Qwen3-4B

The paper presents READ, a novel framework for merging independently trained LoRA adapters by fixing factor coordinates and coupling direction. This approach prevents interference between tasks and allows for efficient, single-model deployment without additional inference costs.

Researchers have developed FlyBy, a framework that allows small language models to overcome knowledge limitations by selectively querying stronger external models. This approach optimizes test-time computation by distinguishing between problems solvable through more thought and those requiring new information.

The researchers present RIDE, a technique that improves language model distillation by extrapolating RL-induced residuals in the hidden-state representation space rather than the logit space. This approach avoids the noise and attenuation caused by the language-model head, allowing student models to match or surpass their RL-trained teachers.

This technical report examines the efficacy of using Random Attention for KV cache eviction to optimize LLM inference throughput. The study shows that this method significantly boosts tokens per second in high-concurrency scenarios without sacrificing model accuracy.

FlowBalance is a novel reinforcement learning method designed to stabilize the self-improvement process in large reasoning models. It effectively integrates sparse verifier feedback with dense, privileged-hindsight self-guidance to prevent model overconfidence and improve mathematical reasoning accuracy.

The paper introduces TTPO, a method designed to optimize LLM reasoning during test-time training by utilizing asymmetric learning signals from pseudo-labels. It effectively mitigates the risks of noisy majority-vote labels by distilling agreeing rollouts and penalizing disagreeing ones.

This technical report examines the efficacy of using Random Attention for KV cache eviction to optimize LLM inference throughput. The study shows that this method significantly boosts tokens per second in high-concurrency scenarios without sacrificing model accuracy.

The paper introduces TTPO, a method designed to optimize LLM reasoning during test-time training by utilizing asymmetric learning signals from pseudo-labels. It effectively mitigates the risks of noisy majority-vote labels by distilling agreeing rollouts and penalizing disagreeing ones.