Meta tags:
author=anil;
description=;

Headings (most frequently used words):

home, recent, updates,

Text of the page (most frequently used words):
accepted (13), was (9), paper (8), for (8), 2025 (8), with (7), unlearning (7), new (6), our (6), and (6), 2026 (6), the (6), models (5), llm (5), 2024 (4), emnlp (4), workshop (4), papers (4), language (3), acl (3), from (3), llms (3), senior (3), scientist (3), recent (3), work (3), privacy (3), home (3), anil (2), titled (2), reasoning (2), main (2), conference (2), challenge (2), semeval (2), attacks (2), neurips (2), amazon (2), detecting (2), towards (2), data (2), icml (2), serving (2), ethics (2), chair (2), meta (2), research (2), two (2), framework (2), visual (2), risk (2), code (2), memorization (2), tree, traversals, zero, shot, algorithm, augmenting, black, box, knowledge, graphs, improving, contrastive, decoding, extrapolating, probabilities, best, nomination, reviewer, benchmark, indirect, prompt, injection, advml, frontiers, nova, family, launched, collaborators, naacl, concluded, aaai, preventing, misinformation, real, sampling, boosting, factuality, diversity, open, ended, generation, via, asymptotic, entropy, appear, transactions, association, computational, linguistics, tacl, safety, agentic, deliberation, policy, embedded, cot, creation, resilient, relearning, sharpness, aware, minimization, perspective, beyond, will, starting, role, findings, selective, lume, constrained, entropic, primal, dual, large, area, arr, march, cycle, excited, join, organizing, committee, one, chairs, were, swan, semantic, watermarking, abstract, meaning, representation, narrow, emergent, misalignment, causes, consequences, containment, hosting, colm, exciting, lineup, speakers, responsibly, enabling, foundation, introduces, taxonomy, eccv, rethinking, compositional, severity, assessment, vlms, measuring, risks, memfm, functional, updates, previously, applied, responsible, team, agi, group, graduated, phd, program, department, computer, science, usc, 2019, where, working, prof, shrikanth, narayanan, ramakrishna,

Text of the page (random words):
anil ramakrishna home cv home home i work as a senior research scientist at meta previously i was a senior applied scientist in the responsible ai team of amazon agi group i graduated from the phd program at the department of computer science of usc in 2019 where i was working with prof shrikanth narayanan recent updates recent work on measuring code memorization risks in llms detecting functional memorization in code language models was accepted at icml 2026 memfm workshop our recent work rethinking visual privacy a compositional privacy risk framework for severity assessment with vlms introduces a new taxonomy for visual privacy risk and was accepted to eccv 2026 we re hosting a new workshop on responsibly enabling data for foundation models at colm 2026 with an exciting lineup of speakers two of our papers were accepted to the acl 2026 main conference from narrow unlearning to emergent misalignment causes consequences and containment in llms and swan semantic watermarking with abstract meaning representation i m excited to join the emnlp 2026 organizing committee as one of the ethics co chairs i m serving as an area chair at arr march 2026 cycle our paper constrained entropic unlearning a primal dual framework for large language models was accepted at neurips 2025 two papers on llm unlearning lume selective unlearning accepted to emnlp 2025 findings starting a new role at meta as a senior research scientist i will be serving as an ethics chair at emnlp 2025 our paper towards llm unlearning resilient to relearning attacks a sharpness aware minimization perspective and beyond was accepted at icml 2025 new paper towards safety reasoning in llms ai agentic deliberation for policy embedded cot data creation was accepted at acl 2025 new paper titled real sampling boosting factuality and diversity of open ended generation via asymptotic entropy to appear in transactions of association for computational linguistics tacl 2025 2 papers accepted at aaai 2025 workshop on preventing and detecting llm misinformation concluded llm unlearning challenge at semeval 3 papers with collaborators accepted at naacl 2025 amazon nova family of models launched our benchmark paper for indirect prompt injection attacks accepted at neurips advml frontiers 2024 workshop our paper on improving contrastive decoding by extrapolating probabilities was accepted at emnlp 2024 with best paper nomination from a reviewer llm unlearning challenge accepted at semeval 2024 new paper titled tree of traversals a zero shot reasoning algorithm for augmenting black box language models with knowledge graphs was accepted to main conference at acl 2024 anil