Recent Large Reasoning Models (LRMs) achieve strong performance by leveraging long-form Chain-of-Thought (CoT) reasoning, but uniformly applying overlong reasoning at inference time incurs substantial and often unnecessary computational cost. To address this, prior work explores various strategies to infer an appropriate reasoning budget from the input. However, such approaches are unreliable in the worst case, as estimating the minimal required reasoning effort is fundamentally difficult, and they implicitly fix the trade-off between reasoning cost and accuracy during training, limiting flexibility under varying deployment scenarios. Motivated by these limitations, we propose ORBIT, a controllable multi-budget reasoning framework with well-separated reasoning modes triggered by input. ORBIT employs multi-stage reinforcement learning to discover Pareto-optimal reasoning behaviors at each effort, followed by on-policy distillation to fuse these behaviors into a single unified model. Experiments show that ORBIT achieves (1) controllable reasoning behavior over multiple modes, (2) competitive reasoning density within each mode, and (3) integration of these frontier policies into a single unified student model while preserving clear mode separation and high per-mode performance.

On-policy distillation (OPD) transfers reasoning behavior by training a student on teacher feedback along student-generated trajectories, but standard full-rollout training ties every update to a costly completion and can over-allocate supervision to late positions with low marginal value for the current student. We revisit this assumption through the useful supervision horizon: student-induced rollouts can drift from teacher-preferred continuations, while aligned prefixes may already preserve the long-horizon OPD update direction. We propose ADWIN, an adaptive-window framework for OPD that treats rollout length as an online admissibility decision, training on short teacher-anchored prefixes while using delayed full-rollout probes to audit prefix–full alignment and adapt the next horizon with staleness control. Across math and code reasoning benchmarks in single-task, multi-task, and strong-to-weak settings, ADWIN improves the accuracy–compute trade-off over full-rollout OPD and prefix-based baselines, reducing end-to-end training cost by up to 4.1× while achieving comparable or better accuracy.

Recent advances in cross-prompt automated essay scoring typically train models jointly on all available source domains, often requiring simultaneous access to unlabeled target domain samples. However, using all sources can lead to suboptimal transfer and high computational cost. Moreover, repeatedly accessing the source essays for continual adaptation raises privacy concerns. We propose a source-free adaptation approach that selectively merges the parameters of individually trained source models without further access to the source datasets. In particular, we mix the task vectors—the parameter updates from fine-tuning—via a weighted sum to efficiently simulate selective joint-training. We use Bayesian optimization to determine the mixing weights using our proposed Prior-encoded Information Maximization (PIM), an unsupervised objective which promotes score discriminability by leveraging useful priors pre-computed from the sources. Experimental results with LLMs on in-dataset and cross-dataset adaptation show that our method (1) consistently outperforms joint-training on all sources, (2) maintains superior robustness compared to other merging methods, (3) excels under severe distribution shifts where recent leading cross-prompt methods struggle, all while retaining computational efficiency.