{"paper":{"title":"MTA: Multi-Granular Trajectory Alignment for Large Language Model Distillation","license":"http://creativecommons.org/licenses/by/4.0/","headline":"Multi-granular trajectory alignment improves knowledge distillation by matching teacher and student representations at word level in lower layers and phrase level in higher layers.","cross_cats":[],"primary_cat":"cs.CL","authors_text":"Linh Ngo Van, Pham Khanh Chi, Quoc Phong Dao, Thanh Hong Nguyen, Thuat Nguyen, Trung Le","submitted_at":"2026-05-02T10:52:49Z","abstract_excerpt":"Knowledge distillation is a key technique for compressing large language models (LLMs), but most existing methods align representations at fixed layers or token-level outputs, ignoring how representations evolve across depth. As a result, the student is only weakly guided to capture the teacher's internal relational structure during distillation, which limits knowledge transfer. To address this limitation, we propose Multi-Granular Trajectory Alignment (MTA), a framework that aligns teacher and student representations along their layer-wise transformation trajectory. MTA adopts a layer-adaptiv"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"MTA consistently outperforms state-of-the-art baselines on standard benchmarks, with ablations confirming the contribution of each component.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That aligning teacher and student representations along their layer-wise transformation trajectory using a layer-adaptive multi-granular strategy (word-level lower, phrase-level higher) will better guide the student to capture the teacher's internal relational structure than fixed-layer or token-level methods.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"MTA improves LLM knowledge distillation by aligning representations along layer-wise trajectories with adaptive granularity from words to phrases using dynamic structural and hidden representation alignment losses.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Multi-granular trajectory alignment improves knowledge distillation by matching teacher and student representations at word level in lower layers and phrase level in higher layers.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"054a12ff34e7f9874c0f1e6053c3c6a0dbd8e11508c2f494a714da33f122e632"},"source":{"id":"2605.01374","kind":"arxiv","version":2},"verdict":{"id":"0aa38e82-e6fa-4fdd-9d6e-d325616a4875","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-09T14:50:59.346561Z","strongest_claim":"MTA consistently outperforms state-of-the-art baselines on standard benchmarks, with ablations confirming the contribution of each component.","one_line_summary":"MTA improves LLM knowledge distillation by aligning representations along layer-wise trajectories with adaptive granularity from words to phrases using dynamic structural and hidden representation alignment losses.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That aligning teacher and student representations along their layer-wise transformation trajectory using a layer-adaptive multi-granular strategy (word-level lower, phrase-level higher) will better guide the student to capture the teacher's internal relational structure than fixed-layer or token-level methods.","pith_extraction_headline":"Multi-granular trajectory alignment improves knowledge distillation by matching teacher and student representations at word level in lower layers and phrase level in higher layers."},"integrity":{"clean":true,"summary":{"advisory":0,"critical":0,"by_detector":{},"informational":0},"endpoint":"/pith/2605.01374/integrity.json","findings":[],"available":true,"detectors_run":[{"name":"ai_meta_artifact","ran_at":"2026-05-20T18:34:27.146335Z","status":"completed","version":"1.0.0","findings_count":0},{"name":"doi_compliance","ran_at":"2026-05-19T17:19:42.068043Z","status":"completed","version":"1.0.0","findings_count":0}],"snapshot_sha256":"9ae7dd55168262a410c47f3b61079f92c9d350a401bbe530ed675e3582b7d7aa"},"references":{"count":0,"sample":[],"resolved_work":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57","internal_anchors":0},"formal_canon":{"evidence_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}