@feifeiM

Computer Science, Photography, Soccer, IT

 VA
cs.virginia.edu/~hw5x/
Joined September 2009
  • Tweets63
  • Following376
  • Followers190
5Media
Lifelong learning is hard, but exciting!
Chengxing Xie
@Chengxing_XieMar 31
Claude Code may be hiding far more than most people realize We analyzed a source snapshot and found evidence of 24+ gated feature systems, including: coordinator mode, forked subagents, agent swarms, Kairos, Auto Dream, and deep MCP/auth infrastructure... The real story is the 3-layer gating architecture behind it. Read: yitianlian.github.io/claude-… #ClaudeCode #AIEngineering #Agents #MCP #AIAgents

Claude Code's Secret Features

What Anthropic keeps for themselves — 24+ hidden features discovered through source code archaeology.

yitianlian.github.io
1
102
(1/2) Will humans be replaced by generative AI #Sora #chatgpt4? Our recent work provides some cues from a game theoretical perspective: arxiv.org/abs/2402.15467
1
1
1
8
2,127
(2a/2) Leave alone cold theoretical results, simulations suggest several interesting phenomena which are predicted by our theorems. For example, Q: who will switch to using generative AI? A: Those with high creation cost.
1
1
342
(2b/2) Q: Would the rise of genAI-powered creators discourage the productivity and utility of human creators? A: The total utility and content generation by human creators decrease, but the average utility and content output per human creator actually improve!
1
222
Really excited about our recent work on RLHF: from a probability matching perspective, we showed that DPO does mean-seeking, while our EXO does mode-seeking!
Haozhe Ji@HaozJi9 Feb 2024
Introducing EXO, a method towards efficient exact optimization of LM alignment. tldr: EXO is more efficient than RLHF by perfoming probability matching, and exact for arbitary policy parametrization compared to DPO. arxiv: arxiv.org/pdf/2402.00856.pdf repo: github.com/cdjhz/exact-optim…
4
357
Hongning Wang retweeted
Haozhe Ji@HaozJi9 Feb 2024
Introducing EXO, a method towards efficient exact optimization of LM alignment. tldr: EXO is more efficient than RLHF by perfoming probability matching, and exact for arbitary policy parametrization compared to DPO. arxiv: arxiv.org/pdf/2402.00856.pdf repo: github.com/cdjhz/exact-optim…
2
14
3
58
11,957
Welcome to our #ICML2023 oral presentation about a game theoretical analysis of the "invisible hands" in recommender systems! Very insightful and interesting results!
Fan Yao@FanYao5567154525 Jul 2023
#ICML2023 Excited to share our ICML oral paper arxiv.org/pdf/2302.01971.pdf . Our result reveals the economic insight of content creator competition in recommendation platforms, which echos the insight of "invisible hand" from Adam Smith. Welcome to join our oral and poster sessions!
10
1,214
Want to use generalized linear bandits in a federated learning setting? Come to our #NeurIPS2022 poster #939 at 11am CST and check out our paper: arxiv.org/abs/2202.01087v1 to find out more details!

Communication Efficient Federated Learning for Generalized Linear Bandits

Contextual bandit algorithms have been recently studied under the federated learning setting to satisfy the demand of keeping data decentralized and pushing the learning of bandit models to the...

arxiv.org
7
Finally arrived at #NeurIPS2022! Amazing location and crowd! Excited about the next three days!
14
Do we really need reinforcement learning for conversational recommendations? We are providing a simpler alternative using decision trees! Going to present it at #CIKM2022 tomorrow arxiv.org/pdf/2208.14614.pdf
2
1
17
An interesting follow-up: right before my talk, the fire alarm was set off and everyone was asked to evacuate immediately. 5 minutes later, we were told it was a false alarm, which in turn brought more people to my talk :)
3
Replying to @huazheng_wang
@wang_huazheng is presenting our private bandit work: arxiv.org/abs/2208.14555, at #recsys2022!

Dynamic Global Sensitivity for Differentially Private Contextual Bandits

Bandit algorithms have become a reference solution for interactive recommendation. However, as such algorithms directly interact with users for improved recommendations, serious privacy concerns...

arxiv.org
1
2
5
Great to see many rooms are fully packed at #KDD2022!
3
Going to talk about our effort in "Learning from a Learning User" at the #KDD2022 Online And Adaptive Recommender Systems workshop: oars-workshop.github.io/. The room is fully packed!
1
3
Heading to #KDD2022 in such a beautiful afternoon!
2
We have a wonderful lineup of 14 full-day and 20 half-day workshops at KDD2022: kdd.org/kdd2022/workshop.htm…, starting from this Sunday morning eastern time. Look forward to meeting everyone at the workshops and the conference!
3
Some contextual bandit instances are intrinsically hard to attack?! Come to our #ICML2022 talk at 4:40pm, Room 301-303. And here is our paper: arxiv.org/abs/2110.09008
Wants to learn from a learning agent to improve yourself? Check out our talk this afternoon at #ICML2022, Online learning and bandit session, at ballroom 3&4! And here is the paper: arxiv.org/abs/2202.01879

Learning from a Learning User for Optimal Recommendations

In real-world recommendation problems, especially those with a formidably large item space, users have to gradually learn to estimate the utility of any fresh recommendations from their experience...

arxiv.org
Driving to #ICML2022 in extreme weather. First in-person conference #NeurIPS2019. Look forward to meeting many new friends!
1
We are giving a tutorial on “Learning by Exploration: New Challenges in Real-World Environments” at KDD’2020, August 23, 11am-3pm EST! More details can be found here: rb.gy/keke2v and rb.gy/q4xo3m. #KDD2020
Load more