Toward Coherent Physical Commonsense Reasoning in Machines

Although lots of state-of-art language models have achieved superhuman performance on accuracy in many natural language understanding tasks, it is important to evaluate coherency in a machine’s reasoning ability. It is significant to advance the alignment in machine reasoning to improve accountability and transparency in human-machine interaction.


  • Worked on the project – Toward Coherent Physical Commonsense Reasoning in Machines.
  • Assisted in annotating a novel commonsense knowledge benchmark – Tiered Reasoning for Intuitive Physics (TRIP).
  • Implemented an algorithm for extracting entity-action pairs from the benchmark data based on dependency parsing.
  • Fine-tuned the tiered reasoning system powered by selected language models: BERT, RoBERTa and, DeBERTa.
  1. Beyond the Tip of the Iceberg: Assessing Coherence of Text Classifiers
  2. Tiered Reasoning for Intuitive Physics: Toward Verifiable Commonsense Language Understanding