evanthebouncy
@evanthebouncy
Followers
1K
Following
614
Media
87
Statuses
584
asst prof @ NTU, ex principal scientist @ autodesk, phd mit 2019. I make programming more communicative 🧠↔️🤖
SG
Joined January 2013
want to get into program synthesis but don't know how to started? I wrote a minimalist intro to modern program synthesis that can help you -- from problem formulation to generating code by fine-tuning llm on huggingface.
2
17
127
So, people are beginning to realize to what programs actually are.
0
0
1
RT @GabrielPoesia: Thrilled to join the UMich faculty in 2026!. I'll also be recruiting PhD students this upcoming cycle. If you're interes….
0
27
0
cleaning the whiteboard in my office, I noticed in the upper left corner, in faded ink, "I'd rather be feared than be loved". some previous prof really wrote that down, and it refused to fade. I don't know what to make of it.
0
0
6
for fun I went through all my papers and tallied how many times I submitted each. note how it follows a geometric distribution, modeling the policy of "yolo submit until accepted". "but each submission's outcome are not independent coin tosses!!" . or is it?. plot yours.
0
0
6
Industry only engages with academia in areas that are:.1) poorly understood 2) tantalizingly profitable 3) threaten extinction. Engage industry through mystique, lust, and fears.
From about 2013-2022, the highest impact thing you could do for AI in the tech industry was publish in academic venues. You didn't have to choose between climbing the ladder and doing open science. Now that world is gone, and I'm still not sure how to navigate this new world.
0
0
2
I have yet to see RL succeeds with negative rewards in high dimensions of observation and action space. Do we know of a method that works well while using ONLY negative rewards? . OTOH, humans learn very well from negative rewards, we just call them 'lessons'. what gives?.
1
0
2
Just look at these multi-modal refinement instructions! How would we ground them into reasonable executions?? joint work with. @wp_mccarthy.@saujasv.@judyefan.@dan_fried.@KarlDD.@JustinMatejka
0
2
5
What would it take to build agents that can similarly follow refinement instructions?. We hope that mrCAD can help, by giving rollouts of successful human-human communications. [8/n]
1
1
3
Evaluating human vs VLMs in the role of the Maker reveals that:. - human consistently makes improvement on the current design towards the target.- VLMs only makes improvement in generation (round 1) but make things worse in refinement (round 2+). [7/n]
1
0
0
Analyzing the instructions from successful rollouts reveals that:. - people used more drawings in generation (round 1) and more texts in refinement (round 2+).- the texts become more "verb like" in refinements.- the drawings become more partial in refinements. [6/n]
1
1
2
We collected successful rollouts of humans playing this game from 1,092 pairs of participants. The collected mrCAD dataset has 3 subsets:. - coverage : 2249 CADs with 1-2 rollouts.- dense : 698 CADs with 3+ rollouts.- very-dense : 27 CADs with 30+ rollouts.[5/n]
1
0
0
We test this hypothesis with a communication game, where 2 players collaborate to recreate target CAD designs. A target design is shown to the Designer, who must communicate how to recreate it to the Maker over several rounds, communicating using drawing + text. [4/n]
1
0
0
generation and refinements are different processes. - generation is nothing to something: {} → x.- refinement is something to something else: x → x’. consequently, we hypothesize that they are communicated and executed differently.[3/n].
1
0
0
Imagine using AI to generate an image or to write code. If the first output doesn’t quite work, instructing the AI to refine it further (tweaking img, debugging code) often leads to nowhere. The mrCAD dataset makes this intuition measurable. [2/n].
1
0
0
new multi-turn instruction grounding dataset with @wp_mccarthy and @saujasv . - multi-modal instruction : drawing + txt.- verifiable execution : 2D CAD gym env.- easy eval : API → score.- baselines : human vs VLMs.- large : 15,163 inst-exe rounds. [1/n]
1
10
28
I've recently started my job as an asst professor at NTU, Singapore. If you are ever in town come say hi :)
28
11
695
I do feel most projected growth trend (in AI, for instance) are logistic functions disguised as exponentials.
0
0
5
$200 to fly from SFO to JFK. $100 to uber from JFK to Manhattan . Something is deeply wrong here.
2
0
9
One of the best presentations of Chinese food categories.
0
0
4
Winner winner chicken dinner.
ARC Prize 2024 Paper Award Winners! 🏆. 1st "Combining Induction and Transduction For Abstract Reasoning", @xu3kev, @HuLillian39250, Carter Larsen, Yuqing Wu, @simon_alford0, Caleb Woo, Spencer M. Dunn, @haotang_ai, Michelangelo Naim, Dat Nguyen, @WeiLongZheng1, @ZennaTavares,.
0
0
11