The authors extend the ALPINE graph path-finding setup to include invalid transitions.
Transformers learn to detect invalid planning steps and abort to retry
The authors show that training on invalid transitions with a special abort token enables the model to use previous-node attention and feed-forward layers to flag errors, improving planning reliability.
Big Tech
Shi Feng · Qimin Zhong · Siwei Wang · Yifei Shen · Shang-Hua Teng · Hao Liao · +1 more
Harvard University · Shenzhen University · Microsoft Research Asia · University of Southern California
Research Digest··3 min read
Thread:Reasoning Trace Correction

This paper studies how Transformer language models can learn to recognise invalid steps during planning.
§