EIBench: A Simulator-Based Benchmark and Turn-Credit RL for Emotion Management

A new benchmark called EIBench evaluates large language models on interactive emotion management, moving beyond static understanding to multi-turn dialogue where a model must improve a user's emotional state, researchers report [1]. The benchmark, introduced in a paper posted to arXiv, contains 2,222 scenarios, with 2,009 designated for training and 213 for held-out testing [1]. The scenarios are organized by a 2x2 taxonomy covering Support, Defense, Repair, and Charm, which together capture forms of support, boundary maintenance, trust repair, and rapport building [1]. In each scenario, an LLM simulator plays the user, updates an emotion-relation state after each turn, and maps the final state to an anchor-based score [1]. This design makes EIBench both an evaluation benchmark and a training environment: the final state gives the outcome reward, while the per-turn state updates provide dense feedback for reinforcement learning [1]. The researchers evaluated 15 open- and closed-source LLMs [1]. Current models performed well on support and rapport-building scenes, but struggled with boundary maintenance under user pressure [1]. To address this gap, the team proposed Centered Turn-Credit GRPO, or CTC-GRPO, a GRPO extension that reuses the simulator's per-turn state updates as dense turn-level feedback while preserving the final outcome reward [1]. CTC-GRPO improved Qwen3-8B from -22.4 to +22.4 on EIBench [1]. The method also showed gains on out-of-distribution evaluations, including SAGE at +12.4 and EQBench3 at +20.9% [1]. The results indicate that simulator-tracked user states can support both evaluation and training for multi-turn emotion management [1]. The work reflects a broader push in AI research to create dynamic, interactive evaluation frameworks rather than relying solely on static benchmarks. While the primary source does not detail comparable historical efforts, the approach aligns with trends in other domains, such as catalysis informatics, where researchers have explored transfer learning across datasets to improve model performance on related tasks [4]. The EIBench paper's emphasis on reusable simulator feedback for reinforcement learning also parallels efforts in other fields to consolidate computational methods into anthological dataset collections for more robust model training [4].