Xinyuan Song

Spectral inheritance allows for a dramatic reduction in training steps while enhancing model accuracy, challenging traditional optimization methods in RLVR.

Small changes in parameters can trigger a dramatic collapse in language agents' world models, revealing a critical phase transition that impacts their long-term decision-making.

Mid-planning environment queries can dramatically reduce world-model errors, transforming how agents interact with their surroundings.

Code generation models may excel at passing traditional benchmarks but falter dramatically when faced with novel algorithmic challenges, revealing a hidden weakness in their reasoning capabilities.

AlgoSkill redefines algorithm design by treating it as a skill scheduling problem, outperforming traditional LLM methods in complex programming tasks.

Sparsity, often viewed as a means for efficiency, actually unlocks deeper, more effective LLMs by taming variance and boosting layer utilization.