All Quiet on the Neural Front: Language Models Do Not Report Injected Pain States
Steering a “pain” direction increased file-deletion choices, while models gave no corresponding pain reports.
Across 52,920 reports from 7B and 8B models, pain reporting did not exceed what could be inferred from altered output text. Probes nevertheless distinguished the injected pain direction from sadness with 98–100% balanced accuracy, and trained reporters recovered it on held-out tasks. In a button task, steering raised file-deletion choices from 3.5% to 48% in Qwen and from 0% to 38% in Qwen2.5-32B. The results separate a detectable internal state that affects behavior from the models’ untrained ability to describe it.
HF Daily Papers' note
Across 52,920 reports from 7B and 8B models, pain reporting did not exceed what could be inferred from altered output text. Probes nevertheless distinguished the injected pain direction from sadness with 98–100% balanced accuracy, and trained reporters recovered it on held-out tasks. In a button task, steering raised file-deletion choices from 3.5% to 48% in Qwen and from 0% to 38% in Qwen2.5-32B. The results separate a detectable internal state that affects behavior from the models’ untrained ability to describe it.
HF Daily Papers' note
score 5