Convergence and Divergence in Measures of Induced Valence: A Causal, Placebo-Controlled Test of Induced Valence in Language Models

We built a way to directly push a language model's internal state up or down along a "valence" direction found in its activations, and then checked whether the model's own account of how it's doing actually tracks that push, across 47 open-weight Qwen and Llama models spanning three years of releases.

TLDR: not really, at least not the way you'd want it to. Take the model's self-report scale and swap which end means "good" and which means "bad." If the number reflected something the model was genuinely introspecting on, flipping the labels should flip the sign of how it responds to the push. Across 29 models, it barely did. The reports move in a way that's much better explained by the model mapping a push direction onto a number line than by it noticing an internal state and describing it.

We checked the same question from a different angle with an "endurance" setup: give the model a reward, then apply a slowly worsening negative push over several turns and see whether it keeps enduring it or asks to stop, with a placebo condition that describes the same worsening push but never actually applies it. Most of what the model said about how bad things were came from being told things were getting worse, not from the manipulation itself.

One thing did track scale: the strength of the push mattered less as models got bigger, even though the direction stayed just as easy to detect. Bigger models weren't harder to read, just harder to move with the same-size nudge.

This was a solo three-day sprint, so we're treating the numbers as a first pass. The main thing we'd want a reader to take from it: if you're going to trust an AI's account of its own state, check it against something other than asking, because asking is the part that turned out to be least trustworthy here.