Find the right supervision interface.

The authors argue that credit allocation within aligned chunks and anchoring unaligned tokens are both necessary, reporting gains across mathematics, science, and code benchmarks.

In the tested pairs, strict-position top-16 supervision matches full shared-vocabulary OPD, while added mismatch-span MSE hurts accuracy. The paper prioritizes supervision reliability over broader coverage.

The paper argues that supervising the combined probability of valid event completions complements token alignment, improving mathematics, code, and science without extra rollouts or vocabulary changes.

The authors propose adapting the teacher to the student's vocabulary and token positions, reporting full-dimensional supervision and gains across six reasoning tasks.

The paper argues that cross-family offsets obscure useful teacher improvements; transferring within-family likelihood shifts improves reasoning in the tested student families.

The paper proposes high-probability anchors and residual regularization to reduce cross-tokenizer alignment noise, reporting reasoning gains in single-teacher and multi-teacher off-policy distillation.

The authors report compact financial agents with reliable reasoning and long-horizon execution from a combined SFT, critical-step OPD, RLVR, and model-merging pipeline.

The authors report that exact text-span alignment and student-reference constraints enable stable transfer of long-context teachers' reasoning capabilities to shorter-context students.

The paper proposes a byte-aligned target over the student vocabulary that preserves teacher probability mass, reporting gains on mathematics and programming benchmarks.

The paper presents a shared verl backend with separate method modules, showing that different OPD settings can retain their objectives within one execution framework.

The paper proposes transmitting token-level OPD signals across model families through decoded-text mapping, reporting better compute efficiency than the evaluated baselines.

The authors claim that hidden-state supervision avoids output-head limitations, and report cross-architecture and cross-tokenizer transfer through frozen low-rank projectors.

The paper argues that semantic verification of response chunks supplies reliable OPD feedback without teacher logits, including feedback from proprietary text-only teachers.

The authors report that restricting rollouts to early response tokens can outperform full-rollout OPD while improving GPU efficiency and training stability.

The paper argues that short jointly tokenizable continuations recover useful supervision lost by exact token matching, improving mathematics and code results in the tested heterogeneous pairs.

The authors argue that prompt-specific rubrics can replace teacher logits for OPD, reporting improved performance and sample efficiency in the evaluated settings.

The authors argue that selectively allocating teacher feedback reduces negative transfer and entropy collapse, and report improved sample efficiency, including cross-vocabulary experiments.

The authors propose separate inference and training backends for teacher and student, reporting faster distributed distillation while supporting on-policy, off-policy, and cross-tokenizer methods.

The authors propose retaining strong student trajectories and replacing weaker segments using verbal teacher feedback, reporting reasoning and efficiency gains with text-only teachers.

The paper proposes learning an adaptive discriminator from teacher text alone, reporting stronger black-box distillation than sequence-level imitation.

The official documentation describes cross-tokenizer distillation using decoded-text alignment, probability merging, and hybrid ULD loss within TRL's experimental trainer.

The paper argues that one draft model can adapt online to target models with different vocabularies, reporting faster speculative decoding through hybrid distillation and adaptive drafting.

The authors propose shared prediction heads and dual-space projection to distill across vocabularies, reporting improvements on instruction following, mathematics, and code generation.