Linear probes detect task format, not reasoning mode in LLMs

Arxiv CS.CL (NLP/language models)

Probing Qwen3-14B hidden states, the paper finds that linear classifiers detect task format (e.g., math vs. common sense) rather than reasoning mode (e.g., deductive vs. inductive). Benchmark spanning the classical trichotomy shows emergent capabilities are not captured by such probes.

llminterpretabilitylinear-probing

Arxiv CS.CL (NLP/language models)·Subramanyam SahooVinija JainAman ChadhaDivya Chaudhary

8 days ago