ZhexiLu commited on
Commit
8bfb76e
·
verified ·
1 Parent(s): b58afcc

Model card: add code link and citation authors

Browse files
Files changed (1) hide show
  1. README.md +4 -3
README.md CHANGED
@@ -18,6 +18,8 @@ tags:
18
  StateAlign-8B is a multi-turn tool-use agent trained from [Qwen3-8B](https://huggingface.co/Qwen/Qwen3-8B) with **StateAlign**, introduced in
19
  *StateAlign: Internalizing State-Aware Decision-Making for Reliable Multi-Turn Tool Agents via Reinforcement Learning*.
20
 
 
 
21
  A tool agent should decide its next action from the current task state: what information is known or missing, whether the available tools can perform the request, and which parts of the task are already done. StateAlign constructs task states from tool-use trajectories (execute, ask the user, retrieve, decline, stop) and pairs states that differ in one condition. Training combines outcome rewards from interaction with a token-level state-contrastive signal that compares the same generated reasoning and actions under paired states.
22
 
23
  ## Results
@@ -87,10 +89,9 @@ Released under Apache 2.0, following the Qwen3-8B license.
87
  ## Citation
88
 
89
  ```bibtex
90
- @article{statealign2026,
91
  title = {StateAlign: Internalizing State-Aware Decision-Making for Reliable Multi-Turn Tool Agents via Reinforcement Learning},
92
- author = {TBD},
93
- journal= {arXiv preprint arXiv:TBD},
94
  year = {2026}
95
  }
96
  ```
 
18
  StateAlign-8B is a multi-turn tool-use agent trained from [Qwen3-8B](https://huggingface.co/Qwen/Qwen3-8B) with **StateAlign**, introduced in
19
  *StateAlign: Internalizing State-Aware Decision-Making for Reliable Multi-Turn Tool Agents via Reinforcement Learning*.
20
 
21
+ **Code:** [github.com/ZhexiLu/StateAlign](https://github.com/ZhexiLu/StateAlign) (training code, example data and the BFCL evaluation script).
22
+
23
  A tool agent should decide its next action from the current task state: what information is known or missing, whether the available tools can perform the request, and which parts of the task are already done. StateAlign constructs task states from tool-use trajectories (execute, ask the user, retrieve, decline, stop) and pairs states that differ in one condition. Training combines outcome rewards from interaction with a token-level state-contrastive signal that compares the same generated reasoning and actions under paired states.
24
 
25
  ## Results
 
89
  ## Citation
90
 
91
  ```bibtex
92
+ @misc{lu2026statealign,
93
  title = {StateAlign: Internalizing State-Aware Decision-Making for Reliable Multi-Turn Tool Agents via Reinforcement Learning},
94
+ author = {Lu, Zhexi and Chaudhury, Subhajit and Pedapati, Tejaswini and Murugesan, Keerthiram and Yu, Lei},
 
95
  year = {2026}
96
  }
97
  ```