Alex Havrilla @Dahoas1
Georgia Tech ML Researcher studying neural network learning theory and LLMs for mathematical reasoning. Intern at FAIR, MSFT Research. Co-founder of CarperAI. dahoas.github.io Joined August 2021-
Tweets150
-
Followers1K
-
Following503
-
Likes958
In my humble opinion the recent Stream of Search paper (arxiv.org/abs/2404.03683) is truly outstanding. Everyone should give it a thorough read.
@Thom_Wolf The 3 key elements of a good dataset: 1. quality 2. diversity 3. quantity You can only easily measure the last one but the performance is a sensitive function of all three. Super interesting topic ty for #longread :)!
I've finally uploaded the thesis on arXiv: arxiv.org/abs/2404.12150 It ties together a bunch of papers exploring some alternatives to RL for finetuning LMs, including pretraining with human preferences and minimizing KL divergences from pre-defined target distributions.
I've finally uploaded the thesis on arXiv: arxiv.org/abs/2404.12150 It ties together a bunch of papers exploring some alternatives to RL for finetuning LMs, including pretraining with human preferences and minimizing KL divergences from pre-defined target distributions. https://t.co/jq03eRcEhK
I am super excited to share our Llama3 preview models (8B and 70B). I am proud to have been a part of this amazing effort over the past 8 months. We still have some super cool stuff coming up in the coming months... until then, enjoy playing with these preview models…
Had a great time during our discussion, thanks again for having me!
Had a great time during our discussion, thanks again for having me!
How to define Diversity in the context of CodeLMs and Programming Languages ? 1. Diversity is positively correlated with Performance in solving a problem. 2. Shortcomings of diversity in small codeLMs. 3. Code Embedding models don't capture semantics. reshinthadithyan.github.io/blog/2023/code…
Super proud to share what we've been cooking with the amazing team at @cohere - a nimble model that's efficient, retrieves docs and gives citations, knows how to use tools and supports 10 languages. Ready to use in your business! Have a chat with it at: coral.cohere.com
Super proud to share what we've been cooking with the amazing team at @cohere - a nimble model that's efficient, retrieves docs and gives citations, knows how to use tools and supports 10 languages. Ready to use in your business! Have a chat with it at: coral.cohere.com
Jim Fan @DrJimFan
229K Followers 3K Following @NVIDIA Sr. Research Manager & Lead of Embodied AI (GEAR Lab). Creating foundation models for Humanoid Robots & Gaming. @Stanford Ph.D. @OpenAI's first intern.SynthLabs @synth_labs
12K Followers 43 Following AI Aligned with Your Vision. We’re doing cutting edge research for transparent, auditable AI alignment.Louis Castricato @lcastricato
3K Followers 477 Following Math @uwaterloo, RLHF @BrownCSDept, Goosefluencer. x-RS @aieleuther, x-Head of LLMs @stabilityai, x-lead @CarperAI. co-founder @synth_labs. We're hiring.Tanishq Mathew Abraha.. @iScienceLuvr
54K Followers 1K Following PhD at 19 | Founder and CEO at @MedARC_AI | Research Director at @StabilityAI | @kaggle Notebooks GM | Biomed. engineer @ 14 | TEDx talk➡https://t.co/xPxwKTq6QbStella Biderman @BlancheMinerva
15K Followers 748 Following Open source LLMs and interpretability research at @BoozAllen and @AiEleuther. My employers disown my tweets. She/herSebastian Raschka @rasbt
267K Followers 906 Following Machine learning & AI researcher writing at https://t.co/A0tXWzG1p5. LLM research engineer @LightningAI. Previously stats professor at UW-Madison.Nathan Benaich @nathanbenaich
51K Followers 32K Following solo member of investment staff @airstreet, brewing ambition @airstreetcafe, next token predictor @airstreetpressEleutherAI @AiEleuther
19K Followers 76 Following A non-profit research lab focused on interpretability, alignment, and ethics of artificial intelligence. Creators of GPT-J, GPT-NeoX, and VQGAN-CLIPDelip 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 🌈Robert Scoble @Scobleizer
504K Followers 68K Following Follow me on my new podcast with AI startups, Unaligned. Tech industry color commentator since 1993. Author/Blogger. Former strategist @Microsoft.Jeremy Howard @jeremyphoward
222K Followers 5K Following 🇦🇺 Co-founder: @AnswerDotAI & @FastDotAI ; Hon Professor: @UQSchoolITEE ; Digital Fellow: @StanfordSharif Shameem @sharifshameem
53K Followers 3K Following founder @LexicaArt • in pursuit of good explanationsNathan Lambert @natolambert
25K Followers 689 Following Figuring out AI @allen_ai, "rl boi" DM me papers. Writes @interconnectsai, talks @retortai Has phd and some credentialsHaofu Liao @LiaoHaofu
2 Followers 73 FollowingBayan Sporle @BSporle54687
83 Followers 5K FollowingElana Unzicker @ElaUnzicke
91 Followers 5K FollowingAl Mamun @al_mamun_sardar
276 Followers 3K Following Looking for NLP/LLM related PhD positions | Research Assistant (NLP) at Jahangirnagar University | MSc (CS)Eyüp Pastırmacı @eyuppastirmaci
15 Followers 128 FollowingHarriett Becky @HarrieBeck
41 Followers 5K FollowingMojtaba Vàlipour @ValipourMojtaba
391 Followers 3K Following CS PhD at @UWaterloo, Founding Engineer at Coastal Carbon and part time Researcher at @huawei Noah’s Arc Lab, prev enjoyed my time at @oraclelabs, & @CVC_UABJanhavee Shinde @SJanhavee
56 Followers 2K FollowingCzentye Levente @lordlewo
23 Followers 3K Following阮小胖 @Pang1231231
0 Followers 7 FollowingSam Kuhn @SamKuhnDev
214 Followers 4K Following Full Stack 3D Web Developer. WebGPU, WebGL/XR, Serverless, CloudVarun Gangal @VarunGangal
724 Followers 5K Following Research Scientist @asapp; #NLProc PhD from Carnegie Mellon LTI @LTIatCMU (2016-22) RTs ≠ endorsements. Views personal, not of employers/institutions.Krishna Srinivasan @krishna2
471 Followers 2K Following I code stuff. I build models. I work in NLP/DL at Google Research.alonet @alonet
94 Followers 1K Following A dreamer. Life is getting full of Prompts.#cloudnative #GenerativeAI #modelontheEdge #petrolhead #gadgets freak #smart device #SaaS Architect #xR #immersiveExpDuttonΦ @duttonphi
105 Followers 478 Following ..aagen (double (2x) agent).. ..previously Wołfram|Ałpha.. ..baeksu.ai.. ..jajangmyeon all day/all night..ZKP @ZKPxyz
557 Followers 5K Following AIxWeb3 hacker and Angel Investor. Web1 and Web2 veteran. Currently exploring the agentic efficient frontier.S. Iqbal @S_Iqbal90
26 Followers 1K FollowingSidney Liquori @LiquoriSidn
50 Followers 5K FollowingFred Bliss @fblissjr
432 Followers 1K Following 🔎 actively pondering the next big adventure, all things Data & AI 🤝 Founder @ Aptitive (acquired Q4 2021 @ ~55 FTEs) 💡 AI/ML/LLM enthusiast, building in openAnton Volnuhin @antonme
2K Followers 3K Following PM who loves to write a bit of code, into 3d printing and some other stuff. @yandex alumni (12 years, head of blog search service, tech marketing, taxi)Blake Camp @blake_camp_1
406 Followers 1K Following AI (PhD), markets, aviation, brains, futbol, pizza, wine, books, films, family, friends, lifeAndrew Dai @andrewdai99
191 Followers 253 Following Researcher @Aleph__Alpha, a Kerryman, LLMs x open-endedness; prev @tcddublin, @FormulaTrinity Autonomous | 🇮🇪tom @0xluciusv
149 Followers 454 Following i like cuda kernels, c++, rust, go, and nvim. (cons e/λx.x 🌎/acc) wrong about a lot of things but trying to learnJosh @JoshPurtell
730 Followers 2K Following ML Researcher. Ex: Cyber microexit, Yale Math. Hiring in ML Ars longaNikita @nikitavoloboev
4K Followers 7K Following Make @LearnAnything_ Learn in public: https://t.co/GbFvuErkYn macOS course: https://t.co/JdbJWru6zG https://t.co/94R8ER7K2h https://t.co/ROkqhyhpEKMarkus Rauhalahti, Ph.. @MRauhalahti
844 Followers 5K Following Independent researcher: molecular design/nanotech, human-AI collab, informetrics, DIY-scihw. Prev compchem/biophys/info phd&postdoc @helsinkiuniRonnie Cotham @RonnieCoth83122
17 Followers 3K FollowingJonathan Balloch @JonathanBalloch
343 Followers 923 Following I mostly tweet about #ai, #robots, #science, and the @packers... Robotics PhD student @GeorgiaTech studying #RL and #AI My thoughts and opinions are my ownShannon Turansky @ShanTurans
80 Followers 5K FollowingKanishk Gandhi @gandhikanishk
921 Followers 692 Following Phd @Stanford CS; w/ Noah Goodman, Dorsa Sadigh | Prev: @LakeBrenden @NYUDataScience, @IITKanpur, @Path_AIWaypoint AI @waypointai
111 Followers 4K Following Automate bug triage & root cause analysis to solve bugs quickly & get back to building! -- Book an appointment: 📅https://t.co/lgRrvWE0gSAaditya ; @Aaditya26082004
524 Followers 7K Following CS'26 • Machine Learning • Open-Source • Web Dev. • Algorithms • Jai Shree Krishna 🦚🪈..still, who is David.. @jimat944
507 Followers 979 Following curious how will we define the public good i.e., 👉 https://t.co/sE02VgqC1t #righttotheurbancity ..don't die in a crosswalk. #imagineeringatlanta #twitterdeveloperCardGamesFanatic @CardmeisterGuy
2K Followers 5K Following Guy interested in transit and politics. Tends to talk too much.Briony Kitanik @kita_bri
38 Followers 5K FollowingYofo Diame @YDiame87519
329 Followers 4K FollowingAlex Murphy @Alxmrphi
491 Followers 2K Following 🙋♂️= NeuroAI Researcher. Postdoc @ Amii / UAlberta studying language, vision & (and in) DNNs and brains. PhD & ex-Google Brain intern (🇮🇪 & 🇬🇧)Fasil Muhammad @pvfasil
13 Followers 56 FollowingAmanda Rendel @rend_aman
36 Followers 5K FollowingMartin Andrews @mdda123
601 Followers 1K Following AI Research / Founder @ Red Dragon AI. Co-organiser of Machine Learning Singapore MeetUp. @GoogleDevExpert (ML). Fixed Income quant in NYC during AI winterAndrej Karpathy @karpathy
978K Followers 904 Following 🧑🍳. Previously Director of AI @ Tesla, founding team @ OpenAI, CS231n/PhD @ Stanford. I like to train large deep neural nets 🧠🤖💥Jim Fan @DrJimFan
229K Followers 3K Following @NVIDIA Sr. Research Manager & Lead of Embodied AI (GEAR Lab). Creating foundation models for Humanoid Robots & Gaming. @Stanford Ph.D. @OpenAI's first intern.Yann LeCun @ylecun
711K Followers 718 Following Professor at NYU. Chief AI Scientist at Meta. Researcher in AI, Machine Learning, Robotics, etc. ACM Turing Award Laureate.SynthLabs @synth_labs
12K Followers 43 Following AI Aligned with Your Vision. We’re doing cutting edge research for transparent, auditable AI alignment.Stability AI @StabilityAI
190K Followers 31 Following We are building the foundation to activate humanity's potential.François Chollet @fchollet
469K Followers 770 Following Deep learning @google. Creator of Keras. Author of 'Deep Learning with Python'. Opinions are my own.Google DeepMind @GoogleDeepMind
943K Followers 275 Following We’re a team of scientists, engineers, ethicists and more, committed to solving intelligence, to advance science and benefit humanity.Louis Castricato @lcastricato
3K Followers 477 Following Math @uwaterloo, RLHF @BrownCSDept, Goosefluencer. x-RS @aieleuther, x-Head of LLMs @stabilityai, x-lead @CarperAI. co-founder @synth_labs. We're hiring.Tanishq Mathew Abraha.. @iScienceLuvr
54K Followers 1K Following PhD at 19 | Founder and CEO at @MedARC_AI | Research Director at @StabilityAI | @kaggle Notebooks GM | Biomed. engineer @ 14 | TEDx talk➡https://t.co/xPxwKTq6QbJohn Carmack @ID_AA_Carmack
1.1M Followers 241 Following AGI at Keen Technologies, former CTO Oculus VR, Founder Id Software and Armadillo AerospaceStella Biderman @BlancheMinerva
15K Followers 748 Following Open source LLMs and interpretability research at @BoozAllen and @AiEleuther. My employers disown my tweets. She/herBojan Tunguz @tunguz
187K Followers 8K Following Machine Learning ex Nvidia. Kaggle Quadruple Grandmaster. Data Scientist. Physicist. Catholic. Husband. Father. Stanford Alum. e/xgb. XGBoost.eth. AMDG.Sebastian Raschka @rasbt
267K Followers 906 Following Machine learning & AI researcher writing at https://t.co/A0tXWzG1p5. LLM research engineer @LightningAI. Previously stats professor at UW-Madison.Eliezer Yudkowsky ⏹.. @ESYudkowsky
175K Followers 89 Following The original AI alignment person. Missing punctuation at the end of a sentence means it's humor. If you're not sure, it's also very likely humor.Eric Jang @ericjang11
69K Followers 3K Following physical AGI at 1X. Author of "AI is Good for You" https://t.co/eFg4WXhg0pOpenAI @OpenAI
3.4M Followers 0 Following OpenAI’s mission is to ensure that artificial general intelligence benefits all of humanity. We’re hiring: https://t.co/dJGr6LgzPAAndrew Dai @andrewdai99
191 Followers 253 Following Researcher @Aleph__Alpha, a Kerryman, LLMs x open-endedness; prev @tcddublin, @FormulaTrinity Autonomous | 🇮🇪Teknium (e/λ) @Teknium1
29K Followers 3K Following Cofounder @NousResearch, prev @StabilityAI Github: https://t.co/LZwHTUFwPq HuggingFace: https://t.co/sN2FFU8PVE Support me on Github SponsorsRonen Eldan @EldanRonen
2K Followers 115 Following Previously doing maths at @WeizmannScience, currently AI researcher at @MSFTResearch. Pretty good at loading a dishwasher.Sebastian Borgeaud @borgeaud_s
994 Followers 259 Following Research Engineer at DeepMind with a focus on Large Language Models and large scale Deep LearningQingfei You @qingfeiyou
30 Followers 199 FollowingNathan Cooper @ncooper57
721 Followers 650 Following The world can be ugly and cruel to the most innocent. Consider donating to help children suffering from one of the worst things: https://t.co/PYZWj8o4OWAlon Albalak @AlbalakAlon
885 Followers 464 Following CS PhD candidate at @ucsbNLP. Research: Data-centric AI, Efficiency in ML, NLP.Song Mei @Song__Mei
1K Followers 550 Following Assistant Professor at UC Berkeley, Department of Statistics and EECS. Researcher working on foundations of generative AI.Brian Hie @BrianHie
5K Followers 402 Following Assistant professor @StanfordEng ChemE and @StanfordData, Innovation Investigator @arcinstitute | Machine learning for biologyAubrey de Grey @aubreydegrey
60K Followers 11 Following I'm spearheading the global crusade to defeat aging. President and CSO of https://t.co/QxFW8fuCt2Shunyu Yao @ShunyuYao12
7K Followers 857 Following Language agents (ReAct, Reflexion, Tree of Thoughts) for digital automation (WebShop, SWE-bench, SWE-agent)Aleksandr Sviridov @_sviridov_
993 Followers 106 Following Longevity activist and software engineering lead. CTO at @GeroSense. Co-Founder of @SayForeverOrg.Tom Goldstein @tomgoldsteincs
23K Followers 2K Following Professor at UMD. AI security & privacy, algorithmic bias, foundations of ML. Follow me for commentary on state-of-the-art AI.Tom Zahavy @TZahavy
2K Followers 318 Following Building agents that discover knowledge and get better at doing so over time. Staff research scientist @GoogleDeepMindTaelin @VictorTaelin
17K Followers 903 Following Founder of @HigherOrderComp Building the massively parallel future of computing Reaching AGI to cure all diseases and suffering is all that mattersAmbrosia Path - Longe.. @ambrosiapath
227 Followers 163 Following Reporting the latest longevity news and interviews Subscribe to our weekly newsletter to get the latest longevity news in your inbox. Join 2,354+ subscribersSakana AI @SakanaAILabs
19K Followers 0 Following We are a Tokyo-based R&D company on a quest to create a new kind of foundational AI model based on nature-inspired intelligence. https://t.co/1q07mb3TzERobin Rombach @robrombach
6K Followers 397 Following Generative enthusiast and long-term PhD Student @LMU_Muenchen. Author of VQGAN, Latent Diffusion, Stable Diffusion.Tianlin @linylinx
6K Followers 579 Following ML Tech Lead @sourceful ⏩: @illumina AI Lab @qualcomm AI, PhD @LSEStatistics 📜 generative models 🤪 joking not jokingAllan Jie @allanjienlp
58 Followers 327 Followingalewkowycz @alewkowycz
3K Followers 173 Following Member of Technical Staff at @inflectionAI. Former Research Scientist @Google. In a previous life, I did String Theory. Language models and Conversational AI.Jacob Austin @jacobaustin132
3K Followers 797 Following @Google @DeepMind researcher. AI for math and science. Coding. Gemini. I also play piano. NYC. Opinions my ownRobert Yang @GuangyuRobert
3K Followers 185 Following Co-founder, CEO at @Altera_AL, Computational Neuroscientist, former Assistant Professor @mitbrainandcog & @MITEECSAlex Gu @minimario1729
2K Followers 2K Following phd @MIT_CSAIL, llm for math and code. intern @MetaAI and analyst @pillar_vc. prev @BigCodeProject, @MITIBMLab, @JaneStreetGroup, @PonyAI_techArc Institute @arcinstitute
22K Followers 24 Following A new scientific institution for curiosity-driven biomedical science and technology.Wojciech Galuba @wgaluba
490 Followers 1K Following Head of Data & Evals @Cohere | prev: Research Eng Lead @MetaAI | founded @Meta’s A/B testing platform and the AI annotation platform | @ICepfl alumnusCognition @cognition_labs
123K Followers 19 Following Makers of Devin, the first AI software engineer. We are an applied AI lab focused on reasoning, and code is just the beginning. Join us: https://t.co/tpfZwEwGiqAlbert Gu @_albertgu
9K Followers 90 Following assistant prof @mldcmu. chief scientist @cartesia_ai. leading the ssm revolution.Jenny Zhang @jennyzhangzt
752 Followers 430 Following PhD student @UBC • Previously at @ASTARsg • Undergrad @imperialcollege • Reinforcement Learning, Open-endedness, AI-GAsCong Lu @cong_ml
635 Followers 864 Following Postdoctoral Research Fellow @UBC_CS in open-endedness, generative models, and deep RL. Prev: PhD @UniofOxford, Research Intern @Waymo, @MSFTResearch!Lior⚡ @AlphaSignalAI
84K Followers 895 Following Covering the latest in AI R&D • ML Engineer • Ex-Mila researcher • MIT Lecturer • Building AlphaSignal, a technical newsletter read by 180,000+ ML experts.nathan lile @NathanThinks
2K Followers 883 Following ceo/cofounder @ https://t.co/bDd3J4Lmzf (we're hiring!) #GenerativeAI recurrent rabbit hole victim. swims in data lakes & pools. nothing great is easy.Julia Bauman @JuliaBauman2
8K Followers 456 Following PhD student at @Stanford Genetics in @LarsMSteinmetz lab | Prev @broadinstitute | Explaining cool biotech to the world here and @ 60_SecondScience on TikTokYuandong Tian @tydsh
16K Followers 801 Following Research Scientist and Senior Manager in Meta AI (FAIR). AI-guided Optimization and Representation Learning. Novelist in spare time. PhD in @CMU_Robotics.Anton Bakhtin @ SF @anton_bakhtin
2K Followers 127 Following MTS at @AnthropicAI, Ex @MetaAI, Ex @Google Three logicians walk into a bar ...Anshul Kundaje (anshu.. @anshulkundaje
22K Followers 2K Following Genomics, Machine Learning, Statistics, Big Data and Football (Soccer, GGMU). Post: @anshulkundaje, Threads: anshulkundajeMichael Poli @MichaelPoli6
2K Followers 279 Following @Stanford @StanfordAILab, Staff Scientist @togethercompute, prev @MSFTResearch. DL, numerics and systems. I like to architect big neural nets that run fast.Eric Nguyen @exnx
2K Followers 324 Following PhD in BioEngineering & AI @stanford @HazyResearch @StanfordAILab @arcinstituteBo Wang @BoWang87
8K Followers 2K Following Assistant Prof. CS,LMP @UofT; CIFAR AI Chair @VectorInst; Chief AI Scientist, @UHN; former PHD, CS @Stanford; opinions my own. #AI #healthcare #combioPaul Scotti @humanscotti
1K Followers 937 Following Head of Neuroimaging @StabilityAI, leading the @MedARC_AI Neuroimaging & AI Lab | Collaborating w/ @PrincetonNeuro @ptoncompmemlabCool stuff, we found a similar result back in December arxiv.org/abs/2312.01037. Kind of upset they didn't cite/link to us tbh.
New Anthropic research: we find that probing, a simple interpretability technique, can detect when backdoored "sleeper agent" models are about to behave dangerously, after they pretend to be safe in training. Check out our first alignment blog post here: anthropic.com/research/probe…
Happening now!
events.nationalacademies.org/42507_04-2024_… Giving a talk on evaluating large language models for mathematics through interactions (work co-lead with @katie_m_collins) on Thursday. In the same session is the one and only @ChrSzegedy!
@alec_helbling This can be understood very nicely from the sampling perspective! The score can be written out as an expectation of the conditional distribution. When time is large this conditioning doesn't change the original distribution!
Cannot agree more. My intuition is that FFN is for storing knowledge (this is why most knowledge editing are on FFNs) and Attention is for implementing algorithms (this is why most mechanistic interpretability, e.g., induction heads, are on Attn). Additionally, it seems that…
not true, especially for language. if you trained a large & deep MLP language model with no self-attention, no matter how much data you'll feed it you'll still be lacking behind a transformer (with much less data). will it get to the same point? i don't think so. your tokens…
@Dahoas1 @stefan_fee This is likely one reason why model-generated rationales can be better than human-written rationales.
Super excited to share that I successfully defended my PhD thesis "Understanding Generalization and Robustness in Modern Deep Learning" today 👨🎓 A huge thanks to the thesis examiners @SebastienBubeck, @zicokolter, and @KrzakalaF, jury president Rachid Guerraoui, and, of course,…
@Thom_Wolf The 3 key elements of a good dataset: 1. quality 2. diversity 3. quantity You can only easily measure the last one but the performance is a sensitive function of all three. Super interesting topic ty for #longread :)!
Introducing AgentKit. AgentKit offers a unified framework for explicitly constructing a complex human "thought process" from simple natural language prompts. The user puts together chains of nodes, like stacking LEGO pieces. (1/3) github.com/Holmeswww/Agen…
I2SB (i2sb.github.io) + (nearly optiaml) flow matching yields 1000x speed-up compared to standard denoising diffusion for generating highly accurate transition states 🧑🔬⚗️🧪 Check out our new preprint👇 arxiv.org/abs/2404.13430 Very fun collab w/ @chenru_duan @YuanqiD!
New paper alert: React-OT: Optimal Transport for Generating Transition State in Chemical Reactions (arxiv.org/abs/2404.13430). React-OT formulates TS search as a transport problem, approaching chemical accuracy while taking only 0.5 seconds in inference on a single GPU. #compchem
Among the coolest projects I helped with at Stanford. The key idea is very simple: a pragmatic response in one context is something you'd rarely say in other contexts. This basic principle lets LMs teach themselves to generally follow constitutions but has many cool implications
Constitutional AI showed LMs can learn to follow constitutions by labeling their own outputs. But why can't we just tell a base model the principles of desired behavior and rely on it to act appropriately? Introducing SAMI: Self-Supervised Alignment with Mutual Information!
updated github.com/balloch/awesom… finally. let me know if I am missing anything obvious/exciting
Join us on the fun LLM agent workshop at CMU!
On May 2-3, we're going to have a big event in Pittsburgh about LLM Agents. We have invited talks from great speakers inside and outside CMU, student research presentations and posters, tutorials and discussions! Come join us at CMU campus, and register at cmu-agent-workshop.github.io
Many LLM fine-tuning methods. Unclear what you should use & why? In our new paper, we did an extensive study of on-policy RL, supervised & offline contrastive methods (DPO, IPO) to answer this... 🧵⬇️ On-policy > offline, mode-seeking > mode-covering understanding-rlhf.github.io
Can LLMs acquire meaning/semantics from just text? Some think it is a priori not possibile, I personally think it's a super interesting philosophical question which needs further investigation! Thoughts? arxiv.org/abs/2404.12145
📜 New preprint! Equipped with our multisense consistency method, we dive deep into an exploration of the semantic understanding of #LLMs. @eliabruni & @_dieuwke_ @metaai #NLProc [1/7]🧵 arxiv.org/abs/2404.12145
On May 2-3, we're going to have a big event in Pittsburgh about LLM Agents. We have invited talks from great speakers inside and outside CMU, student research presentations and posters, tutorials and discussions! Come join us at CMU campus, and register at cmu-agent-workshop.github.io
Crazy finding!!!!! -> ” Without introducing any additional data or advanced training techniques, and merely by reformatting the response, LLaMA-2-13B’s mathematical reasoning ability on GSM8K can be improved from 46.77% to 56.63% in accuracy"
Been diving into some papers on data synthesis lately, especially those about enhancing math reasoning. Most of them seem to miss our work on 'Reformatted Alignment' (arxiv.org/abs/2402.12219)—another approach to boosting data for math reasoning.
I've finally uploaded the thesis on arXiv: arxiv.org/abs/2404.12150 It ties together a bunch of papers exploring some alternatives to RL for finetuning LMs, including pretraining with human preferences and minimizing KL divergences from pre-defined target distributions.
I was very impressed with @tomekkorbak's thesis! Some really nice insights into LLM alignment: 1) RL is not the way --> distribution matching let's us target constraints like "generate as many of these as of those" 2) fine-tuning is not the way --> PHF aligns during pre-training
@moinnadeem I used to study mathematics. Sometimes I would not know how to prove something, so I'd stare at the wall and think until I found a proof. I don't think the visual input of the wall was essential to this process.
A lot of LLM benchmarks don't properly test out of distribution behavior. As uses of LLMs for science increase, we need benchmarks that actually check generalization beyond training data. My experience so far that LLMs are pretty weak outside training distribution, but can have…
Very excited about the release of arena hard, the main benchmark we looked at when selecting the checkpoints for Starling model. It focuses on a subset of very hard prompts from chatbot arena.
Introducing Arena-Hard – a pipeline to build our next generation benchmarks with live Arena data. Highlights: - Significantly better separability than MT-bench (22.6% -> 87.4%) - Highest agreement to Chatbot Arena ranking (89.1%) - Fast & cheap to run ($25) - Frequent update…