{ "schema_version": "paper_public_manifest_v1", "paper_id": "hyperworld_hypergraph_structured_state_serialization_improves_learned_textual_wo_2026_09", "slug": "hyperworld_hypergraph_structured_state_serialization_improves_learned_textual_wo", "title": "HyperWorld:拆开,泛化掉31分", "authors": [], "source": { "arxiv_id": "2609.00002", "pdf_url": "https://arxiv.org/pdf/2609.00002", "project_url": "", "github_url": "", "huggingface_url": "", "original_source": "https://arxiv.org/abs/2609.00002" }, "site": { "post_url": "/posts/hyperworld_hypergraph_structured_state_serialization_improves_learned_textual_wo", "canonical_url": "https://haiguangboy.com/posts/hyperworld_hypergraph_structured_state_serialization_improves_learned_textual_wo", "cover_image": "https://static.haiguangboy.com/papers/hyperworld_hypergraph_structured_state_serialization_improves_learned_textual_wo/cover.webp" }, "taxonomy": { "domain": "embodied_ai", "track": "world_model", "tasks": [ "embodied_ai", "world_model", "action_generation", "robotics", "state_prediction", "具身智能", "世界模型", "状态表示" ], "related_topics": [ { "paper_id": "skild_blog_introducing_s1_in_context_learning_20260817_2026_08", "title": "S1:机器人的GPT-3涌现时刻", "url": "https://haiguangboy.com/posts/skild_blog_introducing_s1_in_context_learning_20260817", "relation": "same_track", "summary": "S1:机器人的GPT-3涌现时刻", "strength": "strong" }, { "paper_id": "sunday_blog_20260717_2026_07", "title": "sunday_blog_20260717", "url": "https://haiguangboy.com/posts/sunday_blog_20260717", "relation": "same_track", "summary": "模型规模越大,结构化输入的优势越被稀释(3B下二元组能追平甚至反超超边);训练数据只有10%时,四种表示方式在分布外表现都差不多,结构优势要等数据量到25%以上才明显显现 validates 预训练规模收敛泛化差距：82→10→8→4→0pp，域外14%→100%", "strength": "strong" }, { "paper_id": "learning_a_thousand_tasks_in_a_day_2026_08", "title": "1 天 1000 任务，靠的是归纳偏置", "url": "https://haiguangboy.com/posts/mt3-thousand-tasks", "relation": "contrast", "summary": "不代表结构化输入在任何数据量下都有用——训练数据只有10%时,四种序列化方式在分布外表现都差不多,结构上的优势在极端小数据场景下不成立 contradicts 关键的反向趋势:分解法在 1–10 条时快速起量但 50 条附近见顶,单体 BC 在 10–50 区间加速追赶", "strength": "medium" }, { "paper_id": "prism_precision_and_contact_rich_real_world_industrial_skill_dataset_with_multim_2026_08", "title": "PRISM:规模能否替代显式建模?", "url": "https://haiguangboy.com/posts/prism_precision_and_contact_rich_real_world_industrial_skill_dataset_with_multim", "relation": "same_track", "summary": "PRISM:规模能否替代显式建模?", "strength": "medium" }, { "paper_id": "joyai_ra_05_scaling_robot_manipulation_learning_via_dual_action_alignment_2026_08", "title": "异构数据的问题不是量，是监督格式不统一", "url": "https://haiguangboy.com/posts/joyai-ra-dual-alignment", "relation": "same_track", "summary": "把学到的世界模型接入一个简单贪心规划器,在30局未见过的真实游戏上测:超边世界模型规划成功率76.7%,远超二元组(56.7%)和句子(53.3%),平均步数也更少 validates 世界模型消融:有无 WM 均分 48.4 对 51.5;换成潜在动作条件版(LAC-WM)后从 87.3 提到 92.1", "strength": "medium" }, { "paper_id": "flex_pi_a_multi_stream_world_action_model_with_compute_flexibility_2026_08", "title": "把RGB、3D几何、物体语义揉进同一个表征——这才是它比π0.5、Fast-WAM更准的根本原因", "url": "https://haiguangboy.com/posts/flex_pi_a_multi_stream_world_action_model_with_compute_flexibility", "relation": "same_track", "summary": "FLEX-π:更全面的联合表征", "strength": "medium" } ] }, "analyst_take": { "type": "author_opinion", "text": "\"结构/方法的优势需要一个数据量门槛才会显现\"这条判断，这次在HyperWorld身上又被验证了一遍——和之前解读过的几篇工作方向一致：Skild自己承认ICL在小数据量区间反而不如传统微调；FLEX-π自称依然\"很吃数据\"；PRISM的预训练收益要在全量数据上才明显。语言模型、机器人操作、文本世界模型，三个差异很大的领域，反复得出同一个模式：不是\"结构好就一定赢\"，是\"数据攒够之后，结构好的方案才能把优势兑现出来\"。另外这篇有个值得记的态度——作者没有把超边包装成全面最优，老老实实报告了它在3B分布内、单纯可行性判定上都不是最强，这种克制比很多\"全胜叙事\"的论文更值得信任。" }, "ruling": { "importance_score": 3.0, "one_sentence": "HyperWorld:拆开,泛化掉31分" }, "asset_base_url": "https://static.haiguangboy.com/papers/hyperworld_hypergraph_structured_state_serialization_improves_learned_textual_wo", "assets": [ { "type": "pdf_screenshot", "object_key": "papers/hyperworld_hypergraph_structured_state_serialization_improves_learned_textual_wo/page_01.webp", "content_type": "image/webp", "upload_status": "uploaded", "bucket": "paper-assets", "url": "https://static.haiguangboy.com/papers/hyperworld_hypergraph_structured_state_serialization_improves_learned_textual_wo/page_01.webp", "role": "paper_first_page", "size_bytes": 165422 }, { "type": "pdf_screenshot", "object_key": "papers/hyperworld_hypergraph_structured_state_serialization_improves_learned_textual_wo/key_figure.webp", "content_type": "image/webp", "upload_status": "uploaded", "bucket": "paper-assets", "url": "https://static.haiguangboy.com/papers/hyperworld_hypergraph_structured_state_serialization_improves_learned_textual_wo/key_figure.webp", "role": "method_figure", "size_bytes": 146082 }, { "type": "cover_image", "object_key": "papers/hyperworld_hypergraph_structured_state_serialization_improves_learned_textual_wo/cover.webp", "content_type": "image/webp", "upload_status": "uploaded", "bucket": "paper-assets", "url": "https://static.haiguangboy.com/papers/hyperworld_hypergraph_structured_state_serialization_improves_learned_textual_wo/cover.webp", "role": "post_cover", "size_bytes": 98278 }, { "type": "public_brief", "object_key": "papers/hyperworld_hypergraph_structured_state_serialization_improves_learned_textual_wo/public_brief.md", "content_type": "text/markdown; charset=utf-8", "upload_status": "uploaded", "bucket": "paper-assets", "url": "https://static.haiguangboy.com/papers/hyperworld_hypergraph_structured_state_serialization_improves_learned_textual_wo/public_brief.md", "role": "public_brief", "size_bytes": 5237 }, { "type": "public_brief", "object_key": "papers/hyperworld_hypergraph_structured_state_serialization_improves_learned_textual_wo/public_brief.en.md", "content_type": "text/markdown; charset=utf-8", "upload_status": "uploaded", "bucket": "paper-assets", "url": "https://static.haiguangboy.com/papers/hyperworld_hypergraph_structured_state_serialization_improves_learned_textual_wo/public_brief.en.md", "role": "public_brief_en", "size_bytes": 6281 }, { "type": "public_manifest", "object_key": "papers/hyperworld_hypergraph_structured_state_serialization_improves_learned_textual_wo/public_manifest.en.json", "content_type": "application/json; charset=utf-8", "upload_status": "uploaded", "bucket": "paper-assets", "url": "https://static.haiguangboy.com/papers/hyperworld_hypergraph_structured_state_serialization_improves_learned_textual_wo/public_manifest.en.json", "role": "public_manifest_en", "size_bytes": 8446 }, { "type": "public_manifest", "object_key": "papers/hyperworld_hypergraph_structured_state_serialization_improves_learned_textual_wo/public_manifest.json", "content_type": "application/json; charset=utf-8", "upload_status": "uploaded", "bucket": "paper-assets", "url": "https://static.haiguangboy.com/papers/hyperworld_hypergraph_structured_state_serialization_improves_learned_textual_wo/public_manifest.json", "role": "public_manifest", "size_bytes": 9179 } ], "published_at": "2026-09-07T20:13:12+08:00", "created_at": "2026-09-07T20:13:12+08:00", "updated_at": "2026-09-07T20:13:26+08:00" }