer, and Padraic Smyth. 2025. What large language models know and what people think they know. Nature Machine Intelligence 7, 2 (Feb. 2025), 221–231. doi:10.1038/s42256-024-00976-7DOI
- [119]Mukund Sundararajan, Ankur Taly, and Qiqi Yan. 2017. Axiomatic Attribution for Deep Networks. In Proceedings of the 34th International Conference on Machine Learning (ICML) (PMLR, Vol. 70). 3319–3328. arXiv:1703.01365 [cs.LG]arxiv.org/abs/1703.01365
- [120]Ivan E. Sutherland. 1963. Sketchpad: A Man-Machine Graphical Communication System. Ph. D. Dissertation. Massachusetts Institute of Technology, Cambridge, MA, USA.
- [121]Daniel Tan, David Chanin, Aengus Lynch, Dimitrios Kanoulas, Brooks Paige, Adria Garriga-Alonso, and Robert Kirk. 2024. Analysing the Generalisation and Reliability of Steering Vectors. Advances in Neural Information Processing Adly Tletum, Tom Conerly, Jonathan Marcus, Jack Lindsey, Trenton Bricken, Brian Chen, Adam Pearce, Craig Citro, Emmanuel Ameisen, Andy Jones, Hoagy Cunningham, Nicholas L Turner, Callum McDougall, Monte MacDiarmid, C. Daniel Freeman, Theodore R. Sumers, Edward Rees, Joshua Batson, Adam Jermyn, Shan Carter, Chris Olah, and Tom Henighan. 2024. Scaling Monosemanticity: Extracting Interpretable Features from Claude 3 Sonnet. Transformer Circuits Thread (2024). https://transformer-circuits.pub/2024/scaling-monosemanticity/index.html
- [123]Curt Tigges, Oskar John Hollinsworth, Atticus Geiger, and Neel Nanda. 2023. Linear representations of sentiment in large language models. arXiv preprint arXiv:2310.15154 (2023).arxiv.org/abs/2310.15154
- [124]Alexander Matt Turner, Lisa Thiergart, Gavin Leech, David Udell, Juan Vazquez, Ulisse Mine, and Monte MacDiarmid. 2023. Steering language models with activation engineering. arXiv preprint arXiv:2308.10248 (2023).arxiv.org/abs/2308.10248
- [125]Mor Vered, Tali Livni, Piers Douglas Lionel Howe, Tim Miller, and Liz Sonenberg. 2023. The Effects of Explanations on Automation Bias. Artificial Intelligence 322 (2023), 103952. doi:10.1016/j.artint.2023.103952DOI
- [126]Brian Vickers. 1968. The Artistry of Shakespeare’s Prose. Methuen, London.
- [127]Fernanda Viégas and Martin Wattenberg. 2023. The system model and the user model: Exploring AI dashboard design. arXiv preprint arXiv:2305.02469 (2023).arxiv.org/abs/2305.02469
- [128]Jesse Vig, Sebastian Gehrmann, Yonatan Belinkov, Sharon Qian, Daniel Nevo, Yaron Singer, and Stuart Shieber. 2020. Investigating gender bias in language models using causal mediation analysis. Advances in neural information processing systems 33 (2020), 12388–12401.
- [129]Haoming Wang et al. 2025. UI-TARS-2 Technical Report: Advancing GUI Agent with Multi-Turn Reinforcement Learning. arXiv preprint arXiv:2509.02544 (2025). https://arxiv.org/abs/2509.02544
- [130]Philipp Wesp, Robbie Holland, Vasiliki Sideri-Lampretsa, and Sergios Gatidis. 2026. Sparse Autoencoders for Interpretable Medical Image Representation Learning. arXiv preprint arXiv:2603.23794 (2026).arxiv.org/abs/2603.23794
- [131]Kyra Wilson and Aylin Caliskan. 2024. Gender, race, and intersectional bias in resume screening via language model retrieval. In Proceedings of the AAAI/ACM Conference on AI, Ethics, and Society, Vol. 7. 1578–1590.DOI
- [132]George T. Wright. 1981. Handbooks and Hamlet. PMLA 96, 2 (1981), 168–193.
- [133]George T. Wright. 1988. Shakespeare’s Metrical Art. University of California Press, Berkeley.DOI
- [134]Daniel Wurgraft, Can Rager, Matthew Kowal, Vasudev Shyam, Sheridan Feucht, Usha Bhalla, Tal Haklay, Eric Bigelow, Raphael Sarfati, Thomas McGrath, et al. 2026. Manifold steering reveals the shared geometry of neural network representation and behavior. arXiv preprint arXiv:2605.05115 (2026).arxiv.org/abs/2605.05115
- [135]Xinyuan Yan, Shusen Liu, Kowshik Thopalli, and Bei Wang. 2025. Visual exploration of Feature Relationships in Sparse Autoencoders with Curated Concepts. arXiv preprint arXiv:2511.06048 (2025).arxiv.org/abs/2511.06048
- [136]Marc Zao-Sanders. 2025. How People Are Really Using Gen AI in 2025. Harvard Business Review (9 April 2025). https://hbr.org/2025/04/how-people-are-really-using-gen-ai-in-2025 Practitioner analysis of public forum discourse; ranks therapy/companionship as the top use case. Not peer-reviewed.
- [137]Lvmir Zhang, Anyi Rao, and Maneesh Agrawala. 2023. Adding Conditional Control to Text-to-Image Diffusion Models. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV).
- [138]Yunfeng Zhang, Q. Vera Liao, and Rachel K. E. Bellamy. 2020. Effect of Confidence and Explanation on Accuracy and Trust Calibration in AI-Assisted Decision Making. In Proceedings of the 2020 Conference on Fairness, Accountability, and Transparency (FAT*), 295–305. doi:10.1145/3351095.3372852 Displaying calibrated confidence improves trust calibration.DOI
- [139]Yutong Zhang, Dora Zhao, Jeffrey T. Hancock, Robert Kraut, and Diyi Yang. 2026. Interaction with AI companions and psychological well-being. Nature Human Behaviour (2026). doi:10.1038/s41562-026-02516-2 Mixed-methods study of 1,131 Character.AI users; asks users their primary motivation (productivity, entertainment, curiosity, relational). Preprint: arXiv:2506.12605.DOI
- [140]Zhehao Zhang, Ryan A. Rossi, Branislav Kveton, Yijia Shao, Diyı Yang, Hamed Zamani, Franck Dernoncourt, Joe Barrow, Tong Yu, Sungchul Kim, Ruiyi Zhang, Jiuxiang Gu, Tyler Derr, Hongjie Chen, Junda Wu, Xiang Chen, Zichao Wang, Subrata Mitra, Nedim Lipka, Nesreen Ahmed, and Yu Wang. 2024. Personalization of Large Language Models: A Survey. arXiv preprint arXiv:2411.00027 (2024). https://arxiv.org/abs/2411.00027
- [141]Andy Zou, Long Phan, Sarah Chen, James Campbell, Phillip Guo, Richard Ren, Alexander Pan, Xuwang Yin, Mantas Mazeika, Ann-Kathrin Dombrowski, et al. 2023. Representation engineering: A top-down approach to ai transparency. arXiv preprint arXiv:2310.01405 (2023).arxiv.org/abs/2310.01405
- [142]Fritz Zwicky. 1969. Discovery, Invention, Research through the Morphological Approach. The Macmillan Company, Toronto.