# Awesome Cross-Tokenizer On-Policy Distillation
> A curated, evidence-backed research map of on-policy distillation across different tokenizers and vocabularies.
[](#scope-and-selection-policy)
[](LICENSE)
[](CONTRIBUTING.md)
[English](README.md) · [日本語](README_ja.md) · [简体中文](README_zh.md) · [한국어](README_ko.md)
[Interactive research catalogue](https://usk-kimura.github.io/awesome-cross-tokenizer-opd/) · [Code availability](#code-availability) · [Machine-readable data](data/papers.yml) · [BibTeX](references.bib)
On-policy distillation (OPD) trains a student with teacher feedback on the
student's own generated trajectories. Standard token-level OPD assumes that
teacher and student distributions share coordinates. That assumption breaks
when their tokenizers or vocabularies differ: token IDs, token boundaries, and
sometimes even the available byte strings no longer correspond.
This list focuses on the interfaces proposed to cross that boundary. It is not
another general OPD bibliography. Every entry is labeled by trajectory source,
comparison interface, supervision coverage, teacher access, probability
semantics, tested model pair, and code availability.
## Scope and selection policy
The snapshot date is **2026-10-14**. “Verified” below means that the stated
claim was checked against public full text; it does **not** mean that we
independently reproduced the experiments.
Each structured record has a `last_verified` date for its full-text claims.
Version/link checks (`metadata_last_verified`) and implementation lookups
(`code_last_checked`) are dated separately. Unchanged earlier full-text audits
retain their original dates; the snapshot date does not imply a new reproduction
or full-text audit of every entry.
A work is included in the **core** map when it has:
1. a public English paper or preprint;
2. teacher supervision on student-generated rollouts or student-visited states;
3. an explicit teacher–student tokenizer/vocabulary mismatch;
4. a concrete interface for passing supervision across that mismatch; and
5. a reported training experiment with a mismatched pair.
We separate papers that only report a cross-vocabulary experiment without
documenting the interface, boundary applications such as speculative decoding,
software/resources, and mostly off-policy enabling work. A different model
family is **not** by itself evidence of a different tokenizer.
Throughout this README, model pairs use **Teacher → Student** order, even when a
paper uses the opposite convention.
## Research map
```mermaid
flowchart TD
A[Teacher feedback on student-generated states] --> B{How is the mismatch handled?}
B --> C[Output-space interface]
B --> D[Vocabulary-bypassing interface]
B --> E[Reported cross-vocabulary result
evidence caveat]
B --> F[Boundary application]
C --> C1[Dual space: DSKDv2]
C --> C2[Candidate continuations: SimCT]
C --> C3[Aligned realized tokens: ESR]
C --> C4[Realized synchronized path: BTB]
C --> C5[Full byte-prefix marginal: BPM]
C --> C6[Strict realized-token subset: SimpleOPD]
C --> C7[Within-family likelihood shift: CompassOPD]
C --> C8[Learned student-vocabulary head: DVLM]
C --> C9[Completion-set probability: ESCD]
C --> C10[Strict shared-vocabulary subset: Rethinking OPD]
C --> C11[Chunk credit and selective anchoring: CAST]
C --> C12[Applied critical-turn pipeline: Mint-Agent]
D --> D1[Adversarial text feedback: GAD]
D --> D2[Verbal rubrics: ROPD]
D --> D3[Semantic sampling: OmniOPD]
E --> E1[REOPOLD: interface under-specified]
E --> E2[OPRD: paper/code inconsistency]
F --> F1[Mixed policy: OVD]
F --> F2[Speculative decoding: OmniDraft]
```
### Start here
Three papers expose the central output-space trade-off especially clearly:
- **[SimCT](https://arxiv.org/abs/2605.07711)** constructs a finite set of
jointly realizable continuations, recovering more local alternatives than
exact shared-token matching.
- **[BTB](https://arxiv.org/abs/2606.09456)** synchronizes the realized teacher
and student token paths and transfers a scalar path-level signal.
- **[BPM](https://arxiv.org/abs/2607.22334)** routes the full teacher next-token
mass into student byte-prefix events plus a residual category.
They should not be read as interchangeable “token alignment” methods: their
comparison events and the amount of teacher distribution they expose are
different.
Read **[Rethinking Cross-Tokenizer OPD](https://arxiv.org/abs/2610.08448)**
alongside these methods. It tests whether broader coverage helps learning,
comparing strict shared-vocabulary supervision with compact top-16 support and
additional span supervision. **[ESCD](https://arxiv.org/abs/2609.34738)** studies
completion events after a teacher token is partially realized, while
**[CAST](https://arxiv.org/abs/2610.13257)** studies where to allocate chunk credit.
They address supervision quality and granularity as well as alignment coverage.
## Core methods
### Output-space interfaces
| Work | Public status | Cross-tokenizer interface | Feedback coverage | Teacher access | Probability semantics | Representative verified pair · code |
|---|---|---|---|---|---|---|
| [DSKD / DSKDv2](https://arxiv.org/abs/2504.11426) (2025) | arXiv v1; under review | Project teacher/student hidden states into the other's representation space, reuse the corresponding LM head, and retain exactly aligned positions with ETA | Dense distributions at retained positions, in both model output spaces | Hidden states and LM heads | Comparable head-space distributions; not a string-probability transport | Qwen2-1.5B → GPT2-120M; Mistral-7B → TinyLLaMA-1.1B · [code](https://github.com/songmzhang/DSKDv2) |
| [SimCT](https://arxiv.org/abs/2605.07711) (2026) | arXiv v2 | Shared surface tokens plus minimal multi-token continuations realizable by both tokenizers | A finite, locally constructed candidate set | White-box next-token scores | Candidate-normalized continuation scores; not a full teacher-induced marginal | Qwen2.5-7B-Instruct → Phi-4-mini-Instruct / Gemma-2-2B-IT; Phi-4-mini-Instruct → Gemma-2-2B-IT · [code](https://github.com/sunjie279/SimCT-) |
| [ESR](https://arxiv.org/abs/2605.27028) (2026) | arXiv v1 | Decode the early student rollout, retokenize it for the teacher, then greedily match text spans and apply reverse KL on aligned tokens | Greedily aligned tokens in an early rollout window | White-box token log-probabilities | Partial aligned-token objective; no full mapped distribution is characterized | Qwen3-4B → Gemma-2-2B (cross-family), plus cross-generation Qwen/Gemma pairs · code not found |
| [BTB](https://arxiv.org/abs/2606.09456) (2026) | arXiv v1 | Dual pointers form minimal decoded-string synchronization chunks; teacher/student canonical-path likelihoods provide chunk credit | The realized student path | Per-token teacher log-probabilities on the realized text | Path likelihood/credit, not a counterfactual full-vocabulary target | Qwen3-8B → Llama-3.1-8B-SFT; DeepSeek-R1 → DeepSeek-R1-Distill-Qwen-7B · [code](https://github.com/ivanniu/On-Policy-Distill) |
| [BPM](https://arxiv.org/abs/2607.22334) (2026) | arXiv v1 | Route each teacher token to the longest student-token byte prefix; aggregate collisions and retain unmatched mass in a residual category | Full student vocabulary plus residual | Full teacher next-token distribution | Mass-preserving; exact byte-prefix marginal at >99% of reported positions, chain-factorized lower bound otherwise | Qwen3-32B / GLM-Z1-9B-0414 / MiniMax-M2.7 → Qwen3.5-2B · [code](https://github.com/K1XE/BPM), released 2026-09-01 |
| [SimpleOPD](https://arxiv.org/abs/2608.14277) (2026) | arXiv v1 | Retokenize student text and match tokens with identical start/end offsets | Realized tokens at strict 1:1 spans | Sampled-token teacher log-probabilities | Partial reverse-KL surrogate; unmatched positions have zero distillation advantage | SU-01 → GLM-4.7-Flash / Gemma-4-26B-A4B; Qwen3.5 students · [code](https://github.com/hhnqqq/SimpleOPD) |
| [CompassOPD](https://arxiv.org/abs/2609.10154) (2026) | arXiv v1 | Align text spans and transfer strong-teacher minus teacher-family-reference likelihood shifts, anchored to the initial student | Realized aligned units | Teacher and teacher-family-reference log-probabilities | Span-averaged shifts, not a transported teacher distribution | Qwen3.5-35B-A3B → Granite4.1-3B / Qwen3-4B / OLMo-3-7B-Instruct-SFT · code not found |
| [DVLM](https://arxiv.org/abs/2609.33816) (2026) | arXiv v1 | Train a student-vocabulary teacher head with parallel-tokenized input and Hybrid-Prefix Attention | Dense logits at student positions | Teacher backbone, hidden states, replacement head | Learned adapted-teacher distribution; not exact original-teacher string marginalization | Qwen3-4B → Llama-3.2-1B-Instruct / OLMo-2-1B-Instruct · code not found |
| [ESCD](https://arxiv.org/abs/2609.34738) (2026) | arXiv v1 | Aggregate prefix-related events; add completion-set supervision at visited child states to byte-aligned root supervision | Total probability of compatible one-step completions | Teacher event probabilities; full vocabulary in experiments | Set-probability loss leaves within-set allocation unconstrained | Qwen3-32B / GLM-Z1-9B → Qwen3.5-2B; Kimi-K2.7-Code → Qwen3.6-35B-A3B · code announced, not located |
| [Rethinking Cross-Tokenizer OPD](https://arxiv.org/abs/2610.08448) (2026) | arXiv v1 | Strict 1:1 span alignment; reverse KL on shared vocabulary or student-selected top-k candidates | Strict positions; top-16 tested | Teacher next-token scores on shared candidates | Subset-renormalized distributions, not full-vocabulary transport | Qwen2.5-7B-Instruct → Llama-3.2-3B-Instruct; Granite-4.1-8B → Phi-4-mini-instruct / Qwen2.5-7B-Base · [paper-linked code](https://anonymous.4open.science/r/Cross-Tokenizer-OPD), access unverified (HTTP 401) |
| [CAST](https://arxiv.org/abs/2610.13257) (2026) | arXiv v1 | Decoded-text alignment, uncertainty/smoothness-weighted chunk credit, and selective initial-policy anchoring | Realized aligned tokens; anchoring at unaligned positions | Teacher log-probabilities on retokenized text | Preserves chunk advantage, not vocabulary probability mass | Qwen3-8B → Mistral-7B / Llama-3.1-8B (SFT initialized) · code not found |
### Coverage and reliability
[Rethinking Cross-Tokenizer OPD, Sections 3–5](https://arxiv.org/html/2610.08448v1)
reports that strict alignment covers 85.57–96.98% of student tokens over the
studied training runs. Student-selected top-16 supervision retains at least 96%
of the improvement from the full shared-vocabulary objective. Adding span
log-probability MSE lowers accuracy in all 18 positive-weight settings tested.
This result concerns those objectives and model pairs; it does not establish
that every form of mismatch supervision is harmful.
[ESCD, Section 3 and Appendix B.4](https://arxiv.org/html/2609.34738v1)
supervises a residual byte constraint through the sum of valid completion-token
probabilities. Tokens may extend beyond the residual boundary, and empty
one-step completion sets receive no child loss. [CAST, Section 3](https://arxiv.org/html/2610.13257v1)
instead allocates realized chunk advantage and constrains unaligned-token drift.
These mechanisms should be compared using their actual events and initialization
regimes, rather than a universal coverage ranking.
### Applied cross-tokenizer OPD
[Mint-Agent](https://arxiv.org/abs/2608.16386) (arXiv v2) documents critical-turn
distillation from DeepSeek-V4-Pro into Qwen3.5-9B and Qwen3.6-27B agentic
specialists. The current student resamples actions at its own failure states;
text-equivalent chunks redistribute teacher path scores to the student tokens.
See [Section 5.2](https://arxiv.org/html/2608.16386v2) for the interface and
Section 6 for evaluation. It meets the core criteria as an applied pipeline,
but final Mint-Cu/Mint-Ag results combine SFT, OPD, RLVR, and merging. Its later
multi-teacher integration uses same-tokenizer specialists and does not isolate
the benefit of the cross-tokenizer stage. [Project](https://mint-fin.github.io/mint-agent/);
no official training implementation located at this snapshot.
### Vocabulary-bypassing interfaces
These methods avoid directly mapping teacher token probabilities into the
student vocabulary.
| Work | Public status | Interface | Feedback coverage | Teacher access | Representative verified pair · code |
|---|---|---|---|---|---|
| [GAD](https://arxiv.org/abs/2511.10643) (2025) | arXiv v3 | A co-evolving discriminator distinguishes fixed teacher responses from current student responses and supplies the student's RL reward | Sequence-level score on current student rollouts | Pre-collected teacher text only | GPT-5-Chat → Qwen2.5-3B/7B/14B-Instruct or Llama-3.x-Instruct · [code](https://github.com/microsoft/LMOps/tree/main/gad) |
| [ROPD](https://arxiv.org/abs/2605.07396) (2026) | arXiv v1 | A rubricator contrasts teacher and current student answers to induce prompt-specific textual criteria; a verifier scores student rollouts against them | Weighted rubric-item pass rate per rollout | Text-only teacher/rubricator/verifier | GPT-5.2-Chat → Qwen3-4B / Gemma-3-4B-IT; Qwen3-30B-A3B → Qwen3-4B · [code](https://github.com/Peregrine123/ROPD_official) |
| [OmniOPD](https://arxiv.org/abs/2606.01476) (2026) | arXiv v2 | Monte Carlo teacher continuations scored by semantic similarity over decoded chunks | Selected high-entropy chunks on student trajectories | Text generation only; supports black-box/API teachers | Qwen3-32B, Claude-4.5-Haiku, or Gemini-2.5-Flash supervising student reasoning · code not found |
### Reported cross-vocabulary results with evidence caveats
| Work | What the paper reports | Why it is not in the core interface table |
|---|---|---|
| [REOPOLD](https://openreview.net/forum?id=NPhEUvT1jH) ([arXiv](https://arxiv.org/abs/2603.11137), 2026) | The previously audited OpenReview version reports Phi-4-Mini-Reasoning → DeepSeek-R1-Distill-Qwen-1.5B as different families and vocabularies. It applies selective token credit on student rollouts. | The audited description does not specify equivalent sampled-token events across vocabularies. The [latest author publication page](https://www.microsoft.com/en-us/research/publication/relaxed-on-policy-distillation-selective-credit-allocation-for-scaling-reasoning-efficiently/) lists **NeurIPS 2026**, correcting our earlier ICML attribution, and links a [different OpenReview record](https://openreview.net/forum?id=l0h1SzRI7H). Browser verification blocked retrieval of the latest full text; the earlier interface caveat has not been re-audited against that version. |
| [OPRD / OPRD-Bridge](https://arxiv.org/abs/2606.06021) (2026) | Frozen low-rank projectors align teacher/student hidden states. The paper labels Phi-4-mini-reasoning → Qwen3-1.7B-Base as “completely disjoint tokenizers.” | At the audited public commit, the code says both models receive the [student tokenizer's input IDs](https://github.com/ShenzhiYang2000/OPRD/blob/a63236c990b1bfa9551e9d020719490d3ecd93e8/scripts/analysis/cross_arch_repr_analysis.py#L5-L8); projector training likewise [loads the student tokenizer](https://github.com/ShenzhiYang2000/OPRD/blob/a63236c990b1bfa9551e9d020719490d3ecd93e8/scripts/analysis/cross_arch_preexp2_train_ps.py#L1564-L1567) and [passes the same IDs to both models](https://github.com/ShenzhiYang2000/OPRD/blob/a63236c990b1bfa9551e9d020719490d3ecd93e8/scripts/analysis/cross_arch_preexp2_train_ps.py#L888-L890). Token IDs from disjoint vocabularies do not denote the same text, so the released implementation does not verify semantic cross-tokenizer alignment. The representation-space idea remains relevant, but the cross-tokenizer result needs clarification or corrected code. |
### Boundary application
| Work | Connection to OPD | Boundary |
|---|---|---|
| [OVD](https://arxiv.org/abs/2601.21968) (arXiv v2) · [code](https://github.com/menik1126/OVD-code) | Verbal scores retain student sub-trajectories or trigger teacher continuations; v2 details query/suffix replacement. | Outcome-based updates train on both retained student tokens and teacher replacements. Some evaluation settings also invoke the teacher at test time. |
| [OmniDraft](https://arxiv.org/abs/2507.02659) (NeurIPS 2025) | An online adaptive cross-vocabulary drafter learns from its own proposals using direct-mapping reverse KL and accepted n-gram supervision. | The target model verifies speculative proposals and determines the final decoding distribution. This is not canonical full-response student-policy OPD, so we keep it outside the core method table. |
## Implementations and research resources
| Resource | What it contributes | Status at snapshot |
|---|---|---|
| [GOLD: Unlocking On-Policy Distillation for Any Model Family](https://huggingface.co/spaces/HuggingFaceH4/on-policy-distillation) · [TRL docs](https://huggingface.co/docs/trl/main/gold_trainer) | Decoded-span alignment and probability merging, exact shared-token matching, and sorted-probability ULD fallback | Non-paper resource in `trl.experimental.gold.GOLDTrainer`; full completion coverage does not imply identity-preserving probability transport |
| [KDFlow](https://arxiv.org/abs/2603.01875) · [code](https://github.com/songmzhang/KDFlow) | A distributed KD framework supporting off-policy, on-policy, and DSKD-based cross-tokenizer paths | arXiv v4; Qwen3-14B → Llama3.2-3B is an **off-policy** experiment (Section 4.6 / Appendix A.1), separate from the on-policy runs |
| [EasyOPD](https://arxiv.org/abs/2607.11012) · [code](https://github.com/lds-ustc/EasyOPD) | A `verl`-based framework with runnable modules for SimCT and other tokenizer-alignment baselines | arXiv v1; framework/resource paper |
## Adjacent cross-tokenizer KD
The following papers provide alignment or distribution-transfer machinery, but
their primary setup is off-policy/general KD, preference learning, or scoring.
They are useful foundations, not automatically evidence for cross-tokenizer OPD.
| Family | Work |
|---|---|
| Distribution reordering or transport | [Universal Logit Distillation](https://openreview.net/forum?id=bwRxXiGO9A), [Multi-Level Optimal Transport](https://ojs.aaai.org/index.php/AAAI/article/view/34543), [Approximate Likelihood Matching](https://openreview.net/forum?id=DxKP2E0xK2) |
| Shared/latent output spaces | [Dual-Space Knowledge Distillation](https://aclanthology.org/2024.emnlp-main.1010/), [Dual-Space KD with Key–Query Matching](https://doi.org/10.1109/ICASSP55912.2026.11464298), [X-Token](https://arxiv.org/abs/2605.21699) |
| Token/span/string alignment | [FuseLLM](https://openreview.net/forum?id=jiDsk12qcz), [Contextual Dynamical Mapping](https://aclanthology.org/2025.findings-acl.419/), [Span Representation Alignment](https://arxiv.org/abs/2605.01205), [DWA-KD](https://arxiv.org/abs/2602.21669) |
| Other objectives/interfaces | [Cross Tokenizer Preference Distillation](https://arxiv.org/abs/2601.11865), [Cross-Tokenizer LLM Distillation through a Byte-Level Interface](https://aclanthology.org/2026.customnlp4u-1.9/) |
| Anchor supervision and residual regularization | [ACTD](https://arxiv.org/abs/2608.29662) · [code](https://github.com/hyhy892/ACTD); Section 4.1 explicitly trains on dataset reference solutions with cached teacher logits |
### String-probability and tokenization foundations
- [Marginal Likelihood over Tokenisations](https://aclanthology.org/2021.emnlp-main.161/)
- [Should You Marginalize over Possible Tokenizations?](https://aclanthology.org/2023.acl-short.1/)
- [Language Models over Canonical Byte-Pair Encodings](https://proceedings.mlr.press/v267/vieira25b.html)
- [Exact Byte-Level Probabilities from Tokenized Language Models](https://openreview.net/forum?id=zGej22CBnS)
- [Cross-Tokenizer Likelihood Scoring Algorithms for Language Model Distillation](https://openreview.net/forum?id=hD69qj15Os)
## How to interpret claims
- **Student rollout is checked directly.** A method that can run on either
teacher or student text is not labeled OPD unless the reported setup uses
student-generated states.
- **Cross-family is not cross-tokenizer.** For example, a student from a named
“DeepSeek distill” family may still inherit Qwen's tokenizer.
- **Support matters.** Full-vocabulary, finite-candidate, aligned-subset, and
realized-path objectives supervise different events; a shared “KL” label does
not make them equivalent.
- **Exactness is claim-specific.** “Exact alignment” of retained positions,
mass preservation, and exact marginal string probability are different
properties.
- **Code status is time-sensitive.** “Not found” only means no official public
implementation was located by the snapshot date.
Examples of scope traps include [Revisiting On-Policy Distillation](https://arxiv.org/abs/2603.25562),
whose reported cross-model reasoning setup does not establish a tokenizer
bridge, and [Concrete Score Distillation](https://arxiv.org/abs/2509.25837),
whose formulation assumes a shared vocabulary. Mentioning tokenizer mismatch in
motivation is not sufficient for inclusion.
## Open research questions
1. **Probability fidelity vs. tractability.** Can a sparse interface preserve
the decisions of a full teacher-induced string distribution without costly
enumeration over tokenization paths?
2. **A common evaluation protocol.** Results currently mix model pairs, rollout
lengths, prompt sources, tasks, and teacher-access regimes. A fixed suite of
same-family/cross-family and high/low boundary-mismatch pairs is still needed.
3. **Special-token semantics.** EOS, BOS, chat templates, whitespace, Unicode,
invalid bytes, and code indentation need explicit treatment and reporting.
4. **Teacher-access tiers.** Full logits, top-k logits, sampled-token
log-probabilities, hidden states, and text-only APIs should be benchmarked as
distinct information budgets.
5. **Coverage diagnostics.** Papers should report how much teacher mass, how many
rollout positions, and which mismatch cases actually receive supervision.
Report gradient agreement and downstream ablations as well: the Rethinking
study shows that larger structural coverage can introduce conflicting signals.
6. **Reproducibility.** Several reported cross-vocabulary results have no code or
omit the exact mapping procedure.
## Related surveys and lists
This repository deliberately complements broader collections:
- [AwesomeOPD](https://github.com/thinkwee/AwesomeOPD)
- [Awesome On-Policy Distillation](https://github.com/chrisliu298/awesome-on-policy-distillation)
- [Awesome LLM On-Policy Distillation](https://github.com/nick7nlp/Awesome-LLM-On-Policy-Distillation)
- [OPDHub](https://github.com/nick7nlp/OPDHub)
- [Awesome Knowledge Distillation of LLMs](https://github.com/Tebmer/Awesome-Knowledge-Distillation-of-LLMs)
- [A Survey of On-Policy Distillation for Large Language Models](https://arxiv.org/abs/2604.00626)
- [A Formula-Driven Survey and Research Agenda for On-Policy Distillation](https://arxiv.org/abs/2606.22793)
## Official code availability
Status is based on official-source lookups, not independent reproduction. “Not located” does not mean no code exists. An announced release and an inaccessible code link are distinct from a verified public implementation. OPRD has public code with an implementation concern described below.
| Work | Scope | Official implementation | Last checked |
|---|---|---|---|
| [DSKD / DSKDv2](https://arxiv.org/abs/2504.11426) | Output-space interfaces | [Official code available](https://github.com/songmzhang/DSKDv2) | 2026-10-14 |
| [SimCT](https://arxiv.org/abs/2605.07711) | Output-space interfaces | [Official code available](https://github.com/sunjie279/SimCT-) | 2026-10-14 |
| [ESR](https://arxiv.org/abs/2605.27028) | Output-space interfaces | Official code not located | 2026-10-14 |
| [GAD](https://arxiv.org/abs/2511.10643) | Vocabulary-bypassing | [Official code available](https://github.com/microsoft/LMOps/tree/main/gad) | 2026-10-14 |
| [ROPD](https://arxiv.org/abs/2605.07396) | Vocabulary-bypassing | [Official code available](https://github.com/Peregrine123/ROPD_official) | 2026-10-14 |
| [OmniOPD](https://arxiv.org/abs/2606.01476) | Vocabulary-bypassing | Official code not located | 2026-10-14 |
| [OPRD / OPRD-Bridge](https://arxiv.org/abs/2606.06021) | Implementation concern | [Official code available](https://github.com/ShenzhiYang2000/OPRD) | 2026-10-14 |
| [BTB](https://arxiv.org/abs/2606.09456) | Output-space interfaces | [Official code available](https://github.com/ivanniu/On-Policy-Distill) | 2026-10-14 |
| [REOPOLD](https://openreview.net/forum?id=NPhEUvT1jH) | Interface under-specified | Official code not located | 2026-10-14 |
| [BPM](https://arxiv.org/abs/2607.22334) | Output-space interfaces | [Official code available](https://github.com/K1XE/BPM) | 2026-10-14 |
| [OVD](https://arxiv.org/abs/2601.21968) | Boundary applications | [Official code available](https://github.com/menik1126/OVD-code) | 2026-10-14 |
| [OmniDraft](https://arxiv.org/abs/2507.02659) | Boundary applications | Official code not located | 2026-10-14 |
| [KDFlow](https://arxiv.org/abs/2603.01875) | Frameworks | [Official code available](https://github.com/songmzhang/KDFlow) | 2026-10-14 |
| [EasyOPD](https://arxiv.org/abs/2607.11012) | Frameworks | [Official code available](https://github.com/lds-ustc/EasyOPD) | 2026-10-14 |
| [GOLD](https://huggingface.co/spaces/HuggingFaceH4/on-policy-distillation) | Implementation resources | [Official code available · experimental](https://github.com/huggingface/trl) | 2026-10-14 |
| [SimpleOPD](https://arxiv.org/abs/2608.14277) | Output-space interfaces | [Official code available](https://github.com/hhnqqq/SimpleOPD) | 2026-10-14 |
| [CompassOPD](https://arxiv.org/abs/2609.10154) | Output-space interfaces | Official code not located | 2026-10-14 |
| [DVLM](https://arxiv.org/abs/2609.33816) | Output-space interfaces | Official code not located | 2026-10-14 |
| [ESCD](https://arxiv.org/abs/2609.34738) | Output-space interfaces | Code announced · not verified | 2026-10-14 |
| [Rethinking Cross-Tokenizer OPD](https://arxiv.org/abs/2610.08448) | Output-space interfaces | [Code linked · access unverified](https://anonymous.4open.science/r/Cross-Tokenizer-OPD) | 2026-10-14 |
| [CAST](https://arxiv.org/abs/2610.13257) | Output-space interfaces | Official code not located | 2026-10-14 |
| [Mint-Agent](https://arxiv.org/abs/2608.16386) | Applied pipelines | Official code not located | 2026-10-14 |
| [ACTD](https://arxiv.org/abs/2608.29662) | Adjacent / off-policy KD | [Official code available](https://github.com/hyhy892/ACTD) | 2026-10-14 |
## Contributing
Corrections and missing papers are welcome. Please read
[CONTRIBUTING.md](CONTRIBUTING.md) and include primary-source evidence for both
the student-rollout criterion and the tokenizer/vocabulary mismatch. The
structured record is in [`data/papers.yml`](data/papers.yml).
## Citation and license
Use [CITATION.cff](CITATION.cff) to cite this living survey. Bibliographic
metadata is released under [CC0 1.0](LICENSE). Papers, figures, linked code, and
external pages remain under their respective licenses.