AI Agents Beyond ChatGPT

Zhou(Jo) Yu Columbia University & Arklex AI

Who supports AI Agents?

Bill Gates

Agents are bringing about the biggest revolution in computing since we went from typing commands to tapping on icons.

Andrew Ng

I think AI agentic workflows will drive massive AI progress this year.

Sam Altman

2025 is when agents will work.

What are AI Agents?

Adapted from Russell& Norvig (2020)

Perception: Multimodal inputs including, text, image, audio, video, touch, etc.

Planning (Inner Monologue): Chain-of-Thought reasoning over tokens that powered by LLMs

Reflection: meta-reasoning in every stop Actions: function/tool calling, embodied actions.

PHASE 1

Level1 Wanna Chat"

PHASE 2

Level 2 VorkAssistant"

Level 3 "Agent-as-a-Service"

Level 5 “Human,holdmybeer”

Level 4 "Autonomous Agents”

#

A completion to the human prompt with its internalknowledge (i.e. no tool use).

An LLM-centric software system for assisting realworld tasks.

A service-centric system with LLMs as core components for automating a predefined workflows.

Slide: Alex Wang @Scale AI

Overview

Model self-improvement with LLMs (Yu et al, NAACL 2024, Outstanding paper)
Eliciting stronger model ability via tree search (Yu et al, EMNLP 2023 )
AI agent self-improvement via tree search (Yu et al, ICLR 2025)

Background: In-Context Self-Improvement

Input:

Background: In-Context Self-Improvement

Input:

Q: Calculate (4 * 1) - (2 * 3) = ?

Self-Improvement Prompting (Madaan, et al, 2023)

Step 1 : (4 * 1) - (2 * 3) = 4 - 6

Ans: -3

Step 2: 4 - 6 = -3

Background: In-Context Self-Improvement

Input:

Q: Calculate (4 * 1) - (2 * 3) = ?

Self-Improvement Prompting (Madaan, et al, 2023)

Step 1 : (4 * 1) - (2 * 3) = 4 - 6
Step 2: 4 - 6 = -3
Ans: -3

prompt feedback

In step 2 the part “4-6=-3” is incorrect. This is because

prompt update

Step 1: (4 * 1) - (2 * 3) = 4 - 6
Step 2: 4 - 6 = -2
Ans: -2

Background: In-Context Self-Improvement

Input:

Q: Calculate (4 * 1) - (2 * 3) = ?

Self-Improvement Prompting (Madaan, et al, 2023)

Step 1 : (4 * 1) - (2 * 3) = 4 - 6
Step 2: 4 - 6 = -3
Ans: -3

prompt feedback

In step 2 the part “4-6=-3” is incorrect. This is because

prompt update

prompt feedback

Step 1: (4 * 1) - (2 * 3) = 4 - 6
Step 2: 4 - 6 = -2
Ans: -2

Background: In-Context Self-Improvement

Problem 1: small LM cannot self-improve via prompting!

In step 1 the part “2*3=6” is incorrect. This is because …

… error propagates!

Background: In-Context Self-Improvement

Problem 2: small LM cannot learn “self-improvement” from LLM demonstrations!

Background: In-Context Self-Improvement

Multistep Arithmetic

Problem 2: small LM cannot learn “self-improvement” from LLM demonstrations!

Q: Calculate 4 - 0 * -1 * 8 + 6 = ?

feedback: …

Motivation

Prior work shows that self-improvement (S.I.) is useful for task performance/generalization (Madaan, et al, 2023)

We find prompt-based S.I./simple distillation methods fails with small LM

Madaan, A. et al. (2023) ‘Self-Refine: Iterative Refinement with Self-Feedback’

Motivation

Prior work shows that self-improvement (S.I.) is useful for task performance/generalization (Madaan, et al, 2023)

We find prompt-based S.I./simple distillation methods fails with small LM

Treat “self-improvement” as a task to learn

(attempt) -> (feedback, update)

Madaan, A. et al. (2023) ‘Self-Refine: Iterative Refinement with Self-Feedback’

Motivation

Prior work shows that self-improvement (S.I.) is useful for task performance/generalization (Madaan, et al, 2023)

We find prompt-based S.I./simple distillation methods fails with small LM

Treat “self-improvement” as a task to learn
But learn “self-improvement” online

consider LLMs/python scripts as teacher model edit models to modify small LM’s attempts

replay this interaction experience to train the small LM

Madaan, A. et al. (2023) ‘Self-Refine: Iterative Refinement with Self-Feedback’

Motivation

Prior work shows that self-improvement (S.I.) is useful for task performance/generalization (Madaan, et al, 2023) We find prompt-based S.I./simple distillation methods fails with small LM

Treat “self-improvement” as a task to learn
But learn “self-improvement” online

consider LLMs/python scripts as teacher model edit models to modify small LM’s attempts replay this interaction experience to train the small LM

Madaan, A. et al. (2023) ‘Self-Refine: Iterative Refinement with Self-Feedback’

TriPosT

1 Interactive trajectory editing

uses LLM/python scripts as edit models

gather interaction records between small LM and LLM

training sample:

Q: Calculate (4 * 1) - (2 * 3) = ?

prompt feedback

Step 1 : (4 * 1) - (2 * 3) = 4 - 6
Step 2: 4 - 6 = -3
Ans: -3

In step 1 the part “2*3=6” is incorrect. This is because …

prompt update

Step 1: …

TriPosT

1 Interactive trajectory editing

uses LLM/python scripts as edit models

gather interaction records between small LM and LLM

training sample:

Q: Calculate (4 * 1) - (2 * 3) = ?

prompt feedback

Step 1 : (4 * 1) - (2 * 3) = 4 - 6
Step 2: 4 - 6 = -3
Ans: -3

In step 1 the part “2*3=6” is incorrect. This is because

In step 2 the part “4-6=-3” is incorrect. This is because …

Step 1: …

TriPosT

1 Interactive trajectory editing

uses LLM/python scripts as edit models

gather interaction records between small LM and LLM

training sample:

Q: Calculate (4 * 1) - (2 * 3) = ?

prompt feedback

Step 1 : (4 * 1) - (2 * 3) = 4 - 6
Step 2: 4 - 6 = -3
Ans: -3

In step 1 the part “2*3=6” is incorrect. This is because

Step 1: …

In step 2 the part “4-6=-3” is incorrect. This is because …

Step 1 : (4 * 1) - (2 * 3) = 4 - 6 Step 2:

TriPosT

1 Interactive trajectory editing

uses LLM/python scripts as edit models

gather interaction records between small LM and LLM

training sample:

Q: Calculate (4 * 1) - (2 * 3) = ?

prompt feedback

Step 1 : (4 * 1) - (2 * 3) = 4 - 6
Step 2: 4 - 6 = -3
Ans: -3

In step 1 the part “2*3=6” is incorrect. This is because …

Step 1: …

In step 2 the part “4-6=-3” is incorrect. This is because …

Step 1 : (4 * 1) - (2 * 3) = 4 - 6
Step 2: 包 4 - 6 = -2
Ans: -2

TriPosT

1 Interactive trajectory editing

2 Data post-processing

reformats interaction data into (attempt, feedback, update) triplet

data filtering and re-balancing

Q: …

attempt 1: …

attempt: …

feedback: …

attempt 2: …

update: …

feedback: …

filter

Q: …

attempt N: …

attempt: …

feedback N: …

feedback: …

TriPosT

1 Interactive trajectory editing

2 Data post-processing

reformats interaction data into (attempt, feedback, update) triplet

data filtering and re-balancing

Q: …

attempt 1: …

attempt: …

feedback: …

attempt 2: …

feedback: …

update: …

filter

Q: …

re-balance

dset!

attempt N: …

attempt: …

feedback N: …

feedback: …

TriPosT

1 Interactive trajectory editing

2 Data post-processing

3 Model training

weighted SFT with more emphasis on feedback and update tokens

“on-policy” data

LLaMA-1 / LLaMA-2

Model self-improvement with LLMs

Main Idea:

Prior work shows that LLMs can be prompted to self-improve

Explicit craft “self-improvement” data with LLMs to train/enhance this ability

Let a weak LLM attempt self-improveme

Q: Calculate (4 * 1) - (2 * 3) = ?

Use a stronger LLM to perform “process supervision”

8 prompt feedback

8 prompt update

Attempt solution…

Feedback: …

Revised solution…

Edited feedback: …

Edited revised solution: …

Model self-improvement with LLMs

Main Idea:

Prior work shows that LLMs can be prompted to self-improve

Explicit craft “self-improvement” data with LLMs to train/enhance this ability

1 Let a weak LLM attempt self-improvement2 Use a stronger LLM to perform “process supervision”

3 Train the LM with improved data

Improved “on-policy” data

LLaMA-1 / LLaMA-2

Model self-improvement with LLMs

Evaluation: Big Bench Hard

tasks where small LM struggles

split tasks into easy (seen) and harder (unseen) subtasks to measure generalization

Dataset	Criterion	Example	seen subtask	unseen subtask
Multistep Arithmetic	nesting depth (d)and number of operands (l)	Q:(22+1)+(31-1)) 1=3,d=2	1={3,4}×d={2}	l={3,4}×d={3}and 1={56}×d={2,3}
Word Sorting	number of words to sort (l)	Q:orange apple banana pear //1=4	1={2,3,.,.7}	1={8,9,..,16}
Date Understanding	number of steps to solve (l)	Q:Today is 01/02,what's the date yesterday?//l=1	1={1,2}	1≥3
Logical Deduction	number of options (l)	Q: John runs ... Who runs fastest?？ Options: (A).(B).(C).//l=3	l={3,5}	1={7}

Can TriPosT improve overall performance?

Evaluation: Big Bench Hard

Method		Multistep Arithmetict			Word Sorting†			Date Understanding			Logical Deduction
Method		seen	unseen	total	seen	unseen	total	seen	unseen	total	seen	unseen	total
sino	LMSI	10.83	0.00	4.33	67.72	5.56	26.83	14.55	9.09	12.99	61.11	20.00	48.10
	ft rationale	39.75	1.48	16.78	73.49	5.82	28.50	33.35	21.21	29.87	62.69	8.67	45.78
	ft SI. demo	29.17	0.00	11.67	53.54	1.98	19.26	27.27	18.18	24.68	54.63	15.00	41.67
	TRIPOST(t = 1)	41.67	0.84	17.17	74.02	5.16	28.23	32.73	13.64	27.27	57.88	22.00	46.52
	TRIPOST(t =2)	49.58	1.39	20.67	74.02	7.14	29.55	35.46	25.00	32.47	58.80	18.00	45.25
	TRIPOST(t =3)	52.50	2.50	22.50	77.17	5.95	29.82	40.00	29.55	37.01	63.89	15.00	48.42

Can TriPosT trained models self-improve?

Evaluation: Big Bench Hard

Xiao Yu, et al. 2024. Teaching Language Models to Self-Improve through Interactive Demonstrations. NAACL 2024 Outstanding Paper.

Interactive (“on-policy”) data is crucial

Ablation Studies:

Model self-improvement with LLMs

Takeaway: improving model performance without human supervision is possible

Limitations: need a strong editor LLM for supervision

Takeaway: improving model performance without human supervision is possible

Figure 3:Improvement demonstrations become more difficult to collect as TRIPosT iteration increases.

prompt edited feedback:

All steps are correct. The final answer is also correct.

Overview

Model self-improvement with LLMs

Eliciting stronger model ability via tree search(Yu et al, EMNLP 2023 )

AI agent self-improvement via tree search

LLM Model performance improves with training compute

Model performance improves with test-time compute (e.g. GPT4-o1)

Jones, Andy L. "Scaling scaling laws with board games." arXiv preprint arXiv:2104.03113 (2021). OpenAI. "Learning to Reason with LLMs" https://openai.com/index/learning-to-reason-with-llms/ (2024)

Performance Improvement via Scaling

Central to these are scaling laws is to improve, without human supervision:

Elicit stronger model behavior beyond CoT Improve model performance with stronger data

Figure1 Language modeling performance improves smoothly a size,and amount of compute² used for training.For optimal perf upintandem.Empirical performancehasapower-lawrelationsh bottlenecked by the other two.

Fig.9.The trade-off between train-time compute and test-time compute.Ea dotted line gives theminimum train-test computerequired foracertain I ona9×9board

o1performance smoothlyimproves with both train-time and testtimecompute

Enhancing Model Capability via Tree Search

Main Idea:

Many dialogue tasks are essentially about decision making

We can use look-ahead search from games like chess to enhance this

Enhanced Model Capability via Search

Enhancing Model Capability via Tree Search

Main Idea:

Many dialogue tasks are essentially about decision making We can use look-ahead search from games like chess, to enhance this

Self-Improvement with LLM

Enhanced Model Capability via Search

Self-Improvement vis Search

Look ahead via tree search

chess: white to move

Look ahead via tree search

chess: white to move

Look ahead via tree search

chess: white to move

Look ahead via tree search

chess: white to move

: simply winning
Hikaru Nakamura, Grand Master

Enhancing Model Capability via Tree Search

chess: white to move

simulate

Dialog decision making as tree search

Main Idea:

Many dialogue tasks are essentially about decision making We can use look-ahead search from games like chess to enhance this

MCTS with Zero-training

search (potentially) promising actions

simulate action outcomes

evaluate action quality

update its estimate of each actions quality

= prompt an LLM to act as π

= prompt an LLM to act as M

= prompt an LLM to act as V

Dialog decision making as tree search

MCTS with Zero-training

search (potentially) promising actions

simulate action outcomes

evaluate action quality

update its estimate of each actions quality

= prompt an LLM to act as π

= prompt an LLM to act as M

= prompt an LLM to act as V

LLM as user simulator

Are you familiar with Save the Children?

It's a charity. Do you want to donate to it?

I am not sure ...

Dialog decision making as tree search

MCTS with Zero-training

search (potentially) promising actions

simulate action outcomes

evaluate action quality

update its estimate of each actions quality

= prompt an LLM to act as π

= prompt an LLM to act as M

= prompt an LLM to act as V

LLM as value function

Are you familiar with Save the Children?

It's a charity. Do you want to donate to it?

Would you like to make a donation?

Prompt: [negative reaction]

MCTS with Zero-training

Open-Loop MCTS for dialogue

considers stochastic transitions from a dialogue state

(traditional) Closed-Loop MCTS

MCTS with Zero-training

Open-Loop MCTS for dialogue

considers stochastic transitions from a dialogue state

Open-Loop MCTS

Enhancing Model Capability via Tree Search

Evaluation: Persuasion Task

PersuasionForGood Dataset: persuade a person to donate to a charity called Save the Children

“what is a good policy?” is subjective -> very hard to train

Enhancing Model Capability via Tree Search

Evaluation: Persuasion Task

PersuasionForGood Dataset: persuade a person to donate to a charity called Save the Children

“what is a good policy?” is subjective -> very hard to train

- Can GDP-Zero produce a more persuasive policy than the base LLM itself?

GDP-Zero improves dialog task success

Evaluation: Persuasion Task

PersuasionForGood Dataset: persuade a person to donate to a charity called Save the Children

“what is a good policy?” is subjective -> very hard to train

- Can GDP-Zero produce a more persuasive policy than the base LLM itself?

Xiao Yu, Maximillian Chen, and Zhou Yu. 2023. Prompt-Based Monte-Carlo Tree Search for Goal-oriented Dialogue Policy Planning. EMNLP 2023.

(Interactive Evaluation)

GDP-Zero improves dialog task success

Evaluation: Persuasion Task

PersuasionForGood Dataset: persuade a person to donate to a charity called Save the Children

“what is a good policy?” is subjective -> very hard to train

- Can GDP-Zero produce a more persuasive policy than the base LLM itself?

The chatbot...	RAP	ChatGPT	GDP-Zero
donation prob.个	0.72±0.38	0.73±0.38	0.79±0.37
inc.donation intent 个	4.08±0.68	3.77±0.90	4.30±0.71**
strong argument 个	3.89±0.97	3.91±0.99	4.28±0.74*
is convincing↑	4.11±0.74	4.10±0.70	4.38±0.66**
diverse strategy ↑	3.98±0.80	3.83±1.03	3.95±0.82
is manipulative↓	2.64±1.36	2.96±1.38	2.29±1.33**
is natural 个	4.25±0.68	4.03±0.65	4.38±0.62**
is relevant 个	4.64±0.54*	4.31±0.86	4.59±0.49
is coherent ↑	4.28±0.65	4.06±0.89	4.42±0.49*

GPD-Zero learns domain knowledge

How did GDP-Zero planning help?

avoids eager “proposition of donation” 22

balanced strategy with emotion and logical appeal

Enhancing Model Capability via Tree Search

Takeaway: tree search as an effective method to directly improve model behavior at test-time

Enhancing Model Capability via Tree Search

Takeaway: tree search as an effective method to directly improve model behavior at test-time

Limitations:

Extension beyond dialogue tasks such as AI agents?

Transfer this improved behavior back to the model via training?

Overview

Model self-improvement with LLMs

Eliciting stronger model ability via tree search

AI agent self-improvement via tree search(Yu et al, ICLR 2025)

2023+

Background: VLM on Computer Tasks

VQA Tasks

Q: What is he doing?

He is performing a skateboard trick…

Computer Tasks

Can you help me clear my shopping cart?

click button [shopping cart] ….

Challenge: extremely difficult as interacting with computer was not part of VLM (pre-)training

Scale test-time compute to improve agent performance

Transfer search knowledge back to VLM via training

Introduce R-MCTS

R-MCTS = explore decision space and self-improve on-the-fly

Scaling test-time compute

Transferring search knowledge

Introduce R-MCTS

R-MCTS = explore decision space and self-improve on-the-fly

INFERENCE

Scaling test-time compute

Transferring search knowledge

Introduce R-MCTS

R-MCTS = MCTS with contrastive self-reflection

REFLECT

Scaling test-time compute

Transferring search knowledge

Introduce R-MCTS

R-MCTS = MCTS with contrastive self-reflection

REFLECT

IMPROVE

Scaling test-time compute

Transferring search knowledge

Introduce R-MCTS

R-MCTS = MCTS with contrastive self-reflection and a multi-agent-debate value function 2

Scaling test-time compute

Transferring search knowledge

Introduce R-MCTS

Within each task, R-MCTS performs a tree search to find the best trajectory

Introduce R-MCTS

Within each task, R-MCTS performs a tree search to find the best trajectory

After each task, R-MCTS performs contrastive self-reflection to improve it future execution

Scaling test-time compute

Transferring search knowledge

R-MCTS Results

Benchmark: VisualWebArena and OSWorld

Realistic and reproducible

Tasks spans multiple domains

VisualWebArena

OSWorld

R-MCTS Results

R-MCTS outperforms other search algorithms (ToT, A*, or MCTS)

Transferring search knowledge

R-MCTS Results

R-MCTS achieves new SOTA on VisualWebArena, and is highly competitive on OSWorld!

Model	Inputs	Success Rate(%)
ExACT(GPT-4o+R-MCTSMAD)	SoM+Caption+Image	33.74
ExACT(GPT-4o+R-MCTSSA)	SoM+Caption+Image	32.53
GPT-4o+MCTS	SoM+Caption+Image	30.22
GPT-40+Search	SoM+Caption+Image	26.40
GPT-40+ICAL	SoM+Caption+Image	23.40
GPT-40	SoM+Caption+Image	19.78
Llama-3-70B+Search	AxTree+Caption	16.70
GPT-4V	SoM+Caption+Image	16.37
GPT-4+BLIP-2-T5XL	AxTree+Caption+Image	15.05
GPT-4	AxTree+Caption	12.75
Gemini-Pro-1.5	SoM+Caption+Image	11.98
Llama-3-7OB-Instruct+BLIP-2-T5XL	AxTree+Caption	9.78
GPT-4	AXTree	7.25
Gemini-Flash-1.5	SoM+Caption+Image	6.59
Gemini-Pro	SoM+Caption+Image	6.04
Gemini-Pro	SoM+Caption+Image	5.71
Gemini-Pro+BLIP-2-T5XL	AxTree+Caption	3.85
GPT-3.5+BLIP-2-T5XL	AxTree+Caption	2.97
GPT-3.5+LLaVa-7B	AxTree+Caption	2.75
GPT-3.5	AXTree	2.20
Gemini-Pro	AXTree	2.20
Mixtral-8x7b+BLIP-2-T5XL	AxTree+Caption	1.87
Mixtral-8x7b	AXTree	1.76
Llama-2-70B	AXTree	1.10
IDEFICS-80B-Instruct	SoM+Caption+Image	0.99
IDEFICS-80B-Instruct	AXTree+Caption+Image	0.77
Llama-2-70B+BLIP-2-T5XL	AxTree+Caption	0.66
CogVLM	SoM+Caption+Image	0.33
CogVLM	AXTree+Caption+Image	0.33

VisualWebArena Leaderboard

Screenshot	A11ytree	Screenshot + A11y tree	Set-of-Mark
Notice: t=temperauretop-p=o-ptof,en=mcoee
Rank	Model		Details	Score
1 Ot 12,2024	leam-by-interactw/Claude-3.5set Anonymous authors UnderReview.'24		t=0	22.5
2 Dec 20.2024	R-MCTS w/GPT-40o (max 60tree nodes) Columbia University&Microsoft Research Yu et al.,24			16.6
3 Mr 20,2024	GPT-4 OpenAl OpenAl,'23		[=1.0,top-p=0.9 len =128k	12.24
4 My20. 2024	GPT-40 OpenAl OpenAl,'24		t=1.0,top-p=0.9 len=32k	11.36
5 Aprl 23.2024	GPT-4 isio 0409) OpenAl OpenAl'23		t=1.0,top-p=0.9 len=32k	10.82
6 OCt 2.2024	leam-byntectw.Ge Anonymousauthors UnderReview.'24		=0	10.3
7 HMay 22, 2024	Qwen-Max Qwen Qwen Team.'24		t=1.0，top-p=0.9 len =32k
8 Au 17,.2024	GPT-40-mni OpenAl OpenAl'24	t=1.0,top-p=0.9 len =128k		6.21
9 MBy3.2024	Gemini-r-1.5 Google Gemini Team,Google,24		t=1.0,top-p=0.9 len=128k
10 Sep 10. 2024	Qwen-vi-Max-0809 Qwen Qwen Team,24		t=1.0,top-=0.9 len=32k

OSWorld Leaderboard

Scaling test-time compute

Transferring search knowledge

Scale test-time compute to improve agent performance

Transfer search knowledge back to VLM via training

Introduce Exploratory Learning

Exploratory Learning = explore, evaluate, and backtrack by training on tree traversals!

Scaling test-time compute

Transferring search knowledge

Exploratory Learning Results

GPT-4o after exploratory learning on R-MCTS trees

exhibits compute scaling properties without augmenting with search algorithm!

Exploratory Learning Results

GPT-4o after exploratory learning on R-MCTS trees

exhibits compute scaling properties without augmenting with search algorithm!

Scaling test-time compute

Transferring search knowledge

Summary

❖ Inference/training methods to improve agent performance

R-MCTS improves agent performance at inference-time

Exploratory Learning improve agent performance at training-time

R-MCTS

Exploratory Learning

Summary

Inference/training methods to improve agent performance

R-MCTS to improve agent performance at inference-time

Exploratory Learning to improve agent performance at training-time

R-MCTS

❖ Future work

RL methods to reduce reliance on tree search

Model predictive control (MPC) methods to reduce expensive environment interactions

Ongoing and Future Work

Arklex: Agent-First Organization Framework

Mixed Control

Enablesagents toaddress diverse
goals driven by user needs and
builderobjectives,fostering
dynamicand goal-oriented
interactions.

Task Composition

Breaksdowncomplexreal-world tasksintomodular,reusable
componentsmanagedby individual
workers,promoting efficiencyand scalability.

Human Intervention

Integrates human oversightand enables interactive refinement, ensuring critical decisionsare accurateanduserpreferencesare prioritized.

Continuous Learning

Allowsagents to evolveand improve
by learning from interactions,
ensuring sustained relevanceand
effectiveness in dynamic
environments.

Join Arklex Open-source Community: Github Documentation Discord

Arklex is Intelligent and Controllable

More Intelligent

Arklex VS. Other Frameworks

Framework	Open-Source	Mixed-Control	Action Graph	NLU	Task	HumanIntervention	ContinualLearning
Framework	Open-Source	Mixed-Control	Action Graph	NLU	Composition	HumanIntervention	ContinualLearning
Traditional Dialog System
DialogFlow	×		rule-based				×
Amazon Lex		×	rule-based				×
RASA		× r	ule-based				×
LLM-Based Agent Framework
LangChain			LangGraph			LangGraph	×
Llamalndex			Workflows			WorkFlows
AG2						Always/Never
CrewAI			Flows			Always/Never	X
AgentForce
OpenAl Swarm						Swarm
Arklex						√

DAPLab on AI Agents

Eugene Wu Data Systems

Zhou Yu NLP

Kostis Kaffes Computer Vision

David Blei ML, Causal Inference

Lydia Chilton Human-AI Interaction

Adam Elmachtoub Operations Research

Junfeng Yang Secure Systems

Baishakhi Ray Software Eng
Daniel Hsu ML Theory
Shipra Agrawal RL

Carl Vondrick Computer Vision

Yunzhu Li Robotics, Agents

Advisory Board

Richard Zemel, ML Columbia University Director of NSF AI Institute

Michael Franklin, CS University of Chicago Founder of AMPLab Co-Director of Chicago’s Data Science Institute

ZhouYuColumbiaAI_Agent_Yale