Meta tags:
author=Rohit Gandikota;
description=Rohit Gandikota is a PhD student at Northeastern University researching how to understand, control, and edit generative models — diffusion models, LLMs, and vision-language models.;

Headings (most frequently used words):

research, advising, collaborations, miscellanea, media, open, source, awards,

Text of the page (most frequently used words):
models (27), and (26), the (25), arxiv (18), diffusion (17), rohit (17), gandikota (17), project (15), page (14), code (13), david (13), bau (13), model (11), from (10), text (9), language (9), this (9), 2025 (8), concept (8), image (8), art (8), research (7), can (7), editing (7), are (7), for (6), sliders (6), knowledge (6), llm (6), not (6), capabilities (6), concepts (6), control (5), output (5), 2026 (5), like (5), your (5), demo (5), antonio (5), torralba (5), hidden (4), unlearning (4), news (4), rigorous (4), topics (4), agents (4), propose (4), what (4), them (4), browser (4), does (4), support (4), video (4), tag (4), neurips (4), work (4), joanna (4), materzyńska (4), they (4), through (4), method (4), visual (4), where (4), iccv (3), india (3), discovering (3), inside (3), sliderspace (3), generative (3), tampering (3), northeastern (3), media (3), precise (3), but (3), with (3), context (3), training (3), just (3), reference (3), vision (3), generation (3), that (3), human (3), diversity (3), gaze (3), heads (3), website (2), space (2), bachelor (2), into (2), community (2), trained (2), civitai (2), adobe (2), blog (2), marktechpost (2), uce (2), evaluations (2), stephen (2), casper (2), far (2), attacks (2), forbidden (2), red (2), teaming (2), autonomous (2), chaos (2), authors (2), modifying (2), weights (2), evaluation (2), under (2), tuning (2), more (2), refusal (2), ipc (2), which (2), applied (2), behaviors (2), chris (2), email (2), identity (2), cross (2), how (2), find (2), algebra (2), some (2), have (2), ripple (2), effects (2), requires (2), dataset (2), ripplebench (2), automatically (2), presents (2), new (2), data (2), release (2), free (2), less (2), than (2), images (2), really (2), erased (2), when (2), tokens (2), aren (2), prompt (2), during (2), causally (2), causal (2), specific (2), their (2), erasing (2), 2024 (2), fast (2), retraining (2), enables (2), attention (2), wacv (2), adaptors (2), outputs (2), itself (2), distilled (2), vlms (2), look (2), describe (2), interpretable (2), understanding (2), worked (2), prof (2), template, adopted, jon, barron, outstanding, reviewer, top, department, full, ride, scholarship, inae, best, thesis, award, awards, integrated, sliderfactory, automatic1111, comfyui, toolkit, open, source, reporter, christina, hager, metaphysic, analytics, magazine, cbs, esd, capability, cited, bengio, chaired, talk, international, safety, report, security, institute, revealing, objectives, khoury, china, economist, youtube, tutorials, hacker, article, unite, jack, clark, anthropic, founder, global, constellation, times, futurism, techrepublic, afp, wire, die, zeit, ieee, spectrum, wired, science, miscellanea, input, testing, standard, way, evaluate, risks, latent, activations, complementary, lens, state, methods, reversed, fine, steps, tmlr, dylan, hadfield, menell, zora, che, enable, refuse, discuss, introduce, discovery, iterated, prefill, crawler, retrieved, tulu, frontier, surfaces, censorship, thought, suppression, wendler, rager, two, week, study, deployed, live, lab, environment, persistent, memory, discord, shell, access, twenty, researchers, probe, benignly, adversarially, documenting, emergent, failures, spoofing, agent, propagation, unsafe, partial, system, takeover, natalie, shapira, transformers, reason, about, variables, whose, meaning, shifts, across, contexts, train, finite, group, arithmetic, three, symbolic, mechanisms, emerge, commutative, copying, recognition, closure, based, cancellation, learned, purely, iclr, jannik, brinkmann, eric, todd, cool, projects, gotten, collaborate, collaborations, llms, unintended, nearby, measuring, building, custom, challenging, our, rag, pipeline, generate, structured, datasets, given, biology, paper, mechinterp, spotlight, igor, shilov, usha, bhalla, roy, rinberg, capturing, leveraging, existing, repositories, paradigm, need, millions, artworks, actually, learn, completely, still, mimic, seeing, hui, ren, opt, learning, styles, only, few, examples, traditionally, erasure, evaluated, externally, analyzing, generated, nuanced, suite, techniques, including, incontext, dynamic, tracing, probes, investigate, answer, niv, cohen, hegde, minh, pham, nicky, kriplani, kevin, flux, holding, quietly, absorb, color, style, then, drive, looks, interventions, trace, surprising, routing, info, flows, instead, directly, noise, tamar, rott, shaham, binding, had, opportunity, advise, most, amazing, students, advising, 2023, removing, while, preserving, overall, employs, closed, form, solution, address, bias, copyright, offensive, content, simultaneously, without, scalable, concurrent, edits, yonatan, belinkov, hadas, orgad, unified, light, weight, attributes, these, very, simple, provide, prompts, winter, weather, old, age, abstract, composed, continuously, controlled, eccv, tingrui, zhou, lora, show, used, classifiers, self, critiques, its, guides, unlearn, essential, retain, fluent, after, samuel, marks, sheridan, feucht, conceptual, decomposes, controllable, understandable, directions, single, framework, users, explore, discover, novel, encoded, any, nick, kolkin, eli, shechtman, richard, zhang, zongze, decomposing, super, lack, samples, ask, why, representations, required, base, don, use, theoretical, analysis, experiments, narrow, down, first, timestep, distilling, small, set, fewer, all, qwen3, track, region, currently, being, narrated, much, eye, speech, redirecting, inference, time, steers, describes, mask, should, extend, intent, getting, there, because, natural, extension, reading, behind, controlling, works, toward, future, behave, black, boxes, renderers, programmable, compositional, github, linkedin, twitter, scholar, previously, scientist, advancing, sensing, towards, better, disaster, management, monitoring, systems, using, neural, networks, indian, organization, phd, student, university, advised, interested, unlocking, also, closely, mit, interned,

Text of the page (random words):
rohit gandikota 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 your browser does not support the video tag gaze heads how vlms look at what they describe rohit gandikota david bau arxiv 2026 project page arxiv code dataset demo where do vlms look when they describe an image we find a small set of gaze heads fewer than 9 of all heads in qwen3 vl 8b that track the region currently being narrated much like human eye gaze during speech redirecting them at inference time causally steers what the model describes no retraining just attention mask editing distilling diversity and control in diffusion models rohit gandikota david bau wacv 2026 project page arxiv code 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 your browser does not support the video tag sliderspace decomposing the visual capabilities of diffusion models rohit gandikota zongze wu richard zhang david bau eli shechtman nick kolkin iccv 2025 project page arxiv code demo media blog 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 erasing conceptual knowledge from language models rohit gandikota sheridan feucht samuel marks david bau neurips 2025 project page arxiv code models 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 lora adaptors for precise control in diffusion models rohit gandikota joanna materzyńska tingrui zhou antonio torralba david bau eccv 2024 project page arxiv code demo 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 unified concept editing in diffusion models rohit gandikota hadas orgad yonatan belinkov joanna materzyńska david bau wacv 2024 project page arxiv code 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 erasing concepts from diffusion models rohit gandikota joanna materzyńska antonio torralba david bau iccv 2023 project page arxiv code demo 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 your browser does not support the video tag vision language binding in in context image generation chris ge rohit gandikota antonio torralba tamar rott shaham arxiv 2026 project page arxiv code flux 2 s text tokens aren t just holding your prompt during image editing they quietly absorb the reference image s color and style which then causally drive what the output looks like through causal interventions we trace this surprising routing reference info flows through text tokens to the output instead of directly from reference to noise when are concepts erased from diffusion models kevin lu nicky kriplani rohit gandikota minh pham david bau c hegde niv cohen neurips 2025 project page arxiv code 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 opt in art learning art styles only from few examples hui ren joanna materzyńska rohit gandikota david bau antonio torralba neurips 2025 project page arxiv code demo 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 ripplebench capturing ripple effects by leveraging existing knowledge repositories roy rinberg usha bhalla igor shilov rohit gandikota mechinterp spotlight neurips 2025 project page paper code 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 your browser does not support the video tag in context algebra eric todd jannik brinkmann rohit gandikota david bau iclr 2026 project page arxiv code 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 agents of chaos natalie shapira rohit gandikota david bau 37 authors arxiv 2026 project page arxiv 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 discovering forbidden topics in language models can rager chris wendler rohit gandikota david bau arxiv 2025 project page arxiv code 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 model tampering attacks enable more rigorous evaluations of llm capabilities zora che stephen casper rohit gandikota dylan hadfield menell 15 authors tmlr 2025 arxiv models 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 media agents of chaos red teaming autonomous llm agents science wired ieee spectrum die zeit afp wire techrepublic futurism times of india constellation research northeastern global news jack clark anthropic co founder concept sliders precise control in diffusion models unite ai marktechpost youtube tutorials 1 2 civitai community article hacker news discovering forbidden topics revealing hidden objectives in language models the economist china media project northeastern khoury news model tampering attacks rigorous llm capability evaluations uk ai security institute cited in the international ai safety report bengio chaired far ai talk by stephen casper esd and uce generative text to image unlearning cbs news by reporter christina hager analytics india magazine metaphysic ai marktechpost sliderspace discovering hidden knowledge inside diffusion models adobe research blog open source concept sliders integrated into ai toolkit comfyui automatic1111 community trained sliders on sliderfactory and civitai awards 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