Detectable empathy directions in LLMs don't guarantee control
Researchers tested a common assumption in interpretability work: that if you can decode a concept, such as "empathy," as a direction inside a language model's internal activations, you can also use that same direction to reliably control the model's behavior. The study focuses on two empathy facets drawn from the EPITOME framework: Recognition, the cognitive facet (recognizing what someone else is feeling), and Resonance, the affective facet (sharing or reflecting that feeling). Both facets were tested across three instruction-tuned language models, Qwen, Llama and Gemma. Every intervention was scored with two separate LLM judges plus a discriminative EPITOME classifier, and each of those three instruments was first checked with a positive control: a simple test using clearly emotional versus neutral text, to confirm the instrument actually registers the thing it is supposed to measure before trusting it on subtler cases.
The positive control passed cleanly for the affective facet (Resonance) across all three scoring instruments. For the cognitive facet (Recognition), the control's range was inconsistent from one instrument to the next, meaning the instruments do not agree on how to register cognitive empathy even before any steering is applied. Both facets stayed decodable even after the researchers removed a sentence-embedding-derived surface score from the signal, which rules out the possibility that the direction was just tracking superficial wording rather than the underlying concept. Steering intervention could also substantially rewrite the generated text. But when the Resonance direction was added into Qwen, the model's automated affective empathy score rose by only +0.29, about 26% of the score's natural gap, a partial shift rather than a full one. A direct contrast between the two facet directions confirmed the shift was facet-specific, meaning the Resonance direction moved the affective score and not the cognitive one, in Qwen and in Llama, but not in Gemma. The study does not establish that this automated-score shift corresponds to a matching change in how a human reader would perceive the text's empathy.
On the cognitive side, additively steering with the Recognition direction produced no measurable change on the automated instruments. A separate within-domain control test showed the cognitive-empathy instrument is too coarse to resolve the size of change that this kind of steering would be expected to produce, so the researchers describe the result as unmeasurable rather than a genuine null effect. Gemma was the exception on the cognitive side: ablating (removing, rather than adding) the Recognition direction lowered the model's classifier-scored cognitive empathy, and that drop held up even after adjusting for response length. The authors conclude that being able to detect a concept's direction inside a model does not imply you can reliably control that concept through global steering interventions, and that any claim about steering cognitive empathy specifically needs an explicit check on whether the measurement instrument is sensitive enough to detect the change being claimed.
Key facts
- The study tests two EPITOME-derived empathy facets, cognitive Recognition and affective Resonance, as steering directions across three instruction-tuned LLMs: Qwen, Llama and Gemma.
- Adding the Resonance (affective) direction into Qwen raised its automated affective empathy score by only +0.29, about 26% of the score's natural gap, a partial rather than full shift.
- A direct contrast between the two directions confirmed the steering effect was facet-specific in Qwen and Llama, but not in Gemma.
- Additive cognitive steering produced no measurable change on the automated instruments; a control test showed the cognitive instrument is too coarse to resolve such changes, so the result counts as unmeasurable rather than a clean null.
- Gemma was the exception: removing (ablating) the Recognition direction lowered its cognitive empathy score even after adjusting for response length.
Why it matters
Interpretability and alignment work increasingly treats a decodable concept direction inside a model's activations, something you can read a concept like empathy off of, as also being a usable lever: add the direction and the model's behavior along that concept should shift accordingly. This study is a direct empirical test of that assumption, using empathy as the test case, and finds the two properties come apart. A direction can be reliably decodable, survive controls for surface wording, and still produce only a partial, inconsistent, or in one model's case unmeasurable shift when used for control. That gap matters for anyone treating steering vectors as a dependable behavior-control technique rather than just a diagnostic one.
Who it affects
Interpretability and alignment researchers who build or evaluate steering-based behavior control, and teams that lean on automated LLM-judge or classifier scores to certify that an intervention worked. The finding is a caution rather than a product change: it does not describe a tool, release or policy that affects end users of these models directly.
How to use it
The study's own recipe is the practical takeaway: before trusting a steering result, run a positive control (a clearly obvious case) on every scoring instrument used, check whether facets are truly separable with a direct between-direction contrast, and when a steering effect comes back null, run a within-domain control to check whether the measurement instrument is even capable of detecting the expected change before calling it a genuine null. The authors explicitly frame this measurement-sensitivity check as something cognitive-empathy claims specifically need.
How solid is it
The design is reasonably rigorous for an automated-metrics study: three separate models, two empathy facets, three independent scoring instruments (two LLM judges plus a discriminative classifier), and a positive control gating each instrument before its results are trusted. The affective facet's control passed consistently across instruments; the cognitive facet's control did not, which weakens confidence in the cognitive-side results specifically before steering even enters the picture. The facet-specificity result replicated in two of the three tested models (Qwen and Llama) but not the third (Gemma), and the whole study rests on automated scores only.
Risks and caveats
No human evaluation was run, so there is no confirmation that the automated score shifts the researchers measured correspond to a human reader actually perceiving more or less empathy in the text. The cognitive-steering null result is explicitly flagged as unmeasurable rather than a confirmed absence of effect, since the instrument used may simply be too coarse. Results also differ by model: Gemma behaves differently from Qwen and Llama on both the facet-specificity test and the cognitive-ablation result, so the findings should not be assumed to generalize across all instruction-tuned models or to other steerable concepts beyond empathy.