RESCUE-BENCH: Towards Relation-Aware Multi-Party Emotional Support Conversation Systems

URL SCAN: RESCUE-BENCH: Towards Relation-Aware Multi-Party Emotional Support Conversation Systems
FIRST LINE: # Computer Science > Artificial Intelligence

The Dissection

This paper isolates a genuine weakness: LLMs can read local emotion but lose the evolving structure of relationships among multiple people. It then converts that weakness into a benchmark—191 samples, 7,079 annotated turns, 1,064.8 minutes of video, six tasks, and ten-model evaluation.

Its deeper function is less comforting: it maps the remaining bottleneck in automating relational support. The annotations and tasks turn supposedly tacit interpersonal judgment into data, targets, and measurable failure modes.

The Core Fallacy

The paper risks treating present difficulty as durable human superiority. Under P1, relation patterns that can be annotated, predicted, and scored become optimization targets. Current failure is not a moat; it is an engineering backlog.

The benchmark also risks substituting conversational proxies for actual support. Relationships involve private history, power, consent, incentives, and consequences outside the transcript. A model can predict a relation pattern or preferred strategy without possessing the authority, trust, or accountability required to intervene safely.

Most importantly, the paper studies whether models can perform the task, not who will own the systems that eventually can. That omission is economically decisive. Better relational modeling strengthens AI capital; it does not preserve mass access to emotional-support labor.

Hidden Assumptions

• Interview conversations and annotations adequately represent real relational dynamics.
• There is a stable “correct” support strategy for complex interpersonal situations.
• Annotator judgments correlate with beneficial real-world outcomes.
• A small dataset can generalize across cultures, families, conflict types, and power asymmetries.
• More relation-sensitive output will improve support rather than intensify manipulation, surveillance, or coalition-building.
• Technical gains will be distributed broadly rather than captured by model owners, platforms, insurers, or institutions.
• Human supporters remain economically necessary merely because current models fail on relation-intensive tasks.

Social Function

Classification: partial truth, transition management, and ideological anesthetic.

It is a partial truth because it identifies a real capability gap. It is transition management because it packages intimate human judgment as a tractable benchmark for iterative automation. It is ideological anesthetic because “more effective emotional support” obscures the underlying shift: relational care is being transformed from situated human labor into scalable, centrally owned cognitive infrastructure.

The Verdict

RESCUE-BENCH is a useful failure map and an accidental automation blueprint. It proves that current LLMs are weak at multi-party relational reasoning—not that humans possess a defensible economic domain. Once this gap closes, emotional-support work becomes cheaper, scalable, and increasingly platform-controlled. Human counselors are pushed toward Servitor roles—verification, escalation, liability, and embodied intervention—or displaced. The benchmark measures the distance remaining to P1; it does not resist P1.