Haau-Sing Li 李 效丞

1/12 📢New paper alert “DOCE: Finding the Sweet Spot for Execution-Based Code Generation” The work is a group effort with @psanfernandes, @IGurevych, and @andre_t_martins. @UKPLab @Lisbon_ELLIS @deep_spin 📰: arxiv.org/pdf/2408.13745

@Xianbao_QIAN Impressive performance when deployed for HTML/SVG services even when just on a 14B model.

New model updates from iquestlab. If you're trying to find an inference model that you can run offline, this is probably the one you're looking for. - 7B and 14B coding models - Optimized for tool use, CLI agents and HTML generation - 128k context length - Explicit and detailed prompting works best - MiT license with requirement of display logo - available on @huggingface

@SinclairWang1 We received very similar comments! We put "software engineering" and "analysis" there for keywords, with justifications already in the paper (intro and more), yet "lacks novelty" still appears for all reviewers. Guess a good rebuttal has to go right? @COLM_conf 🤔

#colm2025 an interesting review i just received: “this work does a lots of job balabala，but lack of novelty” The track we selected is "all about data". So, what is the novelty in the lens of language modeling? 😅😅😅😅 I believe the colm conference is definitely different academic conference and the reviewers are more professional, so I choose to submit my paper to it. However, it has been proven that the reviewers of colm are not always professional and even somewhat view the research results in the field of language models with an outdated perspective under the rapid and drastic iteration.😅 @COLM_conf

Also huge shoutout to my amazing supervisors @IGurevych and @andre_t_martins for their support and wisdom. Wouldn’t be possible without you guys, and more exciting things will come.

Sharing that I will join @tiktok_us AI Innovation Center @BytedanceTalk for research internship. Will still work on Code AI advised by the amazing @sivil_taram. Hope there’ll be luck to us bringing the community nice deliverables! Execution can and will talk. Happy 2025!

2) We have a spotlight poster in the main conference, “Reranking Laws for Language Generation: A Communication-Theoretic Perspective” with @tozefarinhas and @LHaausing (Thursday Dec 12 11:00-14:00).

@a_stadt Was going to say I can review one but seems it’s going well!☺️

I'm in need of NINE emergency reviews for ACL ARR. Over 1/3 of my reviewers are nonresponsive, I think that's a personal record 🙃 Please let me know if you can take on some of these!!

@apsdehal Do you provide visa sponsorship? Would be nice to work with you as I think I work on things super related to what you do but I need sponsorship.

We're hiring both winter and summer research interns at Contextual AI! Apply at contextual.ai/careers/?gh_ji… if you're interested in working on RAG, LLMs, retrieval, alignment, synthetic data, evaluation, and multimodal related topics with our amazing research team.

@gan_chuang DOCE: Finding the Sweet Spot for Execution-Based Code Generation (the first unified inference framework for code gen, I presume strongly correlated to reasoning :) ) arxiv.org/pdf/2408.13745

I’m hiring multiple research interns at the MIT-IBM lab to work on advanced LLM reasoning! The application process is simple—just send me your favorite published paper. Only one😀!!

@LucileSaulnier @MistralAI Hi! I applied for the PhD research internship, can I message you for a discussion about that?

🌟 AI enthusiasts! Join @MistralAI and shape the future of generative AI! 🌟 We're hiring AI Scientists, Research Engineers, and more 🌐 Check out our openings: jobs.lever.co/mistral 🚀 Be part of a brilliant team working on cutting-edge projects. #AIJobs #TechCareers

@mialon_gregoire @MekalaDheeraj Hi Grégoire! I'm interested in this position and have submitted my application. We recently have papers published (at NeurIPS 2024 spotlight)/in submission (x.com/LHaausing/stat…, x.com/LHaausing/stat…), see if that interests you!

I am hiring an intern in our Llama team for 2025! Near the end of PhD completion, willing to be based out of Paris. You will succeed @MekalaDheeraj, work around frontier LLMs, tool use, agents, and more :) Please apply here: metacareers.com/jobs/109555634…

@natolambert 🫤

@LHaausing nope, no PRMs effectively

Yay! There's another reward model evaluation other than RewardBench (they do build on our code :) ). This one is a mix of seeking better correlation with "vibes" evals like ArenaHard and MT Bench + some best of N sampling correlation. RMB: Comprehensively Benchmarking Reward Models in LLM Alignment Zhou et al. Some dataset notes: * Source prompts from WildChat, uses InstructGPT task taxonomy for helpfulness * Also uses 14 model generation pool, wide capabilities * LLM as a judge with human verification of 200 prompts in eval set I would like to see (and am working on): * Better correlation beyond just LLM as a judge evaluators (we know GPT likes itself) * More human data Regardless, great to have options. I do think in future benchmarks for reward models, safety should be separate from capabilities. RewardBench being first was to see where we are at, but now RLHF training pipelines are changing noteably!

@natolambert Quite a lot of times I ever tried to look for seemingly nice reward models/synthetic data as well, the only GitHub page I found is the repo of e.g. vLLM

This is a pipeline we're seeing again and again for curating synthetic data for specific domains. You need: 1. Diversity, 2. Quality responses, and 3. Verification. AI-Assisted Generation of Difficult Math Questions Shah et al. When you do this stuff, plz release the data ;) -- "plan to release" often falls through.

@KempeLab @AIatMeta I have sent an follow-up email to your NYU email address, hope it reaches you well despite potential bothering~ :D

@KempeLab @AIatMeta Hi Julia, I work exactly on test-time compute+reasoning and am super interested in this position, can I write you a follow-up email or DM?

Looking for a PhD intern in my team at @AIatMeta in Paris starting Spring (12-24 weeks, 24 better). With Yann Ollivier, we tackle LLM reasoning/planning via RL training (System 1) with test-time optimization (System 2). Motivated students, please apply: metacareers.com/jobs/173828249…

@ssgrn Interested, and super related to what I’m working on! DM sent

@natolambert During inference specifically

@natolambert A shameless plug, but yeah we also found execution feedback quite important arxiv.org/abs/2408.13745

Meta with another solid looking RLHF paper: RLEF: Grounding Code LLMs in Execution Feedback with Reinforcement Learning This is how big labs improve math etc. Funny because I wrote about "RLCF" in April of 2023. We're slowly plodding along in open RLHF.

@mrdrozdov We have some more exploration using more up-to-date LLMs, with more methods. A shameless plug but we also got quite some interesting findings arxiv.org/abs/2408.13745

Shi et al. 2022 probably deserves more credit for introducing a version of this idea a couple years earlier. arxiv.org/abs/2204.11454

@gh_marjan Thanks for sharing! I have applied.

We are hiring interns for summer 2025 at FAIR. Get involved in cutting-edge projects related to LLM alignment, reasoning, and synthetic data generation for text/multimodal LLMs. Apply now! metacareers.com/jobs/119904986…