Jiachen Li
@JiachenLi11
Followers
484
Following
330
Media
14
Statuses
101
Explore trending content on Musk Viewer
#BUS_KnockKnockKnock_Korat
• 177812 Tweets
#UFC307
• 142680 Tweets
Watch Sant RampalJi YouTube
• 130762 Tweets
Tennessee
• 114600 Tweets
Dodgers
• 112274 Tweets
Miami
• 88066 Tweets
Pereira
• 83005 Tweets
ドジャース
• 68815 Tweets
Khalil
• 54463 Tweets
Arkansas
• 39206 Tweets
Aldo
• 33921 Tweets
Ifigenia Martínez
• 30631 Tweets
Pennington
• 23256 Tweets
京都大賞典
• 23058 Tweets
Poatan
• 14637 Tweets
#PressConMGSongKhla25xENGLOT
• 11943 Tweets
Cam Ward
• 10462 Tweets
Canes
• 10220 Tweets
Pinned Tweet
Thank you
@_akhaliq
for sharing our work! The 4-step generation from our T2V-Turbo outperforms Gen-2 and Pika on VBench🚀
Our project page and codes are provided here:
Project Page:
GitHub Repo:
T2V-Turbo
Breaking the Quality Bottleneck of Video Consistency Model with Mixed Reward Feedback
Diffusion-based text-to-video (T2V) models have achieved significant success but continue to be hampered by the slow sampling speed of their iterative sampling processes. To
2
38
158
2
11
42
How to become a toxic Reviewer💀 for
#NeurIPS
? 🤔
1. List "Lack of technical novelty" as the only weakness.
2. Give the paper a rating 4 and confidence 4.
3. During rebuttal, acknowledge the authors' response but never take them into account. Tell the author that after
20
20
344
Met many folks at ICML this year. We all agree that the key to the success of RLHF in training LLMs is the HF instead of the RL.
Overall, the RL in RLHF only acts as a gradient estimator to address the non-differentiability of the "sampling operation" from a categorical
# RLHF is just barely RL
Reinforcement Learning from Human Feedback (RLHF) is the third (and last) major stage of training an LLM, after pretraining and supervised finetuning (SFT). My rant on RLHF is that it is just barely RL, in a way that I think is not too widely
401
1K
8K
6
24
217
🎉 Exciting news! Our paper "Offline RL with Closed-Form Policy Improvement Operators" just got accepted for
#ICML2023
! 🚀 We solved the policy improvement in closed form by approximating Gaussian Mixture behavior policies!
Check out our paper here:
1
18
79
Finally got my
#ECCV2024
reviews and meta-reviews! I was eager to dive in to understand why my confident rebuttal didn't sway the reviewers. However, I was completely shocked 🤯
Reviewer R3 claimed our methods only apply to a latent diffusion model and can't generalize to a
10
2
72
Super excited that our paper gets accepted to
#neurips2024
@NeurIPSConf
!
Many thanks to my awesome collaborator
@weixi_feng
@tsujuifu
@XinyiWang98
@ Sugato and great supervision from
@WenhuChen
@WilliamWangNLP
If you are interested in our work, stay tuned! We are cooking a
T2V-Turbo
Breaking the Quality Bottleneck of Video Consistency Model with Mixed Reward Feedback
Diffusion-based text-to-video (T2V) models have achieved significant success but continue to be hampered by the slow sampling speed of their iterative sampling processes. To
2
38
158
2
4
40
We just cleaned up and released our training codes. If you want to train your own T2V-Turbo with different teacher and reward models, check out our codes! 👇
T2V-Turbo
Breaking the Quality Bottleneck of Video Consistency Model with Mixed Reward Feedback
Diffusion-based text-to-video (T2V) models have achieved significant success but continue to be hampered by the slow sampling speed of their iterative sampling processes. To
2
38
158
2
6
34
🚀🚀I'm delighted to share that our paper "Reward Guided Latent Consistency Distillation" has been accepted by
@TmlrOrg
#TMLR
with a 𝐅𝐞𝐚𝐭𝐮𝐫𝐞𝐝 𝐂𝐞𝐫𝐭𝐢𝐟𝐢𝐜𝐚𝐭𝐢𝐨𝐧! Kudos to my collaborator
@weixi_feng
and advisors
@WenhuChen
@WilliamWangNLP
We integrate feedback
1
5
35
🔥🚀 Excited to share our latest work 👉
🤖We trained a transformer to tackle robot manipulation with multi-modal prompts (🖼️ + 📝).
🏆Set a new SOTA on VIMA-BENCH by
@YunfanJiang
@DrJimFan
, outperforming baselines by 10% across all 4 eval protocols!🎯
1
5
31
🤔Already know online preference learning is definitely better than learning from an offline dataset, but still hesitate to do it because don't want to annotate new preference data at every step? We got you! 🔥
🚀🚀Introducing 𝓑PO: Supercharging Online Preference Learning by
1
13
28
🎉Exciting news! 🎉Our GauchoAI team
@ucsbNLP
@ucsantabarbara
has made it to the finals of the Alexa Prize SimBot Challenge
@AmazonScience
! So proud of our team's hard work and dedication!
#AlexaPrize
#ConversationalAI
2
10
24
I am excited to share that our 𝓑PO has been accepted by the EMNLP Main Conference. See you in Miama🌴
Congratulations again to my awesome co-authors
@WendaXu2
@WilliamWangNLP
@lileics
🤔Already know online preference learning is definitely better than learning from an offline dataset, but still hesitate to do it because don't want to annotate new preference data at every step? We got you! 🔥
🚀🚀Introducing 𝓑PO: Supercharging Online Preference Learning by
1
13
28
0
3
16
Happy to share that our MIDAS work has been accepted to ICML 2024, See you in Vienna🇦🇹! I'm currently working on the code release with the Amazon team. Please stay tuned!
Updated Arxiv:
🔥🚀 Excited to share our latest work 👉
🤖We trained a transformer to tackle robot manipulation with multi-modal prompts (🖼️ + 📝).
🏆Set a new SOTA on VIMA-BENCH by
@YunfanJiang
@DrJimFan
, outperforming baselines by 10% across all 4 eval protocols!🎯
1
5
31
1
4
11
Excited to attend
#ICML2024
at the beautiful Vienna 🇦🇹If you are interested in text-to-image and text-to-video generation, learning from human feedback and embodied AI, let’s chat!✌️
0
2
10
Thrilled to be part of this amazing video generation benchmark project led by
@weixi_feng
! We’ve crafted some incredibly meaningful prompts that challenge even Kling and
@LumaLabsAI
🤯. Please check out our work at 👇
🚨🚨🚨[New Preprint] Can text-to-video systems like Kling or Dream Machine generate “a pink chameleon turns green”? Can video generation models demonstrate compositionality in time? The answer is No (or not yet)!
🚀⭐️Our new TC-Bench (including new prompts📃, ground truth
1
11
55
0
3
9
Please check out the awesome MMWorld 🌍 led by
@XuehaiH
!!
If you are curious about your Multimodal LLM's expertise in different disciplines and its understanding of the world dynamics🤔, don't hesitate to test it on MMWorld 🌍
MMWorld
Towards Multi-discipline Multi-faceted World Model Evaluation in Videos
Multimodal Language Language Models (MLLMs) demonstrate the emerging abilities of "world models" -- interpreting and reasoning about complex real-world dynamics. To assess these abilities,
2
15
84
0
0
10
@Shalev_lif
DPO is another great example of RL not being the key 🤓 However, we should focus more on the online learning settings, where new human preference data will be gathered during the training, instead of focusing on a static batch of data.
Here is the link to my recent paper that
0
0
8
I'll be presenting our MIDAS work today at Hall C 4-9
#2904
from 11:30 a.m. to 1 p.m. CEST🚀🚀 Let's chat about all exciting opportunities in multimodal learning and embodied AI!
6
2
8
🚀 Excited for
#ICML2023
! 🎉 If you are interested in offline RL, don't miss our poster session at Exhibit Hall 1
#112
on Thu, July 27th, from 10:30 a.m. to 12 p.m. 🗓️ Can't wait to meet everyone in person and exchange perspectives on the future of RL! 🤝 See you there!
🎉 Exciting news! Our paper "Offline RL with Closed-Form Policy Improvement Operators" just got accepted for
#ICML2023
! 🚀 We solved the policy improvement in closed form by approximating Gaussian Mixture behavior policies!
Check out our paper here:
1
18
79
0
0
6
@WenhuChen
Here's my 2 cents :)
My belief: RL objective cannot provide enough supervision for learning good "representation" but can excel with a good representation learned induced by the other training signals.
Evidence: Distributional RL (distRL) developed by
@marcgbellemare
,
@wwdabney
0
0
6
@kimtaehyeon610
I completely understand you! My ICML 2023 paper actually got 7, 7, 6, 5 at NeurIPS 2022 submission (5 was the borderline accept). However, the AC still rejected our paper, and it took me tons of energy to resubmit the paper to the next ICLR and ICML. Even though it finally got
0
0
5
@ZaneGallery
@JeepersMedia
Haha, I tried the same prompt on my open-sourced T2V-Turbo🤣 Here is the result. For sure, Kling is so much better🥲
2
0
5
@YunfanJiang
@DrJimFan
This work is done during my internship at Amazon Alexa AI
@AmazonScience
. Many things to my excellent collaborators:
@qiaozikl
, Michael Johnston,
@GaoSchaffer
,
@XuehaiH
, Hangjie Shi,
@suhailashakiah
, Reza Ghanadan
@WilliamWangNLP
0
0
1
Update on this issue:
The AC allows us to first send them the anonymous website. After verified, the AC will share it to the reviewers : )
Just found out that we are not allowed to include anonymous link in any part of the response for
@NeurIPSConf
rebuttal after spending a whole day making a nice website to present videos generated by our models 😖
Does anyone know what I should do in this case?
2
0
1
0
0
3
This is really mind-blowing 🤯
Chinese new DiT Video AI Generation model 【KLING】
Open access！
Generate 120s Video with FPS30 1080P, Understand Physics Better, Model Complex Motion Accurately
prompt:
Traveling by train, viewing all sorts of landscapes through the window.
52
439
1K
1
0
3
I'm always wondering if we just replace the rebuttal processes of NeurIPS/ICLR/ICML with a data annotation process asking the Top ML/AI-minded "authors" to label data for a fully open-sourced LLM, will we reach AGI sooner than wasting time doing the rebuttal?
@JiachenLi11
It’s not just NeurIPS, recently after being satisfied with the author response and accepting the new changes , the reviewer stuck to the original decision . Makes me question the whole rebuttal process .
0
0
5
0
0
2
@WendaXu2
(who paves the way for me to do research in NLP🤣) and me contributed equally to this work. Many thanks to the great advice from
@WilliamWangNLP
@lileics
0
1
2
Thank you
@camenduru
for sharing our work!
🎬 T2V-Turbo: Breaking the Quality Bottleneck of Video Consistency Model with Mixed Reward Feedback 🚀 Jupyter Notebook +
@LightningAI
+
@runpod_io
+
@tost_ai
🥳
Thanks to
@JiachenLi11
❤
@weixi_feng
❤ Tsu-Jui Fu ❤
@XinyiWang98
❤ Sugato Basu ❤
@WenhuChen
❤
@WilliamWangNLP
0
20
120
0
0
2
As a side note, I've recently been working on generative AI based on human feedback. If you are interested in our video generation model (𝚃𝟸𝚅-𝚃𝚞𝚛𝚋𝚘) and online preference learning algorithm 𝓑PO, come chat with me🤗
0
0
2
@natolambert
@WendaXu2
@ucsbNLP
Thank you
@natolambert
for bringing D2PO to our attention! We will definitely cite it as a related work of online preference learning in our future revision.
To clarify, our work emphasize the importance of constraining the divergence between the learned LLM and the behavior
1
0
2
@YunfanJiang
@DrJimFan
(5/7) 🧐 As the action space is defined to contain the initial pose and target pose of the robot arm, we highlight the importance of modeling the dependency between the initial and target pose using the following example.
1
0
1
Just found out that we are not allowed to include anonymous link in any part of the response for
@NeurIPSConf
rebuttal after spending a whole day making a nice website to present videos generated by our models 😖
Does anyone know what I should do in this case?
2
0
1
@danish_nazir1
R5 criticized the novelty initially. We worked very hard to include additional qualitative and quantitative results to support our novelty, but R5 just ignored our efforts....
1
0
1
@YunfanJiang
@DrJimFan
🏗️We adopt a decoder-only (GPT) architecture and model each action dimension as an individual token to capture the dependencies between the initial and target pose of the robot arm.
1
0
1
@YunfanJiang
@DrJimFan
(3/7)👓 To capture fine-grained visual and textual details from the multimodal prompt, we design our multimodal prompt encoder by augmenting a pretrained language model with a residual connection from thinput visual tokens to the encoded embeddings.
1
0
1
@AlberFuen
@NeurIPSConf
@weixi_feng
@tsujuifu
@XinyiWang98
@WenhuChen
@WilliamWangNLP
You can check out our project page and GitHub repo : )
1
1
1
Kudos to my amazing collaborators: Edwin Zhang
@MingYin_0312
Qinxun Bai
@yuxiangw_cs
@WilliamWangNLP
0
0
1
@JeepersMedia
@ZaneGallery
Haha, thank you for your kind words😆
Our model can generate a video around 10s. If you want to play with our model, here are the links👇
0
0
1
@YunfanJiang
@DrJimFan
(7/7) 🔬We further hold out `Twist` and `Follow Order` from the training data of the original VIMA-BENCH. Then empirically evaluate the superior in-context learning capabilities of the policy produced via our framework.
1
0
1
@YunfanJiang
@DrJimFan
(6/7) 🏆 Tested on VIMA-BENCH by
@DrJimFan
, MIDAS sets a new SOTA, outperforming the baseline methods by a whopping 10% across all 4 evaluation protocols! 🎯
1
0
1
@YunfanJiang
@DrJimFan
(2/7)🌟 Understanding the 🖼️+📝 often requires a robot to infer state transitions from 📝, and actions from 🖼️. Thus, we introduced a 2-stage training pipeline: inverse dynamics pretrain & multi-task finetune. For pretrain, we convert any robot traj. into a motion-following task.
1
0
1