aiqbench analysis
The paper tests whether RL-trained empathetic language models hold up against adversarial users such as gaslighting, escalation, and demands for unconditional validation. Its main finding is an outcome-legibility dissociation: the RLVER-trained model improves emotional outcomes over its baseline, but a cross-family judge finds no difference in how legible the user's emotional state is in the transcript. For readers, the practical lesson is that outcome metrics alone can overstate how well a model understands users, and the authors explicitly avoid claiming their score measures internal understanding.
Can You Break RLVER? Probing Adversarial Robustness of RL-Trained Empathetic Agents
arXiv:2605.07138v2 Announce Type: replace Abstract: Reinforcement learning from verifiable emotion rewards (RLVER) yields strong empathetic language models, but is evaluated with cooperative, honest simulated users. Real users gaslight, escalate, and demand unconditional validation. We introduce the Adversarial Empathy Benchmark (AEB), six adversarial trajectory types whose discriminative rewards penalize formulaic comfort, and the Emotional Consistency Score (ECS), which separates whether a model \emph{improves} the user's emotional state from whether that state stays legible in the transcript. Across eight scenario-matched conditions (480 dialogues; two RLVER and two Qwen2.5 base models, with and without thinking), RLVER-PPO-Think outperforms its untuned same-scale baseline (FS $0.979$ vs.\ $0.761$, $p<10^{-12}$) with no dialogue collapses and 53\% higher hidden-intention detection; the gain is not explained by response length. Yet under an independent cross-family judge (Gemma-3-4b-it), ECS is indistinguishable between the two ($p=0.43$). We call this an outcome--legibility dissociation: RLVER improves emotional outcomes without making the user's state more legible. We do not claim ECS measures internal understanding.
Title:Can You Break RLVER? Probing Adversarial Robustness of RL-Trained Empathetic Agents
Abstract:Reinforcement learning from verifiable emotion rewards (RLVER) yields strong empathetic language models, but is evaluated with cooperative, honest simulated users. Real users gaslight, escalate, and demand unconditional validation. We introduce the Adversarial Empathy Benchmark (AEB), six adversarial trajectory types whose discriminative rewards penalize formulaic comfort, and the Emotional Consistency Score (ECS), which separates whether a model \emph{improves} the user's emotional state from whether that state stays legible in the transcript. Across eight scenario-matched conditions (480 dialogues; two RLVER and two Qwen2.5 base models, with and without thinking), RLVER-PPO-Think outperforms its untuned same-scale baseline (FS $0.979$ vs.\ $0.761$, $p<10^{-12}$) with no dialogue collapses and 53\% higher hidden-intention detection; the gain is not explained by response length. Yet under an independent cross-family judge (Gemma-3-4b-it), ECS is indistinguishable between the two ($p=0.43$). We call this an outcome--legibility dissociation: RLVER improves emotional outcomes without making the user's state more legible. We do not claim ECS measures internal understanding.
Artificial Intelligence (cs.AI); Machine Learning (cs.LG)
arXiv:2605.07138 [cs.AI]
(or arXiv:2605.07138v2 [cs.AI] for this version)
https://doi.org/10.48550/arXiv.2605.07138 Focus to learn more arXiv-issued DOI via DataCite
arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.
Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.
Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.
Full text under the publisher's open licence · 260 words
Related coverage
Robust Driving Control for Autonomous Vehicles: An Intelligent General-sum Constrained Adversarial Reinforcement Learning Approach
arXiv · Jun 8, 2026
Geometry Is Not Robustness: A Trajectory-Level Study of PGD Evaluation
arXiv · Aug 19, 2026
Adversarial Training Improves Generalization Under Distribution Shifts in Bird Sound Classification
arXiv · Aug 25, 2026
CARD: Closing the Teacher-Student Loop for Adversarial Robustness Distillation
arXiv · Oct 15, 2026