When LLM Agents Fail to Read the Room: ReAdapt for Relational Social Reasoning

A social agent's most fundamental decisions (for example, “should I react to this post?”, “who should I reach out to?”) are not purely content problems. The appropriate action can hinge on the latent relationship between people, i.e., tie strength, reciprocity, and mutual connections, rather than on which piece of content is most salient. Yet standard LLM-based agent loops do not explicitly represent how newly observed relational evidence should revise the agent's current social hypothesis, making them vulnerable to surface-obvious choices when relational and content cues diverge. We formalize this failure mode through a relationship-reasoning benchmark: 500 synthetic social worlds containing friendships, directed follows, reaction histories, and feeds, yielding 1,000 queries across two tasks---reaction selection (“which post should I engage with?”) and warm introduction (”who is the best bridge to this person?“). By construction, the surface-obvious candidate differs from the relationship-grounded oracle in approximately 53% of queries, forming an overturn subset in which the agent must use relational evidence to revise an initially plausible choice. We propose ReAdapt Relationship-Adaptive Agent with Policydriven sTate, which augments the ReAct loop with an explicit structured social state $z=(\mathcal{G},\mathcal{B},\mathcal{R},\mathcal{N},\mathcal{D})$ capturing goal, belief, relationship, norm, and disclosure. After each tool observation, ReAdapt performs a typed Adapt step that updates this state and emits a policy operation (continue, switch, abandon, or clarify) before selecting the next action. Evaluated with Gemini-3-Flash on a stratified subset of n=150 queries per task, ReAdapt improves warm-introduction accuracy from 37% to 51% (+14 points) and reaction-selection accuracy from 69% to 77% (+8 points). Oracle regret decreases from 0.260 to 0.152 on warm introduction and from 0.095 to 0.053 on reaction selection. With the same underlying model, tool interface, and social environments, these results suggest that explicit relational-state adaptation helps LLM agents translate retrieved social evidence into revised decisions.