BerkeleyNLP @BerkeleyNLP
We work on natural language processing, machine learning, linguistics, and deep learning. nlp.cs.berkeley.edu Berkeley, California Joined September 2019-
Tweets92
-
Followers5K
-
Following33
-
Likes90
New paper from @berkeley_ai on Autonomous Evaluation and Refinement of Digital Agents! We show that VLM/LLM-based evaluators can significantly improve the performance of agents for web browsing and device control, advancing sotas by 29% to 75%. arxiv.org/abs/2404.06474 [🧵]
Do brain representations of language depend on whether the inputs are pixels or sounds? Our @CommsBio paper studies this question from the perspective of language timescales. We find that representations are highly similar between modalities! rdcu.be/dACh5 1/8
We know LLMs hallucinate, but what governs what they dream up? Turns out it’s all about the “unfamiliar” examples they see during finetuning Our new paper shows that manipulating the supervision on these special examples can steer how LLMs hallucinate arxiv.org/abs/2403.05612 🧵
The final layer of an LLM up-projects from hidden dim —> vocab size. The logprobs are thus low rank, and with some clever API queries, you can recover an LLM’s hidden dimension (or even the exact layer’s weights). Our new paper is out, a collaboration between lot of friends!
The final layer of an LLM up-projects from hidden dim —> vocab size. The logprobs are thus low rank, and with some clever API queries, you can recover an LLM’s hidden dimension (or even the exact layer’s weights). Our new paper is out, a collaboration between lot of friends!
What happens when RAG models are provided with documents that have conflicting information? In our new paper, we study how LLMs answer subjective, contentious, and conflicting queries in real-world retrieval-augmented situations.
What Evidence Do Language Models Find Convincing? Finds that LLMs today rely heavily on the relevance of a website to the query, while largely ignoring stylistic features that humans find important such as whether a text contains scientific references arxiv.org/abs/2402.11782
This is a very general flexible & general framework to automatically discover and explain patterns in image dataset. Could be used for ML models, scientific applications, etc. Check it out if you are interested!!
This is a very general flexible & general framework to automatically discover and explain patterns in image dataset. Could be used for ML models, scientific applications, etc. Check it out if you are interested!!
Honored to share our exciting paper on pacing!🎉 #EMNLP2023 Have you suffered overly verbose or vague LLM outputs? 👺 ✨Pacing is vital!✨ We try to improve pacing in long-form story planning.📚 All applause and thanks to my mentor @kevinyang41 first! [1/11]
LLMs can facilitate student cheating, spread misinformation on the web, and even poison future training datasets. Today, we’re releasing Ghostbuster, a state-of-the-art method for detecting LLM-generated text. Paper: arxiv.org/abs/2305.15047 Try it: ghostbuster.app
How should humans supervise AI🤖 if the gold answer is hard to directly verify? My paper on Scalable Oversight has been accepted to EMNLP 2023: "Labeling Programs with Non-Programmers Indirectly via Active Examples: A Case Study with Text-to-SQL" 🧵below
We got fascinating results in this work! * we reverse engineer the training set for Copilot/Codex * we show that data deduplication can sometimes hurt privacy * we reveal the tokenizer of black-box LLMs * we reveal other users' test inputs when adv ex defenses are used
We got fascinating results in this work! * we reverse engineer the training set for Copilot/Codex * we show that data deduplication can sometimes hurt privacy * we reveal the tokenizer of black-box LLMs * we reveal other users' test inputs when adv ex defenses are used
When analyzing ML security and privacy you need to study 𝐬𝐲𝐬𝐭𝐞𝐦𝐬, not just models! Our new paper shows that privacy is way worse when models are deployed in systems that use data cleaners, output filters, etc. Paper: arxiv.org/abs/2309.05610 Blog: spylab.ai/blog/side-chan…
Excited to present SILO, a new nonparametric LM that * excludes copyrighted data from parameters❌ * instead stores it in a datastore and retrieves it at inference time✨ * achieves performance that is close to the model trained on all data🚀 📄arxiv.org/abs/2308.04430
Excited to present SILO, a new nonparametric LM that * excludes copyrighted data from parameters❌ * instead stores it in a datastore and retrieves it at inference time✨ * achieves performance that is close to the model trained on all data🚀 📄arxiv.org/abs/2308.04430
Paper link is here arxiv.org/abs/2308.04430 and this work was led by @ssgrn and @sewon__min!
Feel risky to train your language model on copyrighted data? Check out our new LM called SILO✨, with co-lead @sewon__min Recipe: collect public domain & permissively licensed text data, fit parameters on it, and use the rest of the data in an inference-time-only datastore.
Copyright and legal risks are big open issues in today’s LLMs! In our new paper, we: - curate a pre-training corpus thats legally-permissive - analyze challenges w/ using public domain data - train permissive LMs - propose nonparametric “silos” for data of different legal risks
Copyright and legal risks are big open issues in today’s LLMs! In our new paper, we: - curate a pre-training corpus thats legally-permissive - analyze challenges w/ using public domain data - train permissive LMs - propose nonparametric “silos” for data of different legal risks
How can agents understand the world from diverse language? 🌎 Excited to introduce Dynalang, an agent that learns to understand language by 𝙢𝙖𝙠𝙞𝙣𝙜 𝙥𝙧𝙚𝙙𝙞𝙘𝙩𝙞𝙤𝙣𝙨 𝙖𝙗𝙤𝙪𝙩 𝙩𝙝𝙚 𝙛𝙪𝙩𝙪𝙧𝙚 with a multimodal world model!
Excited to share our new preprint on simulating RLHF preference data more effectively: "RLCD: Reinforcement Learning from Contrast Distillation for Language Model Alignment"! RLCD outperforms strong baselines on three alignment tasks across multiple LLaMA scales. 1/7
[1/9] Large Language Models (LLMs) can mimic humans to explain human decisions. But can they explain THEMSELVEs? How to evaluate explanations along this axis? Check out our work “Do Models Explain Themselves? Counterfactual Simulatability of Natural Language Explanations”!
Excited to announce our ACL Findings 2023 paper (w/ @KevinYa33964384 and Dan Klein): "PREADD: Prefix-Adaptive Decoding for Controlled Text Generation"! PREADD is a prompting-only controlled text generation method, allowing *variable control strength* by contrasting two prompts.
Akari Asai @AkariAsai
11K Followers 650 Following Ph.D. student @uwcse & @uwnlp. NLP. IBM Ph.D. fellow (2022-2023). Meta student researcher (2023-) . ☕️ 🐕 🏃♀️🧗♀️🍳Sam Bowman @sleepinyourhat
35K Followers 3K Following AI alignment + LLMs at NYU & Anthropic. Views not employers'. No relation to @s8mb. I think you should join @givingwhatwecan.Delip Rao e/σ @deliprao
46K Followers 5K Following Busy inventing the shipwreck. @Penn. Past: @johnshopkins, @UCSC, @Amazon, @Twitter ||Art: #NLProc, Vision, Speech, #DeepLearning || Life: 道元, improv, running 🌈Yoav Artzi @yoavartzi
13K Followers 163 Following Research/prof @cs_cornell + @cornell_tech🚡 / https://t.co/9YnWry7yHs / https://t.co/3VmRSyYm2d / asso. faculty director @arxiv / building https://t.co/f9QkzO5kaCDanish Pruthi @danish037
7K Followers 628 Following Faculty at Indian Institute of Science, Bangalore. PhD from @LTIatCMU.Kayo Yin @kayo_yin
8K Followers 560 Following PhD student @berkeley_ai @berkeleynlp working on interpretability and signed languages. Former @msftresearch @deepmind @carnegiemellon @polytechnique. 🇫🇷🇯🇵Bill Yuchen Lin 🤖 @billyuchenlin
6K Followers 2K Following Research @allen_ai. I evaluate (multi-modal) LLMs, build agents, and study the science of LLMs. Previously: @GoogleAI & @MetaAI FAIR @nlp_uscShruti Rijhwani @shrutirij
4K Followers 499 Following * Research Scientist @GoogleDeepMind * #NLProc research * PhD from @LTIatCMU * Amateur woodworker, scuba diver, foosball playerTim Dettmers @Tim_Dettmers
29K Followers 821 Following PhD Student at @UW. I blog about deep learning and PhD life at https://t.co/Y78KDJJFE7.Ofir Press @OfirPress
10K Followers 3K Following I build tough benchmarks for LMs and then I get the LMs to solve them. Postdoc @Princeton. PhD from @nlpnoah @UW. Ex-visiting researcher @MetaAI & @MosaicML.Naomi Saphra @nsaphra
7K Followers 1K Following Waiting on a robot body. ML/NLP. All opinions are universal and held by both employers and family. Same username on every lifeboat off this sinking ship.Christopher Potts @ChrisGPotts
11K Followers 620 Following Stanford Professor of Linguistics and, by courtesy, of Computer Science, and member of @stanfordnlp and @StanfordAILab. He/Him/His.Sebastian Ruder @seb_ruder
80K Followers 1K Following Multilingual LLMs @cohere • Prev: @GoogleDeepMind • Newsletter: https://t.co/7JGh2qpG98Mike Lewis @ml_perception
6K Followers 227 Following Llama3 pre-training lead. Partially to blame for things like the Cicero Diplomacy bot, BART, RoBERTa, kNN-LM, top-k sampling & Deal Or No Deal.Shaily @shaily99
5K Followers 2K Following PhD @LTIatCMU Prev: @GoogleAI @MSFTResearch. Working on ethics and evaluation in #NLProc. Usually ranting, often about research & DEI. 📚 @readsndrantsXin Eric Wang @xwang_lk
7K Followers 1K Following Multimodal and Embodied AI Researcher / Professor @UCSC. Director of https://t.co/Y4swOBag21. AI for Humanity in the long run. he/himWeijia Shi @WeijiaShi2
5K Followers 968 Following PhD student @uwcse @uwnlp | Visiting Researcher @MetaAI | Undergrad @CS_UCLA | https://t.co/eLBQmgkvymNils Reimers @Nils_Reimers
10K Followers 434 Following Director of Machine Learning @Cohere | ex-huggingface | Creator of SBERT (https://t.co/MKKOMfuQ4C)Facetointerface @facetointerface
0 Followers 4K Following The lion has come🫱🌐 #facetointerface @facetointerfaceD S K @DSK9919
63 Followers 4K Following Believe me anything is achievable if you are ready to die for it . #TECH SAVVY DATA HUNTER . PASSIONATE FOR NUMBERS AND POLITICS BEHIND IT .Guangyuan Jiang @jiang_gy
123 Followers 751 Following Computational Cognition & CogAI 🤖 Undergrad in AI @PKU1898 Peking University 🤯 Concept Learning & Abstraction 👋 Visiting Student @MITCoCoSciYujie Qian @Yujie_Qian
262 Followers 193 Following Founding Research Scientist @ Voyage AI; PhD @ MIT NLP GroupAbdulrahman Tabaza @embed_dim
4 Followers 798 Following enjoyer of various vector spaces, encoders and modalitiesMohammadreza @Mohammadre71127
1 Followers 362 FollowingMuaz Alemdar @mzalmdar
362 Followers 724 Following Bogazici Linguistics BA | 29 Mayıs Felsefe YL AI ve Klasik Türk Müziği ilgi alanlarıHarsh Bindal @HarshBindal16
37 Followers 181 Following Student of .....(still trying to figure it out)Himbo Mathématique @lhmccabe
462 Followers 1K Following cs phd student | data scientist | likes, follows, etc. not representative of personal viewsElrondex @elrondex
263 Followers 4K Following Elixir library to interact with Elrond Blockchain ⚡ $EGLD, Arwen, WASM, DeFI, SC, ESDTs, NFTs, SFTs, $MEX, DEX, AMM https://t.co/yPL9XXZguTSenaBeren @findingmerit
286 Followers 3K FollowingZhenwei @zenwill_ai
3 Followers 44 FollowingGerasimos Lampouras @glampouras_NLP
134 Followers 226 Following Call me Makis :) Team Leader of the London NLP Team at Huawei Noah's Ark Lab. Geek extra-ordinaire. My views are my own but they can be yours too!Eve Fleisig @enfleisig
375 Followers 332 Following PhD student @Berkeley_EECS | Princeton ‘21 | NLP, ethical + equitable AI, and linguistics enthusiastYingjian Fu @yingjianfu
17 Followers 496 Followingxcvxger @IrisLuo9
0 Followers 76 Followingoidestio @oidestio
2 Followers 327 Following On Twitter to learn about AI research and some related topicsMing-Bin Chen (Bryan) @chenbryan2103
5 Followers 68 Following Poetic coding monkey for journalism and AI.Aniket Pramanick @aniket_prama
81 Followers 366 Following PhD student at @UKPLab / @CS_TUDarmstadt | prev. @iiscbangalore | opinions are my own | he/himLost Epsi @quardepsi
0 Followers 15 FollowingShiqi Lou @lou_shiqi60535
10 Followers 119 FollowingLeon Engländer @LeonEnglaender
57 Followers 85 Following Co-maintainer @AdapterHub | #NLP research @UKPLab | Student @TUDarmstadtDazhi Peng @DazhiP
4 Followers 91 Followingfeifei bliu @BliuFeifei
16 Followers 53 FollowingSuvrakamal Das @subhrokomol
327 Followers 5K Following ML Research @Woxsen | Scholarship @HackTheNorth '23 | Building https://t.co/knTeHpyF0HEric Huang @EricHuang4312
2 Followers 52 Followinglin yu @linyu61852547
0 Followers 19 FollowingYao Tang @tyao923
19 Followers 191 Following Undergrad @SJTU1986 CS, working on RL & Decision MakingArda @ArdaYl37
83 Followers 409 FollowingQihui Lyu @QihuiL
157 Followers 222 Following Assistant Professor of Medical Physics @UCSF Radiation Oncology | @UCLA alumnaGood @Good86950654951
125 Followers 1K FollowingDonghong Ji @dhji_Jeff
2 Followers 54 FollowingAnikait Singh @Anikait_Singh_
125 Followers 264 Following PhD Student @StanfordAILab, Previously Student Researcher @GoogleDeepMind, Undergraduate @Berkeley_AI Deep Learning, Reinforcement Learning, Robotics.Deepankur John @jdeepankur
8 Followers 31 FollowingWeijun Qin @qinweijun99
12 Followers 48 FollowingHesam Asadollahzadeh @HesamAsdz
82 Followers 495 Following Trustworthy & Reliable AI/ML Researcher @MLL_SharifAI & @sangerinstitute; CSE BSc Student @UnivOfTehran; Header created by DALL·E 3 :)Percy Liang @percyliang
49K Followers 408 Following Associate Professor in computer science @Stanford @StanfordHAI @StanfordCRFM @StanfordAILab @stanfordnlp | cofounder @togethercompute | PianistJacob Andreas @jacobandreas
14K Followers 958 Following Teaching computers to read. Assoc. prof @MITEECS / @MIT_CSAIL (he/him). https://t.co/5kCnXHjtlY https://t.co/2A3qF5vdJwKayo Yin @kayo_yin
8K Followers 560 Following PhD student @berkeley_ai @berkeleynlp working on interpretability and signed languages. Former @msftresearch @deepmind @carnegiemellon @polytechnique. 🇫🇷🇯🇵Stanford NLP Group @stanfordnlp
145K Followers 179 Following Computational Linguists—Natural Language—Machine Learning @chrmanning @jurafsky @percyliang @ChrisGPotts @tatsu_hashimoto @MonicaSLam @Diyi_Yang @StanfordAILabAlane Suhr / suhr @ s.. @alsuhr
2K Followers 560 Following still kinda here while this site slowly falls apart. I like language, birds, cats, trains, buses, long walks, cities, and other things 🌻 ?/?Eve Fleisig @enfleisig
375 Followers 332 Following PhD student @Berkeley_EECS | Princeton ‘21 | NLP, ethical + equitable AI, and linguistics enthusiastZineng Tang @ZinengTang
1K Followers 569 Following PhD in @Berkeley_ai and @BerkeleyNLP. Previously @UNCNLP and @MSFTResearch.Alexander Wan @alexwan55
473 Followers 944 Following CS at Berkeley; @BerkeleyML @BerkeleyNLP; NLP researchSanjay Subramanian @sanjayssub
746 Followers 532 Following Building/analyzing NLP and vision models. PhD student @berkeley_ai. Formerly: @allen_ai, @pennArnav Gudibande @arnavg_
471 Followers 298 Following Research Engineer @perplexity_ai | prev MS @berkeley_ai @berkeleyNLPVanL @VanL
2K Followers 331 Following IP and Open Source Lawyer at @TaylorEnglish. Founder and CEO of @OSPOCO. Tweets are my own.Charlie Snell @sea_snell
4K Followers 5K Following PhD @berkeley_ai & student researcher @GoogleDeepMind. My friend told me to tweet more. I stare at my computer a lot and make thingsJessy Lin @realJessyLin
2K Followers 726 Following PhD @Berkeley_AI | interactive language agents 🤖 💬Jane Wakefield @janewakefield
8K Followers 1K Following I write about tech and have done for two decades. I also make pods, speak at conferences, and offer media training and consultancy under my 🍌🦔 brand.Kevin Yang @kevinyang41
455 Followers 176 Following 4th year PhD student at UC Berkeley working with Dan Klein, interested in controlled generation and long-form story generation.Kevin Lin 林冠言 @nlpkevinl
421 Followers 332 Following phd student @berkeleynlp @ucbrise, formerly @ai2_allennlpDaniel Fried @dan_fried
3K Followers 797 Following Assistant prof. @LTIatCMU @SCSatCMU, working on NLP: language interfaces, applied pragmatics, language-to-code, grounding. 🐘: @[email protected]Jonathan K. Kummerfel.. @jkkummerfeld
2K Followers 404 Following NLP faculty - University of Sydney he/him (this account is for professional topics only)David Hall @dlwh
2K Followers 1K Following Research Engineering Lead at @StanfordCRFM . Previously co-founder at Semantic Machines ⟶ MSFT. Lead developer of Levanter, Breeze. he/him @[email protected]Ruiqi Zhong @ZhongRuiqi
2K Followers 698 Following 5th Year Ph.D. @BerkeleyNLP, Columbia'19. part time working for @AnthropicAI . Supervising machines to do what I can't do.Rodolfo (Rudy) Corona @_rodolfocorona_
306 Followers 499 Following PhD student at @berkeley_ai and @BerkeleyNLP| Interested in language, embodiment, abstraction, and compositionality | 🇲🇽Berkeley AI Research @berkeley_ai
149K Followers 190 Following We're graduate students, postdocs, faculty and scientists at the cutting edge of artificial intelligence research.Lucy Li @lucy3_li
4K Followers 2K Following @UCBerkeley PhD student + @allen_ai. Human-centered #NLProc, computational social science, AI fairness. she/her. https://t.co/rtSSUhWQnLTaylor Berg-Kirkpatri.. @BergKirkpatrick
550 Followers 235 FollowingMohit Bansal @mohitban47
9K Followers 651 Following Parker Distinguished Professor, UNC Chapel Hill (@unc). Director https://t.co/5qlPVgnrlN (@uncnlp). Prev: @Berkeley_AI, @TTIC_Connect @IITKanpur #NLP, #CV, #AI, #MLGreg Durrett @gregd_nlp
6K Followers 752 Following CS professor at UT Austin. I do NLP most of the time. he/himJason Eisner @adveisner
8K Followers 547 Following Professor of CS at Johns Hopkins University, Director of Research at Microsoft Semantic Machines, ACL Fellow. My tweets speak only for me.Nicholas Tomlin @NickATomlin
693 Followers 619 Following PhD Student @Berkeley_EECS. Natural language processing. He/him.Eric Wallace @Eric_Wallace_
6K Followers 1K Following Researcher at OpenAI working to make language models more trustworthy, secure, and private.Do brain representations of language depend on whether the inputs are pixels or sounds? Our @CommsBio paper studies this question from the perspective of language timescales. We find that representations are highly similar between modalities! rdcu.be/dACh5 1/8
We know LLMs hallucinate, but what governs what they dream up? Turns out it’s all about the “unfamiliar” examples they see during finetuning Our new paper shows that manipulating the supervision on these special examples can steer how LLMs hallucinate arxiv.org/abs/2403.05612 🧵
Google presents: Stealing Part of a Production Language Model - Extracts the projection matrix of OpenAI’s ada and babbage LMs for <$20 - Confirms that their hidden dim is 1024 and 2048, respectively - Also recovers the exact hidden dim size of gpt-3.5-turbo…
The final layer of an LLM up-projects from hidden dim —> vocab size. The logprobs are thus low rank, and with some clever API queries, you can recover an LLM’s hidden dimension (or even the exact layer’s weights). Our new paper is out, a collaboration between lot of friends!
Google presents: Stealing Part of a Production Language Model - Extracts the projection matrix of OpenAI’s ada and babbage LMs for <$20 - Confirms that their hidden dim is 1024 and 2048, respectively - Also recovers the exact hidden dim size of gpt-3.5-turbo…
Future LLMs---whether they be RAG models, chatbots, or agents--will have to sift through misinformation, SEO text, and conflicting opinions when reading text. Alex led an interesting analysis of how current LLMs handle such conflicts. TLDR: LLMs love relevance, not style.
What happens when RAG models are provided with documents that have conflicting information? In our new paper, we study how LLMs answer subjective, contentious, and conflicting queries in real-world retrieval-augmented situations.
We also find that simply prefixing a website's text with "the following text is about [query]" can significantly improve its win-rate. On the other hand, perturbations that target stylistic features of the website, like adding scientific references, have a much weaker effect.
What happens when RAG models are provided with documents that have conflicting information? In our new paper, we study how LLMs answer subjective, contentious, and conflicting queries in real-world retrieval-augmented situations.
What Evidence Do Language Models Find Convincing? Finds that LLMs today rely heavily on the relevance of a website to the query, while largely ignoring stylistic features that humans find important such as whether a text contains scientific references arxiv.org/abs/2402.11782
This is a very general flexible & general framework to automatically discover and explain patterns in image dataset. Could be used for ML models, scientific applications, etc. Check it out if you are interested!!
[1/5] Introducing VisDiff - an #AI tool that describes differences in image sets with natural language. VisDiff can summarize model failures, compare models, find nuanced dataset differences, discover what makes an image memorable, and so much more! …derstanding-visual-datasets.github.io/VisDiff-websit…
Honored to share our exciting paper on pacing!🎉 #EMNLP2023 Have you suffered overly verbose or vague LLM outputs? 👺 ✨Pacing is vital!✨ We try to improve pacing in long-form story planning.📚 All applause and thanks to my mentor @kevinyang41 first! [1/11]
We got fascinating results in this work! * we reverse engineer the training set for Copilot/Codex * we show that data deduplication can sometimes hurt privacy * we reveal the tokenizer of black-box LLMs * we reveal other users' test inputs when adv ex defenses are used
When analyzing ML security and privacy you need to study 𝐬𝐲𝐬𝐭𝐞𝐦𝐬, not just models! Our new paper shows that privacy is way worse when models are deployed in systems that use data cleaners, output filters, etc. Paper: arxiv.org/abs/2309.05610 Blog: spylab.ai/blog/side-chan…
When analyzing ML security and privacy you need to study 𝐬𝐲𝐬𝐭𝐞𝐦𝐬, not just models! Our new paper shows that privacy is way worse when models are deployed in systems that use data cleaners, output filters, etc. Paper: arxiv.org/abs/2309.05610 Blog: spylab.ai/blog/side-chan…
Paper link is here arxiv.org/abs/2308.04430 and this work was led by @ssgrn and @sewon__min!
SILO Language Models: Isolating Legal Risk In a Nonparametric Datastore paper page: huggingface.co/papers/2308.03… The legality of training language models (LMs) on copyrighted or otherwise restricted data is under intense debate. However, as we show, model performance significantly…
Feel risky to train your language model on copyrighted data? Check out our new LM called SILO✨, with co-lead @sewon__min Recipe: collect public domain & permissively licensed text data, fit parameters on it, and use the rest of the data in an inference-time-only datastore.
Excited to present SILO, a new nonparametric LM that * excludes copyrighted data from parameters❌ * instead stores it in a datastore and retrieves it at inference time✨ * achieves performance that is close to the model trained on all data🚀 📄arxiv.org/abs/2308.04430
Feel risky to train your language model on copyrighted data? Check out our new LM called SILO✨, with co-lead @sewon__min Recipe: collect public domain & permissively licensed text data, fit parameters on it, and use the rest of the data in an inference-time-only datastore.
Copyright and legal risks are big open issues in today’s LLMs! In our new paper, we: - curate a pre-training corpus thats legally-permissive - analyze challenges w/ using public domain data - train permissive LMs - propose nonparametric “silos” for data of different legal risks
Feel risky to train your language model on copyrighted data? Check out our new LM called SILO✨, with co-lead @sewon__min Recipe: collect public domain & permissively licensed text data, fit parameters on it, and use the rest of the data in an inference-time-only datastore.
Excited to share our new preprint on simulating RLHF preference data more effectively: "RLCD: Reinforcement Learning from Contrast Distillation for Language Model Alignment"! RLCD outperforms strong baselines on three alignment tasks across multiple LLaMA scales. 1/7
I’m at #ICML2023 this week presenting work on poisoning LLMs and analyzing model pre-training data! Would love to chat about all things LLMs, especially on aspects like robustness/memorization/security/privacy. Feel free to DM or email.
We analyze 14 lang pairs to find when translation requires context. Our thematic analysis identifies 5 discourse phenomena, and we build the MuDA benchmark to automatically tag them. An exciting new resource to evaluate document-level MT on any data, check it out @aclmeeting!
Document-level context is essential to close the gap between MT and humans. But which words require context to be translated? And do models translate them well? Check our MuDA benchmark, for 14 LPs! To appear at #ACL2023! (co-lead @kayo_yin) arxiv.org/abs/2109.07446 1/15