Tag
A paper finds that reasoning models and humans struggle with similar commonsense problems, and that chain-of-thought token count correlates with human response time when averaged across multiple runs, making it a more reliable signal of problem difficulty.
A new paper introduces SaliTrap, a benchmark of 1,145 prompts revealing that LLMs know impossible tasks but still optimize for explicit details, a failure called salience bias. Even the best models avoid traps only ~55% of the time, and awareness often doesn't prevent compliance.
This paper introduces SaliTrap, a benchmark revealing that large language models suffer from salience bias—being distracted by explicit but irrelevant conditions—in commonsense reasoning. It shows the failure is mostly due to knowledge suppression rather than absence, and that simple prompting can close the gap.
TouchThinker introduces a million-scale tactile reasoning dataset and benchmark to scale tactile commonsense reasoning to open-world settings, using action-aware representation for efficient reasoning.
This paper identifies a critical failure mode in LLM agents where they fail to update personalized memories when new evidence conflicts with prior beliefs. It introduces the STALE benchmark and a three-dimensional probing framework, revealing that even the best models achieve only 55.2% accuracy, and proposes CUPMem as a prototype for robust memory revision.