TL;DR: When a student trajectory drifts too far from the teacher, on-policy distillation becomes unstable. DTDD detects the drift via segment-level divergence and hands off to teacher recovery. Across long-reasoning math, base-model distillation, and ALFWorld agents, it cuts the peak gradient norm from 12.0 to 1.4, reaches OPD-level accuracy in about 2.5× fewer updates, and raises success from 76.5% to 81.3%.

TL;DR: Which experiences should an agent learn from? We select post-training data by the agent's own prediction gaps: for SFT, traces whose execution outcomes the agent mispredicts yield 12.8% higher mean reward than random selection; for RL, tasks whose rollout rewards it mispredicts improve final reward by 33.3%, and cheap predicted-reward proxies work without extra verification.

TL;DR: Multi-harness RL for coding agents both trains across several harnesses and compares their rewards inside one GRPO group. Replaying identical Qwen3-8B trajectories from Aider, OpenHands, Qwen Code, and SWE-agent under per-harness vs. cross-harness grouping, we find the evaluation harness moves SWE-bench Verified solve rate by 4.3×, while the grouping rule has no detectable effect on an unseen held-out harness.

TL;DR: VFIG is the first large-scale ecosystem for scientific figure-to-SVG generation, introducing a 66K figure–SVG dataset, a new benchmark, and a family of VLMs trained with SFT+RL that achieve state-of-the-art open-source performance and parity with GPT-5.2.

TL;DR: When a coding agent fails, should it attempt cheap recovery or escalate to a stronger model? ARCHER casts this as a routing problem over heterogeneous recovery actions trained on execution rollouts, with a conformal risk control layer that adapts to any budget without retraining—exceeding always-escalate solve rates at 35% of its recovery cost.

TL;DR: Large language models (LLMs) generate rationales that improve reasoning and interpretability, but existing evaluations using binary human or LLM preferences are limited and opaque. We introduce an attribute-based evaluation framework that defines key rationale qualities, explains human preferences, and enables more nuanced model comparisons through attribute-specific analysis.

TL;DR: This paper introduces a weighted controlled risk approach to causal mediation analysis, addressing scenarios where traditional fixed-level interventions violate the positivity assumption by targeting subpopulations with a high probability of attaining specific mediator levels.

TL;DR: We proposed a framework for using placebo samples in observational studies to detect and correct for unmeasured confounding bias. It develops identification assumptions and estimation methods—including regression, weighting, and doubly robust approaches—and validates them through simulations and an applied case study on tax credits and infant health.

TL;DR: We developed a potential outcomes framework to estimate the ITT effect of an active control versus placebo in active-controlled trials, using historical placebo-controlled data. Our method enables ITT estimation when the placebo arm is unavailable and accounts for unmeasured confounders using instrumental variables.

TL;DR: We introduced a new criterion for individualized treatment rules based on the adjusted probability of longer survival, offering a clear and clinically relevant objective. Our method, optimal adjusted probability learning, constructs the best treatment rule by maximizing this nonparametric survival benefit.