Tag
Researchers from Bridgewater AIA Labs, UIUC, and MIT prove the first non-vacuous generalization bounds for reasoning LLMs trained with RLVR, providing provable accuracy lower bounds on unseen data to guide safe deployment.
The paper introduces On-Policy Delta Distillation (OPD^2), a new distillation reward called the delta signal that captures the difference between a teacher model and its base model before reasoning tuning, providing a more direct signal for transferring reasoning capabilities. Experiments across math, science, and code benchmarks show OPD^2 consistently outperforms conventional on-policy distillation.
This paper investigates semantic context drift in reasoning-class LLMs within hybrid decision support systems, proposing a mathematical model and a stability metric. A two-month experiment reveals latent goal-targeting drift and formulates engineering recommendations for control stability.
PluraMath extends the PolyMath dataset to 18 underrepresented languages, providing a human-validated benchmark for evaluating multilingual mathematical reasoning in LLMs. The paper reveals a persistent performance gap between high-resource and low-resource languages across 27 models.
Introduces STATEWITNESS, an activation explainer for auditing deception in reasoning LLMs, achieving significant improvements over existing monitors and providing human-inspectable evidence.
Study reveals that answer tokens in thinking LLMs follow a structured self-reading pattern—forward drift plus focus on key anchors—during quantitative reasoning, and proposes a training-free SRQ steering method to exploit this for accuracy gains.