Research
Recent work on speech dialogue, multimodal reasoning, and learning from rubric feedback.
Speech-to-speech
Steerable Duplex Speech Dialogue Models
NeurIPS 2026 Workshop RTCA | Under submission ARR
Steerable full-duplex speech dialogue through supervised fine-tuning and two-stage reinforcement learning, with control over vocal delivery and conversational behavior.
Reinforcement learning
Not Every Rubric Teaches Equally: Policy-Aware Rubric Rewards for RLVR
NeurIPS 2026
Policy-aware rubric rewards prioritize criteria that distinguish a model’s current responses, reaching comparable performance in 2.5–4× fewer training steps.
Multimodal reasoning
Benchmarking Intuitive Visual Reasoning in Multimodal Models
ICLR 2027 (under submission)
A benchmark for intuitive visual reasoning across temporal, spatial, social, and abstract tasks, examining the gap between multimodal models and human judgment.
Post-training
A Simple Way to Mitigate Reward Hacking in Rubric-as-Reward RL
ICLR 2027 (under submission)
Randomly masks rubric criteria during reinforcement learning to reduce overfitting to a fixed reward signal and mitigate reward hacking.
Post-training
Post-Training Without Rubric Verifiers
ICLR 2027 (under submission)
Distills a rubric-conditioned teacher into a prompt-only student using token-level feedback, removing external verifier calls from the training loop.
Spoken dialogue evaluation
A Multi-Turn Evaluation of Spoken Dialogue Systems on Natural Human Interaction
ACL 2026 · Best Resource Paper Award
Tests spoken dialogue systems on natural multi-turn interaction, including memory, instruction retention, self-coherence, and voice editing.
Conference papers and recent preprints. Submission statuses are indicated below.