text stringlengths 1.71k 5.38k |
|---|
Preprint Agent Kit: Flow Engineering with Graphs, not Coding Yue Wu1,3/envel⌢pe, Yewen Fan1, So Yeon Min1, Shrimai Prabhumoye2,4, Stephen Mc Aleer1, Yonatan Bisk1, Ruslan Salakhutdinov1, Yuanzhi Li1,3, Tom Mitchell1 1Carnegie Mellon University,2NVIDIA,3Microsoft,4Boston University Abstract We propose an intuitive LLM p... |
Preprint Figure 1: A user breaks down a task into sub-tasks (nodes) representing a “thought process” and creates prompts for the sub-tasks (nodes). Sub-tasks (nodes) in Agent Kit can be designed and assembled in different ways to achieve diverse functionalities, similar to LEGO pieces. intuitively instructs an LLM to f... |
Preprint Figure 2: Each node in Agent Kit takes outputs from its dependencies and outputs a string to complete a predefined subtask. The orange components (After-query) are optional and can be further customized with minimal programming through the Agent Kit API. Left: The evaluation process inside a node consists of c... |
Preprint “intentions of other road users”, may need to be post-processed into Json format and added to the database, as illustrated in Algorithm 1 and Figure 2. Algorithm 1 After-query operation example Av← M LLM(Cv) ▷initialize Avwith the LLM response result←Parse( Av) ifresult does not match format then raise After Q... |
Preprint Figure 3: Node names are abbreviated for space. (a)At every step in the game, three summary nodes (green) ns-obs,ns-plan,ns-action summarize the observation, plan, and action of the current step. (b)At step T, all planner nodes (blue) take o T-1,o Tand manual Ias input, and output 3 subgoals and a skill s T. n... |
Preprint Method Score Reward Costs1,3 Human Experts 50. 5 ±6. 8% 14. 3 ±2. 3 N/A Agent Kit (Model mix∗) 20. 64%‡12. 8±2. 1 $138 SPRING (GPT-4) (Wu et al., 2023b) 27. 3±1. 2% 12. 3±0. 7 $243† Dreamer V3 (Hafner et al., 2023) 14. 5 ±1. 6% 11. 7±1. 9 1M steps EDE (Jiang et al., 2022) 11. 7 ±1. 0% N/A 1M steps Dreamer V2 (... |
Preprint Figure 4: Left three columns: an example trajectory in Crafter. Different nodes on planning, reflection, feedback, knolwedge discovery work together to complete the first 11 steps and successfully crafting the table. Through environment interactions and error identification/-correction, the agent discovered tw... |
Preprint 4. 1. 2 Planning, Reflection, and Learning from Interactions We show an illustrated example of the first 11 steps taken by our Crafter agent in Figure 4. 1. 1. The agent started empty handed in the environment (step 0) and decided to approach the challenge n0 cof “Resource Collection”. Given the challenge, it ... |
Preprint 5. 2 LLM Agents In a more interactive setting, the LLM typically has access to some status information of the agent, and can react to environment signals. Yao et al. (2022b) solves simple natural language tasks through LLM interaction with Chain-of-thought prompts. Shinn et al. (2023) further adds the capabili... |
Preprint References Michael Ahn, Anthony Brohan, Noah Brown, Yevgen Chebotar, Omar Cortes, Byron David, Chelsea Finn, Chuyuan Fu, Keerthana Gopalakrishnan, Karol Hausman, Alex Herzog, Daniel Ho, Jasmine Hsu, Julian Ibarz, Brian Ichter, Alex Irpan, Eric Jang, Rosario Jauregui Ruano, Kyle Jeffrey, Sally Jesmonth, Nikhil ... |
Preprint Matteo Hessel, Joseph Modayil, Hado Van Hasselt, Tom Schaul, Georg Ostrovski, Will Dab-ney, Dan Horgan, Bilal Piot, Mohammad Azar, and David Silver. Rainbow: Combining improvements in deep reinforcement learning. In Thirty-second AAAI conference on artificial intelligence, 2018. Wenlong Huang, Pieter Abbeel, D... |
Preprint Jupinder Parmar, Shrimai Prabhumoye, Joseph Jennings, Mostofa Patwary, Sandeep Subra-manian, Dan Su, Chen Zhu, Deepak Narayanan, Aastha Jhunjhunwala, Ayush Dattagupta, Vibhu Jawa, Jiwei Liu, Ameya Mahabaleshwarkar, Osvald Nitski, Annika Brundyn, James Maki, Miguel Martinez, Jiaxuan You, John Kamalu, Patrick Le... |
Preprint Yue Wu, Xuan Tang, Tom Mitchell, and Yuanzhi Li. Smartplay: A benchmark for llms as intelligent agents. In The Twelfth International Conference on Learning Representations, 2024b. Shunyu Yao, Howard Chen, John Yang, and Karthik Narasimhan. Webshop: Towards scalable real-world web interaction with grounded lang... |
Preprint A Kahn's Algorithm Algorithm 2 Kahn's Algorithm (Kahn, 1962) for topological transverse 1:L←empty list that will contain the sorted elements 2:F←set of all nodes with no incoming edges ▷Initialize the frontier F 3:in Degree ←map of node to in-degree 4:while Fis not empty do 5: remove a node nfrom F ▷Retrieve t... |
Preprint L i s t of desired i n t e r a c t i o n s :-avoid zombies, skeletons, and spiders.-c o l l e c t s a p l i n g s.-c r a f t a wood pickaxe.-c o l l e c t wood.-c r a f t a stone pickaxe.-c o l l e c t stone.-c r a f t a furnace.-c o l l e c t coal.-c o l l e c t iron.-c r a f t an iron pickaxe.-c o l l e c t ... |
Preprint L i s t of a l l a c t i o n s and t h e i r requirements : 1. Move West : F l a t ground to the west of the agent. 2. Move East : F l a t ground to the e a s t of the agent. 3. Move North : F l a t ground to the north of the agent. 4. Move South : F l a t ground to the south of the agent. 5. Do : Facing c r e... |
Preprint-t r e e 3 s tep s to north-west, 2N 1 W-t a b l e 5 s tep s to south-east, 2S 3E-zombie 1 s tep s to west, 1 W-zombie 1 s tep s to east, 1E *Further to the north : 25 grass ( s ), 2 t r e e ( s ). *V i t a l s :-health : 1/9-food : 7/9-drink : 6/9-energy : 9/9 *Inventory :-sapling : 1-coal : 1-iron : 1-wood_pi... |
Preprint "Proximity_Requirements_for_Do_Action": "Must face and be near tree for 'Do'action. ", "Wood_Sword_Crafting_Requirements": "Wood and table needed for wood sword. ", "Crafting_Location_Specifics_for_Wood_Sword": "Near table, facing not specified. ", "Crafting_Process_Details_for_Wood_Sword": " 'Do'action near t... |
Preprint }, 's-obs-vitals ':{ 'prompt ':"In one sentence, describe the current vitals. ", 'dep':['obs_vit '], }, 's-action ':{ 'prompt ':"""Output a Json dictionary of the following format: ``` { "action": $action, # The most recent action, including the direction if it 's a movement action "repeats": $repeats # The nu... |
Preprint "mistake": $ANSWER, # [yes/no] Did the player make any mistakes? "correction_planned": $ANSWER, # [yes/no] Was correction planned at current step for the mistake? "confused": $ANSWER, # [yes/no] Does the player seem confused? "top_subgoal_completed": $ANSWER, # [yes/no] Is the most recent top subgoal complete ... |
Preprint 'prompt ':"""Merge the 'unknown/missing information 'in the previous answer and the 'Previous unknown information and details 'into a single Json dictionary., Give each item a concise but precise name as the key, and a dictionary of answers to the following inquiries as the value: ``` "item_name": { # if appli... |
Preprint First, identify any hazards in the observation/surrounding, and identify any obstacles that may interfere with achieving the target.-Plan-sketch: $db. action_summary. plan-sketch$-Plan details: $db. action_summary. details$ Discuss how to spatially address or evade the hazards and obstacles based on analysis o... |
Preprint Use 'TBD'for the action corresponding to the current step. Be precise with the chronological order of each step. Third, describe the current webpage in less than two sentences, drawing connection to the task: ``` $db. environment. refined_task$ ``` Then, reason with the browsing summary, the current webpage, a... |
README.md exists but content is empty.
- Downloads last month
- 4