{"id":"arxiv:2512.24601","title":"Recursive Language Models","date":"2025-12-31","section":"LLMs: architectures, context, training","quality":0.153868,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ac41983c934a3432ce7b6531e367b9e639bddbed93d4c6513781e76407d9a8bd/","source_url":"https://arxiv.org/abs/2512.24601","text_available":true,"authors":["Alex L. Zhang","Tim Kraska","Omar Khattab"],"outgoing":[{"source":"arxiv:2512.24601","target":"arxiv:2404.07143","evidence":"id","id":"arxiv:2404.07143","title":"Leave No Context Behind: Efficient Infinite Context Transformers with Infini-attention","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/fcdbb3ceff27039e0eeeeac187ebf2570e2dc8b6f58ed5ee278f4259e88cfab1/"}],"incoming":[{"source":"arxiv:2601.21557","target":"arxiv:2512.24601","evidence":"id","id":"arxiv:2601.21557","title":"Meta Context Engineering via Agentic Skill Evolution","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/bedaeecadf58846f4728c50c03396fc96ba18429ecf14052b4c5e177394a246e/"},{"source":"arxiv:2603.19461","target":"arxiv:2512.24601","evidence":"id","id":"arxiv:2603.19461","title":"Hyperagents","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/17eec8a15debc212b5530f117dd8ba8bad339590729cbbc4f75a47755ebcb544/"},{"source":"arxiv:2603.28052","target":"arxiv:2512.24601","evidence":"id","id":"arxiv:2603.28052","title":"Meta-Harness: End-to-End Optimization of Model Harnesses","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8bb0b87384dfdd42c419cf26cc7e52f1f0d89f63afa5d5968e246b73a44c99cb/"},{"source":"arxiv:2607.07847","target":"arxiv:2512.24601","evidence":"id","id":"arxiv:2607.07847","title":"When Does Continual Learning Require Learning","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/86e88c6cf21a66879eab5c156913fc85d14eb69c394c78c33c0e81d144e528eb/"},{"source":"url:alexzhang13.github.io/blog/2026/harness","target":"arxiv:2512.24601","evidence":"url","id":"url:alexzhang13.github.io/blog/2026/harness","title":"Language model harnesses are compositional generalizers (Alex Zhang)","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/d724c392cf5eb2db98d9a15e510d13b4220810c062a249c6a02ab86ee1895bb5/"}],"text_parts":[{"start":0,"end":15811,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ac41983c934a3432ce7b6531e367b9e639bddbed93d4c6513781e76407d9a8bd/part-1.html"},{"start":15811,"end":31526,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ac41983c934a3432ce7b6531e367b9e639bddbed93d4c6513781e76407d9a8bd/part-2.html"},{"start":31526,"end":47080,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ac41983c934a3432ce7b6531e367b9e639bddbed93d4c6513781e76407d9a8bd/part-3.html"},{"start":47080,"end":62808,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ac41983c934a3432ce7b6531e367b9e639bddbed93d4c6513781e76407d9a8bd/part-4.html"},{"start":62808,"end":67338,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ac41983c934a3432ce7b6531e367b9e639bddbed93d4c6513781e76407d9a8bd/part-5.html"}],"contents":[{"title":"Recursive Language Models","offset":0,"level":1,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ac41983c934a3432ce7b6531e367b9e639bddbed93d4c6513781e76407d9a8bd/part-1.html#text"},{"title":"Abstract","offset":29,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ac41983c934a3432ce7b6531e367b9e639bddbed93d4c6513781e76407d9a8bd/part-1.html#text"},{"title":"1 Introduction","offset":1214,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ac41983c934a3432ce7b6531e367b9e639bddbed93d4c6513781e76407d9a8bd/part-1.html#text"},{"title":"2 Recursive Language Models","offset":6378,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ac41983c934a3432ce7b6531e367b9e639bddbed93d4c6513781e76407d9a8bd/part-1.html#text"},{"title":"3 Scaling Long Context Tasks","offset":11315,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ac41983c934a3432ce7b6531e367b9e639bddbed93d4c6513781e76407d9a8bd/part-1.html#text"},{"title":"3.1 Tasks","offset":12438,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ac41983c934a3432ce7b6531e367b9e639bddbed93d4c6513781e76407d9a8bd/part-1.html#text"},{"title":"3.2 Methods and Baselines","offset":14707,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ac41983c934a3432ce7b6531e367b9e639bddbed93d4c6513781e76407d9a8bd/part-1.html#text"},{"title":"4 Results and Discussion","offset":18304,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ac41983c934a3432ce7b6531e367b9e639bddbed93d4c6513781e76407d9a8bd/part-2.html#text"},{"title":"5 Analyses of RLM Trajectories","offset":25062,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ac41983c934a3432ce7b6531e367b9e639bddbed93d4c6513781e76407d9a8bd/part-2.html#text"},{"title":"6 Related Works","offset":27104,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ac41983c934a3432ce7b6531e367b9e639bddbed93d4c6513781e76407d9a8bd/part-2.html#text"},{"title":"7 Limitations and Future Work","offset":29206,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ac41983c934a3432ce7b6531e367b9e639bddbed93d4c6513781e76407d9a8bd/part-2.html#text"},{"title":"8 Conclusion","offset":30609,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ac41983c934a3432ce7b6531e367b9e639bddbed93d4c6513781e76407d9a8bd/part-2.html#text"},{"title":"Appendix A Additional Training Details","offset":31483,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ac41983c934a3432ce7b6531e367b9e639bddbed93d4c6513781e76407d9a8bd/part-2.html#text"},{"title":"Appendix B Negative Results: Things We Tried That Did Not Work.","offset":35123,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ac41983c934a3432ce7b6531e367b9e639bddbed93d4c6513781e76407d9a8bd/part-3.html#text"},{"title":"Appendix C Additional Methods and Baseline Details","offset":37655,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ac41983c934a3432ce7b6531e367b9e639bddbed93d4c6513781e76407d9a8bd/part-3.html#text"},{"title":"C.1 Prompts for Experiments","offset":37710,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ac41983c934a3432ce7b6531e367b9e639bddbed93d4c6513781e76407d9a8bd/part-3.html#text"},{"title":"C.2 Summary agent baseline","offset":39479,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ac41983c934a3432ce7b6531e367b9e639bddbed93d4c6513781e76407d9a8bd/part-3.html#text"},{"title":"C.3 LongCoT-mini experiment.","offset":40547,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ac41983c934a3432ce7b6531e367b9e639bddbed93d4c6513781e76407d9a8bd/part-3.html#text"},{"title":"Appendix D Additional Benchmark Details","offset":41607,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ac41983c934a3432ce7b6531e367b9e639bddbed93d4c6513781e76407d9a8bd/part-3.html#text"},{"title":"D.1 OOLONG-Pairs Benchmark","offset":41734,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ac41983c934a3432ce7b6531e367b9e639bddbed93d4c6513781e76407d9a8bd/part-3.html#text"},{"title":"D.2 Scaling Huge Document Corpora in BrowseComp+","offset":54780,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ac41983c934a3432ce7b6531e367b9e639bddbed93d4c6513781e76407d9a8bd/part-4.html#text"},{"title":"Appendix E Additional RLM Trajectories","offset":56165,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ac41983c934a3432ce7b6531e367b9e639bddbed93d4c6513781e76407d9a8bd/part-4.html#text"},{"title":"E.1 RLM(GPT-5) on BrowseComp-Plus-Query_74","offset":57207,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ac41983c934a3432ce7b6531e367b9e639bddbed93d4c6513781e76407d9a8bd/part-4.html#text"},{"title":"E.2 RLM(Qwen3-Coder) on OOLONG-Pairs-Query_3","offset":58608,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ac41983c934a3432ce7b6531e367b9e639bddbed93d4c6513781e76407d9a8bd/part-4.html#text"},{"title":"E.3 RLM(Qwen3-Coder) on OOLONG-Query_212","offset":62004,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ac41983c934a3432ce7b6531e367b9e639bddbed93d4c6513781e76407d9a8bd/part-4.html#text"},{"title":"E.4 RLM(GPT-5) on CodeQA-Query_44","offset":63653,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ac41983c934a3432ce7b6531e367b9e639bddbed93d4c6513781e76407d9a8bd/part-5.html#text"},{"title":"Appendix F Additional Quantitative Results","offset":64385,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ac41983c934a3432ce7b6531e367b9e639bddbed93d4c6513781e76407d9a8bd/part-5.html#text"},{"title":"F.1 Additional Quantitative Analysis of Main Results","offset":64432,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ac41983c934a3432ce7b6531e367b9e639bddbed93d4c6513781e76407d9a8bd/part-5.html#text"},{"title":"F.2 Additional Runtime and Cost Analysis of RLMs","offset":65291,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ac41983c934a3432ce7b6531e367b9e639bddbed93d4c6513781e76407d9a8bd/part-5.html#text"},{"title":"Instructions for reporting errors","offset":66374,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ac41983c934a3432ce7b6531e367b9e639bddbed93d4c6513781e76407d9a8bd/part-5.html#text"}],"text_code_points":67338,"source_markdown":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ac41983c934a3432ce7b6531e367b9e639bddbed93d4c6513781e76407d9a8bd/source.md"}