DualOPSD: Adaptive Privileged Teachers

CLAIM Lab, Clemson University · 2026

On-policy self-distillation uses a privileged copy of the student model to provide dense token-level supervision, but prior approaches keep that teacher fixed while the student distribution and output style change. DualOPSD closes this loop: the student learns conservatively from privileged supervision, and the teacher then adapts toward the updated student distribution on the same student trajectory.

The asymmetric alternating update changes future supervision without requiring an additional rollout or a correctness verifier. Experiments across Qwen3 models also reveal a scale-dependent benefit: teacher adaptation helps at 4B and 8B, while the smallest 1.7B model does not improve.

Key result: On Qwen3-8B, DualOPSD raises avg@12 over OPSD by 23.61 points on AIME 2024, 13.89 points on AIME 2025, and 10.00 points on HMMT 2025. It also reduces truncation across all three tested model scales.

CoScale-RL

Multimedia Signal and Intelligent Information Processing Lab, Tsinghua University · Aug 2025–Jan 2026

Training large reasoning models can be unstable on difficult problems or with weaker foundation models. We scale the number of solutions per problem to improve solvability, then scale rollout computation to stabilize reinforcement learning. A model-merging technique called re-distillation helps sustain efficiency as computation grows.

Key result: An average 3.76× improvement in accuracy across four benchmarks, extending the model's capability boundary without an extensive supervised fine-tuning dataset.

Small-Scale Fine-Tuning in R1-Style RL

Multimedia Signal and Intelligent Information Processing Lab, Tsinghua University · Feb–Jul 2025

We developed an analytical framework based on sample effect to compare supervised fine-tuning and reinforcement learning. The resulting re-distillation method samples from an RL-trained policy to make small-scale distillation more effective across Qwen and Llama models.

Key result: On the K&K dataset, the re-distilled Qwen-2.5-1.5B model surpassed DeepSeek-V3-0324 using only 1K supervised fine-tuning examples.

Expressive Speech-Driven Facial Animation

Speech and Audio Technology Lab · Jan–Jul 2023

We introduced an emotion controller that learns the relationship between variations in emotion type and intensity and facial expression parameters. The system generates expressive facial animation while retaining accurate lip synchronization and lets users continuously adjust the target expression.

Contribution: Coding, experiment design, and the initial manuscript draft under the supervision of Wei-Qiang Zhang and Junhong Zhao.