Guan Zhe Hong, Sumeet Ramesh Motwani, Liang Chen, Yangchen Pan, Junchi Yu, Jialin Yu, Philip Torr, Yi Ren
On-policy distillation (OPD) is emerging as an important paradigm for language-model post-training, combining dense teacher supervision with training on student-generated trajectories. This creates a feedback loop where each update changes the student and thereby the training distribution that generates subsequent updates. Yet, how this feedback shapes learning and generalization remains poorly understood. We conduct a mechanistic analysis of OPD's learning dynamics, distinguishing the two feedback loop stages: conversion of the distillation signal into parameter updates, and broadcast through which these updates change predictions across prefixes and reshape the rollout distribution. This framework helps us understand both limited student responsiveness despite substantial teacher-student disagreement, and degeneration when updates that improve local teacher agreement also increase the probability of unsuccessful continuations, exposing the student to unreliable supervision in subsequent rollouts. Finally, we show that offline distillation can prepare a student for OPD without significantly improving immediate held-out accuracy. While OPD alone may degrade generation, it can yield substantial accuracy gains after sufficient offline preparation. Together, our results connect teacher supervision, student geometry, and the evolving rollout distribution in determining the learning dynamics of OPD and its effects on downstream performance.
Sean Wu, Aneesh Pappu, Jiaru Zou, Yubin Kim, Yuncong Mao, Hongjian Zhou, Junchi Yu, Jieneng Chen 等 9 位作者
Progress in AI is often measured by what a single model can do on its own. Yet teams of AI agents are starting to make breakthroughs and discoveries in science, medicine, mathematics and software engineering. As AI begins to work in teams of agents and people, the ability to collaborate becomes a core part of their intelligence. Research on human teams has measured this ability for decades. Inspired by this, we introduce the Collaborative Intelligence Benchmark, which measures how well an AI agent works as a teammate. It contains 240 tasks across 8 dimensions of teamwork, from refusing an unsafe shortcut to checking the work that a weaker model leaves behind. Each task comes from a real record or an existing benchmark, from aviation accident reports and clinical cases to coding tasks and negotiations, and has a programmatically verifiable outcome. To isolate teamwork from individual capability, models are scored only on tasks that they can solve by themselves when they have full information. We evaluate 18 closed- and open-weight models and find that every model, across sizes and families, has clear weaknesses as a teammate, with the best model scoring only 82.2 out of 100. Frontier models almost always refuse unsafe shortcuts and report misconduct, yet they cave under interpersonal pressure, reaching the best overall deal in only 14% of conflicts, and their pass rate drops from 97% to 81% when they take over a task in which a weaker model has cheated. Crucially, general capability does not guarantee good teamwork, as models with nearly the same capability score differ by up to 11 points as teammates. The benchmark includes 160 public tasks with code and data, along with 80 held-out tasks for evaluation. As AI agents increasingly work alongside other agents and people, the ability to collaborate effectively is becoming a crucial capability to measure and develop.
LLM agents increasingly support research planning and execution, where judgments about current and anticipated research activity inform decisions. Yet whether they can infer changes across a whole field from the partial evidence they retrieve through online scholarly search remains unclear. We introduce Research Attention Prediction (RAP), a literature search benchmark for evaluating this ability. We build a local search environment from 356,357 dated arXiv papers and construct two complementary evaluation tasks. The benchmark comprises a Core set of 2,160 questions across 54 AI/ML fields and an Extended set covering 269 fields. Our evaluation and analysis of LLM agents show that search substantially improves stronger agents' recognition of recent changes, while forecasting gains remain limited or negative. Agents update forecasts across dates and adjust their investigations to forecasting goals, but these adaptations do not reliably yield accurate predictions of future shifts. Even for recent changes, final judgments can diverge from what the retrieved papers indicate. These findings show that agents can recognise some recent changes, yet reliably anticipating future research activity remains difficult despite changes in their answers and investigations.