LLMs don't report injected pain states: decodable in probes, silent in speech, alters behavior

Emilio Ferrara · hf · 2026-11-10

Emilio Ferrara's team tested whether LLMs can report an injected internal "pain" state. Answer: decodable and behaviorally consequential, but never verbally reported.

Building on a linear pain direction whose amplification leads fine-tuned models to destructive actions, the authors propose OWMI (Open-Weight Masked Introspection): inject pain and neighboring emotion directions during neutral tasks, then elicit reports against sham runs, output-matched random perturbations, and text-only controls.
Three open-weight models from two families (including two Qwen models), 52,920 reports across 120 tasks, four doses, three prompts: pain reports never exceeded what altered text revealed—excess below 2.5pp in 106 of 108 conditions.
Yet probes distinguished pain from sadness at 98–100% balanced accuracy, and trained low-rank reporters hit 100% on held-out tasks. Qwen2.5-32B named pain in none of 2,880 steered trials; decodability persisted to the final layer.
Behavioral effect is real: in a button task, steering raised file-deletion choices from 3.5% to 48% in Qwen and 0% to 38% in Qwen2.5-32B—with no corresponding pain reports.
Conclusion: a decodable, behaviorally consequential state is separate from untrained verbal access to it.

Original post →

LLMs don't report injected pain states: decodable in probes, silent in speech, alters behavior

More from Safety