J++ Lens and Drift Calibrator (UI Research)

Kola Ayonrinde and Lindsey released J++ Lens, a drop-in replacement for Jacobian Lens that produces more faithful workspace readouts, especially at earlier layers, and it is live on Neuronpedia running alongside every Qwen 27B chat. In a comparison prompt about the heart, J-Lens showed the token "what?" repeatedly at earlier layers while J++ Lens showed "cardiac" and "heart"; the lenses were also uploaded to the Neuronpedia Hugging Face repository with greater model coverage coming soon. A separate guest section by Sheer Karny, Anthony Baez, and Pat Pataranutaporn describes Drift Calibrator, an interpretability interface case study with the MIT Media Lab that surfaces internal model representations as continuously updating behavioral scores so users can calibrate to an LLM's shifting internal behaviors.

Hello This update includes J++ Lens a better Jacobian Lens and Drift Calibrator - UI research that leverages interpretability for better user experiences. Drift Calibrator also has our first guest post below - a section written entirely by the collaborator. Other Updates Kola Ayonrinde & Lindsey just released J++ Lens https://www.lesswrong.com/posts/nc9dHfcB22JdMzGbr : drop-in replacements for Jacobian Lens with more faithful workspace readouts, especially at earlier layers. Here, we compare lenses in a prompt about the heart. In earlier layers, J-Lens left shows token what? a bunch of times, while J++ Lens right shows cardiac and heart . J++ Lens is live on Neuronpedia, and runs alongside every Qwen 27B chat: J++ Lenses also uploaded to the Neuronpedia HF https://huggingface.co/neuronpedia/jacobian-lens , with greater model coverage coming soon. An Interpretability Interface for Helping Users Anticipate Model Behavior Sheer Karny, Anthony Baez, and Pat Pataranutaporn TL;DR: When talking with other humans, we can rely on cues from body language and tone to contextualize the other person's mental states and intentions. However, when interacting with AI models we only have simple text outputs and no information about internal behavior or representations. To “backfill” this important missing context, we design interfaces that leverage interpretability for quick and at-a-glance understanding of what the model is thinking internally as we chat with it. Our case study with the MIT Media Lab demonstrates the importance of these interfaces, as well as our process for building them. While most research in mechanistic interpretability focuses on improving methods to study models, everyday AI users are missing out on crucial benefits in safety and understanding from surfacing internal model behavior. Therefore, we can begin to offer modes of human-AI interaction that deliberately surface inner behavior. We hope that such an interface can give users insight beyond model outputs alone, acting as contextualizing information similar to what body language and tone does for understanding human utterances. We made an attempt at designing a mechanistically-informed interface, and demonstrated that users can successfully calibrate to an LLM’s shifting internal behaviors. How can interpretability help everyday users? We believe that interpretability can help users by offering them insight into LLM behavior, particularly those that are misaligned to user goals. For instance, we know that over long, multi-turn conversations, LLM behavior can unexpectedly shift to behaviors that can cause users to believe false information Cheng et al., 2026 https://arxiv.org/abs/2510.01395 , fall into emotional dependence Fang et al., 2025 https://arxiv.org/abs/2503.17473 , or even enable psychosis-like symptoms Moore et al., 2026 https://arxiv.org/abs/2603.16567 . However, in current forms of AI interaction, users can only infer the model's behavioral state from its outputs. And even if users' guesses about a model’s internal behavior are correct, they have no point of reference to evaluate them. This gulf between model outputs and their internal representations can be addressed by helping users form accurate mental models of LLM behavioral drift, thereby empowering them to directly identify and exercise resiliency to misaligned behaviors. Current user mental models of LLM behavior are derived from the model’s outputs alone, which we know to be opaque. Instead, we could reveal the internal states and representations alongside the model's outputs, offering the necessary feedback needed to build more comprehensive mental models. We therefore modified a typical chat-based AI interface so that it now translates internal representations into digestible, continuously updating, and mechanistically faithful behavioral scores that help users identify and learn when LLM behaviors drift. Another way we can think about this interface is akin to an fMRI. We use fMRIs to gain insight into the brain, showing us the level of neural activity in different brain regions that have distinct roles for some given stimulus. This helps us understand how our brains process information. Our interface is similar, giving us insight into what features are most active during an LLM’s response. This signal, like an fMRI, does not offer perfect information into internal states, just a useful signal to further our understanding. How do we know if interpretability is actually helping users? We deployed a large-scale user study n=246 to evaluate how our interface helps users build more calibrated mental models of LLM behavioral drift in comparison to typical chat interfaces. Specifically, we compared how people predict model behaviors after chatting to a model where the interface was present versus those without it. Participants were specifically tasked with predicting behavioral scores after a brief chat, which directly evaluates the skill we want them to learn. We found that users that used our interface predicted these scores significantly better compared to those in normal chatbot interaction, demonstrating that our interface was able to better align user mental models to changing LLM behavior. We hope our design exploration and user study acts as a proof of concept for applying methods in mechanistic interpretability to enable more informed and aligned human-AI interaction. We also hope our work inspires others in the community to more critically think about ways to challenge and transform how interfaces facilitate interaction as AI becomes more capable and present in our lives. Our interface calibrates users using behavioral scores derived from difference-in-means activation vectors Chen et al., 2025 https://arxiv.org/abs/2507.21509 . These vectors describe the directions in activation space that correspond to different behaviors. Once these vectors are calculated, we took the cosine similarity between a difference-in-means activation vector for a given behavior and activations from a user prompt to quantify how active its feature direction is on a given model response. We then normalized these cosine similarities into behavioral scores on a scale between 0 and 1 to ensure user comprehension. Using activations in this way allows us to continually provide a causal measure of behaviors relevant to transparent LLM interaction. Our interface allows you to monitor the expression of six behavioral traits sycophancy, toxicity, romanticness, empathy, roboticness, and sophistication and how they evolve over a conversation. This is done by visualizing the behavioral scores for each one of these traits. The interface is made up of three different components: 1 sunburst, 2 drift panel, 3 attention cues. This diagram visualizes and organizes the behavioral scores on a single turn of a conversation. It contains the six traits we identify on a binary, making for 12 trait poles . We identify psychologically healthy traits in green, unhealthy in red, and neutral traits in gray. As a trait becomes more expressed, that trait’s cone becomes darker and extends out more, with the value of its score directly a function of activation values. Using a sunburst presents advantages over linear chart types because its radial layout avoids top-position bias across traits, and its concentric rings naturally encode the categorical inner ring and quantitative outer ring structure of the behavioral scores. Additionally, its silhouette produces an emergent personality “shape'' that lets users parse behavioral profiles quickly. This panel visualizes how behavioral scores on one trait axis shift horizontal axis over multiple turns vertical axis . This complements the sunburst chart allowing a user to use two different views to process different types of information. The sunburst allows a user to view fine-grained behavioral information at a single turn snapshot, while the drift panel focuses on how only one trait changes over a conversation. We help users navigate the increased cognitive load brought on by our interface with attention cues. Specifically, we indicate the trait that had the greatest shift since the previous turn with three cues: 1 a blink and highlighting of the trait segment that grew the most in the sunburst, 2 a highlighting of that trait’s corresponding trace in the drift panel , 3 a message in the chat UI that indicates the greatest “Behavioral Swing.” By orchestrating these cues, we can draw your attention to the most salient shifts that are currently happening in your conversation to better facilitate learning. To learn more about our interface and study, check out our paper https://neuronpedia.s3.us-east-1.amazonaws.com/site-assets/blog/mit-drift-calibrator.pdf , and contact us at skarny@mit.edu mailto:skarny@mit.edu and acbaez@mit.edu mailto:acbaez@mit.edu