{"paper":{"title":"KLong: Training LLM Agent for Extremely Long-horizon Tasks","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"KLong shows that trajectory-splitting SFT followed by progressive RL lets a 106B agent outperform a 1T model on extremely long-horizon tasks.","cross_cats":["cs.CL"],"primary_cat":"cs.AI","authors_text":"Bryan Hooi, Flood Sung, Jiaheng Zhang, Xinlong Yang, Yanhao Li, Yibo Miao, Yingwei Ma, Yuchong Xie, Yue Liu, Zhiyuan Hu","submitted_at":"2026-02-19T17:01:08Z","abstract_excerpt":"This paper introduces KLong, an open-source LLM agent trained to solve extremely long-horizon tasks. The principle is to first cold-start the model via trajectory-splitting SFT, then scale it via progressive RL training. Specifically, we first activate basic agentic abilities of a base model with a comprehensive SFT recipe. Then, we introduce Research-Factory, an automated pipeline that generates high-quality training data by collecting research papers and constructing evaluation rubrics. Using this pipeline, we build thousands of long-horizon trajectories distilled from Claude 4.5 Sonnet (Thi"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"our proposed KLong (106B) surpasses Kimi K2 Thinking (1T) by 11.28% on PaperBench, and the performance improvement generalizes to other coding benchmarks like SWE-bench Verified and MLE-bench.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That trajectories distilled from Claude 4.5 Sonnet via the Research-Factory pipeline are high-quality and that trajectory-splitting SFT plus progressive RL stages preserve sufficient context and capability for true long-horizon generalization without hidden failure modes.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"KLong trains a 106B LLM agent using trajectory-splitting SFT and progressive RL to outperform a 1T model by 11.28% on PaperBench with generalization to coding benchmarks.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"KLong shows that trajectory-splitting SFT followed by progressive RL lets a 106B agent outperform a 1T model on extremely long-horizon tasks.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"672e21a8ba139cfd1517a415d54751d0de2d8bc6c4efe0145f7f084dda03212d"},"source":{"id":"2602.17547","kind":"arxiv","version":3},"verdict":{"id":"31e98931-e61b-4679-8934-ce8d62549855","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-15T20:45:41.123186Z","strongest_claim":"our proposed KLong (106B) surpasses Kimi K2 Thinking (1T) by 11.28% on PaperBench, and the performance improvement generalizes to other coding benchmarks like SWE-bench Verified and MLE-bench.","one_line_summary":"KLong trains a 106B LLM agent using trajectory-splitting SFT and progressive RL to outperform a 1T model by 11.28% on PaperBench with generalization to coding benchmarks.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That trajectories distilled from Claude 4.5 Sonnet via the Research-Factory pipeline are high-quality and that trajectory-splitting SFT plus progressive RL stages preserve sufficient context and capability for true long-horizon generalization without hidden failure modes.","pith_extraction_headline":"KLong shows that trajectory-splitting SFT followed by progressive RL lets a 106B agent outperform a 1T model on extremely long-horizon tasks."},"integrity":{"clean":true,"summary":{"advisory":0,"critical":0,"by_detector":{},"informational":0},"endpoint":"/pith/2602.17547/integrity.json","findings":[],"available":true,"detectors_run":[],"snapshot_sha256":"c28c3603d3b5d939e8dc4c7e95fa8dfce3d595e45f758748cecf8e644a296938"},"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"}