Jiaheng Liu

Looping the middle layers of MoE Transformers can save up to 18% in training FLOPs while boosting performance beyond validation loss predictions.

LLMs can create effective execution infrastructures for certain tasks, but struggle to match human-engineered solutions in critical areas like coding and search.

Transforming pre-training data with reasoning annotations boosts model performance by over 2 percentage points without altering the training objective.

Self-improvement in LLMs is not automatic; the pathway to effective learning depends critically on the task structure and the nature of the feedback received.

Vague goals can misdirect model evolution, revealing that agents often overfit to narrow self-assessments, hindering broader learning outcomes.

Procedura achieves the sharpest edges and the most editable 3D models yet, transforming how we generate and interact with 3D geometry.

Transforming scientific papers into multi-turn generation trajectories not only doubles the training data but also boosts academic writing benchmarks while maintaining reasoning skills.

Despite advances in AI, even top models struggle with real-world tasks, achieving only 30% success on a benchmark grounded in market-validated workflows.

LLMs can be made better software engineers by pre-training them to reconstruct the messy, iterative development process that led to the final, clean code in repositories.

Ditch the deep thought: this new agentic search framework slashes reasoning steps by 70% while boosting accuracy by prioritizing parallel evidence gathering.