{"slug": "boosting-deepresearch-and-longcontext-ability-with-self-generated-deepresearch", "title": "Boosting Deepresearch and LongContext Ability with Self-Generated Deepresearch Rollouts Traces", "summary": "A new arXiv paper (2609.20844v1) reports that DLD-RL, a three-stage training method that converts Deepresearch reinforcement-learning rollout trajectories into long-context QA instances at zero annotation cost, outperforms standard DR-RL by 7.3% on three Deepresearch benchmarks and by 13.5% on three long-context benchmarks. The authors observed that 61.6% of the model's remaining prediction errors after DR-RL stem from insufficient long-context understanding, including long-context hallucination and failures in cross-document evidence integration. The method, DR Rollouts to LongContext-QA (DR-to-Long), replaces compact snippets and webpage summaries in each trajectory with the full contents of their corresponding URLs to produce longer multi-document contexts while preserving original evidence relationships.", "body_md": "arXiv:2609.20844v1 Announce Type: new \nAbstract: Deepresearch (DR) agents interact with real-world web environments through multi-turn search and visit, causing their contexts to grow rapidly over time. We observe that, even after DR Agentic Reinforcement Learning (DR-RL), 61.6% of the model's remaining prediction errors can still be attributed to insufficient long-context understanding, including longcontext hallucination and failures in cross-document evidence integration. It motivates us to further break the bottleneck of DR-RL by strengthening the model's long-context ability. However, effective LongContext training requires more than simply increasing context length. To bridge the data gap, we propose `DR Rollouts to LongContext-QA (DR-to-Long)'. The method repurposes DR-RL trajectories, which naturally contain search histories, visited webpages, evidence snippets, and final-answer supervision. It then replaces the compact snippets and webpage summaries in each trajectory with the full contents of their corresponding URLs, producing substantially longer multi-document contexts while preserving the original evidence relationships. Building on DR-to-Long, we introduce DLD (DR -> LongQA -> DR)-RL. DLD-RL first performs a short DR-RL stage to collect rollout trajectories, which are then converted into LongQA instances at zero annotation cost. The model is subsequently optimized with LongQA-RL to strengthen LongContext ability, followed by full DR-RL to continue improving its DR capability. Experiments show that DLD-RL outperforms standard DR-RL by 7.3% on three Deepresearch benchmarks and improves performance by 13.5% on three long-context benchmarks.", "url": "https://wpnews.pro/news/boosting-deepresearch-and-longcontext-ability-with-self-generated-deepresearch", "canonical_source": "https://arxiv.org/abs/2609.20844", "published_at": "2026-09-21 04:00:00+00:00", "updated_at": "2026-09-21 04:23:43.622739+00:00", "lang": "en", "topics": ["ai-research", "large-language-models", "ai-agents", "machine-learning"], "entities": ["arXiv", "DLD-RL", "DR-RL", "DR-to-Long", "Deepresearch"], "alternates": {"html": "https://wpnews.pro/news/boosting-deepresearch-and-longcontext-ability-with-self-generated-deepresearch", "markdown": "https://wpnews.pro/news/boosting-deepresearch-and-longcontext-ability-with-self-generated-deepresearch.md", "text": "https://wpnews.pro/news/boosting-deepresearch-and-longcontext-ability-with-self-generated-deepresearch.txt", "jsonld": "https://wpnews.pro/news/boosting-deepresearch-and-longcontext-ability-with-self-generated-deepresearch.jsonld"}}