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" - 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" - 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" - 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" - 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" - 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" - 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" - 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" - 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" - 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" - 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" - 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" - 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" - 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" - 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" - 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." - 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." - 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." - 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." - 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." - 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." - 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." - 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."