Henry Hung Le

35 posts

Henry Hung Le banner
Henry Hung Le

Henry Hung Le

@LHung1610

Member of Technical Staff - stealth

San Francisco Bay Area Katılım Nisan 2013
123 Takip Edilen155 Takipçiler
Henry Hung Le
Henry Hung Le@LHung1610·
OpenReview should have a new feature to find and track author-AC confidential comments more easily...At this point, I have >230 all replies in my batch. With LLM -augmented reviews, authors seem to flag concerns more often too. The only way I can track the confidential comments now is via emails 😂
English
0
0
0
52
Henry Hung Le
Henry Hung Le@LHung1610·
Have a record number of 16 papers in my ICML AC batch. 10/16 already had all 4/4 reviews and the rest with 3/4 reviews (with only 1 review I considered not so high quality). I am impressed by the reviewer pool this year (that or AI-augmented review getting too good)!
English
0
0
4
391
Alex Dimakis
Alex Dimakis@AlexGDimakis·
Very excited that the Terminal Bench paper is out. The TB and OpenThoughts community has been building the leading benchmark for coding agents. Also we are starting to work on Terminal Bench 3- reach out if you want to help.
Mike A. Merrill@Mike_A_Merrill

The Terminal-Bench paper is here! Read it to learn where frontier models still fail and the secrets of how we sourced hundreds of high quality environments from our open source community. 🧵

English
2
6
68
11.4K
Henry Hung Le retweetledi
Salesforce AI Research
Salesforce AI Research@SFResearch·
Problem: LLMs excel at code generation, but outputs often contain security blindspots. Fine-tuning alone can't keep pace with sophisticated attacks. Solution: Enter INDICT - our new framework that empowers LLMs with Internal Dialogues of Critiques, boosting code safety by >80% in tests. Discover this new paradigm for #AI #CodeSafety: 👩‍💻 Code: tinyurl.com/yphubk85 🔖 Blog: tinyurl.com/bderyhv4 📚 Paper: arxiv.org/abs/2407.02518
GIF
English
0
4
18
1.1K
Henry Hung Le retweetledi
Caiming Xiong
Caiming Xiong@CaimingXiong·
Generating code with LLMs poses risks like security vulnerabilities, logical errors, and context misinterpretations. Critical for developers to scrutinize and validate AI-generated code to ensure safety and correctness. We introduce #INDICT, a novel multi-agent cooperative framework that enhances #LLMs for secure & helpful code generation. Utilizing dual critics for safety and helpfulness, INDICT leverages external tools for grounded feedback, significantly improving code security across diverse programming languages. #AI #CyberSecurity #CodeGeneration For more details, read the full blog: blog.salesforceairesearch.com/indict-code-ge… paper: arxiv.org/abs/2407.02518… code: github.com/SalesforceAIRe…
Caiming Xiong tweet media
English
2
19
63
12.4K
Henry Hung Le
Henry Hung Le@LHung1610·
👉Generated sub-modules are then extracted from potentially correct solutions and grouped into different semantic clusters. The cluster centroids are selected as representative sub-modules. The model is then instructed to 🔃reuse/adapt these modules into its revised solutions.
Henry Hung Le tweet media
English
1
0
1
133
Henry Hung Le
Henry Hung Le@LHung1610·
1⃣First, the model is required to outline sub-modules needed, each of which consists of a function header and docstring describing the intended use. 2⃣Subsequently, the model implements each module fully in code and integrates them as parts of the complete final solution.
Henry Hung Le tweet media
English
1
0
1
99
Henry Hung Le
Henry Hung Le@LHung1610·
Language models are well known for their strong performance in NLP. What about competitive programming problems e.g. Codeforces? Check out our work "CodeChain: Towards Modular Code Generation Through Chain of Self-revisions with Representative Sub-modules" accepted to #ICLR2024!
Salesforce AI Research@SFResearch

Check out our #ICLR2024 Accepted Papers. Congratulations to all of our authors!

English
2
1
8
1.1K
Henry Hung Le
Henry Hung Le@LHung1610·
👉We proposed a simple yet powerful generation method for LLMs, incorporating both ✅ Chain of Thought (CoT) and ✅ Self-revision. For example, using CoT prompting: 👇
English
1
0
1
89
Henry Hung Le
Henry Hung Le@LHung1610·
Check out CodeT5+, our latest code LLMs with SOTA results on many code tasks 🎉🎉 Paper: arxiv.org/abs/2305.07922 Code: github.com/salesforce/Cod… Blog: blog.salesforceairesearch.com/codet5-open-co… Thanks to my team at Salesforce Research @ayueei @AkhileshGotmare @QuocNghi91 @LiJunnan0409 @stevenhoi
Steven Hoi@stevenhoi

Introducing 🔥CodeT5+🔥, a new family of open-source code LLMs for both code understanding and generation, achieved new SoTA code generation performance on HumanEval, surpassing all the open-source code LLMs. Paper: arxiv.org/pdf/2305.07922… Code: github.com/salesforce/Cod… (1/n)

English
0
5
18
2K
Michaël Trazzi
Michaël Trazzi@MichaelTrazzi·
Aran Komatsuzaki giving walkthroughs of the codeRL paper before the author arrives. After 10 minutes of SBFing his way into answering poster questions he revealed he was not the author and everyone lost their mind (Poster 138 #NeurIPS2022)
English
9
33
568
0
Henry Hung Le
Henry Hung Le@LHung1610·
Check out this nice blog post blog.salesforceairesearch.com/coderl/ about our code generation research work CodeRL @SFResearch and more details on our paper arxiv.org/abs/2207.01780 and code github.com/salesforce/Cod…
Salesforce AI Research@SFResearch

CodeRL advances program synthesis by integrating pretrained language models + deep reinforcement learning. Using unit test feedback in model training and inference + an improved CodeT5 model, it achieves SOTA results on competition-level programming tasks. blog.salesforceairesearch.com/coderl

English
0
0
5
0
Henry Hung Le
Henry Hung Le@LHung1610·
Check out our new research work for code generation using pretrained LMs! Very happy to achieve this work with my amazing fellow researchers at Salesforce @ayueei @AkhileshGotmare @silviocinguetta @stevenhoi
Steven Hoi@stevenhoi

Excited to introduce CodeRL: a novel code generation framework as a whole new way of building SOTA AI Coding systems by combining Pretrained Models and Deep Reinforcement Learning (RL). Code & models are open-source: Code: github.com/salesforce/Cod… Paper: arxiv.org/abs/2207.01780

English
1
0
5
0
Henry Hung Le
Henry Hung Le@LHung1610·
We will publish the codes and models related to these papers soon!
English
0
0
1
0
Henry Hung Le
Henry Hung Le@LHung1610·
2. VGNMN: Video-grounded Neural Module Networks: an interpretable approach that decomposes video-grounded dialogue utterances into modular steps as a reasoning process Many thanks to my co-authors and advisors @stevenhoi and @nancyfchen1
English
1
0
2
0
Henry Hung Le
Henry Hung Le@LHung1610·
Very excited to have 2 papers accepted to NAACL! 1. Multimodal Dialogue State Tracking: a new machine learning task that tracks the information states of visual objects mentioned in the dialogue context
Salesforce AI Research@SFResearch

Check out our #NAACL2022 accepted papers! Congrats to the authors! We hope everyone enjoys the conference! @EhsanHAsl @owenhaoliu @CaimingXiong @murakhovska @jasonwu0731 @alexfabbri4 @mrnt0810 @jesse_vig @iam_wkr @semih__yavuz @yingbozhou_ai @LHung1610 @stevenhoi @PhilippeLaban

English
1
0
10
0

Keşfet

@AlexGDimakis @abeirami @zhouyingbo @caimingxiong @silviocinguetta @doyensahoo @SFResearch @HailinChen3