{"id":"arxiv:2606.06021","title":"OPRD: On-Policy Representation Distillation","date":"2026-06-04","section":"Post-training","quality":0.460296,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a16a9d7ea78c5748acc7f651c9a96663e0200d72e51d6a52c9528e1739a27d8d/","source_url":"https://arxiv.org/abs/2606.06021","text_available":true,"authors":["Shenzhi Yang","Guangcheng Zhu","Bowen Song","Haobo Wang"],"outgoing":[{"source":"arxiv:2606.06021","target":"arxiv:2104.14294","evidence":"title","id":"arxiv:2104.14294","title":"Emerging Properties in Self-Supervised Vision Transformers","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/4372186d141ca514b08e93bb3b1e26fdbcba9ce62187fb772c14bf223160c06d/"},{"source":"arxiv:2606.06021","target":"arxiv:2405.07987","evidence":"id","id":"arxiv:2405.07987","title":"The Platonic Representation Hypothesis","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/6c7695929f8bbcea43ee98da9cd8f2fe806f2ba292f0b4e3e7fea2b64254afb4/"},{"source":"arxiv:2606.06021","target":"arxiv:2503.14476","evidence":"title","id":"arxiv:2503.14476","title":"DAPO: An Open-Source LLM Reinforcement Learning System at Scale","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/548e934b912ef84691ace63bf1cfd670e7b0257d9d6986dde926f569e7b2f8b7/"},{"source":"arxiv:2606.06021","target":"arxiv:2601.18734","evidence":"id","id":"arxiv:2601.18734","title":"Self-Distilled Reasoner: On-Policy Self-Distillation for Large Language Models","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8a352e37fa79840a4f35f86ebc39661cfc34fbfc4391cd020a69af158c63485a/"},{"source":"arxiv:2606.06021","target":"arxiv:2601.19897","evidence":"id","id":"arxiv:2601.19897","title":"Self-Distillation Enables Continual Learning","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/cc146df3e9ca40c6d111a5264e08710100c1b34bd0eabd9052e8d55d13589efe/"},{"source":"arxiv:2606.06021","target":"arxiv:2601.20802","evidence":"id","id":"arxiv:2601.20802","title":"Reinforcement Learning via Self-Distillation","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/6073ec3d17b9a932a68e79ae078928f4635e40fc79cfcbc31709715d1f1e1ffa/"},{"source":"arxiv:2606.06021","target":"arxiv:2603.24472","evidence":"id","id":"arxiv:2603.24472","title":"Why Does Self-Distillation (Sometimes) Degrade the Reasoning Capability of LLMs?","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/d99edb6fac2f4ad6d63e98ec0a2952187ed2ae190604fbed3507b3ed56d57315/"},{"source":"arxiv:2606.06021","target":"arxiv:2603.25562","evidence":"id","id":"arxiv:2603.25562","title":"Revisiting On-Policy Distillation: Empirical Failure Modes and Simple Fixes","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a66707a05c8aa09328e4a1f322a57800a3e12262a85a9a1cf195b41dc1926b32/"},{"source":"arxiv:2606.06021","target":"arxiv:2604.02288","evidence":"id","id":"arxiv:2604.02288","title":"Unifying Group-Relative and Self-Distillation Policy Optimization via Sample Routing","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/3d1160029d321f2069016f7702d4ebebee0de6cd31a8f296ee220abd6f69c99b/"},{"source":"arxiv:2606.06021","target":"arxiv:2604.13016","evidence":"id","id":"arxiv:2604.13016","title":"Rethinking On-Policy Distillation of Large Language Models: Phenomenology, Mechanism, and Recipe","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/d6d7ff1e2e7b3b9c528f19f7144098920bf573f2cb411fcac5879d5d2cac93e0/"}],"incoming":[],"text_parts":[{"start":0,"end":15480,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a16a9d7ea78c5748acc7f651c9a96663e0200d72e51d6a52c9528e1739a27d8d/part-1.html"},{"start":15480,"end":31269,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a16a9d7ea78c5748acc7f651c9a96663e0200d72e51d6a52c9528e1739a27d8d/part-2.html"},{"start":31269,"end":47239,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a16a9d7ea78c5748acc7f651c9a96663e0200d72e51d6a52c9528e1739a27d8d/part-3.html"},{"start":47239,"end":62317,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a16a9d7ea78c5748acc7f651c9a96663e0200d72e51d6a52c9528e1739a27d8d/part-4.html"},{"start":62317,"end":78225,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a16a9d7ea78c5748acc7f651c9a96663e0200d72e51d6a52c9528e1739a27d8d/part-5.html"},{"start":78225,"end":93947,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a16a9d7ea78c5748acc7f651c9a96663e0200d72e51d6a52c9528e1739a27d8d/part-6.html"},{"start":93947,"end":109232,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a16a9d7ea78c5748acc7f651c9a96663e0200d72e51d6a52c9528e1739a27d8d/part-7.html"},{"start":109232,"end":124603,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a16a9d7ea78c5748acc7f651c9a96663e0200d72e51d6a52c9528e1739a27d8d/part-8.html"},{"start":124603,"end":133588,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a16a9d7ea78c5748acc7f651c9a96663e0200d72e51d6a52c9528e1739a27d8d/part-9.html"}],"contents":[{"title":"OPRD: On-Policy Representation Distillation","offset":0,"level":1,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a16a9d7ea78c5748acc7f651c9a96663e0200d72e51d6a52c9528e1739a27d8d/part-1.html#text"},{"title":"Abstract","offset":47,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a16a9d7ea78c5748acc7f651c9a96663e0200d72e51d6a52c9528e1739a27d8d/part-1.html#text"},{"title":"1 Introduction","offset":1983,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a16a9d7ea78c5748acc7f651c9a96663e0200d72e51d6a52c9528e1739a27d8d/part-1.html#text"},{"title":"2 Background and Problem Setup","offset":13348,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a16a9d7ea78c5748acc7f651c9a96663e0200d72e51d6a52c9528e1739a27d8d/part-1.html#text"},{"title":"2.1 Notation","offset":13795,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a16a9d7ea78c5748acc7f651c9a96663e0200d72e51d6a52c9528e1739a27d8d/part-1.html#text"},{"title":"2.2 The On-Policy Distillation Framework","offset":15420,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a16a9d7ea78c5748acc7f651c9a96663e0200d72e51d6a52c9528e1739a27d8d/part-1.html#text"},{"title":"Setup.","offset":15466,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a16a9d7ea78c5748acc7f651c9a96663e0200d72e51d6a52c9528e1739a27d8d/part-1.html#text"},{"title":"Objective.","offset":16618,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a16a9d7ea78c5748acc7f651c9a96663e0200d72e51d6a52c9528e1739a27d8d/part-2.html#text"},{"title":"2.3 Three Output-Space Variants","offset":17733,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a16a9d7ea78c5748acc7f651c9a96663e0200d72e51d6a52c9528e1739a27d8d/part-2.html#text"},{"title":"(a) Sampled-token OPD ( S t = { y ^ t } S_{t}=\\{\\hat{y}_{t}\\} ).","offset":18103,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a16a9d7ea78c5748acc7f651c9a96663e0200d72e51d6a52c9528e1739a27d8d/part-2.html#text"},{"title":"(b) Full-vocabulary OPD ( S t = 𝒱 S_{t}=\\mathcal{V} ).","offset":19237,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a16a9d7ea78c5748acc7f651c9a96663e0200d72e51d6a52c9528e1739a27d8d/part-2.html#text"},{"title":"(c) Top- k k OPD ( S t = TopK ⁡ ( p t , k ) S_{t}=\\mathrm{TopK}(p_{t},k) ).","offset":19932,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a16a9d7ea78c5748acc7f651c9a96663e0200d72e51d6a52c9528e1739a27d8d/part-2.html#text"},{"title":"2.4 A Shared Structural Limitation","offset":21166,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a16a9d7ea78c5748acc7f651c9a96663e0200d72e51d6a52c9528e1739a27d8d/part-2.html#text"},{"title":"3 On-Policy Representation Distillation","offset":22614,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a16a9d7ea78c5748acc7f651c9a96663e0200d72e51d6a52c9528e1739a27d8d/part-2.html#text"},{"title":"3.1 The OPRD Objective","offset":23113,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a16a9d7ea78c5748acc7f651c9a96663e0200d72e51d6a52c9528e1739a27d8d/part-2.html#text"},{"title":"3.2 Why OPRD Works","offset":26447,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a16a9d7ea78c5748acc7f651c9a96663e0200d72e51d6a52c9528e1739a27d8d/part-2.html#text"},{"title":"Theorem 1 (Zero-variance gradient) .","offset":26745,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a16a9d7ea78c5748acc7f651c9a96663e0200d72e51d6a52c9528e1739a27d8d/part-2.html#text"},{"title":"Comparison with top- k k : no truncation bias.","offset":27945,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a16a9d7ea78c5748acc7f651c9a96663e0200d72e51d6a52c9528e1739a27d8d/part-2.html#text"},{"title":"Theorem 2 (Hidden-state information beyond the LM head) .","offset":29026,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a16a9d7ea78c5748acc7f651c9a96663e0200d72e51d6a52c9528e1739a27d8d/part-2.html#text"},{"title":"3.3 OPRD-Bridge: Cross-Architecture Extension","offset":31218,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a16a9d7ea78c5748acc7f651c9a96663e0200d72e51d6a52c9528e1739a27d8d/part-2.html#text"},{"title":"3.3.1 The Shared Low-Rank Structure","offset":33227,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a16a9d7ea78c5748acc7f651c9a96663e0200d72e51d6a52c9528e1739a27d8d/part-3.html#text"},{"title":"3.3.2 Stage 1: Bridge Construction","offset":35159,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a16a9d7ea78c5748acc7f651c9a96663e0200d72e51d6a52c9528e1739a27d8d/part-3.html#text"},{"title":"3.3.3 Stage 2: Bridge-Based Distillation","offset":38282,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a16a9d7ea78c5748acc7f651c9a96663e0200d72e51d6a52c9528e1739a27d8d/part-3.html#text"},{"title":"3.3.4 Design Philosophy","offset":39682,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a16a9d7ea78c5748acc7f651c9a96663e0200d72e51d6a52c9528e1739a27d8d/part-3.html#text"},{"title":"Theorem 3 (Optimality and bias–variance trade-off of the low-rank bridge, informal) .","offset":43194,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a16a9d7ea78c5748acc7f651c9a96663e0200d72e51d6a52c9528e1739a27d8d/part-3.html#text"},{"title":"4 Experiments","offset":45439,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a16a9d7ea78c5748acc7f651c9a96663e0200d72e51d6a52c9528e1739a27d8d/part-3.html#text"},{"title":"4.1 OPRD-Vanilla: Same-Architecture Distillation","offset":45606,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a16a9d7ea78c5748acc7f651c9a96663e0200d72e51d6a52c9528e1739a27d8d/part-3.html#text"},{"title":"4.1.1 Experimental Setup","offset":46194,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a16a9d7ea78c5748acc7f651c9a96663e0200d72e51d6a52c9528e1739a27d8d/part-3.html#text"},{"title":"Models.","offset":46225,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a16a9d7ea78c5748acc7f651c9a96663e0200d72e51d6a52c9528e1739a27d8d/part-3.html#text"},{"title":"Training data.","offset":46920,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a16a9d7ea78c5748acc7f651c9a96663e0200d72e51d6a52c9528e1739a27d8d/part-3.html#text"},{"title":"Distillation objectives.","offset":47239,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a16a9d7ea78c5748acc7f651c9a96663e0200d72e51d6a52c9528e1739a27d8d/part-4.html#text"},{"title":"Optimization.","offset":48306,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a16a9d7ea78c5748acc7f651c9a96663e0200d72e51d6a52c9528e1739a27d8d/part-4.html#text"},{"title":"Evaluation.","offset":48667,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a16a9d7ea78c5748acc7f651c9a96663e0200d72e51d6a52c9528e1739a27d8d/part-4.html#text"},{"title":"4.1.2 Main Results","offset":49153,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a16a9d7ea78c5748acc7f651c9a96663e0200d72e51d6a52c9528e1739a27d8d/part-4.html#text"},{"title":"4.1.3 Training Dynamics","offset":51312,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a16a9d7ea78c5748acc7f651c9a96663e0200d72e51d6a52c9528e1739a27d8d/part-4.html#text"},{"title":"Accuracy curves: OPRD climbs monotonically, OPD plateaus.","offset":51717,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a16a9d7ea78c5748acc7f651c9a96663e0200d72e51d6a52c9528e1739a27d8d/part-4.html#text"},{"title":"Behavioural view: OPRD produces shorter, more efficient reasoning.","offset":52970,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a16a9d7ea78c5748acc7f651c9a96663e0200d72e51d6a52c9528e1739a27d8d/part-4.html#text"},{"title":"Internal view: OPRD’s own loss is being optimised end-to-end.","offset":53828,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a16a9d7ea78c5748acc7f651c9a96663e0200d72e51d6a52c9528e1739a27d8d/part-4.html#text"},{"title":"4.1.4 Efficiency","offset":54857,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a16a9d7ea78c5748acc7f651c9a96663e0200d72e51d6a52c9528e1739a27d8d/part-4.html#text"},{"title":"4.1.5 Mechanistic Analysis","offset":57792,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a16a9d7ea78c5748acc7f651c9a96663e0200d72e51d6a52c9528e1739a27d8d/part-4.html#text"},{"title":"Composing OPD with OPRD ( μ \\mu sweep).","offset":58550,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a16a9d7ea78c5748acc7f651c9a96663e0200d72e51d6a52c9528e1739a27d8d/part-4.html#text"},{"title":"Where does the student diverge from the teacher? (motivation for last- k k supervision).","offset":59959,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a16a9d7ea78c5748acc7f651c9a96663e0200d72e51d6a52c9528e1739a27d8d/part-4.html#text"},{"title":"(a) Policy-gradient loss: OPRD accelerates distillation and validates the bottleneck theory.","offset":62217,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a16a9d7ea78c5748acc7f651c9a96663e0200d72e51d6a52c9528e1739a27d8d/part-4.html#text"},{"title":"(b) Top- 16 16 overlap: hidden-state alignment propagates to next-token agreement.","offset":63519,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a16a9d7ea78c5748acc7f651c9a96663e0200d72e51d6a52c9528e1739a27d8d/part-5.html#text"},{"title":"(c) Predictive entropy.","offset":65006,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a16a9d7ea78c5748acc7f651c9a96663e0200d72e51d6a52c9528e1739a27d8d/part-5.html#text"},{"title":"4.2 OPRD-Bridge: Cross-Architecture and Cross-Tokenizer Distillation","offset":66397,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a16a9d7ea78c5748acc7f651c9a96663e0200d72e51d6a52c9528e1739a27d8d/part-5.html#text"},{"title":"4.2.1 Experimental Setup","offset":66842,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a16a9d7ea78c5748acc7f651c9a96663e0200d72e51d6a52c9528e1739a27d8d/part-5.html#text"},{"title":"Models.","offset":66873,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a16a9d7ea78c5748acc7f651c9a96663e0200d72e51d6a52c9528e1739a27d8d/part-5.html#text"},{"title":"Bridge construction (Stage 1).","offset":67764,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a16a9d7ea78c5748acc7f651c9a96663e0200d72e51d6a52c9528e1739a27d8d/part-5.html#text"},{"title":"Distillation (Stage 2).","offset":68427,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a16a9d7ea78c5748acc7f651c9a96663e0200d72e51d6a52c9528e1739a27d8d/part-5.html#text"},{"title":"Baselines.","offset":69048,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a16a9d7ea78c5748acc7f651c9a96663e0200d72e51d6a52c9528e1739a27d8d/part-5.html#text"},{"title":"Evaluation.","offset":69383,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a16a9d7ea78c5748acc7f651c9a96663e0200d72e51d6a52c9528e1739a27d8d/part-5.html#text"},{"title":"4.2.2 Main Results","offset":70206,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a16a9d7ea78c5748acc7f651c9a96663e0200d72e51d6a52c9528e1739a27d8d/part-5.html#text"},{"title":"4.2.3 Ablation Studies","offset":75647,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a16a9d7ea78c5748acc7f651c9a96663e0200d72e51d6a52c9528e1739a27d8d/part-5.html#text"},{"title":"Effect of bridge rank r r .","offset":75771,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a16a9d7ea78c5748acc7f651c9a96663e0200d72e51d6a52c9528e1739a27d8d/part-5.html#text"},{"title":"4.2.4 Cross-Tokenizer Distillation","offset":77578,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a16a9d7ea78c5748acc7f651c9a96663e0200d72e51d6a52c9528e1739a27d8d/part-5.html#text"},{"title":"5 Discussion","offset":79742,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a16a9d7ea78c5748acc7f651c9a96663e0200d72e51d6a52c9528e1739a27d8d/part-6.html#text"},{"title":"From representational alignment to cross-architecture distillation.","offset":79759,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a16a9d7ea78c5748acc7f651c9a96663e0200d72e51d6a52c9528e1739a27d8d/part-6.html#text"},{"title":"High-value application 1: multi-model RL merging.","offset":80496,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a16a9d7ea78c5748acc7f651c9a96663e0200d72e51d6a52c9528e1739a27d8d/part-6.html#text"},{"title":"High-value application 2: on-policy self-distillation (OPSD).","offset":81599,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a16a9d7ea78c5748acc7f651c9a96663e0200d72e51d6a52c9528e1739a27d8d/part-6.html#text"},{"title":"Toward alignment-aware pre-training.","offset":82306,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a16a9d7ea78c5748acc7f651c9a96663e0200d72e51d6a52c9528e1739a27d8d/part-6.html#text"},{"title":"6 Related Work","offset":83790,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a16a9d7ea78c5748acc7f651c9a96663e0200d72e51d6a52c9528e1739a27d8d/part-6.html#text"},{"title":"Output-Space Knowledge Distillation.","offset":84143,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a16a9d7ea78c5748acc7f651c9a96663e0200d72e51d6a52c9528e1739a27d8d/part-6.html#text"},{"title":"On-Policy Distillation.","offset":85226,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a16a9d7ea78c5748acc7f651c9a96663e0200d72e51d6a52c9528e1739a27d8d/part-6.html#text"},{"title":"Feature / Intermediate-Representation Distillation.","offset":86593,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a16a9d7ea78c5748acc7f651c9a96663e0200d72e51d6a52c9528e1739a27d8d/part-6.html#text"},{"title":"Hint Learning, Auxiliary Losses, and Distribution Matching.","offset":88753,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a16a9d7ea78c5748acc7f651c9a96663e0200d72e51d6a52c9528e1739a27d8d/part-6.html#text"},{"title":"7 Conclusion and Future Work","offset":89645,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a16a9d7ea78c5748acc7f651c9a96663e0200d72e51d6a52c9528e1739a27d8d/part-6.html#text"},{"title":"Future Work.","offset":92057,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a16a9d7ea78c5748acc7f651c9a96663e0200d72e51d6a52c9528e1739a27d8d/part-6.html#text"},{"title":"Appendix A Notation Summary","offset":95424,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a16a9d7ea78c5748acc7f651c9a96663e0200d72e51d6a52c9528e1739a27d8d/part-7.html#text"},{"title":"Appendix B Formal Theoretical Guarantees","offset":95456,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a16a9d7ea78c5748acc7f651c9a96663e0200d72e51d6a52c9528e1739a27d8d/part-7.html#text"},{"title":"B.1 Setup and Assumptions","offset":96567,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a16a9d7ea78c5748acc7f651c9a96663e0200d72e51d6a52c9528e1739a27d8d/part-7.html#text"},{"title":"Definition 1 (Stochastic gradient estimators) .","offset":96598,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a16a9d7ea78c5748acc7f651c9a96663e0200d72e51d6a52c9528e1739a27d8d/part-7.html#text"},{"title":"Assumption 1 (Standard regularity) .","offset":97837,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a16a9d7ea78c5748acc7f651c9a96663e0200d72e51d6a52c9528e1739a27d8d/part-7.html#text"},{"title":"B.2 Variance of Sampled-Token OPD","offset":99076,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a16a9d7ea78c5748acc7f651c9a96663e0200d72e51d6a52c9528e1739a27d8d/part-7.html#text"},{"title":"Lemma 1 (Score-function decomposition of OPD gradient) .","offset":99115,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a16a9d7ea78c5748acc7f651c9a96663e0200d72e51d6a52c9528e1739a27d8d/part-7.html#text"},{"title":"Proof.","offset":99606,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a16a9d7ea78c5748acc7f651c9a96663e0200d72e51d6a52c9528e1739a27d8d/part-7.html#text"},{"title":"Theorem 4 (OPD gradient variance lower bound) .","offset":100979,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a16a9d7ea78c5748acc7f651c9a96663e0200d72e51d6a52c9528e1739a27d8d/part-7.html#text"},{"title":"Proof.","offset":102227,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a16a9d7ea78c5748acc7f651c9a96663e0200d72e51d6a52c9528e1739a27d8d/part-7.html#text"},{"title":"B.3 OPRD Gradient Is Deterministic","offset":105010,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a16a9d7ea78c5748acc7f651c9a96663e0200d72e51d6a52c9528e1739a27d8d/part-7.html#text"},{"title":"Theorem 5 (OPRD has zero conditional variance) .","offset":105050,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a16a9d7ea78c5748acc7f651c9a96663e0200d72e51d6a52c9528e1739a27d8d/part-7.html#text"},{"title":"Proof.","offset":105562,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a16a9d7ea78c5748acc7f651c9a96663e0200d72e51d6a52c9528e1739a27d8d/part-7.html#text"},{"title":"Corollary 1 (Variance gap) .","offset":107052,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a16a9d7ea78c5748acc7f651c9a96663e0200d72e51d6a52c9528e1739a27d8d/part-7.html#text"},{"title":"B.4 Signal-to-Noise Ratio Collapse of Sampled-Token OPD","offset":107761,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a16a9d7ea78c5748acc7f651c9a96663e0200d72e51d6a52c9528e1739a27d8d/part-7.html#text"},{"title":"Definition 2 (Signal-to-noise ratio) .","offset":108139,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a16a9d7ea78c5748acc7f651c9a96663e0200d72e51d6a52c9528e1739a27d8d/part-7.html#text"},{"title":"Theorem 6 (SNR collapse for OPD, SNR stability for OPRD) .","offset":108627,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a16a9d7ea78c5748acc7f651c9a96663e0200d72e51d6a52c9528e1739a27d8d/part-7.html#text"},{"title":"Proof.","offset":109217,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a16a9d7ea78c5748acc7f651c9a96663e0200d72e51d6a52c9528e1739a27d8d/part-7.html#text"},{"title":"Remark 1 (Interpretation: late-stage stagnation of pure OPD) .","offset":112450,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a16a9d7ea78c5748acc7f651c9a96663e0200d72e51d6a52c9528e1739a27d8d/part-8.html#text"},{"title":"Sub-summary (Perspective 1).","offset":113492,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a16a9d7ea78c5748acc7f651c9a96663e0200d72e51d6a52c9528e1739a27d8d/part-8.html#text"},{"title":"B.5 Formal Results for Theorem 2 : LM-Head Information Bottleneck","offset":113870,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a16a9d7ea78c5748acc7f651c9a96663e0200d72e51d6a52c9528e1739a27d8d/part-8.html#text"},{"title":"Definition 3 (Effective null space of the LM head) .","offset":115418,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a16a9d7ea78c5748acc7f651c9a96663e0200d72e51d6a52c9528e1739a27d8d/part-8.html#text"},{"title":"Lemma 2 (Softmax kernel) .","offset":115776,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a16a9d7ea78c5748acc7f651c9a96663e0200d72e51d6a52c9528e1739a27d8d/part-8.html#text"},{"title":"Proof.","offset":116099,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a16a9d7ea78c5748acc7f651c9a96663e0200d72e51d6a52c9528e1739a27d8d/part-8.html#text"},{"title":"Theorem 7 (Null-direction identity; formal version of ( 9 )) .","offset":116633,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a16a9d7ea78c5748acc7f651c9a96663e0200d72e51d6a52c9528e1739a27d8d/part-8.html#text"},{"title":"Proof.","offset":116922,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a16a9d7ea78c5748acc7f651c9a96663e0200d72e51d6a52c9528e1739a27d8d/part-8.html#text"},{"title":"Lemma 3 (Local Lipschitzness of output-space losses) .","offset":117763,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a16a9d7ea78c5748acc7f651c9a96663e0200d72e51d6a52c9528e1739a27d8d/part-8.html#text"},{"title":"Proof sketch.","offset":118456,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a16a9d7ea78c5748acc7f651c9a96663e0200d72e51d6a52c9528e1739a27d8d/part-8.html#text"},{"title":"Theorem 8 (Spectral gap; formal version of ( 10 )) .","offset":119170,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a16a9d7ea78c5748acc7f651c9a96663e0200d72e51d6a52c9528e1739a27d8d/part-8.html#text"},{"title":"Proof.","offset":119966,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a16a9d7ea78c5748acc7f651c9a96663e0200d72e51d6a52c9528e1739a27d8d/part-8.html#text"},{"title":"Remark 2 (Intermediate layers) .","offset":121520,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a16a9d7ea78c5748acc7f651c9a96663e0200d72e51d6a52c9528e1739a27d8d/part-8.html#text"},{"title":"Sub-summary ( Theorem 2 ).","offset":122390,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a16a9d7ea78c5748acc7f651c9a96663e0200d72e51d6a52c9528e1739a27d8d/part-8.html#text"},{"title":"Overall summary of § B .","offset":123083,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a16a9d7ea78c5748acc7f651c9a96663e0200d72e51d6a52c9528e1739a27d8d/part-8.html#text"},{"title":"B.6 Formal Results for Theorem 3 : Optimality of the Low-Rank Bridge","offset":123695,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a16a9d7ea78c5748acc7f651c9a96663e0200d72e51d6a52c9528e1739a27d8d/part-8.html#text"},{"title":"B.6.1 Setup","offset":123942,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a16a9d7ea78c5748acc7f651c9a96663e0200d72e51d6a52c9528e1739a27d8d/part-8.html#text"},{"title":"Assumption 2 (Monotone alignment decay) .","offset":125319,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a16a9d7ea78c5748acc7f651c9a96663e0200d72e51d6a52c9528e1739a27d8d/part-9.html#text"},{"title":"B.6.2 Rate–Distortion Optimality","offset":125861,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a16a9d7ea78c5748acc7f651c9a96663e0200d72e51d6a52c9528e1739a27d8d/part-9.html#text"},{"title":"Theorem 9 (PCA bridge is rate–distortion optimal) .","offset":125900,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a16a9d7ea78c5748acc7f651c9a96663e0200d72e51d6a52c9528e1739a27d8d/part-9.html#text"},{"title":"Proof.","offset":126616,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a16a9d7ea78c5748acc7f651c9a96663e0200d72e51d6a52c9528e1739a27d8d/part-9.html#text"},{"title":"B.6.3 Bias–Variance Decomposition of Distillation Error","offset":127497,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a16a9d7ea78c5748acc7f651c9a96663e0200d72e51d6a52c9528e1739a27d8d/part-9.html#text"},{"title":"Definition 4 (Distillation error at rank r r ) .","offset":127699,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a16a9d7ea78c5748acc7f651c9a96663e0200d72e51d6a52c9528e1739a27d8d/part-9.html#text"},{"title":"Theorem 10 (Bias–variance decomposition) .","offset":128138,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a16a9d7ea78c5748acc7f651c9a96663e0200d72e51d6a52c9528e1739a27d8d/part-9.html#text"},{"title":"Proof.","offset":129495,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a16a9d7ea78c5748acc7f651c9a96663e0200d72e51d6a52c9528e1739a27d8d/part-9.html#text"},{"title":"Instructions for reporting errors","offset":132624,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a16a9d7ea78c5748acc7f651c9a96663e0200d72e51d6a52c9528e1739a27d8d/part-9.html#text"}],"text_code_points":133588,"source_markdown":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a16a9d7ea78c5748acc7f651c9a96663e0200d72e51d6a52c9528e1739a27d8d/source.md"}