Selected work

Representative Papers

When Robots Say No: The Empathic Ethical Disobedience Benchmark

Dec 20, 2025

Robots must balance instruction-following with adherence to safety and social norms, yet existing safety reinforcement learning benchmarks emphasize physical risks, while human-robot trust studies suffer from limited scale and poor reproducibility. Method: We propose the Empathic Ethical Disobedience (EED) benchmark and introduce EED Gym—a standardized, multi-role, multi-scenario testbed enabling systematic evaluation of compliance, refusal, clarification, and alternative-action decisions. Contribution/Results: We jointly quantify refusal behavior along three dimensions: safety, user trust, and empathy. We integrate empirically grounded blame/trust models and a personified role framework, and define verifiable credibility tiers for constructive, empathic, and other refusal styles. Experiments show that action masking eliminates unsafe compliance; explanatory refusals preserve trust; constructive refusals achieve highest credibility scores, while empathic refusals yield highest empathy scores; safety-aware RL improves robustness but often induces excessive conservatism.

0 citationsRead paper

Knowledge Transfer in Model-Based Reinforcement Learning Agents for Efficient Multi-Task Learning

Jan 09, 2025

To address the poor generalization and deployment challenges of multi-task model-based reinforcement learning (MBRL) in resource-constrained settings, this paper proposes a knowledge transfer framework tailored for lightweight agents. We first compress a high-capacity multi-task world model (317M parameters) into an ultra-compact agent (1M parameters) via knowledge distillation, then apply FP16 post-training quantization to further reduce model size by 50%. Evaluated on the MT30 multi-task benchmark, our method achieves a normalized score of 28.45—improving upon the original 1M baseline by 50.5% and establishing a new state-of-the-art (SOTA). The model size is reduced by 317× with zero performance degradation. Our core contribution lies in establishing an efficient, fidelity-preserving knowledge distillation and compression paradigm that transfers capabilities from large-scale multi-task world models to extremely lightweight MBRL agents—significantly enhancing multi-task generalization and deployment feasibility on edge devices.

0 citationsRead paper

Recent publications

Latest Papers

When Robots Say No: The Empathic Ethical Disobedience Benchmark

Dec 20, 2025

Robots must balance instruction-following with adherence to safety and social norms, yet existing safety reinforcement learning benchmarks emphasize physical risks, while human-robot trust studies suffer from limited scale and poor reproducibility. Method: We propose the Empathic Ethical Disobedience (EED) benchmark and introduce EED Gym—a standardized, multi-role, multi-scenario testbed enabling systematic evaluation of compliance, refusal, clarification, and alternative-action decisions. Contribution/Results: We jointly quantify refusal behavior along three dimensions: safety, user trust, and empathy. We integrate empirically grounded blame/trust models and a personified role framework, and define verifiable credibility tiers for constructive, empathic, and other refusal styles. Experiments show that action masking eliminates unsafe compliance; explanatory refusals preserve trust; constructive refusals achieve highest credibility scores, while empathic refusals yield highest empathy scores; safety-aware RL improves robustness but often induces excessive conservatism.

0 citationsRead paper

Knowledge Transfer in Model-Based Reinforcement Learning Agents for Efficient Multi-Task Learning

Jan 09, 2025

To address the poor generalization and deployment challenges of multi-task model-based reinforcement learning (MBRL) in resource-constrained settings, this paper proposes a knowledge transfer framework tailored for lightweight agents. We first compress a high-capacity multi-task world model (317M parameters) into an ultra-compact agent (1M parameters) via knowledge distillation, then apply FP16 post-training quantization to further reduce model size by 50%. Evaluated on the MT30 multi-task benchmark, our method achieves a normalized score of 28.45—improving upon the original 1M baseline by 50.5% and establishing a new state-of-the-art (SOTA). The model size is reduced by 317× with zero performance degradation. Our core contribution lies in establishing an efficient, fidelity-preserving knowledge distillation and compression paradigm that transfers capabilities from large-scale multi-task world models to extremely lightweight MBRL agents—significantly enhancing multi-task generalization and deployment feasibility on edge devices.

0 citationsRead paper