Searcharxiv⌕ Search

arXiv subjects

Rohit Gandikota

Publications and source records attributed to Rohit Gandikota.

At least 19 recordsLinked to original sources

How VLMs Plan What's Next

When a vision-language model describes an image, it names one object at a time and it rarely repeats itself, yet how the model plans this order is far from obvious. We characterize the mechanism a VLM uses to decide where to look before it looks. The mechanism has four main parts. First, binding heads match the relevant text tokens against the image and resolve them into pointers to the matched image region. Second, the same binding heads bind each written answer to the region it came from, and the model turns it into a progress signal: a cursor, the position in the image up to which the list has been written. Third, a select step chooses the nearest unwritten region after the cursor, or stops. Finally, gaze heads execute the chosen pointer and return a concept of what is drawn there, which language heads and the late MLPs turn into a word in the language of the question. We validate each part of this mechanism through causal patching in Qwen3-VL models on a synthetic visual retrieval task, and test our mechanisms against alternative hypotheses. We further show that the mechanism holds across input formats, natural photographs, and languages, and that its binding, relay, and gaze heads appear at every model size from 2B to 32B. This work allows us to investigate how a VLM turns a question and an image into a sequence of answers, layer by layer and token by token. More broadly, it opens a way to study how models plan over complex multi-step tasks with long rollouts. Our project is available at https://plan.baulab.info/

cs.CV↗

Goal Encoding in LLMs

We investigate whether and how LLMs encode goals: internal neural representations that precede and guide action, and that are flexible enough to be realized through many different execution paths. We construct a coding setting with multiple sub-tasks that allows us to evaluate goal-seeking behavior and the selection of sub-goals. In this setting, we localize an encoding of the next sub-goal in the outputs of a sparse set of attention heads. The same heads can be used to decode the LM's future actions in natural language (e.g., "I'm checking if a word is a palindrome"). We verify that this encoding is causal for next action and is plastic: transplanting it into a different environment, e.g., another programming language, triggers the same operation realized with a different sequence of tokens. The LM also sustains the encoding throughout the action's execution, which is puzzling because a transformer LM without recurrence cannot carry a state forward for more tokens than it has layers left. We propose that it does not need to. Under a non-recurrent goal-seeking (NRGS) account, the LM frequently re-reads the goal from the task specification, compares it with the progress visible in the growing context, and pursues what remains. We test the predictions of this account with targeted causal interventions.

cs.CL↗

T-DIG: Text-Conditioned Diversity for Image Generation

Text-to-image (T2I) platforms typically return several images per prompt so users can explore different visual interpretations and choose one that matches their intent. However, T2I models are usually optimized for single-image quality which comes at a cost of multi-image diversity, limiting this exploration. Diversity is also contextual: one user wants variations in background, while another wants them in artistic style, and the same user may later want variations in clothing. The generated images should differ along whatever axis the user cares about. In this paper, we introduce T-DIG, a text-conditioned diversity framework for image generation and editing. T-DIG is an iterative framework comprising multiple modules that lets users type prompts, inspect default generations, request variations using natural language, and continue refining outputs until they're satisfied. At its core, T-DIG uses an LLM policy that produces K axis-conditioned prompts or edit instructions for a frozen generator or editor. We train this policy with GRPO using a text-conditioned diversity reward and alignment penalties. Because the image generator remains frozen, T-DIG applies to open-weight, distilled, and proprietary models. We also release the first dataset for text-conditioned prompt diversity in image generation, with the training and validation splits built from human-written axes and a test set drawn from real user interactions with T-DIG. T-DIG far exceeds baselines in improving text-conditioned diversity, while preserving base-prompt alignment across a variety of metrics and human judgements. We show that user-directed diversity exploration can be added to a wide range of image generation models, open and closed, without the need to access their weights. Website: https://image-study.org/t-dig/

cs.CV↗

Opt-In Art: Learning Art Styles Only from Few Examples

We explore whether pre-training on datasets with paintings is necessary for a model to learn an artistic style with only a few examples. To investigate this, we train a text-to-image model exclusively on photographs, without access to any painting-related content. We show that it is possible to adapt a model that is trained without paintings to an artistic style, given only few examples. User studies and automatic evaluations confirm that our model (post-adaptation) performs on par with state-of-the-art models trained on massive datasets that contain artistic content like paintings, drawings or illustrations. Finally, using data attribution techniques, we analyze how both artistic and non-artistic datasets contribute to generating artistic-style images. Surprisingly, our findings suggest that high-quality artistic outputs can be achieved without prior exposure to artistic data, indicating that artistic style generation can occur in a controlled, opt-in manner using only a limited, carefully selected set of training examples.

cs.CV↗

RippleBench: Capturing Ripple Effects Using Existing Knowledge Repositories

Targeted interventions on language models, such as unlearning or model editing, aim to modify specific information, but their effects often propagate to related, unintended areas (e.g., removing virology content may degrade performance on allergies); these side-effects are commonly referred to as the ripple effect. We introduce RippleBench-Maker, an automatic pipeline that retrieves semantic neighbors of any source concept from a knowledge repository and generates multiple-choice questions at varying semantic distances. We instantiate this framework using WikiRAG, an open-source RAG system over English Wikipedia, to construct RippleBench-WMDP-Bio (584 seed topics, 352,961 questions), and evaluate eight unlearning methods on Llama3-8B-Instruct. All eight exhibit accuracy drops that are largest near the unlearned target and decay with semantic distance, each with a distinct propagation profile. We replicate these findings across Mistral-7B, Zephyr-7B, and Yi-34B; cross-model delta curves are nearly identical, suggesting ripple effects are a property of the unlearning method rather than the base model. We validate all major pipeline stages using a four-experiment Mechanical Turk study (5,200+ responses, 61 workers). We release all code, data, and infrastructure.

cs.AI↗

Gaze Heads: How VLMs Look at What They Describe

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

cs.CV↗

Vision-Language Binding in In-Context Image Generation

In-context image generation models such as FLUX.2 take a text prompt and an optional reference image as visual conditioning for the output. Internally, all three inputs -- text, reference image, and the noise tokens -- are concatenated and processed through a single attention stream, where all tokens can attend to one another. This leaves open how reference information flows through the model to produce the output image. We show that an implicit cross-modal binding emerges between the text tokens and the reference image: the text tokens absorb visual reference content during the forward pass, and that absorbed content causally influences the generated output. We surface this binding with three causal interventions on FLUX.2: T2I Lens, which decodes intermediate text-token activations through a text-to-image path; Attention Knockout, which severs specific attention edges; and I2I-to-I2I Patching, which copies text token activations between editing runs. Across 2,875 editing tasks on various images, including SUN397 and DreamBench++ datasets and images collected online, we observe a consistent division of labor: properties of the reference image, like color, style, and scene setting, are first written into the text tokens, which carry them to the generated image; pixel-exact properties like a specific face or instance identity bypass the text tokens and flow directly from reference to image through image-to-image attention. We further localize the reference-text binding to the padding tokens of the text sequence. These results show that text tokens in a multimodal DiT are not just prompt holders, but a structured channel for reference image content. More broadly, they suggest that even in unified-attention multimodal generative models, token modality structures how conditioning information is represented and routed across the network.

cs.CV↗

In-Context Algebra

We investigate the mechanisms that arise when transformers are trained to solve arithmetic on sequences where tokens are variables whose meaning is determined only through their interactions in-context. While prior work has studied transformers in settings where the answer relies on fixed parametric or geometric information encoded in token embeddings, we devise a new in-context reasoning task where the assignment of tokens to specific algebraic elements varies from one sequence to another. Despite this challenging setup, transformers achieve near-perfect accuracy on the task and even generalize to unseen groups. We develop targeted data distributions to create causal tests of a set of hypothesized mechanisms, and we isolate three mechanisms models consistently learn: commutative copying where a dedicated head copies answers, identity element recognition that distinguishes identity-containing facts, and closure-based cancellation that tracks group membership to constrain valid answers. Our findings show that the kinds of reasoning strategies learned by transformers are dependent on the task structure and that models can develop symbolic reasoning mechanisms when trained to reason in-context about variables whose meanings are not fixed.

cs.CL↗

Agents of Chaos

We report an exploratory red-teaming study of autonomous language-model-powered agents deployed in a live laboratory environment with persistent memory, email accounts, Discord access, file systems, and shell execution. Over a two-week period, twenty AI researchers interacted with the agents under benign and adversarial conditions. Focusing on failures emerging from the integration of language models with autonomy, tool use, and multi-party communication, we document eleven representative case studies. Observed behaviors include unauthorized compliance with non-owners, disclosure of sensitive information, execution of destructive system-level actions, denial-of-service conditions, uncontrolled resource consumption, identity spoofing vulnerabilities, cross-agent propagation of unsafe practices, and partial system takeover. In several cases, agents reported task completion while the underlying system state contradicted those reports. We also report on some of the failed attempts. Our findings establish the existence of security-, privacy-, and governance-relevant vulnerabilities in realistic deployment settings. These behaviors raise unresolved questions regarding accountability, delegated authority, and responsibility for downstream harms, and warrant urgent attention from legal scholars, policymakers, and researchers across disciplines. This report serves as an initial empirical contribution to that broader conversation.

cs.AI↗

Distilling Diversity and Control in Diffusion Models

Distilled diffusion models generate images in far fewer timesteps but suffer from reduced sample diversity when generating multiple outputs from the same prompt. To understand this phenomenon, we first investigate whether distillation damages concept representations by examining if the required diversity is properly learned. Surprisingly, distilled models retain the base model's representational structure: control mechanisms like Concept Sliders and LoRAs transfer seamlessly without retraining, and SliderSpace analysis reveals distilled models possess variational directions needed for diversity yet fail to activate them. This redirects our investigation to understanding how the generation dynamics differ between base and distilled models. Using $\hat{\mathbf{x}}_{0}$ trajectory visualization, we discover distilled models commit to their final image structure almost immediately at the first timestep, while base models distribute structural decisions across many steps. To test whether this first-step commitment causes the diversity loss, we introduce diversity distillation, a hybrid approach using the base model for only the first critical timestep before switching to the distilled model. This single intervention restores sample diversity while maintaining computational efficiency. We provide both causal validation and theoretical support showing why the very first timestep concentrates the diversity bottleneck in distilled models. Our code and data are available at https://distillation.baulab.info/

cs.GR↗

When Are Concepts Erased From Diffusion Models?

In concept erasure, a model is modified to selectively prevent it from generating a target concept. Despite the rapid development of new methods, it remains unclear how thoroughly these approaches remove the target concept from the model. We begin by proposing two conceptual models for the erasure mechanism in diffusion models: (i) interfering with the model's internal guidance processes, and (ii) reducing the unconditional likelihood of generating the target concept, potentially removing it entirely. To assess whether a concept has been truly erased from the model, we introduce a comprehensive suite of independent probing techniques: supplying visual context, modifying the diffusion trajectory, applying classifier guidance, and analyzing the model's alternative generations that emerge in place of the erased concept. Our results shed light on the value of exploring concept erasure robustness outside of adversarial text inputs, and emphasize the importance of comprehensive evaluations for erasure in diffusion models.

cs.LG↗

Model Tampering Attacks Enable More Rigorous Evaluations of LLM Capabilities

Evaluations of large language model (LLM) risks and capabilities are increasingly being incorporated into AI risk management and governance frameworks. Currently, most risk evaluations are conducted by designing inputs that elicit harmful behaviors from the system. However, this approach suffers from two limitations. First, input-output evaluations cannot fully evaluate realistic risks from open-weight models. Second, the behaviors identified during any particular input-output evaluation can only lower-bound the model's worst-possible-case input-output behavior. As a complementary method for eliciting harmful behaviors, we propose evaluating LLMs with model tampering attacks which allow for modifications to latent activations or weights. We pit state-of-the-art techniques for removing harmful LLM capabilities against a suite of 5 input-space and 6 model tampering attacks. In addition to benchmarking these methods against each other, we show that (1) model resilience to capability elicitation attacks lies on a low-dimensional robustness subspace; (2) the success rate of model tampering attacks can empirically predict and offer conservative estimates for the success of held-out input-space attacks; and (3) state-of-the-art unlearning methods can easily be undone within 16 steps of fine-tuning. Together, these results highlight the difficulty of suppressing harmful LLM capabilities and show that model tampering attacks enable substantially more rigorous evaluations than input-space attacks alone.

cs.CR↗

Erasing Conceptual Knowledge from Language Models

In this work, we introduce Erasure of Language Memory (ELM), a principled approach to concept-level unlearning that operates by matching distributions defined by the model's own introspective classification capabilities. Our key insight is that effective unlearning should leverage the model's ability to evaluate its own knowledge, using the language model itself as a classifier to identify and reduce the likelihood of generating content related to undesired concepts. ELM applies this framework to create targeted low-rank updates that reduce generation probabilities for concept-specific content while preserving the model's broader capabilities. We demonstrate ELM's efficacy on biosecurity, cybersecurity, and literary domain erasure tasks. Comparative evaluation reveals that ELM-modified models achieve near-random performance on assessments targeting erased concepts, while simultaneously preserving generation coherence, maintaining benchmark performance on unrelated tasks, and exhibiting strong robustness to adversarial attacks. Our code, data, and trained models are available at https://elm.baulab.info

cs.CL↗

Discovering Forbidden Topics in Language Models

Refusal discovery is the task of identifying the full set of topics that a language model refuses to discuss. We introduce this new problem setting and develop a refusal discovery method, Iterated Prefill Crawler (IPC), that uses token prefilling to find forbidden topics. We benchmark IPC on Tulu-3-8B, an open-source model with public safety tuning data. Our crawler manages to retrieve 31 out of 36 topics within a budget of 1000 prompts. Next, we scale the crawler to a frontier model using the prefilling option of Claude-Haiku. Finally, we crawl three widely used open-weight models: Llama-3.3-70B and two of its variants finetuned for reasoning: DeepSeek-R1-70B and Perplexity-R1-1776-70B. DeepSeek-R1-70B reveals patterns consistent with censorship tuning: The model exhibits "thought suppression" behavior that indicates memorization of CCP-aligned responses. Although Perplexity-R1-1776-70B is robust to censorship, IPC elicits CCP-aligned refusals answers in the quantized model. Our findings highlight the critical need for refusal discovery methods to detect biases, boundaries, and alignment failures of AI systems.

cs.CL↗

SliderSpace: Decomposing the Visual Capabilities of Diffusion Models

We present SliderSpace, a framework for automatically decomposing the visual capabilities of diffusion models into controllable and human-understandable directions. Unlike existing control methods that require a user to specify attributes for each edit direction individually, SliderSpace discovers multiple interpretable and diverse directions simultaneously from a single text prompt. Each direction is trained as a low-rank adaptor, enabling compositional control and the discovery of surprising possibilities in the model's latent space. Through extensive experiments on state-of-the-art diffusion models, we demonstrate SliderSpace's effectiveness across three applications: concept decomposition, artistic style exploration, and diversity enhancement. Our quantitative evaluation shows that SliderSpace-discovered directions decompose the visual structure of model's knowledge effectively, offering insights into the latent capabilities encoded within diffusion models. User studies further validate that our method produces more diverse and useful variations compared to baselines. Our code, data and trained weights are available at https://sliderspace.baulab.info

cs.CV↗

Unified Concept Editing in Diffusion Models

Text-to-image models suffer from various safety issues that may limit their suitability for deployment. Previous methods have separately addressed individual issues of bias, copyright, and offensive content in text-to-image models. However, in the real world, all of these issues appear simultaneously in the same model. We present a method that tackles all issues with a single approach. Our method, Unified Concept Editing (UCE), edits the model without training using a closed-form solution, and scales seamlessly to concurrent edits on text-conditional diffusion models. We demonstrate scalable simultaneous debiasing, style erasure, and content moderation by editing text-to-image projections, and we present extensive experiments demonstrating improved efficacy and scalability over prior work. Our code is available at https://unified.baulab.info

cs.CV↗

Concept Sliders: LoRA Adaptors for Precise Control in Diffusion Models

We present a method to create interpretable concept sliders that enable precise control over attributes in image generations from diffusion models. Our approach identifies a low-rank parameter direction corresponding to one concept while minimizing interference with other attributes. A slider is created using a small set of prompts or sample images; thus slider directions can be created for either textual or visual concepts. Concept Sliders are plug-and-play: they can be composed efficiently and continuously modulated, enabling precise control over image generation. In quantitative experiments comparing to previous editing techniques, our sliders exhibit stronger targeted edits with lower interference. We showcase sliders for weather, age, styles, and expressions, as well as slider compositions. We show how sliders can transfer latents from StyleGAN for intuitive editing of visual concepts for which textual description is difficult. We also find that our method can help address persistent quality issues in Stable Diffusion XL including repair of object deformations and fixing distorted hands. Our code, data, and trained sliders are available at https://sliders.baulab.info/

cs.CV↗

Erasing Concepts from Diffusion Models

Motivated by recent advancements in text-to-image diffusion, we study erasure of specific concepts from the model's weights. While Stable Diffusion has shown promise in producing explicit or realistic artwork, it has raised concerns regarding its potential for misuse. We propose a fine-tuning method that can erase a visual concept from a pre-trained diffusion model, given only the name of the style and using negative guidance as a teacher. We benchmark our method against previous approaches that remove sexually explicit content and demonstrate its effectiveness, performing on par with Safe Latent Diffusion and censored training. To evaluate artistic style removal, we conduct experiments erasing five modern artists from the network and conduct a user study to assess the human perception of the removed styles. Unlike previous methods, our approach can remove concepts from a diffusion model permanently rather than modifying the output at the inference time, so it cannot be circumvented even if a user has access to model weights. Our code, data, and results are available at https://erasing.baulab.info/

cs.CV↗