{ "schema_version": 1, "last_verified": "2026-10-14", "conventions": { "verification_dates": "Top-level last_verified dates the refresh; per-work last_verified dates full-text claims, metadata_last_verified dates version/link checks, and code_last_checked dates implementation lookup.", "pair_direction": "teacher_to_student", "verified_meaning": "Checked against public full text; not independently reproduced.", "core_criteria": [ "Public English paper or preprint", "Teacher supervision on student-generated rollouts or student-visited states", "Explicit tokenizer or vocabulary mismatch", "Concrete interface across the mismatch", "Reported training experiment with a mismatched pair" ] }, "classifications": { "core_output_space": "Core OPD method that constructs comparable output-space events or distributions.", "core_application": "Applied pipeline meeting the core criteria, without isolating a new tokenizer bridge.", "adjacent_off_policy": "Cross-tokenizer KD with dataset/reference trajectories rather than current student rollouts.", "core_vocabulary_bypassing": "Core OPD method that bypasses direct vocabulary-distribution mapping.", "evaluation_interface_under_specified": "Cross-vocabulary OPD result is reported, but the bridge is not documented sufficiently.", "paper_claim_implementation_concern": "The paper reports cross-tokenizer OPD, but the audited public implementation does not preserve text semantics across tokenizers.", "boundary_application": "Uses student/drafter-generated states in an adjacent application, but is not canonical full-response OPD.", "framework": "Paper describing software support rather than a new cross-tokenizer OPD interface.", "implementation_resource": "Non-paper implementation or research article." }, "works": [ { "id": "dskdv2", "record_kind": "paper", "classification": "core_output_space", "title": "A Dual-Space Framework for General Knowledge Distillation of Large Language Models", "authors": [ "Xue Zhang", "Songming Zhang", "Yunlong Liang", "Fandong Meng", "Yufeng Chen", "Jinan Xu", "Jie Zhou" ], "first_public_date": "2025-04-15", "checked_version": "arXiv v1", "publication": "arXiv preprint; marked under review", "urls": { "paper": "https://arxiv.org/abs/2504.11426", "code": "https://github.com/songmzhang/DSKDv2", "project": null }, "student_policy_states": "yes", "mismatch_evidence": "Different-vocabulary on-policy extension is explicit; cross-tokenizer experiments use ETA.", "interface": "Dual representation-space projectors, shared prediction heads, and Exact Token Alignment (ETA).", "feedback_coverage": "Dense distributions at exactly retained/aligned positions in student and teacher output spaces.", "teacher_access": "Hidden states, LM heads, and output distributions.", "probability_semantics": "Head-space distributions are comparable after projection; this is not string-event probability transport.", "evaluated_pairs": [ { "teacher": "Qwen2-1.5B", "student": "GPT2-120M", "mismatch": "different vocabularies" }, { "teacher": "Mistral-7B", "student": "TinyLLaMA-1.1B", "mismatch": "different vocabularies" } ], "code_status": "available", "evidence_locations": { "rollout": "Section III-C", "mismatch_interface": "Sections III-B and III-C", "experiment": "Section IV / Table II" }, "caveats": [ "ETA retains exact alignments rather than defining a full common string-event distribution at every position." ], "last_verified": "2026-08-15", "metadata_last_verified": "2026-10-14", "code_last_checked": "2026-10-14", "name": "DSKD / DSKDv2", "access_group": "hidden", "bibtex": "@article{zhang2025dualspace,\n title = {A Dual-Space Framework for General Knowledge Distillation of Large Language Models},\n author = {Zhang, Xue and Zhang, Songming and Liang, Yunlong and Meng, Fandong and Chen, Yufeng and Xu, Jinan and Zhou, Jie},\n journal = {arXiv preprint arXiv:2504.11426},\n year = {2025},\n eprint = {2504.11426},\n archiveprefix = {arXiv},\n primaryclass = {cs.CL},\n url = {https://arxiv.org/abs/2504.11426}\n}", "summary": { "en": "Dual representation-space projectors, shared prediction heads, and Exact Token Alignment (ETA).", "ja": "教師と生徒の隠れ状態を、それぞれ相手の表現空間に射影し、同じ出力ヘッドで分布を比較します。ETAで厳密に対応する位置を選んで学習させます。", "zh": "将隐藏状态投影到对方表示空间，复用对应 LM head；ETA 保留严格对齐位置。", "ko": "은닉 상태를 상대 표현 공간에 투영하고 해당 LM head 사용; ETA로 정확히 대응하는 위치 유지." }, "claim": { "en": "The authors propose shared prediction heads and dual-space projection to distill across vocabularies, reporting improvements on instruction following, mathematics, and code generation.", "ja": "教師と生徒で同じ出力ヘッドを使い、語彙が異なるモデル間でも蒸留できる方法を提案しています。指示追従・数学・コード生成の実験では、既存手法を上回ったと報告しています。", "zh": "作者提出通过共享预测头与双空间投影实现跨词表蒸馏，并报告指令遵循、数学与代码生成上的提升。", "ko": "저자들은 예측 헤드 공유와 이중 공간 투영으로 서로 다른 어휘 간 증류를 제안하며, 지시 수행·수학·코드 생성의 개선을 보고합니다." }, "claim_source": "https://arxiv.org/abs/2504.11426", "japanese": { "coverage": "厳密に対応した位置で、教師と生徒それぞれの出力分布を学習に使います。", "access": "教師の隠れ状態、出力ヘッド、出力分布が必要です。", "semantics": "共通の出力ヘッド空間で分布を比較します。文字列事象の確率を輸送する方法ではありません。" } }, { "id": "simct", "record_kind": "paper", "classification": "core_output_space", "title": "SimCT: Recovering Lost Supervision for Cross-Tokenizer On-Policy Distillation", "authors": [ "Jie Sun", "Mao Zheng", "Mingyang Song", "Qiyong Zhong", "Yilin Cheng", "Bichuan Feng", "Pengfei Liu", "Junfeng Fang", "Xiang Wang" ], "first_public_date": "2026-05-08", "checked_version": "arXiv v2", "publication": "arXiv preprint", "urls": { "paper": "https://arxiv.org/abs/2605.07711", "code": "https://github.com/sunjie279/SimCT-", "project": null }, "student_policy_states": "yes", "mismatch_evidence": "All three main teacher-student pairs use heterogeneous model families and tokenizers.", "interface": "Shared surface tokens plus minimal multi-token continuations that both tokenizers can realize.", "feedback_coverage": "Finite, locally constructed candidate set including mismatch units.", "teacher_access": "White-box next-token scores along candidate continuations.", "probability_semantics": "Continuation scores are normalized over the constructed candidate set, not the full teacher-induced string marginal.", "evaluated_pairs": [ { "teacher": "Qwen2.5-7B-Instruct", "student": "Phi-4-mini-Instruct", "mismatch": "different tokenizers" }, { "teacher": "Qwen2.5-7B-Instruct", "student": "Gemma-2-2B-IT", "mismatch": "different tokenizers" }, { "teacher": "Phi-4-mini-Instruct", "student": "Gemma-2-2B-IT", "mismatch": "different tokenizers" } ], "code_status": "available", "evidence_locations": { "rollout": "Section 3", "mismatch_interface": "Section 3", "experiment": "Section 4.1 / Table 1" }, "caveats": [ "Finite candidate support should not be interpreted as full-vocabulary mass preservation." ], "last_verified": "2026-08-15", "metadata_last_verified": "2026-10-14", "code_last_checked": "2026-10-14", "name": "SimCT", "access_group": "logits", "bibtex": "@article{sun2026simct,\n title = {{SimCT}: Recovering Lost Supervision for Cross-Tokenizer On-Policy Distillation},\n author = {Sun, Jie and Zheng, Mao and Song, Mingyang and Zhong, Qiyong and Cheng, Yilin and Feng, Bichuan and Liu, Pengfei and Fang, Junfeng and Wang, Xiang},\n journal = {arXiv preprint arXiv:2605.07711},\n year = {2026},\n eprint = {2605.07711},\n archiveprefix = {arXiv},\n primaryclass = {cs.CL},\n url = {https://arxiv.org/abs/2605.07711}\n}", "summary": { "en": "Shared surface tokens plus minimal multi-token continuations that both tokenizers can realize.", "ja": "両方のトークナイザーで表せる短い続きを候補に加えます。これにより、共通トークンだけでは利用できない学習信号も使えるようにします。", "zh": "共享表面 token 与双方 tokenizer 可实现的最短多 token continuation。", "ko": "공통 표면 token과 양쪽 tokenizer가 실현할 수 있는 최소 다중 token continuation." }, "claim": { "en": "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.", "ja": "両方のトークナイザーで表せる短い続きを使えば、完全一致だけでは失われる学習信号を利用できるとしています。評価した異種モデル対で、数学とコード生成の性能向上を報告しています。", "zh": "论文认为，双方均能表示的短续写可恢复精确匹配丢失的监督，并在测试的异构模型对上改善数学与代码表现。", "ko": "논문은 두 토크나이저가 모두 표현할 수 있는 짧은 후속 문자열로 정확한 매칭에서 잃는 감독을 복원하며, 평가한 이종 모델 쌍의 수학·코드 성능이 개선된다고 주장합니다." }, "claim_source": "https://arxiv.org/abs/2605.07711", "japanese": { "coverage": "不一致区間も含め、局所的に構成した有限個の続きの候補を学習に使います。", "access": "続きの候補を生成するときの、教師の次トークンスコアが必要です。", "semantics": "構成した候補集合の中でスコアを正規化します。教師が誘導する文字列分布全体を周辺化する方法ではありません。" } }, { "id": "esr", "record_kind": "paper", "classification": "core_output_space", "title": "Less is More: Early Stopping Rollout for On-Policy Distillation", "authors": [ "Zhou Ziheng", "Jiaqi Li", "Huacong Tang", "Ying Nian Wu", "Demetri Terzopoulos" ], "first_public_date": "2026-05-26", "checked_version": "arXiv v1", "publication": "arXiv preprint", "urls": { "paper": "https://arxiv.org/abs/2605.27028", "code": null, "project": null }, "student_policy_states": "yes", "mismatch_evidence": "The setup states that all cross-generation and cross-family pairs use different tokenizers.", "interface": "Decode student rollout, retokenize for the teacher, then greedily match text spans and apply reverse KL on aligned tokens.", "feedback_coverage": "Greedily aligned tokens within an early rollout window (normally 100 response tokens).", "teacher_access": "Teacher token log-probabilities.", "probability_semantics": "Partial aligned-token objective; no complete mapped distribution or mass-preservation property is reported.", "evaluated_pairs": [ { "teacher": "Qwen3-4B", "student": "Gemma-2-2B", "mismatch": "cross-family, different tokenizers" }, { "teacher": "Gemma-3-4B", "student": "Gemma-2-2B", "mismatch": "cross-generation, different tokenizers" }, { "teacher": "Qwen3-1.7B", "student": "Qwen2.5-Math-1.5B", "mismatch": "cross-generation, different tokenizers" } ], "code_status": "not_found", "evidence_locations": { "rollout": "Sections 3 and 4.1", "mismatch_interface": "Section 4.1", "experiment": "Tables 1 and 2" }, "caveats": [ "The paper's tables write student-to-teacher order; pairs here are normalized to teacher-to-student order." ], "last_verified": "2026-08-15", "metadata_last_verified": "2026-10-14", "code_last_checked": "2026-10-14", "name": "ESR", "access_group": "scores", "bibtex": "@article{zhou2026esr,\n title = {Less is More: Early Stopping Rollout for On-Policy Distillation},\n author = {{Zhou Ziheng} and Li, Jiaqi and Tang, Huacong and Wu, Ying Nian and Terzopoulos, Demetri},\n journal = {arXiv preprint arXiv:2605.27028},\n year = {2026},\n eprint = {2605.27028},\n archiveprefix = {arXiv},\n primaryclass = {cs.LG},\n url = {https://arxiv.org/abs/2605.27028}\n}", "summary": { "en": "Decode student rollout, retokenize for the teacher, then greedily match text spans and apply reverse KL on aligned tokens.", "ja": "生徒が生成した応答の冒頭を、教師のトークナイザーで再分割します。対応する位置を学習に使い、長い応答の後半は対象から除きます。", "zh": "将学生 rollout 重新编码为教师 token，贪心对齐文本 span。", "ko": "학생 rollout을 교사 tokenizer로 재인코딩한 뒤 텍스트 span을 탐욕적으로 대응." }, "claim": { "en": "The authors report that restricting rollouts to early response tokens can outperform full-rollout OPD while improving GPU efficiency and training stability.", "ja": "生徒の応答を冒頭だけで打ち切り、その部分を学習に使う手法です。応答全体を使うOPDより性能が高く、GPUの利用効率と学習の安定性も向上したと報告しています。", "zh": "作者报告，仅生成并监督回复前段即可超过完整轨迹 OPD，同时提高 GPU 效率与训练稳定性。", "ko": "저자들은 응답의 앞부분만 생성하고 감독해도 전체 응답 OPD보다 높은 성능과 GPU 효율, 학습 안정성을 얻을 수 있다고 보고합니다." }, "claim_source": "https://arxiv.org/abs/2605.27028", "japanese": { "coverage": "応答の冒頭（通常は100トークン）で、貪欲法による照合で対応したトークンを学習に使います。", "access": "教師のトークン対数確率が必要です。", "semantics": "対応した部分だけで目的関数を計算します。分布全体の写像や、確率質量を保存する性質は報告されていません。" } }, { "id": "gad", "record_kind": "paper", "classification": "core_vocabulary_bypassing", "title": "Black-Box On-Policy Distillation of Large Language Models", "authors": [ "Tianzhu Ye", "Li Dong", "Zewen Chi", "Xun Wu", "Shaohan Huang", "Furu Wei" ], "first_public_date": "2025-11-13", "checked_version": "arXiv v3", "publication": "arXiv preprint", "urls": { "paper": "https://arxiv.org/abs/2511.10643", "code": "https://github.com/microsoft/LMOps/tree/main/gad", "project": "https://ytianzhu.github.io/Generative-Adversarial-Distillation/" }, "student_policy_states": "yes", "mismatch_evidence": "The paper motivates GAD with incompatible tokenizers and evaluates a proprietary GPT-5 teacher against Qwen and Llama students.", "interface": "A discriminator co-evolves to distinguish fixed teacher responses from current student responses and supplies an RL reward.", "feedback_coverage": "Sequence-level scalar on current student rollouts.", "teacher_access": "Pre-collected teacher text only.", "probability_semantics": "No token-probability mapping; supervision is a learned adversarial sequence score.", "evaluated_pairs": [ { "teacher": "GPT-5-Chat", "student": "Qwen2.5-3B/7B/14B-Instruct", "mismatch": "black-box, tokenizer-agnostic" }, { "teacher": "GPT-5-Chat", "student": "Llama-3.2-3B-Instruct or Llama-3.1-8B-Instruct", "mismatch": "black-box, tokenizer-agnostic" } ], "code_status": "available", "evidence_locations": { "rollout": "Sections 2.1-2.3 / Algorithm 1", "mismatch_interface": "Introduction and Section 2", "experiment": "Section 3.1 / Table 2" }, "caveats": [ "Teacher responses are collected once per prompt; the teacher does not directly score each evolving student prefix." ], "last_verified": "2026-08-15", "metadata_last_verified": "2026-10-14", "code_last_checked": "2026-10-14", "name": "GAD", "access_group": "text", "bibtex": "@article{ye2025gad,\n title = {Black-Box On-Policy Distillation of Large Language Models},\n author = {Ye, Tianzhu and Dong, Li and Chi, Zewen and Wu, Xun and Huang, Shaohan and Wei, Furu},\n journal = {arXiv preprint arXiv:2511.10643},\n year = {2025},\n eprint = {2511.10643},\n archiveprefix = {arXiv},\n primaryclass = {cs.CL},\n url = {https://arxiv.org/abs/2511.10643}\n}", "summary": { "en": "A discriminator co-evolves to distinguish fixed teacher responses from current student responses and supplies an RL reward.", "ja": "固定した教師の回答と、現在の生徒の回答を見分ける識別器を、生徒とともに更新します。識別器の評価を、応答全体に対する報酬として使います。", "zh": "共同更新 discriminator，以区分固定教师回答与当前学生回答的序列级 reward 训练学生。", "ko": "고정 교사 응답과 현재 학생 응답을 구분하는 discriminator를 공동 갱신하고 시퀀스 단위 reward로 학습합니다." }, "claim": { "en": "The paper proposes learning an adaptive discriminator from teacher text alone, reporting stronger black-box distillation than sequence-level imitation.", "ja": "教師の回答テキストだけを使い、生徒の変化に合わせて識別器を学習する方法を提案しています。教師の内部情報を取得せずにOPDを行い、回答列を模倣する学習より高い性能を得たと報告しています。", "zh": "论文提出仅用教师文本学习自适应判别器，实现黑盒 OPD，并报告优于序列级模仿的结果。", "ko": "논문은 교사의 텍스트만으로 적응형 판별기를 학습해 블랙박스 OPD를 수행하며, 응답 시퀀스 모방보다 나은 결과를 보고합니다." }, "claim_source": "https://arxiv.org/abs/2511.10643", "japanese": { "coverage": "現在の生徒の応答全体に対する、1つのスカラー評価を学習に使います。", "access": "事前に収集した教師の回答テキストを使います。", "semantics": "トークン確率は対応づけません。学習した識別器の評価を、応答全体に対する報酬として使います。" } }, { "id": "ropd", "record_kind": "paper", "classification": "core_vocabulary_bypassing", "title": "Rubric-based On-policy Distillation", "authors": [ "Junfeng Fang", "Zhepei Hong", "Mao Zheng", "Mingyang Song", "Gengsheng Li", "Houcheng Jiang", "Dan Zhang", "Haiyun Guo", "Xiang Wang", "Tat-Seng Chua" ], "first_public_date": "2026-05-08", "checked_version": "arXiv v1", "publication": "arXiv preprint", "urls": { "paper": "https://arxiv.org/abs/2605.07396", "code": "https://github.com/Peregrine123/ROPD_official", "project": null }, "student_policy_states": "yes", "mismatch_evidence": "The paper explicitly states cross-architecture distillation without tokenizer alignment and evaluates GPT-5.2 against Qwen and Gemma students.", "interface": "A rubricator contrasts teacher and current student responses to induce textual criteria; a verifier scores each student rollout against them.", "feedback_coverage": "Weighted rubric-item pass rate at the response level.", "teacher_access": "Text-only generation and verification.", "probability_semantics": "No token-probability mapping; rubric scores become GRPO rewards.", "evaluated_pairs": [ { "teacher": "GPT-5.2-Chat", "student": "Qwen3-4B", "mismatch": "black-box cross-architecture" }, { "teacher": "GPT-5.2-Chat", "student": "Gemma-3-4B-IT", "mismatch": "black-box cross-architecture without tokenizer alignment" }, { "teacher": "Qwen3-30B-A3B", "student": "Qwen3-4B", "mismatch": "white-box control; same family" } ], "code_status": "available", "evidence_locations": { "rollout": "Section 2.2 / Algorithm 1", "mismatch_interface": "Introduction and Section 3.5", "experiment": "Sections 3.2-3.5 / Tables 1-3" }, "caveats": [ "The signal is semantic and response-level, so it is not a replacement for experiments that require calibrated token probabilities." ], "last_verified": "2026-08-15", "metadata_last_verified": "2026-10-14", "code_last_checked": "2026-10-14", "name": "ROPD", "access_group": "text", "bibtex": "@article{fang2026ropd,\n title = {Rubric-based On-policy Distillation},\n author = {Fang, Junfeng and Hong, Zhepei and Zheng, Mao and Song, Mingyang and Li, Gengsheng and Jiang, Houcheng and Zhang, Dan and Guo, Haiyun and Wang, Xiang and Chua, Tat-Seng},\n journal = {arXiv preprint arXiv:2605.07396},\n year = {2026},\n eprint = {2605.07396},\n archiveprefix = {arXiv},\n primaryclass = {cs.LG},\n url = {https://arxiv.org/abs/2605.07396}\n}", "summary": { "en": "A rubricator contrasts teacher and current student responses to induce textual criteria; a verifier scores each student rollout against them.", "ja": "教師と生徒の回答を比較し、質問ごとに採点基準を作ります。検証モデルがその基準に従って生徒の応答を採点します。", "zh": "对比教师和学生回答生成 prompt 专属 rubric，由 verifier 评分学生 rollout。", "ko": "교사와 학생 응답을 비교해 prompt별 rubric을 생성하고 verifier로 학생 rollout을 채점합니다." }, "claim": { "en": "The authors argue that prompt-specific rubrics can replace teacher logits for OPD, reporting improved performance and sample efficiency in the evaluated settings.", "ja": "質問ごとの採点基準を、教師のlogitsに代わる学習信号として使う方法を提案しています。評価した設定では、性能とサンプル効率が向上したと報告しています。", "zh": "作者认为，按提示生成的评分标准可替代教师 logits，并报告所评估设置中的性能与样本效率提升。", "ko": "저자들은 프롬프트별 평가 기준이 교사 logits를 대체할 수 있다고 주장하며, 평가한 설정의 성능과 표본 효율 개선을 보고합니다." }, "claim_source": "https://arxiv.org/abs/2605.07396", "japanese": { "coverage": "回答ごとに、重み付きの採点項目の合格率を計算します。", "access": "テキストの生成と検証が必要です。", "semantics": "トークン確率の写像は使いません。採点結果をGRPOの報酬にします。" } }, { "id": "omniopd", "record_kind": "paper", "classification": "core_vocabulary_bypassing", "title": "OmniOPD: Logit-Free On-Policy Distillation via Speculative Verification", "authors": [ "Yuhang Zhou", "Lizhu Zhang", "Yifan Wu", "Mingyi Wang", "Bo Peng", "Jiayi Liu", "Xiangjun Fan", "Zhuokai Zhao" ], "first_public_date": "2026-05-31", "checked_version": "arXiv v2", "publication": "arXiv preprint", "urls": { "paper": "https://arxiv.org/abs/2606.01476", "code": null, "project": null }, "student_policy_states": "yes", "mismatch_evidence": "The method is explicitly tokenizer-invariant and evaluates text-only proprietary teachers whose token logits/vocabularies are unavailable.", "interface": "Monte Carlo teacher continuations and semantic similarity over decoded multi-token chunks.", "feedback_coverage": "Chunks selected at peaks of student predictive entropy.", "teacher_access": "Text generation only; black-box/API compatible.", "probability_semantics": "No cross-vocabulary probability transport; a Bayesian-smoothed semantic verification signal is used.", "evaluated_pairs": [ { "teacher": "Qwen3-32B", "student": "Qwen3-1.7B or Qwen3-4B", "mismatch": "open-weight control; same family" }, { "teacher": "Claude-4.5-Haiku", "student": "Qwen3-4B", "mismatch": "tokenizer-agnostic black-box setting" }, { "teacher": "Gemini-2.5-Flash", "student": "Qwen3-4B", "mismatch": "tokenizer-agnostic black-box setting" } ], "code_status": "not_found", "evidence_locations": { "rollout": "Section 3.2.1", "mismatch_interface": "Section 4.4", "experiment": "Section 5 / Table 1" }, "caveats": [ "The exact vocabularies of proprietary teachers are not exposed; mismatch robustness follows from the text-only interface rather than a measurable token map." ], "last_verified": "2026-08-15", "metadata_last_verified": "2026-10-14", "code_last_checked": "2026-10-14", "name": "OmniOPD", "access_group": "text", "bibtex": "@article{zhou2026omniopd,\n title = {{OmniOPD}: Logit-Free On-Policy Distillation via Speculative Verification},\n author = {Zhou, Yuhang and Zhang, Lizhu and Wu, Yifan and Wang, Mingyi and Peng, Bo and Liu, Jiayi and Fan, Xiangjun and Zhao, Zhuokai},\n journal = {arXiv preprint arXiv:2606.01476},\n year = {2026},\n eprint = {2606.01476},\n archiveprefix = {arXiv},\n primaryclass = {cs.LG},\n url = {https://arxiv.org/abs/2606.01476}\n}", "summary": { "en": "Monte Carlo teacher continuations and semantic similarity over decoded multi-token chunks.", "ja": "教師に続きを複数生成させ、復号した応答区間の意味的な類似度を使って生徒を学習させます。教師のトークン確率は不要です。", "zh": "对教师 continuation 做 Monte Carlo sampling，用解码 chunk 的语义相似度监督。", "ko": "교사 continuation의 Monte Carlo 표집과 디코딩 chunk의 의미 유사도로 감독합니다." }, "claim": { "en": "The paper argues that semantic verification of response chunks supplies reliable OPD feedback without teacher logits, including feedback from proprietary text-only teachers.", "ja": "応答区間の意味的な類似度を使えば、教師のlogitsなしでも信頼できる学習信号を得られるとしています。テキストだけを返す非公開モデルも教師として利用できると報告しています。", "zh": "论文认为，回复片段的语义验证可在不访问教师 logits 的情况下提供可靠监督，包括使用仅输出文本的闭源教师。", "ko": "논문은 응답 구간의 의미 검증으로 교사 logits 없이 신뢰할 수 있는 OPD 신호를 얻으며, 텍스트만 제공하는 비공개 교사도 활용할 수 있다고 주장합니다." }, "claim_source": "https://arxiv.org/abs/2606.01476", "japanese": { "coverage": "生徒の予測エントロピーが高い位置で、応答区間を選んで学習に使います。", "access": "テキスト生成だけで利用できます。内部の確率を取得できないAPIの教師にも対応します。", "semantics": "語彙間で確率を輸送する方法ではありません。意味的な検証から得た信号を、ベイズ的に平滑化します。" } }, { "id": "oprd", "record_kind": "paper", "classification": "paper_claim_implementation_concern", "title": "OPRD: On-Policy Representation Distillation", "authors": [ "Shenzhi Yang", "Guangcheng Zhu", "Bowen Song", "Haobo Wang", "Mingxuan Xia", "Xing Zheng", "Yingfan Ma", "Zhongqi Chen", "Weiqiang Wang", "Junbo Zhao", "Gang Chen" ], "first_public_date": "2026-06-04", "checked_version": "arXiv v4", "publication": "arXiv preprint", "urls": { "paper": "https://arxiv.org/abs/2606.06021", "code": "https://github.com/ShenzhiYang2000/OPRD", "project": null }, "student_policy_states": "yes", "mismatch_evidence": "The paper reports completely disjoint Phi and Qwen tokenizers in its cross-tokenizer experiment.", "interface": "OPRD-Bridge aligns selected hidden layers through frozen low-rank projectors.", "feedback_coverage": "Selected positions and teacher-student layer pairs on the same student rollout.", "teacher_access": "Teacher hidden states.", "probability_semantics": "Vocabulary probability mapping is bypassed; the loss is representation-level.", "evaluated_pairs": [ { "teacher": "Phi-4-mini-reasoning", "student": "Qwen3-1.7B-Base", "mismatch": "completely different tokenizers" } ], "code_status": "available", "implementation_audit": { "checked_commit": "a63236c990b1bfa9551e9d020719490d3ecd93e8", "finding": "Public scripts load the student tokenizer and forward the same student-token IDs through both student and teacher; those IDs have unrelated semantics in disjoint vocabularies.", "evidence": [ "https://github.com/ShenzhiYang2000/OPRD/blob/a63236c990b1bfa9551e9d020719490d3ecd93e8/scripts/analysis/cross_arch_repr_analysis.py#L5-L8", "https://github.com/ShenzhiYang2000/OPRD/blob/a63236c990b1bfa9551e9d020719490d3ecd93e8/scripts/analysis/cross_arch_preexp2_train_ps.py#L1564-L1567", "https://github.com/ShenzhiYang2000/OPRD/blob/a63236c990b1bfa9551e9d020719490d3ecd93e8/scripts/analysis/cross_arch_preexp2_train_ps.py#L888-L890" ], "last_checked": "2026-10-14" }, "evidence_locations": { "rollout": "Section 3", "mismatch_interface": "Section 3.3", "experiment": "Section 4.2.4 / Table 6" }, "caveats": [ "Requires internal hidden-state access and a staged bridge construction procedure.", "The released implementation does not retokenize the same decoded text for the teacher, so it does not currently verify the paper's semantic cross-tokenizer claim." ], "last_verified": "2026-08-15", "metadata_last_verified": "2026-10-14", "code_last_checked": "2026-10-14", "name": "OPRD / OPRD-Bridge", "access_group": "hidden", "bibtex": "@article{yang2026oprd,\n title = {{OPRD}: On-Policy Representation Distillation},\n author = {Yang, Shenzhi and Zhu, Guangcheng and Song, Bowen and Wang, Haobo and Xia, Mingxuan and Zheng, Xing and Ma, Yingfan and Chen, Zhongqi and Wang, Weiqiang and Zhao, Junbo and Chen, Gang},\n journal = {arXiv preprint arXiv:2606.06021},\n year = {2026},\n eprint = {2606.06021},\n archiveprefix = {arXiv},\n primaryclass = {cs.LG},\n url = {https://arxiv.org/abs/2606.06021}\n}", "summary": { "en": "OPRD-Bridge aligns selected hidden layers through frozen low-rank projectors.", "ja": "固定した低ランク射影を使い、教師と生徒の隠れ状態を対応づけます。ただし公開実装では、両モデルに同じ生徒トークンIDを入力している点に懸念があります。", "zh": "用 projector 对齐隐藏表示，但公开实现对不同词表的两模型输入相同的学生 token ID，存在语义对应疑问。", "ko": "Projector로 은닉 표현을 대응하지만 공개 구현은 다른 어휘의 두 모델에 같은 학생 token ID를 입력하므로 의미적 대응에 우려가 있습니다." }, "claim": { "en": "The authors claim that hidden-state supervision avoids output-head limitations, and report cross-architecture and cross-tokenizer transfer through frozen low-rank projectors.", "ja": "隠れ状態を使って学習させることで、出力ヘッドの制約を回避できるとしています。固定した低ランク射影を使い、構造やトークナイザーが異なるモデル間でも知識を転移できたと報告しています。", "zh": "作者主张隐藏状态监督可绕过输出头的限制，并报告通过固定低秩投影实现跨架构与跨分词器知识迁移。", "ko": "저자들은 은닉 상태 감독이 출력 헤드의 제약을 피한다고 주장하며, 고정된 저랭크 투영을 통한 서로 다른 구조·토크나이저 간 지식 이전을 보고합니다." }, "claim_source": "https://arxiv.org/abs/2606.06021", "japanese": { "coverage": "同じ生徒の生成系列から位置を選び、教師と生徒の層の組み合わせごとに学習させます。", "access": "教師の隠れ状態が必要です。", "semantics": "語彙の確率写像を使わず、表現に対する損失を計算します。ただし、公開実装での語彙の対応には懸念があります。" } }, { "id": "btb", "record_kind": "paper", "classification": "core_output_space", "title": "Breaking the Tokenizer Barrier: On-Policy Distillation across Model Families", "authors": [ "Yifan Niu", "Han Xiao", "Dongyi Liu", "Zelong Wang", "Dihong Gong", "Yasheng Wang", "Jia Li" ], "first_public_date": "2026-06-08", "checked_version": "arXiv v1", "publication": "arXiv preprint", "urls": { "paper": "https://arxiv.org/abs/2606.09456", "code": "https://github.com/ivanniu/On-Policy-Distill", "project": null }, "student_policy_states": "yes", "mismatch_evidence": "The main Qwen-to-Llama experiment explicitly uses different model families and tokenizers.", "interface": "Dual-pointer decoded-string synchronization chunks and teacher/student canonical-path likelihoods.", "feedback_coverage": "The realized student path.", "teacher_access": "Per-token teacher log-probabilities on retokenized realized text.", "probability_semantics": "Chunk/path likelihood credit; it does not construct a counterfactual full-vocabulary target.", "evaluated_pairs": [ { "teacher": "Qwen3-8B", "student": "Llama-3.1-8B-SFT", "mismatch": "different families and tokenizers" }, { "teacher": "DeepSeek-R1", "student": "DeepSeek-R1-Distill-Qwen-7B", "mismatch": "reported cross-tokenizer setting" } ], "code_status": "available", "evidence_locations": { "rollout": "Section 3", "mismatch_interface": "Section 3", "experiment": "Section 4.1 / Table 1" }, "caveats": [ "Scalar realized-path supervision should not be compared as if it were a dense vocabulary transform." ], "last_verified": "2026-08-15", "metadata_last_verified": "2026-10-14", "code_last_checked": "2026-10-14", "name": "BTB", "access_group": "scores", "bibtex": "@article{niu2026btb,\n title = {Breaking the Tokenizer Barrier: On-Policy Distillation across Model Families},\n author = {Niu, Yifan and Xiao, Han and Liu, Dongyi and Wang, Zelong and Gong, Dihong and Wang, Yasheng and Li, Jia},\n journal = {arXiv preprint arXiv:2606.09456},\n year = {2026},\n eprint = {2606.09456},\n archiveprefix = {arXiv},\n primaryclass = {cs.LG},\n url = {https://arxiv.org/abs/2606.09456}\n}", "summary": { "en": "Dual-pointer decoded-string synchronization chunks and teacher/student canonical-path likelihoods.", "ja": "教師と生徒のトークン列を、復号した文字列が一致する最小区間ごとに対応づけます。その区間の生成経路の尤度から、生徒の学習信号を計算します。", "zh": "双指针构造解码文本同步的最小 chunk。", "ko": "이중 포인터로 디코딩 문자열이 동기화되는 최소 chunk 구성." }, "claim": { "en": "The paper proposes transmitting token-level OPD signals across model families through decoded-text mapping, reporting better compute efficiency than the evaluated baselines.", "ja": "復号した文字列を対応づければ、異なるモデル系列の間でもトークン単位のOPDを行えるとしています。評価した既存手法より計算効率が高かったと報告しています。", "zh": "论文提出通过解码文本映射跨模型系列传递 token 级 OPD 信号，并报告优于测试基线的计算效率。", "ko": "논문은 디코딩한 텍스트의 대응으로 서로 다른 모델 계열에 토큰 단위 OPD 신호를 전달하며, 평가한 기준 방법보다 높은 계산 효율을 보고합니다." }, "claim_source": "https://arxiv.org/abs/2606.09456", "japanese": { "coverage": "生徒が実際に生成した経路を学習に使います。", "access": "生成テキストを教師のトークナイザーで再分割したときの、各トークンの対数確率が必要です。", "semantics": "区間や生成経路の尤度から、学習信号を計算します。生成しなかった全語彙の候補に対する分布は作りません。" } }, { "id": "reopold", "record_kind": "paper", "classification": "evaluation_interface_under_specified", "title": "Relaxed On-Policy Distillation: Selective Credit Allocation for Scaling Reasoning Efficiently", "authors": [ "Jongwoo Ko", "Sara Abdali", "Young Jin Kim", "Tianyi Chen", "Pashmina Cameron" ], "first_public_date": "2026-03-11", "checked_version": "OpenReview version audited 2026-08-15; latest author-linked version not retrieved", "publication": "NeurIPS 2026 (listed by author publication page; latest OpenReview text not retrieved)", "urls": { "paper": "https://openreview.net/forum?id=NPhEUvT1jH", "arxiv": "https://arxiv.org/abs/2603.11137", "author_publication": "https://www.microsoft.com/en-us/research/publication/relaxed-on-policy-distillation-selective-credit-allocation-for-scaling-reasoning-efficiently/", "latest_openreview": "https://openreview.net/forum?id=l0h1SzRI7H", "code": null, "project": null }, "student_policy_states": "yes", "mismatch_evidence": "Appendix F / Table 11 labels Phi-to-DeepSeek-Qwen as different model families and vocabularies.", "interface": "Not specified for the mismatched vocabularies; the objective uses teacher/student likelihood ratios for sampled tokens.", "feedback_coverage": "Selected sampled tokens, using credit clipping and entropy-based sampling.", "teacher_access": "Teacher sampled-token likelihoods according to the method description.", "probability_semantics": "Cross-vocabulary event correspondence was not documented in the previously audited OpenReview version.", "evaluated_pairs": [ { "teacher": "Phi-4-Mini-Reasoning", "student": "DeepSeek-R1-Distill-Qwen-1.5B", "mismatch": "paper states different vocabularies" } ], "code_status": "not_found", "evidence_locations": { "rollout": "Method section", "mismatch_interface": "Appendix F / Table 11; mapping not described", "experiment": "Appendix F / Table 11" }, "caveats": [ "The arXiv v1 predates the final appendix and uses the earlier title 'Scaling Reasoning Efficiently via Relaxed On-Policy Distillation'.", "Do not treat this row as a reproducible cross-tokenizer alignment method without further implementation details.", "The author publication page now lists NeurIPS 2026 and a different OpenReview record; browser verification prevented retrieval, so its interface was not re-audited. The earlier ICML/PMLR attribution has been removed." ], "last_verified": "2026-08-15", "metadata_last_verified": "2026-10-14", "code_last_checked": "2026-10-14", "name": "REOPOLD", "access_group": "scores", "bibtex": "@misc{ko2026reopold,\n title = {Relaxed On-Policy Distillation: Selective Credit Allocation for Scaling Reasoning Efficiently},\n author = {Ko, Jongwoo and Abdali, Sara and Kim, Young Jin and Chen, Tianyi and Cameron, Pashmina},\n howpublished = {Author publication page lists NeurIPS 2026},\n year = {2026},\n url = {https://www.microsoft.com/en-us/research/publication/relaxed-on-policy-distillation-selective-credit-allocation-for-scaling-reasoning-efficiently/},\n note = {arXiv:2603.11137 v1 has an earlier title; the latest author-linked OpenReview full text was not retrieved at the 2026-10-14 refresh}\n}", "summary": { "en": "Not specified for the mismatched vocabularies; the objective uses teacher/student likelihood ratios for sampled tokens.", "ja": "生徒の生成経路に対する学習信号を選択・調整します。以前に確認したバージョンでは、異なる語彙間でトークンが表す事象をどう対応づけるか、説明が不十分でした。", "zh": "选择学生 rollout 的 token credit；旧审计版本对跨词表 token 事件的对应说明不足。", "ko": "학생 rollout의 token credit을 선택합니다. 이전 감사 버전은 교차 어휘 token 사건의 대응 설명이 부족합니다." }, "claim": { "en": "The authors argue that selectively allocating teacher feedback reduces negative transfer and entropy collapse, and report improved sample efficiency, including cross-vocabulary experiments.", "ja": "教師から得た学習信号をトークンごとに選択して配分すると、負の転移やエントロピー崩壊を抑えられるとしています。語彙が異なるモデルを含む設定で、サンプル効率の改善を報告しています。", "zh": "作者认为选择性分配教师反馈可减少负迁移与熵坍缩，并报告包括跨词表实验在内的样本效率提升。", "ko": "저자들은 교사 피드백을 선택적으로 배분하면 부정적 전이와 엔트로피 붕괴를 줄일 수 있다고 주장하며, 서로 다른 어휘를 포함한 실험의 표본 효율 개선을 보고합니다." }, "claim_source": "https://www.microsoft.com/en-us/research/publication/relaxed-on-policy-distillation-selective-credit-allocation-for-scaling-reasoning-efficiently/", "japanese": { "coverage": "標本トークンをエントロピーに基づいて選び、クリッピングした学習信号を使います。", "access": "論文の説明では、教師が標本トークンに与える尤度を使います。", "semantics": "以前に確認したバージョンでは、異なる語彙間で同じ確率事象をどう定めるか、記載されていませんでした。" } }, { "id": "bpm", "record_kind": "paper", "classification": "core_output_space", "title": "Cross-Tokenizer On-Policy Distillation via Byte-Prefix Marginalization", "authors": [ "Hao Wang", "Kun Yuan", "Wenlin Zhong", "Minglei Zhang", "Han Xiao", "Ming Sun", "Honggang Qi" ], "first_public_date": "2026-07-24", "checked_version": "arXiv v1", "publication": "arXiv preprint", "urls": { "paper": "https://arxiv.org/abs/2607.22334", "code": "https://github.com/K1XE/BPM", "project": "https://bpm-opd.github.io/" }, "student_policy_states": "yes", "mismatch_evidence": "Three teacher tokenizers with increasing distance are evaluated against a fixed Qwen3.5 student.", "interface": "Teacher mass is routed to the longest student-token byte prefix; collisions are aggregated and unmatched mass enters a residual category.", "feedback_coverage": "Full student vocabulary plus residual.", "teacher_access": "Full teacher next-token distribution.", "probability_semantics": "Mass-preserving; exact byte-prefix marginal at more than 99% of reported positions and a mass-preserving chain-factorized lower bound otherwise.", "evaluated_pairs": [ { "teacher": "Qwen3-32B", "student": "Qwen3.5-2B", "mismatch": "cross-generation tokenizer mismatch" }, { "teacher": "GLM-Z1-9B-0414", "student": "Qwen3.5-2B", "mismatch": "cross-family tokenizer mismatch" }, { "teacher": "MiniMax-M2.7", "student": "Qwen3.5-2B", "mismatch": "cross-family tokenizer mismatch" } ], "code_status": "available", "evidence_locations": { "rollout": "Section 3", "mismatch_interface": "Sections 3.1-3.3", "experiment": "Section 4.1 / Table 1" }, "caveats": [ "The paper reports a whitespace-only code target failure and uses a precomputed mask.", "Official code was released on 2026-09-01; availability was checked, but experiments were not independently reproduced." ], "last_verified": "2026-08-15", "metadata_last_verified": "2026-10-14", "code_last_checked": "2026-10-14", "name": "BPM", "access_group": "logits", "bibtex": "@article{wang2026bpm,\n title = {Cross-Tokenizer On-Policy Distillation via Byte-Prefix Marginalization},\n author = {Wang, Hao and Yuan, Kun and Zhong, Wenlin and Zhang, Minglei and Xiao, Han and Sun, Ming and Qi, Honggang},\n journal = {arXiv preprint arXiv:2607.22334},\n year = {2026},\n eprint = {2607.22334},\n archiveprefix = {arXiv},\n primaryclass = {cs.LG},\n url = {https://arxiv.org/abs/2607.22334}\n}", "summary": { "en": "Teacher mass is routed to the longest student-token byte prefix; collisions are aggregated and unmatched mass enters a residual category.", "ja": "教師トークンの確率を、バイト列の接頭辞が対応する生徒トークンへ集約します。対応先がない確率質量は、残余カテゴリに保持します。", "zh": "把教师 token 概率分配给其字节序列的最长学生 token 前缀。", "ko": "교사 token의 바이트열을 가장 긴 학생 token 바이트 접두사에 할당." }, "claim": { "en": "The paper proposes a byte-aligned target over the student vocabulary that preserves teacher probability mass, reporting gains on mathematics and programming benchmarks.", "ja": "バイト列を基準に、教師の確率質量を保ったまま、生徒の語彙に対応する目標分布を作る方法を提案しています。数学とプログラミングの評価では、既存手法からの性能向上を報告しています。", "zh": "论文提出保持教师概率质量、覆盖学生词表的字节对齐目标，并报告数学与编程基准上的提升。", "ko": "논문은 교사 확률 질량을 보존하는 바이트 정렬 목표를 학생 어휘 위에 구성하며, 수학·프로그래밍 벤치마크의 개선을 보고합니다." }, "claim_source": "https://arxiv.org/abs/2607.22334", "japanese": { "coverage": "生徒の全語彙と、対応先がない確率を保持する残余カテゴリを対象にします。", "access": "教師の次トークン分布全体が必要です。", "semantics": "確率質量を保存します。論文の報告では、99%を超える位置で厳密なバイト接頭辞周辺分布になります。それ以外の位置では、質量を保存する連鎖分解による下界を使います。" } }, { "id": "ovd", "record_kind": "paper", "classification": "boundary_application", "title": "OVD: On-policy Verbal Distillation", "authors": [ "Jing Xiong", "Hui Shen", "Shansan Gong", "Yuxin Cheng", "Jianghan Shen", "Chaofan Tao", "Haochen Tan", "Haoli Bai", "Lifeng Shang", "Ngai Wong" ], "first_public_date": "2026-01-29", "checked_version": "arXiv v2", "publication": "arXiv preprint", "urls": { "paper": "https://arxiv.org/abs/2601.21968", "code": "https://github.com/menik1126/OVD-code", "project": "https://menik1126.github.io/ovd-project-page/" }, "student_policy_states": "qualified: student-generated sub-trajectories are retained or replaced before mixed-policy optimization", "mismatch_evidence": "The method explicitly avoids token-level alignment and evaluates Qwen teachers/agents with both Qwen and Llama students.", "interface": "Verbal scores select student sub-trajectories; rejected suffixes or queries are replaced by teacher continuations.", "feedback_coverage": "Selected sub-trajectories; outcome-based advantages optimize both retained student tokens and teacher replacements.", "teacher_access": "Teacher verbal scores and generated replacements; no full-vocabulary teacher logits required.", "probability_semantics": "Retain-and-replace mixture; teacher-distribution approximation is conditional on acceptance calibration and replacement error.", "evaluated_pairs": [ { "teacher": "Qwen2.5-7B teacher agent", "student": "Qwen2.5-3B-Base / Llama-3.2-3B-Base", "mismatch": "text-only interface; includes cross-family Llama student" } ], "code_status": "available", "evidence_locations": { "rollout": "Sections 2.1 and 2.3", "mismatch_interface": "Sections 1.1-2.1; Appendix B", "experiment": "Section 3 / Table 1; Appendix B / Table 9" }, "caveats": [ "The effective training distribution mixes retained student segments and teacher replacements.", "Some reported Web Q&A results include teacher intervention at test time; separate these from student-only inference." ], "last_verified": "2026-10-14", "metadata_last_verified": "2026-10-14", "code_last_checked": "2026-10-14", "name": "OVD", "access_group": "text", "bibtex": "@article{xiong2026ovd,\n title = {{OVD}: On-policy Verbal Distillation},\n author = {Xiong, Jing and Shen, Hui and Gong, Shansan and Cheng, Yuxin and Shen, Jianghan and Tao, Chaofan and Tan, Haochen and Bai, Haoli and Shang, Lifeng and Wong, Ngai},\n journal = {arXiv preprint arXiv:2601.21968},\n year = {2026},\n eprint = {2601.21968},\n archiveprefix = {arXiv},\n primaryclass = {cs.CL},\n url = {https://arxiv.org/abs/2601.21968},\n note = {Technical report}\n}", "summary": { "en": "Verbal scores select student sub-trajectories; rejected suffixes or queries are replaced by teacher continuations.", "ja": "教師の言葉による評価で、生徒が生成した軌跡の一部を選びます。評価が低い部分を教師の続きで置き換え、生徒と教師の生成が混在する軌跡を学習に使います。", "zh": "用 verbal score 筛选学生子轨迹，在学生与教师 continuation 混合的轨迹上训练。", "ko": "Verbal score로 학생 부분 trajectory를 선택하고 교사 continuation과 섞인 trajectory에서 학습합니다." }, "claim": { "en": "The authors propose retaining strong student trajectories and replacing weaker segments using verbal teacher feedback, reporting reasoning and efficiency gains with text-only teachers.", "ja": "教師の言葉による評価を使い、生徒が生成した軌跡のうち評価が高い部分を残し、低い部分を教師の生成で置き換える手法です。テキストだけを返す教師を使い、推論性能と学習効率が向上したと報告しています。", "zh": "作者提出用教师的语言化评分保留较好学生轨迹、替换较弱片段，并报告文本教师下的推理与效率提升。", "ko": "저자들은 교사의 언어 점수로 좋은 학생 경로를 유지하고 약한 구간만 교체하며, 텍스트 교사를 활용한 추론 성능과 학습 효율 개선을 보고합니다." }, "claim_source": "https://arxiv.org/abs/2601.21968", "japanese": { "coverage": "選択した軌跡の一部を学習に使います。残した生徒トークンと、教師が置き換えたトークンの両方を更新します。", "access": "教師の言葉によるスコアと、置き換え用の生成テキストが必要です。全語彙のlogitsは不要です。", "semantics": "生徒と教師が生成した軌跡の混合分布で学習します。教師分布を近似できるかどうかは、受理確率の較正と置換誤差の条件によります。" } }, { "id": "omnidraft", "record_kind": "paper", "classification": "boundary_application", "title": "OmniDraft: A Cross-vocabulary, Online Adaptive Drafter for On-device Speculative Decoding", "authors": [ "Ramchalam Kinattinkara Ramakrishnan", "Zhaocong Yuan", "Shaojie Zhuo", "Chen Feng", "Yicheng Lin", "Chenzheng Su", "Xiaopeng Zhang" ], "first_public_date": "2025-07-03", "checked_version": "arXiv v3", "publication": "NeurIPS 2025", "urls": { "paper": "https://arxiv.org/abs/2507.02659", "code": null, "project": null }, "student_policy_states": "qualified: online drafter-generated proposals", "mismatch_evidence": "A single Llama-68M drafter is paired with target models from several tokenizer families.", "interface": "Direct token mapping plus accepted n-gram supervision in online speculative decoding.", "feedback_coverage": "Draft proposals accepted or scored by the target model.", "teacher_access": "Target-model verification and probability information.", "probability_semantics": "Application-specific hybrid distillation; final decoding is governed by speculative verification.", "evaluated_pairs": [ { "teacher": "Vicuna-7B, Qwen2-7B, or Llama3-8B target", "student": "Llama-68M drafter", "mismatch": "cross-vocabulary target/drafter" } ], "code_status": "not_found", "evidence_locations": { "rollout": "Method section", "mismatch_interface": "Method section", "experiment": "Experiments section" }, "caveats": [ "Not canonical full-response student-policy OPD." ], "last_verified": "2026-08-15", "metadata_last_verified": "2026-10-14", "code_last_checked": "2026-10-14", "name": "OmniDraft", "access_group": "logits", "bibtex": "@inproceedings{ramakrishnan2025omnidraft,\n title = {{OmniDraft}: A Cross-vocabulary, Online Adaptive Drafter for On-device Speculative Decoding},\n author = {Ramakrishnan, Ramchalam Kinattinkara and Yuan, Zhaocong and Zhuo, Shaojie and Feng, Chen and Lin, Yicheng and Su, Chenzheng and Zhang, Xiaopeng},\n booktitle = {Advances in Neural Information Processing Systems},\n year = {2025},\n url = {https://arxiv.org/abs/2507.02659}\n}", "summary": { "en": "Direct token mapping plus accepted n-gram supervision in online speculative decoding.", "ja": "ドラフトモデルが提案したトークンを使ってオンライン学習する、異なる語彙間の推測デコーディング手法です。最終出力は対象モデルによる検証で決まります。", "zh": "从 drafter 自身 proposal 在线学习的跨词表推测解码。", "ko": "Drafter 자신의 proposal에서 온라인 학습하는 교차 어휘 추측 디코딩입니다." }, "claim": { "en": "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.", "ja": "1つのドラフトモデルを、語彙が異なる複数の対象モデルにオンラインで適応させる手法です。蒸留と適応的なドラフト生成を組み合わせ、推測デコーディングを高速化したと報告しています。", "zh": "论文主张单个草稿模型可在线适应不同词表的目标模型，并通过混合蒸馏与自适应起草加速推测解码。", "ko": "논문은 하나의 초안 모델을 서로 다른 어휘의 대상 모델에 온라인으로 적응시키고, 혼합 증류와 적응형 초안 생성으로 추측 디코딩을 가속한다고 주장합니다." }, "claim_source": "https://arxiv.org/abs/2507.02659", "japanese": { "coverage": "対象モデルが受理または評価した、ドラフトモデルの提案を学習に使います。", "access": "対象モデルの検証結果と確率情報が必要です。", "semantics": "推測デコーディング用の目的関数を組み合わせています。最終出力分布は対象モデルの検証で決まります。" } }, { "id": "kdflow", "record_kind": "paper", "classification": "framework", "title": "KDFlow: A User-Friendly and Efficient Knowledge Distillation Framework for Large Language Models", "authors": [ "Songming Zhang", "Xue Zhang", "Tong Zhang", "Bojie Hu", "Yufeng Chen", "Jinan Xu" ], "first_public_date": "2026-03-02", "checked_version": "arXiv v4", "publication": "arXiv preprint", "urls": { "paper": "https://arxiv.org/abs/2603.01875", "code": "https://github.com/songmzhang/KDFlow", "project": null }, "student_policy_states": "supported by framework", "mismatch_evidence": "Framework supports DSKDv2-style cross-tokenizer KD.", "interface": "Software integration of existing KD algorithms; no new tokenizer bridge isolated.", "feedback_coverage": null, "teacher_access": "Depends on selected method.", "probability_semantics": "Depends on selected method.", "evaluated_pairs": [ { "teacher": "Qwen3-14B", "student": "Llama3.2-3B", "mismatch": "different tokenizers; off-policy experiment" } ], "code_status": "available", "evidence_locations": { "rollout": "Section 4.4 / Appendix A.2 (separate on-policy configurations)", "mismatch_interface": "Sections 3.4 and 4.6", "experiment": "Section 4.6 / Table 5; Appendix A.1 (off-policy)" }, "caveats": [ "The v4 cross-tokenizer experiment is off-policy (Appendix A.1); on-policy framework support does not make this experiment cross-tokenizer OPD." ], "last_verified": "2026-10-14", "metadata_last_verified": "2026-10-14", "code_last_checked": "2026-10-14", "name": "KDFlow", "access_group": "varies", "bibtex": "@article{zhang2026kdflow,\n title = {{KDFlow}: A User-Friendly and Efficient Knowledge Distillation Framework for Large Language Models},\n author = {Zhang, Songming and Zhang, Xue and Zhang, Tong and Hu, Bojie and Chen, Yufeng and Xu, Jinan},\n journal = {arXiv preprint arXiv:2603.01875},\n year = {2026},\n eprint = {2603.01875},\n archiveprefix = {arXiv},\n primaryclass = {cs.CL},\n url = {https://arxiv.org/abs/2603.01875}\n}", "summary": { "en": "Software integration of existing KD algorithms; no new tokenizer bridge isolated.", "ja": "既存の蒸留手法を分散実行するフレームワークです。異なるトークナイザー間で報告された実験は、固定データを使うoff-policy設定で行われています。", "zh": "支持 on/off-policy 与 DSKD 类跨 tokenizer KD 的分布式框架；报告的跨 tokenizer 实验是 off-policy。", "ko": "On/off-policy와 DSKD 계열 교차 tokenizer KD를 지원하는 분산 프레임워크이며, 보고된 교차 tokenizer 실험은 off-policy입니다." }, "claim": { "en": "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.", "ja": "教師の推論と生徒の学習を別々の実行基盤で処理し、分散蒸留を高速化したと報告しています。on-policy・off-policyの蒸留と、異なるトークナイザー間の既存手法に対応しています。", "zh": "作者提出分别为教师推理与学生训练配置后端，报告分布式蒸馏提速，并支持 on-policy、off-policy 与跨分词器方法。", "ko": "저자들은 교사 추론과 학생 학습에 별도 백엔드를 사용해 분산 증류를 가속하며, on-policy·off-policy·교차 토크나이저 방법을 지원한다고 보고합니다." }, "claim_source": "https://arxiv.org/abs/2603.01875", "japanese": { "coverage": "選択した学習手法によります。", "access": "選択した学習手法によります。", "semantics": "選択した学習手法によります。報告された異なるトークナイザー間の実験は、off-policy設定です。" } }, { "id": "easyopd", "record_kind": "paper", "classification": "framework", "title": "EasyOPD: An Easy-to-use On-Policy Distillation Framework for Large Language Models", "authors": [ "Jie Sun", "Mao Zheng", "Mingyang Song", "Qiyong Zhong", "Gengsheng Li", "Zhepei Hong", "Chang Wu", "Pengfei Liu", "Junfeng Fang", "Xiang Wang" ], "first_public_date": "2026-07-13", "checked_version": "arXiv v1", "publication": "arXiv preprint", "urls": { "paper": "https://arxiv.org/abs/2607.11012", "code": "https://github.com/lds-ustc/EasyOPD", "project": null }, "student_policy_states": "supported by framework", "mismatch_evidence": "Cross-tokenizer OPD is one of three implemented settings.", "interface": "Software modules for SimCT and cross-tokenizer baselines on a shared verl backend.", "feedback_coverage": "Depends on selected method.", "teacher_access": "Depends on selected method.", "probability_semantics": "Depends on selected method.", "evaluated_pairs": [ { "teacher": "Qwen2.5-7B-Instruct", "student": "Phi-4-mini-Instruct", "mismatch": "different tokenizers" } ], "code_status": "available", "evidence_locations": { "rollout": "Framework overview", "mismatch_interface": "Cross-tokenizer OPD module", "experiment": "Experiments section" }, "caveats": [ "Resource/framework paper, not a new alignment objective." ], "last_verified": "2026-08-15", "metadata_last_verified": "2026-10-14", "code_last_checked": "2026-10-14", "name": "EasyOPD", "access_group": "varies", "bibtex": "@article{sun2026easyopd,\n title = {{EasyOPD}: An Easy-to-use On-Policy Distillation Framework for Large Language Models},\n author = {Sun, Jie and Zheng, Mao and Song, Mingyang and Zhong, Qiyong and Li, Gengsheng and Hong, Zhepei and Wu, Chang and Liu, Pengfei and Fang, Junfeng and Wang, Xiang},\n journal = {arXiv preprint arXiv:2607.11012},\n year = {2026},\n eprint = {2607.11012},\n archiveprefix = {arXiv},\n primaryclass = {cs.CL},\n url = {https://arxiv.org/abs/2607.11012}\n}", "summary": { "en": "Software modules for SimCT and cross-tokenizer baselines on a shared verl backend.", "ja": "verl上で、SimCTなどの異なるトークナイザー間の蒸留手法を比較・実行するフレームワークです。語彙の対応方法を新たに提案するものではありません。", "zh": "基于 verl 的框架，提供 SimCT 等跨 tokenizer OPD 方法的可运行模块。", "ko": "Verl에서 SimCT 등 교차 tokenizer OPD 방법의 실행 모듈을 제공하는 프레임워크입니다." }, "claim": { "en": "The paper presents a shared verl backend with separate method modules, showing that different OPD settings can retain their objectives within one execution framework.", "ja": "手法ごとの学習信号の処理と、共通のverl実行基盤を分離したフレームワークです。異なるOPD設定を1つのフレームワークで実行・拡張できると示しています。", "zh": "论文将方法模块与共享 verl 后端分离，展示多种 OPD 设置可在统一框架中保留各自目标并执行。", "ko": "논문은 방법별 모듈을 공통 verl 백엔드와 분리해, 다양한 OPD 설정을 각자의 목적함수를 유지하며 하나의 실행 프레임워크에서 다룰 수 있음을 보여줍니다." }, "claim_source": "https://arxiv.org/abs/2607.11012", "japanese": { "coverage": "選択した学習手法によります。", "access": "選択した学習手法によります。", "semantics": "選択した学習手法によります。既存の対応方法を実行するための基盤です。" } }, { "id": "gold", "record_kind": "research_article_and_implementation", "classification": "implementation_resource", "title": "Unlocking On-Policy Distillation for Any Model Family", "authors": [ "Carlos Miguel Patiño", "Kashif Rasul", "Quentin Gallouédec", "Ben Burtenshaw", "Sergio Paniego", "Vaibhav Srivastav", "Thibaud Frere", "Ed Beeching", "Lewis Tunstall", "Leandro von Werra", "Thomas Wolf" ], "first_public_date": "2025-10-29", "checked_version": "web resource checked 2026-10-14", "publication": "Hugging Face research article and experimental TRL trainer; not a paper", "urls": { "paper": null, "article": "https://huggingface.co/spaces/HuggingFaceH4/on-policy-distillation", "documentation": "https://huggingface.co/docs/trl/main/gold_trainer", "code": "https://github.com/huggingface/trl", "project": null }, "student_policy_states": "yes", "mismatch_evidence": "The resource is explicitly designed for teacher and student models with different tokenizers.", "interface": "Incremental decoded-span alignment with exact/shared-token supervision where possible and ULD fallback.", "feedback_coverage": "Aligned spans plus fallback handling for mismatches.", "teacher_access": "Teacher logits/probabilities; separate teacher tokenizer.", "probability_semantics": "Hybrid implementation objective; consult the trainer documentation for current behavior.", "evaluated_pairs": [], "code_status": "available_experimental", "evidence_locations": { "rollout": "Research article and TRL documentation", "mismatch_interface": "Research article and TRL documentation", "experiment": "Research article" }, "caveats": [ "Not counted as a paper in the core literature table.", "GOLDTrainer is in trl.experimental.gold; APIs may change. Full span coverage does not imply identity-preserving probability transport." ], "last_verified": "2026-10-14", "metadata_last_verified": "2026-10-14", "code_last_checked": "2026-10-14", "name": "GOLD", "access_group": "logits", "bibtex": "@misc{patino2025gold,\n title = {Unlocking On-Policy Distillation for Any Model Family},\n author = {Patiño, Carlos Miguel and Rasul, Kashif and Gallouédec, Quentin and Burtenshaw, Ben and Paniego, Sergio and Srivastav, Vaibhav and Frere, Thibaud and Beeching, Ed and Tunstall, Lewis and von Werra, Leandro and Wolf, Thomas},\n year = {2025},\n howpublished = {Hugging Face research article and TRL implementation},\n url = {https://huggingface.co/spaces/HuggingFaceH4/on-policy-distillation},\n note = {Non-paper resource}\n}", "summary": { "en": "Incremental decoded-span alignment with exact/shared-token supervision where possible and ULD fallback.", "ja": "復号した文字列が同じ区間を対応づけ、確率を併合します。共通トークンは直接比較し、対応しない部分にはULDを使います。TRLの実験的な実装です。", "zh": "对齐解码 span 并合并概率；直接比较共享 token，对未对应 token 使用 ULD fallback。实验性 TRL trainer。", "ko": "디코딩 span 대응과 확률 병합, 공통 token 직접 비교, 불일치 token의 ULD fallback을 사용하는 실험적 TRL trainer입니다." }, "claim": { "en": "The official documentation describes cross-tokenizer distillation using decoded-text alignment, probability merging, and hybrid ULD loss within TRL's experimental trainer.", "ja": "公式文書では、復号した文字列の対応づけ、確率の併合、ULD損失を組み合わせる方法を紹介しています。TRLの実験的なtrainerで、異なるトークナイザー間の蒸留を実行できると説明しています。", "zh": "官方文档介绍在 TRL 实验性 trainer 中，通过解码文本对齐、概率合并与混合 ULD 损失执行跨分词器蒸馏。", "ko": "공식 문서는 TRL의 실험적 trainer에서 디코딩 텍스트 정렬, 확률 병합, 혼합 ULD 손실로 교차 토크나이저 증류를 수행하는 기능을 설명합니다." }, "claim_source": "https://huggingface.co/docs/trl/main/gold_trainer", "japanese": { "coverage": "対応する文字列区間を学習に使い、不一致区間にはフォールバック処理を適用します。", "access": "教師のlogits・確率と、教師のトークナイザーが必要です。", "semantics": "複数の処理を組み合わせて損失を計算します。現在の動作は、実験的なtrainerの公式文書で確認できます。" } }, { "id": "simpleopd", "record_kind": "paper", "classification": "core_output_space", "title": "SimpleOPD: Simple Tokenizer-Agnostic On-Policy Distillation for Long-Context Reasoning", "authors": [ "Haonan He", "Haodi Lei", "Yun Luo", "Haoran Zhang", "Shunkai Zhang", "Yizhuo Li", "Shengji Tang", "Zhilin Wang", "Runzhe Zhan", "Lei Bai", "Ganqu Cui", "Fangchen Yu", "Yafu Li", "Peng Ye", "Ning Ding", "Yu Cheng" ], "first_public_date": "2026-08-14", "last_verified": "2026-10-14", "metadata_last_verified": "2026-10-14", "code_last_checked": "2026-10-14", "checked_version": "arXiv v1", "publication": "arXiv preprint", "urls": { "paper": "https://arxiv.org/abs/2608.14277", "full_text": "https://arxiv.org/html/2608.14277v1", "code": "https://github.com/hhnqqq/SimpleOPD", "project": "https://hhnqqq.github.io/SimpleOPD-project-page/" }, "student_policy_states": "yes", "mismatch_evidence": "The full text explicitly describes tokenizer mismatch and evaluates the listed heterogeneous pairs.", "interface": "Retokenize student text with the teacher template; retain tokens with identical start/end offsets.", "feedback_coverage": "Realized tokens at exact 1:1 text spans; unmatched positions receive zero distillation advantage.", "teacher_access": "Teacher sampled-token log-probabilities.", "probability_semantics": "Partial reverse-KL surrogate; no dense target.", "evaluated_pairs": [ { "teacher": "SU-01 (Qwen3-30B-A3B-based)", "student": "GLM-4.7-Flash / Gemma-4-26B-A4B", "mismatch": "different vocabularies and segmentation" }, { "teacher": "SU-01", "student": "Qwen3.5-4B / Qwen3.5-35B-A3B", "mismatch": "cross-generation tokenizer mismatch" } ], "code_status": "available", "evidence_locations": { "rollout": "Section 2 / Eq. 1", "mismatch_interface": "Sections 2.1-2.2 / Appendix A", "experiment": "Sections 3.2 and 4.2-4.3 / Table 2 / Figure 7" }, "caveats": [ "Termination-token advantage masking and student-reference KL are part of the stabilized recipe.", "Gemma improves on ProofBench but declines on AnswerBench." ], "name": "SimpleOPD", "access_group": "scores", "bibtex": "@article{he2026simpleopd,\n title = {SimpleOPD: Simple Tokenizer-Agnostic On-Policy Distillation for Long-Context Reasoning},\n author = {He, Haonan and Lei, Haodi and Luo, Yun and Zhang, Haoran and Zhang, Shunkai and Li, Yizhuo and Tang, Shengji and Wang, Zhilin and Zhan, Runzhe and Bai, Lei and Cui, Ganqu and Yu, Fangchen and Li, Yafu and Ye, Peng and Ding, Ning and Cheng, Yu},\n journal = {arXiv preprint arXiv:2608.14277},\n year = {2026},\n eprint = {2608.14277},\n archiveprefix = {arXiv},\n primaryclass = {cs.CL},\n url = {https://arxiv.org/abs/2608.14277}\n}", "summary": { "en": "Retokenize student text with the teacher template; retain tokens with identical start/end offsets.", "ja": "生徒の応答を教師のテンプレートで再符号化し、始点と終点が一致するトークンを学習に使います。一致しない位置の蒸留信号はゼロです。", "zh": "使用教师自己的模板重新编码学生文本，匹配起止位置完全相同的 token。", "ko": "교사 고유 template로 학생 텍스트를 재인코딩하고 시작／끝 위치가 같은 token 대응." }, "claim": { "en": "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.", "ja": "文字列区間の厳密な対応づけと、生徒の初期モデルを基準とした制約を組み合わせる手法です。長い文脈を扱う教師の推論能力を、短い文脈の生徒へ安定して転移できたと報告しています。", "zh": "作者报告，精确文本区间对齐与学生参考约束可稳定地将长上下文教师的推理能力迁移给短上下文学生。", "ko": "저자들은 정확한 텍스트 구간 정렬과 학생 기준 모델 제약으로 긴 문맥 교사의 추론 능력을 짧은 문맥 학생에게 안정적으로 이전한다고 보고합니다." }, "claim_source": "https://arxiv.org/abs/2608.14277", "japanese": { "coverage": "生成したトークンのうち、始点と終点が厳密に1対1で対応するものを学習に使います。不一致位置の蒸留信号はゼロです。", "access": "生成したトークンに対する、教師の対数確率が必要です。", "semantics": "対応した部分で逆KLの代理目的を計算します。全語彙にわたる密な目標分布は作りません。" } }, { "id": "compassopd", "record_kind": "paper", "classification": "core_output_space", "title": "CompassOPD: Cross-Family On-Policy Distillation via Within-Family Likelihood Shifts", "authors": [ "Naibin Gu", "Qingyi Si", "Chenxu Yang", "Chuanyu Qin", "Junhao Zhou", "Peng Fu", "Zheng Lin", "Weiping Wang" ], "first_public_date": "2026-09-09", "last_verified": "2026-10-14", "metadata_last_verified": "2026-10-14", "code_last_checked": "2026-10-14", "checked_version": "arXiv v1", "publication": "arXiv preprint", "urls": { "paper": "https://arxiv.org/abs/2609.10154", "full_text": "https://arxiv.org/html/2609.10154v1", "code": null, "project": null }, "student_policy_states": "yes", "mismatch_evidence": "The full text explicitly describes tokenizer mismatch and evaluates the listed heterogeneous pairs.", "interface": "Shared-boundary span alignment; transfer strong-teacher minus teacher-family-reference scores, with a frozen student anchor.", "feedback_coverage": "Realized aligned units; inconsistent boundaries are masked.", "teacher_access": "Teacher and teacher-family-reference log-probabilities.", "probability_semantics": "Span-averaged likelihood shifts; not teacher vocabulary transport.", "evaluated_pairs": [ { "teacher": "Qwen3.5-35B-A3B (reference: Qwen3.5-0.8B)", "student": "Granite4.1-3B / Qwen3-4B / OLMo-3-7B-Instruct-SFT", "mismatch": "cross-family/cross-generation aligned-text setting" } ], "code_status": "not_found", "evidence_locations": { "rollout": "Sections 2 and 4", "mismatch_interface": "Appendix A.4", "experiment": "Section 5.2 / Tables 1-2" }, "caveats": [ "Requires a teacher-family reference and SFT initialization; reduced expert activation provides an MoE self-reference option." ], "name": "CompassOPD", "access_group": "scores", "bibtex": "@article{gu2026compassopd,\n title = {CompassOPD: Cross-Family On-Policy Distillation via Within-Family Likelihood Shifts},\n author = {Gu, Naibin and Si, Qingyi and Yang, Chenxu and Qin, Chuanyu and Zhou, Junhao and Fu, Peng and Lin, Zheng and Wang, Weiping},\n journal = {arXiv preprint arXiv:2609.10154},\n year = {2026},\n eprint = {2609.10154},\n archiveprefix = {arXiv},\n primaryclass = {cs.LG},\n url = {https://arxiv.org/abs/2609.10154}\n}", "summary": { "en": "Shared-boundary span alignment; transfer strong-teacher minus teacher-family-reference scores, with a frozen student anchor.", "ja": "対応する文字列区間について、強い教師と同系列の参照モデルの尤度差を学習信号として使います。生徒の初期モデルを基準に、更新の大きさを調整します。", "zh": "在对齐文本 span 上，传递强教师相对同家族 reference 的似然变化。", "ko": "대응 텍스트 span에서 강한 교사와 같은 계열 reference 간 우도 변화 전달." }, "claim": { "en": "The paper argues that cross-family offsets obscure useful teacher improvements; transferring within-family likelihood shifts improves reasoning in the tested student families.", "ja": "モデル系列間のずれが、教師の能力向上に対応する学習信号を隠してしまうと指摘しています。同系列内での尤度の変化を使うことで、評価した生徒系列の推論精度が向上したと報告しています。", "zh": "论文认为模型系列间的偏移会掩盖教师能力提升的信号，传递系列内部的似然变化可改善所测试学生系列的推理。", "ko": "논문은 모델 계열 간 차이가 교사의 능력 향상 신호를 가리며, 같은 계열 내 우도 변화를 전달하면 평가한 학생 계열의 추론 성능이 개선된다고 주장합니다." }, "claim_source": "https://arxiv.org/abs/2609.10154", "japanese": { "coverage": "実際に生成した対応区間を学習に使います。境界を一致させられない位置は除きます。", "access": "教師と、教師と同系列の参照モデルの対数確率が必要です。", "semantics": "区間内で平均した尤度の変化を学習信号として使います。教師の語彙分布を輸送する方法ではありません。" } }, { "id": "dvlm", "record_kind": "paper", "classification": "core_output_space", "title": "Dual-Vocabulary Language Model for Cross-Tokenizer Distillation", "authors": [ "Kedi Chen", "Chen Lin", "Yutao Sun", "Wei Zhang" ], "first_public_date": "2026-09-27", "last_verified": "2026-10-14", "metadata_last_verified": "2026-10-14", "code_last_checked": "2026-10-14", "checked_version": "arXiv v1", "publication": "arXiv preprint", "urls": { "paper": "https://arxiv.org/abs/2609.33816", "full_text": "https://arxiv.org/html/2609.33816v1", "code": null, "project": null }, "student_policy_states": "yes", "mismatch_evidence": "The full text explicitly describes tokenizer mismatch and evaluates the listed heterogeneous pairs.", "interface": "Train a student-vocabulary teacher head using parallel-tokenized input and Hybrid-Prefix Attention.", "feedback_coverage": "Dense student-vocabulary logits at student positions.", "teacher_access": "Teacher backbone, hidden states, and trainable replacement LM head.", "probability_semantics": "Learned adapted-teacher distribution; not exact original-teacher string marginalization.", "evaluated_pairs": [ { "teacher": "Qwen3-4B", "student": "Llama-3.2-1B-Instruct / OLMo-2-1B-Instruct", "mismatch": "different tokenizers" } ], "code_status": "not_found", "evidence_locations": { "rollout": "Section 3.2 / Eq. 7", "mismatch_interface": "Section 3.1 / Eqs. 2-6", "experiment": "Sections 4.1-4.2 / Table 1" }, "caveats": [ "Requires a separate head-training stage and modified teacher attention; Stage II uses forward KL." ], "name": "DVLM", "access_group": "hidden", "bibtex": "@article{chen2026dvlm,\n title = {Dual-Vocabulary Language Model for Cross-Tokenizer Distillation},\n author = {Chen, Kedi and Lin, Chen and Sun, Yutao and Zhang, Wei},\n journal = {arXiv preprint arXiv:2609.33816},\n year = {2026},\n eprint = {2609.33816},\n archiveprefix = {arXiv},\n primaryclass = {cs.AI},\n url = {https://arxiv.org/abs/2609.33816}\n}", "summary": { "en": "Train a student-vocabulary teacher head using parallel-tokenized input and Hybrid-Prefix Attention.", "ja": "教師に、生徒の語彙に対応する出力ヘッドを学習させます。並列にトークン化した入力列と専用の注意機構を使い、生徒の位置に対応する分布を出力します。", "zh": "平行 tokenization 与 Hybrid-Prefix Attention，训练面向学生词表的教师 head。", "ko": "병렬 tokenization과 Hybrid-Prefix Attention으로 학생 어휘용 교사 head 학습." }, "claim": { "en": "The authors propose adapting the teacher to the student's vocabulary and token positions, reporting full-dimensional supervision and gains across six reasoning tasks.", "ja": "教師を生徒の語彙とトークン位置に合わせ、全次元の出力分布を学習に使う方法を提案しています。6つの推論タスクで、生徒の性能が向上したと報告しています。", "zh": "作者提出将教师适配到学生词表与 token 位置，实现全维度监督，并报告六项推理任务上的提升。", "ko": "저자들은 교사를 학생 어휘와 토큰 위치에 적응시켜 전체 차원의 출력 분포로 감독하며, 여섯 추론 과제의 성능 개선을 보고합니다." }, "claim_source": "https://arxiv.org/abs/2609.33816", "japanese": { "coverage": "生徒のトークン位置で、生徒の語彙全体のlogitsを学習に使います。", "access": "教師の本体、隠れ状態、別途学習する出力ヘッドが必要です。", "semantics": "学習によって適応させた教師の分布を使います。元の教師の文字列確率を厳密に周辺化する方法ではありません。" } }, { "id": "escd", "record_kind": "paper", "classification": "core_output_space", "title": "Beyond Token Alignment: Event Completion for Cross-Tokenizer On-Policy Distillation", "authors": [ "Jiacheng Liu", "Jingwei Song", "Qituan Zhang", "Siheng Chen", "Linfeng Zhang" ], "first_public_date": "2026-09-28", "last_verified": "2026-10-14", "metadata_last_verified": "2026-10-14", "code_last_checked": "2026-10-14", "checked_version": "arXiv v1", "publication": "arXiv preprint", "urls": { "paper": "https://arxiv.org/abs/2609.34738", "full_text": "https://arxiv.org/html/2609.34738v1", "code": null, "project": null }, "student_policy_states": "yes", "mismatch_evidence": "The full text explicitly describes tokenizer mismatch and evaluates the listed heterogeneous pairs.", "interface": "Aggregate prefix-related teacher events; add a completion-set loss at visited child states to byte-aligned root supervision.", "feedback_coverage": "Total probability of compatible one-step completions.", "teacher_access": "Teacher event probabilities (full vocabulary in experiments).", "probability_semantics": "Mass-weighted set probability; within-set allocation is unconstrained.", "evaluated_pairs": [ { "teacher": "Qwen3-32B / GLM-Z1-9B", "student": "Qwen3.5-2B", "mismatch": "cross-generation/cross-family tokenizer mismatch" }, { "teacher": "Qwen3.5-397B-A17B", "student": "Qwen3-30B-A3B-Thinking", "mismatch": "cross-generation tokenizer mismatch" }, { "teacher": "Kimi-K2.7-Code", "student": "Qwen3.6-35B-A3B", "mismatch": "heterogeneous MoE tokenizer mismatch; SFT initialization" } ], "code_status": "announced_not_available", "evidence_locations": { "rollout": "Section 3.3 / Appendix B.6 / Algorithm 1", "mismatch_interface": "Sections 3.1-3.2 / Appendix B.4", "experiment": "Section 4 / Tables 1-2" }, "caveats": [ "Empty one-step completion sets get no child loss.", "Prefix aggregation coarsens events; completions may extend beyond residual boundaries." ], "name": "ESCD", "access_group": "logits", "bibtex": "@article{liu2026escd,\n title = {Beyond Token Alignment: Event Completion for Cross-Tokenizer On-Policy Distillation},\n author = {Liu, Jiacheng and Song, Jingwei and Zhang, Qituan and Chen, Siheng and Zhang, Linfeng},\n journal = {arXiv preprint arXiv:2609.34738},\n year = {2026},\n eprint = {2609.34738},\n archiveprefix = {arXiv},\n primaryclass = {cs.CL},\n url = {https://arxiv.org/abs/2609.34738}\n}", "summary": { "en": "Aggregate prefix-related teacher events; add a completion-set loss at visited child states to byte-aligned root supervision.", "ja": "教師トークンが表すバイト列の一部を生成した後、残りを完成できる生徒トークンの集合を学習に使います。集合の合計確率を扱い、候補ごとの確率配分は固定しません。", "zh": "教师字节事件被部分生成后，在已访问的子状态监督完成集合。", "ko": "교사 바이트 사건을 일부 생성한 뒤, 방문한 자식 상태에서 완성 집합 감독." }, "claim": { "en": "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.", "ja": "有効な補完候補の合計確率を使い、トークンの対応づけを補う学習信号を計算する手法です。追加の生成や語彙の変更なしで、数学・コード・科学の性能向上を報告しています。", "zh": "论文认为有效补全集合的总概率监督可补充 token 对齐，无需额外轨迹或词表修改即可改善数学、代码与科学推理。", "ko": "논문은 유효한 완성 후보 집합의 총확률 감독이 토큰 정렬을 보완하며, 추가 생성이나 어휘 변경 없이 수학·코드·과학 추론을 개선한다고 주장합니다." }, "claim_source": "https://arxiv.org/abs/2609.34738", "japanese": { "coverage": "残りを1トークンで完成できる候補集合の、合計確率を学習に使います。", "access": "教師の事象確率が必要です。実験では全語彙を使います。", "semantics": "教師の確率質量で重みづけした、集合の合計確率を学習に使います。集合内の候補ごとの確率配分は固定しません。" } }, { "id": "rethinking", "record_kind": "paper", "classification": "core_output_space", "title": "Rethinking Cross-Tokenizer On-Policy Distillation: From Alignment Coverage to Supervision Reliability", "authors": [ "Bingxi Hou", "Guochao Jiang", "Guofeng Quan", "Weiqing Li", "Wenfeng Feng", "Guohua Liu", "Yuewei Zhang" ], "first_public_date": "2026-10-06", "last_verified": "2026-10-14", "metadata_last_verified": "2026-10-14", "code_last_checked": "2026-10-14", "checked_version": "arXiv v1", "publication": "arXiv preprint", "urls": { "paper": "https://arxiv.org/abs/2610.08448", "full_text": "https://arxiv.org/html/2610.08448v1", "code": "https://anonymous.4open.science/r/Cross-Tokenizer-OPD", "project": null }, "student_policy_states": "yes", "mismatch_evidence": "The full text explicitly describes tokenizer mismatch and evaluates the listed heterogeneous pairs.", "interface": "Strict 1:1 span alignment; reverse KL on renormalized shared vocabulary or a student-selected top-k subset.", "feedback_coverage": "Strict positions; top-16 tested.", "teacher_access": "Teacher next-token scores on shared candidates.", "probability_semantics": "Subset-renormalized distribution; not full-vocabulary transport.", "evaluated_pairs": [ { "teacher": "Qwen2.5-7B-Instruct", "student": "Llama-3.2-3B-Instruct", "mismatch": "different tokenizers" }, { "teacher": "Granite-4.1-8B", "student": "Phi-4-mini-instruct / Qwen2.5-7B-Base", "mismatch": "different tokenizers" } ], "code_status": "linked_access_unverified", "evidence_locations": { "rollout": "Sections 2.2-2.3 / Eq. 5", "mismatch_interface": "Section 2.3 / Eqs. 2-4", "experiment": "Sections 3-5 / Tables 1-3 / Appendix C" }, "caveats": [ "Negative span-MSE findings concern the tested objective and pairs, not all mismatch supervision.", "Paper-linked anonymous code returned HTTP 401 during this check; implementation contents were not audited." ], "name": "Rethinking Cross-Tokenizer OPD", "access_group": "logits", "bibtex": "@article{hou2026rethinking,\n title = {Rethinking Cross-Tokenizer On-Policy Distillation: From Alignment Coverage to Supervision Reliability},\n author = {Hou, Bingxi and Jiang, Guochao and Quan, Guofeng and Li, Weiqing and Feng, Wenfeng and Liu, Guohua and Zhang, Yuewei},\n journal = {arXiv preprint arXiv:2610.08448},\n year = {2026},\n eprint = {2610.08448},\n archiveprefix = {arXiv},\n primaryclass = {cs.CL},\n url = {https://arxiv.org/abs/2610.08448}\n}", "summary": { "en": "Strict 1:1 span alignment; reverse KL on renormalized shared vocabulary or a student-selected top-k subset.", "ja": "厳密に1対1で対応する位置で、共有語彙に絞って分布を比較します。上位16候補を使う学習と、不一致区間に学習信号を追加する場合の効果を検証しています。", "zh": "严格 1:1 span 对齐，在共享词表或学生选择的 top-k 子集上计算 reverse KL。", "ko": "엄격한 1:1 span 대응과 공통 어휘／학생 선택 top-k 후보에서 reverse KL." }, "claim": { "en": "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.", "ja": "学習信号を与える範囲の広さより、その信号の信頼性を重視すべきだとしています。評価したモデル対では、厳密に対応する位置の上位16候補で、共有語彙全体を使う場合と同等の精度を得ました。一方、不一致区間にMSEの学習信号を加えると精度が低下したと報告しています。", "zh": "测试中，严格对齐位置的 top-16 监督达到完整共享词表 OPD 的精度，而额外的不匹配区间 MSE 降低精度。论文强调监督可靠性优先于覆盖率。", "ko": "평가한 모델 쌍에서 정확히 정렬된 위치의 top-16 감독은 전체 공유 어휘 OPD와 비슷한 정확도를 얻지만, 불일치 구간의 추가 MSE는 정확도를 낮춥니다. 논문은 감독 범위보다 신뢰성을 강조합니다." }, "claim_source": "https://arxiv.org/abs/2610.08448", "japanese": { "coverage": "厳密に対応する位置を学習に使います。生徒が選ぶ上位16候補についても評価しています。", "access": "共通の候補に対する、教師の次トークンスコアが必要です。", "semantics": "部分集合の中で再正規化した分布を比較します。全語彙の確率を輸送する方法ではありません。" } }, { "id": "cast", "record_kind": "paper", "classification": "core_output_space", "title": "CAST: Credit Alignment for Cross-Tokenizer On-Policy Distillation", "authors": [ "Jingru Yang", "Jinhang Wu", "Minjing Shi", "Shengfeng He", "Xuyan Li", "Man Tan", "Biye Li" ], "first_public_date": "2026-10-09", "last_verified": "2026-10-14", "metadata_last_verified": "2026-10-14", "code_last_checked": "2026-10-14", "checked_version": "arXiv v1", "publication": "arXiv preprint", "urls": { "paper": "https://arxiv.org/abs/2610.13257", "full_text": "https://arxiv.org/html/2610.13257v1", "code": null, "project": null }, "student_policy_states": "yes", "mismatch_evidence": "The full text explicitly describes tokenizer mismatch and evaluates the listed heterogeneous pairs.", "interface": "Decoded-text alignment followed by uncertainty/smoothness-weighted chunk credit and selective initial-policy anchoring.", "feedback_coverage": "Realized aligned tokens; initial-policy penalty at unaligned positions.", "teacher_access": "Per-token teacher log-probabilities on retokenized student text.", "probability_semantics": "Preserves total chunk advantage, not full-vocabulary probability mass.", "evaluated_pairs": [ { "teacher": "Qwen3-8B", "student": "Mistral-7B / Llama-3.1-8B (SFT initialized)", "mismatch": "different tokenizers" } ], "code_status": "not_found", "evidence_locations": { "rollout": "Section 3.1", "mismatch_interface": "Sections 3.2-3.3 / Appendix A.1", "experiment": "Section 4.2 / Table 1" }, "caveats": [ "Credit chunks differ from tokenizer alignment chunks.", "Teacher-score smoothness is a structural prior, not a correctness estimate." ], "name": "CAST", "access_group": "scores", "bibtex": "@article{yang2026cast,\n title = {CAST: Credit Alignment for Cross-Tokenizer On-Policy Distillation},\n author = {Yang, Jingru and Wu, Jinhang and Shi, Minjing and He, Shengfeng and Li, Xuyan and Tan, Man and Li, Biye},\n journal = {arXiv preprint arXiv:2610.13257},\n year = {2026},\n eprint = {2610.13257},\n archiveprefix = {arXiv},\n primaryclass = {cs.LG},\n url = {https://arxiv.org/abs/2610.13257}\n}", "summary": { "en": "Decoded-text alignment followed by uncertainty/smoothness-weighted chunk credit and selective initial-policy anchoring.", "ja": "文字列区間の学習信号を、生徒の不確実性と教師スコアの滑らかさに応じて各トークンに配分します。対応しない位置では、初期モデルからの変化を抑えます。", "zh": "解码文本对齐后，按学生不确定性与教师分数平滑度分配 chunk credit。", "ko": "텍스트 대응 후 학생 불확실성과 교사 점수 평활도에 따라 chunk credit 배분." }, "claim": { "en": "The authors argue that credit allocation within aligned chunks and anchoring unaligned tokens are both necessary, reporting gains across mathematics, science, and code benchmarks.", "ja": "対応した区間内で学習信号をどう配分するかと、対応しないトークンで初期モデルからの変化をどう抑えるかが重要だとしています。数学・科学・コード生成で、既存手法からの性能向上を報告しています。", "zh": "作者认为对齐区间内的信号分配和未对齐 token 的初始策略约束均很重要，并报告数学、科学与代码基准上的提升。", "ko": "저자들은 정렬된 구간 안의 학습 신호 배분과 정렬되지 않은 토큰의 초기 정책 제약이 모두 중요하다고 주장하며, 수학·과학·코드 벤치마크의 개선을 보고합니다." }, "claim_source": "https://arxiv.org/abs/2610.13257", "japanese": { "coverage": "実際に生成した対応トークンを学習に使います。不一致位置には、初期モデルからの変化を抑える罰則を加えます。", "access": "生徒のテキストを教師のトークナイザーで再分割したときの、各トークンの対数確率が必要です。", "semantics": "各区間の学習信号の総量を保ちます。全語彙の確率質量を保存する性質とは異なります。" } }, { "id": "mintagent", "record_kind": "paper", "classification": "core_application", "title": "Mint-Agent: Introducing Finance-Native Agentic Foundation Models", "authors": [ "Mint-Agent Team", "Kun Wang", "Gavin Zhang", "Yaze Geng", "Lei Tang", "Yaoyang Yi", "Zonghan Wu", "Yifan Hu", "Qingsong Wen", "Yilei Shao" ], "first_public_date": "2026-08-17", "last_verified": "2026-10-14", "metadata_last_verified": "2026-10-14", "code_last_checked": "2026-10-14", "checked_version": "arXiv v2", "publication": "arXiv preprint", "urls": { "paper": "https://arxiv.org/abs/2608.16386", "full_text": "https://arxiv.org/html/2608.16386v2", "code": null, "project": "https://mint-fin.github.io/mint-agent/" }, "student_policy_states": "yes", "mismatch_evidence": "The full text explicitly describes tokenizer mismatch and evaluates the listed heterogeneous pairs.", "interface": "At critical student turns, decoded-text chunks redistribute teacher path scores proportionally to student token log-probabilities.", "feedback_coverage": "Resampled student actions at selected failure states.", "teacher_access": "DeepSeek-V4-Pro scores on student continuations.", "probability_semantics": "Realized-path credit; no dense vocabulary mapping.", "evaluated_pairs": [ { "teacher": "DeepSeek-V4-Pro", "student": "Qwen3.5-9B / Qwen3.6-27B agentic specialists (Mint-Cu / Mint-Ag pipeline)", "mismatch": "explicitly different tokenizers" } ], "code_status": "not_found", "evidence_locations": { "rollout": "Section 5.2 (Critical-Step OPD)", "mismatch_interface": "Section 5.2 (Critical-Step OPD)", "experiment": "Section 6.1 / Table 2" }, "caveats": [ "Reports a combined SFT/OPD/RLVR/merging pipeline; final scores do not isolate the cross-tokenizer stage.", "The subsequent multi-teacher integration uses same-tokenizer specialists." ], "name": "Mint-Agent", "access_group": "scores", "bibtex": "@article{mintagent2026,\n title = {Mint-Agent: Introducing Finance-Native Agentic Foundation Models},\n author = {{Mint-Agent Team} and Wang, Kun and Zhang, Gavin and Geng, Yaze and Tang, Lei and Yi, Yaoyang and Wu, Zonghan and Hu, Yifan and Wen, Qingsong and Shao, Yilei},\n journal = {arXiv preprint arXiv:2608.16386},\n year = {2026},\n eprint = {2608.16386},\n archiveprefix = {arXiv},\n primaryclass = {cs.CL},\n url = {https://arxiv.org/abs/2608.16386}\n}", "summary": { "en": "At critical student turns, decoded-text chunks redistribute teacher path scores proportionally to student token log-probabilities.", "ja": "生徒が失敗した軌跡の重要な行動を再生成します。教師の区間尤度から計算した学習信号を、生徒の各トークンに配分します。最終的な成績は、蒸留・強化学習・モデル統合を組み合わせた結果です。", "zh": "在学生失败轨迹的关键 turn 重新生成，将教师 chunk 似然分配给学生 token 的应用训练流程。", "ko": "학생 실패 trajectory의 중요 turn에서 다시 생성하고 교사 chunk 우도를 학생 token에 배분하는 실제 학습 pipeline입니다." }, "claim": { "en": "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.", "ja": "SFT・重要ステップでのOPD・RLVR・モデル統合を組み合わせた学習方法です。金融分野での推論の信頼性と、長期タスクを実行する能力を備えた小型エージェントを構築できたと報告しています。", "zh": "作者报告，通过 SFT、关键步骤 OPD、RLVR 与模型合并的完整流程，构建兼具可靠推理和长程执行能力的小型金融智能体。", "ko": "저자들은 SFT·중요 단계 OPD·RLVR·모델 병합을 결합한 전체 과정으로 신뢰할 수 있는 추론과 장기 실행 능력을 갖춘 소형 금융 에이전트를 구축했다고 보고합니다." }, "claim_source": "https://arxiv.org/abs/2608.16386", "japanese": { "coverage": "選択した失敗状態で、生徒が再生成した行動を学習に使います。", "access": "生徒が生成した続きに対する、DeepSeek-V4-Proのスコアが必要です。", "semantics": "実際に生成した経路に対する学習信号を使います。語彙全体の密な写像分布は作りません。" } }, { "id": "actd", "record_kind": "paper", "classification": "adjacent_off_policy", "title": "ACTD: Anchor-Based Cross-Tokenizer Distillation with Residual Regularization", "authors": [ "Huiyi Zhang", "Zijian Li", "Xiaocheng Feng", "Weitao Ma", "Xiaoliang Yang", "Yichong Huang", "Bing Qin" ], "first_public_date": "2026-08-30", "last_verified": "2026-10-14", "metadata_last_verified": "2026-10-14", "code_last_checked": "2026-10-14", "checked_version": "arXiv v1", "publication": "arXiv preprint", "urls": { "paper": "https://arxiv.org/abs/2608.29662", "full_text": "https://arxiv.org/html/2608.29662v1", "code": "https://github.com/hyhy892/ACTD", "project": null }, "student_policy_states": "no: dataset-provided reference solutions", "mismatch_evidence": "The full text explicitly describes tokenizer mismatch and evaluates the listed heterogeneous pairs.", "interface": "Vocabulary/sequence alignment with high-probability anchor loss and residual regularization.", "feedback_coverage": "Anchors and residuals on fixed reference solutions.", "teacher_access": "Cached teacher top-128 logits.", "probability_semantics": "Mapped anchor distribution; off-policy.", "evaluated_pairs": [ { "teacher": "DeepSeek-R1-Distill-Llama-8B / GLM-4-9B-Chat", "student": "Qwen3-1.7B-Base", "mismatch": "different tokenizers" } ], "code_status": "available", "evidence_locations": { "rollout": "Section 4.1 / Appendix C (reference solutions)", "mismatch_interface": "Section 3 / Appendix A", "experiment": "Section 4 / Table 1" }, "caveats": [ "Section 4.1 explicitly uses off-policy reference solutions, so this is not core OPD evidence." ], "name": "ACTD", "access_group": "logits", "bibtex": "@article{zhang2026actd,\n title = {ACTD: Anchor-Based Cross-Tokenizer Distillation with Residual Regularization},\n author = {Zhang, Huiyi and Li, Zijian and Feng, Xiaocheng and Ma, Weitao and Yang, Xiaoliang and Huang, Yichong and Qin, Bing},\n journal = {arXiv preprint arXiv:2608.29662},\n year = {2026},\n eprint = {2608.29662},\n archiveprefix = {arXiv},\n primaryclass = {cs.CL},\n url = {https://arxiv.org/abs/2608.29662}\n}", "summary": { "en": "Vocabulary/sequence alignment with high-probability anchor loss and residual regularization.", "ja": "固定された正解の解法を使い、確率の高い候補を学習させ、残余部分には正則化を加えます。生徒の生成を使うOPDとは異なる、off-policyの蒸留手法です。", "zh": "在固定 reference solution 上使用 anchor 与 residual regularization 的 off-policy KD。", "ko": "고정 reference solution에서 anchor와 residual regularization을 사용하는 off-policy KD입니다." }, "claim": { "en": "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.", "ja": "確率の高い候補の学習と、残余部分の正則化を組み合わせ、異なるトークナイザー間の対応ノイズを減らす手法です。単一・複数教師のoff-policy蒸留で、推論性能が向上したと報告しています。", "zh": "论文提出高概率锚点与残余正则化以减少跨分词器对齐噪声，并报告单教师及多教师 off-policy 蒸馏中的推理提升。", "ko": "논문은 고확률 앵커와 잔여 정규화로 교차 토크나이저 정렬 잡음을 줄이며, 단일·다중 교사의 off-policy 증류에서 추론 성능 개선을 보고합니다." }, "claim_source": "https://arxiv.org/abs/2608.29662", "japanese": { "coverage": "固定された正解の解法について、確率の高い候補と残余部分を学習に使います。", "access": "保存した教師の上位128候補のlogitsを使います。", "semantics": "対応づけた高確率候補の分布を使うoff-policy学習です。生徒の生成データを使うOPDとは区別しています。" } } ] }