{"paper":{"title":"Dynamic Latent Routing","license":"http://creativecommons.org/licenses/by-sa/4.0/","headline":"Dynamic Latent Routing recovers globally optimal policies by composing learned sub-policies to improve low-data language model fine-tuning.","cross_cats":["cs.AI","cs.CL"],"primary_cat":"cs.LG","authors_text":"Amir Abdullah, Fangyuan Yu, Xin Su","submitted_at":"2026-05-14T03:35:46Z","abstract_excerpt":"We investigate the temporal concatenation of sub-policies in Markov Decision Processes (MDP) with time-varying reward functions. We introduce General Dijkstra Search (GDS), and prove that globally optimal goal-reaching policies can be recovered through temporal composition of intermediate optimal sub-policies. Motivated by the \"search, select, update\" principle underlying GDS, we propose Dynamic Latent Routing (DLR), a language-model post-training method that jointly learns discrete latent codes, routing policies, and model parameters through dynamic search in a single training stage. In low-d"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"In low-data fine-tuning settings, DLR matches or outperforms supervised fine-tuning across four datasets and six models, achieving a mean gain of +6.6 percentage points, while prior discrete-latent baselines consistently underperform SFT.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That the optimality guarantees and search principle from General Dijkstra Search in MDPs transfer effectively to the non-stationary, high-dimensional setting of language model post-training without introducing hidden biases or optimization instabilities.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"Dynamic Latent Routing jointly learns discrete latent codes, routing policies, and model parameters via dynamic search to match or exceed supervised fine-tuning by 6.6 points on average in low-data settings across four datasets and six models.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Dynamic Latent Routing recovers globally optimal policies by composing learned sub-policies to improve low-data language model fine-tuning.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"f8e4659fe691fbe54d74086337cf1ec7da3a4a2afd57db837e1e2232949e86d2"},"source":{"id":"2605.14323","kind":"arxiv","version":1},"verdict":{"id":"16c71107-db8a-4430-acef-4ad0a8557dd1","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-15T01:45:27.516595Z","strongest_claim":"In low-data fine-tuning settings, DLR matches or outperforms supervised fine-tuning across four datasets and six models, achieving a mean gain of +6.6 percentage points, while prior discrete-latent baselines consistently underperform SFT.","one_line_summary":"Dynamic Latent Routing jointly learns discrete latent codes, routing policies, and model parameters via dynamic search to match or exceed supervised fine-tuning by 6.6 points on average in low-data settings across four datasets and six models.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That the optimality guarantees and search principle from General Dijkstra Search in MDPs transfer effectively to the non-stationary, high-dimensional setting of language model post-training without introducing hidden biases or optimization instabilities.","pith_extraction_headline":"Dynamic Latent Routing recovers globally optimal policies by composing learned sub-policies to improve low-data language model fine-tuning."},"references":{"count":70,"sample":[{"doi":"","year":2017,"title":"Hindsight experience replay","work_id":"b4ff437e-eec1-45ab-9ee2-e70f0e4c45a4","ref_index":1,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2026,"title":"Introducing Claude Opus 4.7","work_id":"93e7fbe6-fe60-46cb-9cca-d936e25cd4c7","ref_index":2,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2017,"title":"The option-critic architecture","work_id":"eca5b255-3362-4279-8e47-baa1fa132831","ref_index":3,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2017,"title":"Successor features for transfer in reinforcement learning","work_id":"0c3776bc-e342-4c21-adda-1ff1b22ce73c","ref_index":4,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2022,"title":"Probing classifiers: Promises, shortcomings, and advances.Computational Linguistics, 48(1):207–219, 2022","work_id":"14d0b945-57f3-41b0-b491-72316d4c06b5","ref_index":5,"cited_arxiv_id":"","is_internal_anchor":false}],"resolved_work":70,"snapshot_sha256":"de5cc2c99271703570f8a5c460fae39d8ff7b0ade229e6cc096927452a5384a8","internal_anchors":4},"formal_canon":{"evidence_count":2,"snapshot_sha256":"f1863e9f4efa9195501d90122a4c64020c051095898c9a1e196df5aca7dda042"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}