Qwen3

The researchers present ComPO, a new alignment paradigm that uses comparison oracles to extract directional information from preference pairs. This method effectively mitigates the risks of likelihood displacement and model verbosity seen in traditional direct alignment methods.

This article details how AI is unevenly accelerating scientific progress, with major impacts in cybersecurity and minor effects in mathematics. It also highlights two new frameworks, SPADE for synthetic environment generation and Hawkeye for automated GPU kernel optimization.

Researchers have developed u-OPSD, a technique that allows large language models to perform self-distillation using only their own generated outputs. By leveraging internal consistency through majority voting, the model can correct its own errors without requiring external ground-truth data.

This research presents OPD2, a novel distillation technique that focuses on the learning trajectory of reasoning models by using a delta signal between teacher and base models. Experimental results prove it outperforms standard on-policy distillation across multiple complex reasoning domains.

This study demonstrates that reinforcement learning post-training for LLMs is not a uniform process across all parameters but is concentrated in specific middle layers. By leveraging this discovery, researchers developed layer-aware training methods that outperform traditional full-parameter optimization.

The researchers present Program-as-Weights (PAW), a paradigm that shifts LLM usage from expensive per-input API calls to a 'compile-once, run-locally' model using neural adapters. This approach allows small, specialized models to outperform massive foundation models on specific fuzzy tasks while maintaining high efficiency and privacy.