Guojun Yin

UniMem achieves a remarkable 4.0 EM point improvement in task execution fidelity by seamlessly integrating episodic and parametric memory without the need for explicit task labels.

Jointly training MTP and RL doesn't have to hurt: a simple coefficient calibration scheme unlocks performance gains on mathematical reasoning tasks.

Overcome the prohibitive cost of ground-truth labels in reinforcement learning by actively acquiring labels for only the most valuable samples, leading to stable training and improved performance even with limited annotation budgets.

LLMs can reason better and generate more diverse outputs by projecting negative samples onto a positive subspace during reinforcement learning.

Self-play can be dramatically improved by exploiting the "question construction path" it generates as privileged information for self-distillation, leading to 2-3x faster learning.

Current VLMs, despite excelling at general reasoning, still fail to accurately identify food and estimate nutrition, even when given multiple views and chain-of-thought prompting.

Forget noisy, biased LLM evaluators: CDRRM distills preference insights into compact rubrics, letting a frozen judge model leapfrog fully fine-tuned baselines with just 3k training samples.

Predict how well your LLM will transfer to a new domain *before* fine-tuning, by using sparse autoencoders to spot tell-tale signs of domain shift in the model's representations.

Forget static agent communication graphs: AgentConductor uses RL to dynamically rewire agent interactions based on task difficulty, slashing token costs by up to 68% while boosting code generation accuracy.