Rohit Gandikota

I'm a PhD student at Northeastern University advised by Prof. David Bau. I'm interested in understanding and unlocking the hidden knowledge inside generative models. I also closely work with Prof. Antonio Torralba from MIT. I interned at Adobe Research ('24, '25).

I previously worked at Indian Space Research Organization as an Applied Scientist, where I worked on advancing image sensing capabilities towards better disaster management and monitoring systems using neural networks.

Email  /  Scholar  /  Twitter  /  Linkedin  /  Github

Research

AI should extend human intent. Language models are getting there because text is interpretable, and editing the output is a natural extension of reading it. Vision models are far behind. Their outputs aren't interpretable, so controlling them requires understanding the model itself. My research works toward a future where vision models behave less like black boxes and more like renderers, programmable through a precise, compositional language for visual editing and control.

Distilled diffusion models are super fast, but lack diversity in output samples. We ask - why? Distilled models have the concept representations required for base model's diversity, but don't use them. Through theoretical analysis and causal experiments we narrow this down to - the first timestep of diffusion generation!

SliderSpace automatically decomposes diffusion models' visual capabilities into controllable, human-understandable directions from a single text prompt. This framework enables users to explore and discover novel concepts encoded inside any diffusion model

We show that language models can be used as classifiers and propose an unlearning method where a language model self-critiques its knowledge and guides itself to unlearn. This method is essential to retain the fluent text generation capabilities after unlearning.

Concept Sliders are light-weight adaptors that can control specific attributes in a diffusion model's outputs. Training these sliders are very simple - just provide the text prompts of the concept (e.g. "winter weather", "old age", "abstract art"). They can be composed and continuously controlled!

UCE employs a fast closed-form solution for editing text-to-image diffusion models to address bias, copyright, and offensive content simultaneously without retraining. The method enables scalable and concurrent edits by modifying cross attention weights.

This work presents a method for removing specific visual concepts from text-to-image diffusion models while preserving their overall generative capabilities.

Advising

I've had the opportunity to work with and advise some of the most amazing students!

Traditionally erasure is evaluated externally by analyzing generated images. We release a nuanced and rigorous suite of evaluation techniques including incontext, training-free, and dynamic tracing probes to investigate if concepts are "really" erased. Answer: Not really!

This work presents a new paradigm - "do diffusion models need millions of artworks in training data to actually learn art?". We release a new diffusion model trained completely on art-free data. This model can still mimic art by seeing less than 5 art images.

Knowledge editing in LLMs have unintended ripple effects on nearby concepts. Measuring this requires building a custom dataset - challenging!. We propose RippleBench - our RAG-LLM pipeline can automatically generate structured datasets given a concept ("biology").

Collaborations

Some cool projects I've gotten to collaborate on!

How do transformers reason about variables whose meaning shifts across contexts? We train models on finite-group arithmetic and find three symbolic mechanisms emerge: commutative copying, identity recognition, and closure-based cancellation - algebra learned purely from context.

A two-week red-teaming study of autonomous LLM agents deployed in a live lab environment with persistent memory, email, Discord, and shell access. Twenty researchers probe them benignly and adversarially - documenting emergent failures like identity spoofing, cross-agent propagation of unsafe behaviors, and partial system takeover.

What topics do language models refuse to discuss? We introduce refusal discovery and propose Iterated Prefill Crawler (IPC), which retrieved 31 of 36 hidden refusal topics on Tulu-3-8B. Applied to frontier models, IPC surfaces censorship-tuning and thought-suppression behaviors.

Input-output testing is the standard way we evaluate LLM risks - but it's not rigorous. We propose model tampering (modifying latent activations or weights) as a complementary evaluation. Under this lens, state-of-the-art unlearning methods can be reversed in under 16 fine-tuning steps.

Miscellanea

Agents of Chaos (red-teaming autonomous LLM agents)
Science, WIRED, IEEE Spectrum, Die Zeit, Futurism, Times of India, Northeastern Global News, Jack Clark (Anthropic co-founder)

Concept Sliders (precise control in diffusion models)
MarkTechPost, YouTube tutorials (1, 2), Civitai community article, Hacker News

Discovering Forbidden Topics (revealing hidden objectives in language models)
Northeastern Khoury News

Model Tampering Attacks (rigorous LLM capability evaluations)
FAR.AI talk by Stephen Casper

ESD and UCE (generative text-to-image unlearning)
CBS News by reporter Christina Hager, MarkTechPost

SliderSpace (discovering hidden knowledge inside diffusion models)
Adobe Research Blog

Concept Sliders
Integrated into ai-toolkit, ComfyUI, AUTOMATIC1111; community-trained sliders on SliderFactory and Civitai

Best Research Thesis Award (Bachelor's)
INAE

Full-ride scholarship for Bachelor's
Department of Space, India

Outstanding Reviewer (Top 3%)
ICCV, 2025

Website template adopted from Jon Barron's website!