Does theory of mind enhancement help AI across all task types?

Testing the same four ToM enhancement techniques (Foresee and Reflect, Perspective Taking, supervised fine-tuning, and reinforcement learning) on GPT-4o and Llama-3.1-8B, the paper's authors find the techniques split sharply by task type: "ToM enhancements boost empathy and user experience, although fail to support goal achievement." The discussion states this as a standalone takeaway: "ToM enhancement techniques fail to consistently improve goal-oriented task performance across diverse domains," while the same techniques help in experience-oriented settings like counseling. The authors also flag a safety cost specific to the weaker base model: fine-tuning methods "may introduce statistically meaningful safety and ethical trade-offs" in that base model, "whereas stronger models such as GPT-4o appear more robust."

The authors' explanation rests on what each task type actually demands of ToM. Drawing on Bales's (1950) split between task and socio-emotional group processes, they argue experience-oriented tasks are "largely defined by their focus on interpersonal dynamics and responding to affective states like emotions and desires," which is exactly what ToM enhancement (reflecting on predicted mental states, taking the user's perspective) targets. Goal-oriented tasks instead "require understanding users' intention progress and underlying knowledge states for task accomplishment" — a different, more procedural kind of user-modeling that the tested enhancement techniques were not built to improve. The authors conclude "ToM proficiency in one type of task may not guarantee success in the other, as each emphasizes different aspects of user understanding."

This task-type split functions as the mechanism behind Do theory of mind improvements transfer from benchmarks to real conversation?: the benchmark-to-interaction gap documented there is not uniform but concentrated in goal-oriented tasks specifically. It also parallels Does theory of mind predict who thrives in AI collaboration?, which separated individual from collaborative ability rather than treating ToM as one undifferentiated skill; here the split runs along task structure (affective support versus procedural goal-tracking) instead of user trait, but both resist a single ToM score as a complete account of HAI performance. It also extends Can language models solve ToM benchmarks without real reasoning?: the same SFT and RL methods questioned there for exploiting benchmark structure are shown here to fail at generalizing to goal-oriented interaction even when they succeed at experience-oriented interaction, adding a second axis of doubt about what the fine-tuning gains represent.

The excerpt does not give the specific accuracy figures or statistical tests behind the goal-oriented failure, and it does not explain the mechanism by which fine-tuning raises safety and ethical risk in the weaker base model while sparing GPT-4o. The implication, at the strength the excerpt supports, is that a single ToM-enhancement technique should not be assumed to transfer across a product's full task mix: an assistant that handles both task completion and emotional support may need separate tuning, or separate handling, for each.