{"card_id": "new_fact_111", "paper_id": "orca_2026_07", "type": "method_summary", "claim": "Orca核心架构:统一世界潜空间+三目标预训练,backbone冻结只训readout", "evidence_ids": ["Orca.ev_0007", "Orca.ev_0008", "Orca.ev_0025"], "interpretation": "Orca用预训练VLM(基于Qwen)作backbone,通过两种互补学习范式构建统一世界潜空间:无意识学习(unconscious learning,观察-only状态转移,用连续视频做密集监督)+有意识学习(conscious learning,事件条件状态转移+VQA,用文本条件学习稀疏的、与决策/任务结果相关的状态转移)。backbone始终冻结,针对下游任务(文本/图像/动作)只训练对应的readout模块,不动backbone本身。", "confidence": "high", "tags": ["embodied_ai", "world_model", "vla", "action_generation", "robotics", "latent_state", "state_prediction", "具身智能", "世界模型"], "meta": {"paper": "Orca", "type": "fact", "title": "Orca核心架构:统一世界潜空间+三目标预训练,backbone冻结只训readout", "content": "Orca用预训练VLM(基于Qwen)作backbone,通过两种互补学习范式构建统一世界潜空间:无意识学习(unconscious learning,观察-only状态转移,用连续视频做密集监督)+有意识学习(conscious learning,事件条件状态转移+VQA,用文本条件学习稀疏的、与决策/任务结果相关的状态转移)。backbone始终冻结,针对下游任务(文本/图像/动作)只训练对应的readout模块,不动backbone本身。", "source": "new_paper", "lifecycle": "active", "prior_confidence": 0.7, "review_due": "2026-12-31", "posterior_log": [], "heuristic": "冻结backbone只训readout,验证潜空间本身质量", "falsification": "若后续消融显示解冻backbone联合微调能显著超越冻结方案,则'冻结backbone即可验证潜空间有效性'这一实验设计假设的说服力被削弱。", "connections": [{"type": "validates", "target": "new_fact_032", "votes": "2/3", "auto": true}, {"type": "validates", "target": "caip_fact_004", "votes": "2/3", "auto": true}, {"type": "validates", "target": "new_fact_088", "votes": "2/3", "auto": true}, {"type": "validates", "target": "new_boundary_031", "votes": "2/3", "auto": true}], "original_type": "fact"}}
{"card_id": "new_fact_112", "paper_id": "orca_2026_07", "type": "method_summary", "claim": "专门构建125K小时视频+1.6亿事件标注,但当前训练只用了十分之一", "evidence_ids": ["Orca.ev_0007", "Orca.ev_0068"], "interpretation": "为支撑Orca的学习范式,专门构建了包含125K小时视频和1.6亿事件标注的大规模数据集(inventory data),覆盖第一人称交互、第三人称操作、无动作机器人执行、事件级转移等。但受限于模型容量,当前训练实际只用了这份数据的十分之一。", "confidence": "high", "tags": ["embodied_ai", "world_model", "vla", "action_generation", "robotics", "latent_state", "state_prediction", "具身智能", "世界模型"], "meta": {"paper": "Orca", "type": "fact", "title": "专门构建125K小时视频+1.6亿事件标注,但当前训练只用了十分之一", "content": "为支撑Orca的学习范式,专门构建了包含125K小时视频和1.6亿事件标注的大规模数据集(inventory data),覆盖第一人称交互、第三人称操作、无动作机器人执行、事件级转移等。但受限于模型容量,当前训练实际只用了这份数据的十分之一。", "source": "new_paper", "lifecycle": "active", "prior_confidence": 0.7, "review_due": "2026-12-31", "posterior_log": [], "heuristic": "数据规模与模型容量要匹配,单独扩大数据不够", "falsification": "若后续用更大模型规模把这份数据全部用上,性能没有随数据利用率提升而明显改善,则'当前受限于模型容量而非数据规模'这一判断需要重新考虑,可能存在其他瓶颈。", "connections": [], "isolated": {"searched_n": 20, "method": "bge-m3+deepseek", "votes": 3, "date": "2026-07-05"}, "original_type": "fact"}}
{"card_id": "new_fact_113", "paper_id": "orca_2026_07", "type": "method_summary", "claim": "动作生成对比:让Qwen3.5从0%成功率突破,且追平大规模机器人数据预训练的π0.5", "evidence_ids": ["Orca.ev_0055", "Orca.ev_0056", "Orca.ev_0057"], "interpretation": "在相同的'Action Expert从零训练'设置下(V-JEPA 2.1/Qwen3.5/Orca的backbone均冻结,只有Action Expert可训练),Orca在所有OOD设置上让Qwen3.5的表现从0%成功率实现突破;同时Orca的表现与π0.5(在大规模机器人数据上预训练过的强VLA基线)相当。这说明Orca学到的世界潜空间本身包含了对动作生成有效的信息,不需要像π0.5那样依赖大规模机器人专属数据预训练。", "confidence": "high", "tags": ["embodied_ai", "world_model", "vla", "action_generation", "robotics", "latent_state", "state_prediction", "具身智能", "世界模型"], "meta": {"paper": "Orca", "type": "fact", "title": "动作生成对比:让Qwen3.5从0%成功率突破,且追平大规模机器人数据预训练的π0.5", "content": "在相同的'Action Expert从零训练'设置下(V-JEPA 2.1/Qwen3.5/Orca的backbone均冻结,只有Action Expert可训练),Orca在所有OOD设置上让Qwen3.5的表现从0%成功率实现突破;同时Orca的表现与π0.5(在大规模机器人数据上预训练过的强VLA基线)相当。这说明Orca学到的世界潜空间本身包含了对动作生成有效的信息,不需要像π0.5那样依赖大规模机器人专属数据预训练。", "source": "new_paper", "lifecycle": "active", "prior_confidence": 0.65, "review_due": "2026-12-31", "posterior_log": [], "heuristic": "世界模型预训练可能是替代大规模机器人数据预训练的路径", "falsification": "若独立第三方在更多样的真机OOD任务集合上复现,发现Orca与π0.5的差距明显大于本文报告的'相当'水平,则这一结论的适用范围需要限定为本文测试的特定任务集合。", "connections": [{"type": "validates", "target": "new_fact_025", "votes": "3/3", "auto": true}, {"type": "validates", "target": "caip_fact_004", "votes": "2/3", "auto": true}, {"type": "validates", "target": "new_fact_014", "votes": "2/3", "auto": true}, {"type": "validates", "target": "new_boundary_031", "votes": "3/3", "auto": true}], "original_type": "fact"}}
{"card_id": "new_fact_114", "paper_id": "orca_2026_07", "type": "experiment_result", "claim": "文本生成对比:Orca-4B相对Qwen3.5-4B提升非均匀,空间关系几乎持平", "evidence_ids": ["Orca.ev_0115"], "interpretation": "跨基准通用能力对比(Orca-4B vs Qwen3.5-4B):状态转移+12.27%、常识推理+5.19%、空间关系+0.57%(几乎持平)、动态运动+8.52%。四个维度提升幅度差异明显,不是均匀提升,空间关系这一项的收益微乎其微。", "confidence": "high", "tags": ["embodied_ai", "world_model", "vla", "action_generation", "robotics", "latent_state", "state_prediction", "具身智能", "世界模型"], "meta": {"paper": "Orca", "type": "fact", "title": "文本生成对比:Orca-4B相对Qwen3.5-4B提升非均匀,空间关系几乎持平", "content": "跨基准通用能力对比(Orca-4B vs Qwen3.5-4B):状态转移+12.27%、常识推理+5.19%、空间关系+0.57%(几乎持平)、动态运动+8.52%。四个维度提升幅度差异明显,不是均匀提升,空间关系这一项的收益微乎其微。", "source": "new_paper", "lifecycle": "active", "prior_confidence": 0.75, "review_due": "2026-12-31", "posterior_log": [], "heuristic": "均值宣称要拆解到各维度分别核实", "falsification": "若独立复现显示空间关系这一维度的提升幅度与其他三个维度相近(而非接近于零),则'非均匀提升'这一具体判断需要修正。", "connections": [{"type": "validates", "target": "fe0_fact_004", "votes": "2/3", "auto": true}, {"type": "validates", "target": "fe0_fact_006", "votes": "2/3", "auto": true}], "original_type": "fact"}}
{"card_id": "new_fact_115", "paper_id": "orca_2026_07", "type": "experiment_result", "claim": "图像预测基准:小模型(0.8B)不敌专用图像生成基线,大模型(4B)反超", "evidence_ids": ["Orca.ev_0116"], "interpretation": "PRICE-V0.1图像预测基准(多个大模型作裁判打分,数值越高越好):专用图像生成基线OmniGen2得39.6±10.2、FLUX.1-Kontext得40.9±13.5、FLUX.2[klein]得56.1±18.1。Orca-0.8B仅得34.5±15.3,不敌全部三个专用基线;但Orca-4B得59.8±10.9,反超FLUX.2[klein]。同一套方法,仅因模型规模从0.8B升到4B,排名从垫底变成第一。", "confidence": "high", "tags": ["embodied_ai", "world_model", "vla", "action_generation", "robotics", "latent_state", "state_prediction", "具身智能", "世界模型"], "meta": {"paper": "Orca", "type": "fact", "title": "图像预测基准:小模型(0.8B)不敌专用图像生成基线,大模型(4B)反超", "content": "PRICE-V0.1图像预测基准(多个大模型作裁判打分,数值越高越好):专用图像生成基线OmniGen2得39.6±10.2、FLUX.1-Kontext得40.9±13.5、FLUX.2[klein]得56.1±18.1。Orca-0.8B仅得34.5±15.3,不敌全部三个专用基线;但Orca-4B得59.8±10.9,反超FLUX.2[klein]。同一套方法,仅因模型规模从0.8B升到4B,排名从垫底变成第一。", "source": "new_paper", "lifecycle": "active", "prior_confidence": 0.75, "review_due": "2026-12-31", "posterior_log": [], "heuristic": "模型规模是决定性变量,小规模实验结果不能直接外推到大规模", "falsification": "若后续更大规模(如10B+)的Orca版本,图像预测得分相对4B版本没有进一步提升甚至下降,则'规模是决定性变量'这一判断需要修正,可能存在规模收益递减或饱和点。", "connections": [], "isolated": {"searched_n": 20, "method": "L1-coarse", "date": "2026-07-05"}, "original_type": "fact"}}
{"card_id": "new_fact_116", "paper_id": "orca_2026_07", "type": "method_summary", "claim": "三种下游读出机制:文本直出、图像接SD3.5解码器、动作接Action Expert", "evidence_ids": ["Orca.ev_0026", "Orca.ev_0027"], "interpretation": "Orca把学到的世界潜空间暴露为三种下游读出:文本读出——直接过LM head生成语言,不接额外解码器;图像读出——潜空间经MLP adaptor后接预训练的Stable Diffusion 3.5做多步去噪解码,训练时只训MLP adaptor和LoRA参数;动作读出——接Action Expert。三种读出都建立在同一个冻结的世界潜空间之上。", "confidence": "high", "tags": ["embodied_ai", "world_model", "vla", "action_generation", "robotics", "latent_state", "state_prediction", "具身智能", "世界模型"], "meta": {"paper": "Orca", "type": "fact", "title": "三种下游读出机制:文本直出、图像接SD3.5解码器、动作接Action Expert", "content": "Orca把学到的世界潜空间暴露为三种下游读出:文本读出——直接过LM head生成语言,不接额外解码器;图像读出——潜空间经MLP adaptor后接预训练的Stable Diffusion 3.5做多步去噪解码,训练时只训MLP adaptor和LoRA参数;动作读出——接Action Expert。三种读出都建立在同一个冻结的世界潜空间之上。", "source": "new_paper", "lifecycle": "active", "prior_confidence": 0.7, "review_due": "2026-12-31", "posterior_log": [], "heuristic": "统一潜空间+轻量级readout,而非每个任务单独重训整个模型", "falsification": "若消融显示不同readout之间存在明显的负迁移(训练某个readout会损害另一个readout的表现),则'三种读出可以共享同一个冻结潜空间'这一设计假设需要打折。", "connections": [], "isolated": {"searched_n": 20, "method": "bge-m3+deepseek", "votes": 3, "date": "2026-07-05"}, "original_type": "fact"}}
{"card_id": "new_fact_117", "paper_id": "orca_2026_07", "type": "method_summary", "claim": "训练基础设施优化:显存高效损失函数+激活重计算,吞吐提升约4.4倍", "evidence_ids": ["Orca.ev_0031", "Orca.ev_0097"], "interpretation": "采用Chunked Cross-Entropy Loss避免计算完整logits时的显存爆炸,并应用激活重计算(activation recomputation)。这些优化使训练吞吐从0.66提升到2.91 samples/sec/GPU,相对具身社区常用的StarVLA基线,加速约4.4倍。", "confidence": "high", "tags": ["embodied_ai", "world_model", "vla", "action_generation", "robotics", "latent_state", "state_prediction", "具身智能", "世界模型"], "meta": {"paper": "Orca", "type": "fact", "title": "训练基础设施优化:显存高效损失函数+激活重计算,吞吐提升约4.4倍", "content": "采用Chunked Cross-Entropy Loss避免计算完整logits时的显存爆炸,并应用激活重计算(activation recomputation)。这些优化使训练吞吐从0.66提升到2.91 samples/sec/GPU,相对具身社区常用的StarVLA基线,加速约4.4倍。", "source": "new_paper", "lifecycle": "active", "prior_confidence": 0.7, "review_due": "2026-12-31", "posterior_log": [], "heuristic": "工程基础设施优化是规模化训练的前提条件", "falsification": "若独立复现在相同硬件配置下无法达到接近4.4倍的加速比,则该数字的可信度应下调,需怀疑是否存在评测口径差异。", "connections": [], "isolated": {"searched_n": 20, "method": "bge-m3+deepseek", "votes": 3, "date": "2026-07-05"}, "original_type": "fact"}}
{"card_id": "new_claim_118", "paper_id": "orca_2026_07", "type": "core_claim", "claim": "世界模型预训练可能是替代大规模机器人数据预训练的路径", "evidence_ids": ["Orca.ev_0055", "Orca.ev_0056", "Orca.ev_0057"], "interpretation": "综合new_fact_113的对比结果(从零训练动作专家+Orca潜空间,追平大规模机器人数据预训练的π0.5),推断:高质量的通用世界潜空间预训练,可能是一条不需要依赖海量机器人专属数据、就能达到接近水平动作生成能力的路径。这是跨证据的推断,论文本身未明确以这种方式定性表述,是从对比数据反推出的判断。", "confidence": "medium", "tags": ["embodied_ai", "world_model", "vla", "action_generation", "robotics", "latent_state", "state_prediction", "具身智能", "世界模型"], "meta": {"paper": "Orca", "type": "claim", "title": "世界模型预训练可能是替代大规模机器人数据预训练的路径", "content": "综合new_fact_113的对比结果(从零训练动作专家+Orca潜空间,追平大规模机器人数据预训练的π0.5),推断:高质量的通用世界潜空间预训练,可能是一条不需要依赖海量机器人专属数据、就能达到接近水平动作生成能力的路径。这是跨证据的推断,论文本身未明确以这种方式定性表述,是从对比数据反推出的判断。", "source": "new_paper", "lifecycle": "active", "prior_confidence": 0.45, "review_due": "2026-12-31", "posterior_log": [], "heuristic": "跨对比数据反推的路线判断,非论文直接陈述,置信度打折", "falsification": "若后续在更大规模、更多样化的真机部署场景中,世界模型预训练路线的动作生成能力显著落后于大规模机器人数据预训练路线(而非'相当'),则此claim被削弱,说明当前'追平'结果可能只是特定任务集合下的个例,不构成路线替代的普遍证据。", "connections": [{"type": "validates", "target": "legacy_fact_017", "votes": "3/3", "auto": true}, {"type": "validates", "target": "new_claim_076", "votes": "2/3", "auto": true}, {"type": "validates", "target": "legacy_fact_008", "votes": "3/3", "auto": true}, {"type": "validates", "target": "new_fact_014", "votes": "3/3", "auto": true}, {"type": "refines", "target": "new_boundary_080", "votes": "2/3", "auto": true}, {"type": "validates", "target": "new_fact_026", "votes": "3/3", "auto": true}, {"type": "validates", "target": "new_fact_032", "votes": "3/3", "auto": true}, {"type": "validates", "target": "new_fact_025", "votes": "3/3", "auto": true}, {"type": "validates", "target": "new_fact_075", "votes": "3/3", "auto": true}], "original_type": "claim"}}
{"card_id": "new_claim_119", "paper_id": "orca_2026_07", "type": "core_claim", "claim": "作者展望:统一状态转移世界表征未来可扩展到AI for science等复杂系统", "evidence_ids": ["Orca.ev_0079"], "interpretation": "论文Future Works部分展望:逐步把具身智能扩展到AI for science、微观量子力学、宏观宇宙学、生命科学等复杂系统,用统一的状态转移世界表征支撑科学发现和认知边界的扩展。这是纯前瞻性展望,当前论文没有任何实验支撑这一跨域扩展,是作者对研究方向的期望,不是已验证结果。", "confidence": "low", "tags": ["embodied_ai", "world_model", "vla", "action_generation", "robotics", "latent_state", "state_prediction", "具身智能", "世界模型"], "meta": {"paper": "Orca", "type": "claim", "title": "作者展望:统一状态转移世界表征未来可扩展到AI for science等复杂系统", "content": "论文Future Works部分展望:逐步把具身智能扩展到AI for science、微观量子力学、宏观宇宙学、生命科学等复杂系统,用统一的状态转移世界表征支撑科学发现和认知边界的扩展。这是纯前瞻性展望,当前论文没有任何实验支撑这一跨域扩展,是作者对研究方向的期望,不是已验证结果。", "source": "new_paper", "lifecycle": "active", "prior_confidence": 0.2, "review_due": "2026-12-31", "posterior_log": [], "heuristic": "纯前瞻性展望,无当前实验支撑,置信度应低", "falsification": "若后续确实有工作把类似的统一状态转移世界表征方法应用到量子力学/宇宙学/生命科学等非具身领域并取得有意义的结果,则此展望应被上调置信度;若长期(3年以上)无任何跟进工作尝试这一方向,应视为未兑现的空头支票。", "connections": [{"type": "validates", "target": "new_claim_076", "votes": "3/3", "auto": true}, {"type": "extends", "target": "new_boundary_080", "votes": "2/3", "auto": true}, {"type": "extends", "target": "fe0_boundary_001", "votes": "2/3", "auto": true}, {"type": "validates", "target": "claim_unitachand_20260627_001", "votes": "2/3", "auto": true}], "original_type": "claim"}}
{"card_id": "new_boundary_120", "paper_id": "orca_2026_07", "type": "negative_boundary", "claim": "模型规模不够,语言/图像/动作三种读出能力之间存在此消彼长的权衡", "evidence_ids": ["Orca.ev_0068"], "interpretation": "作者在Limitation章节明确承认:受资源限制,当前实验主要在4B和0.8B规模进行,不足以充分整合更多世界知识、更多模态、更多数据。4B模型在预训练过程中表现出语言、图像、动作读出性能之间的权衡,这种权衡在0.8B模型上更明显。尽管构建了125K小时视频和1.6亿事件标注,当前训练只用了十分之一的库存数据——说明世界学习不仅受数据规模限制,也需要足够的模型容量。", "confidence": "high", "tags": ["embodied_ai", "world_model", "vla", "action_generation", "robotics", "latent_state", "state_prediction", "具身智能", "世界模型"], "meta": {"paper": "Orca", "type": "boundary", "title": "模型规模不够,语言/图像/动作三种读出能力之间存在此消彼长的权衡", "content": "作者在Limitation章节明确承认:受资源限制,当前实验主要在4B和0.8B规模进行,不足以充分整合更多世界知识、更多模态、更多数据。4B模型在预训练过程中表现出语言、图像、动作读出性能之间的权衡,这种权衡在0.8B模型上更明显。尽管构建了125K小时视频和1.6亿事件标注,当前训练只用了十分之一的库存数据——说明世界学习不仅受数据规模限制,也需要足够的模型容量。", "source": "new_paper", "lifecycle": "active", "prior_confidence": 0.8, "review_due": "2026-12-31", "posterior_log": [], "heuristic": "作者自述局限性优先级最高(诚实信号)", "falsification": "若后续用更大规模模型(如10B+)训练,语言/图像/动作三种读出能力之间的权衡明显缓解或消失,则此boundary可以降级为'小规模阶段性局限,已解决'。", "connections": [{"type": "validates", "target": "new_boundary_016", "votes": "3/3", "auto": true}, {"type": "validates", "target": "fe0_fact_005", "votes": "2/3", "auto": true}], "not_claim": "Orca已经证明了在任意模型规模下都能同时兼顾语言/图像/动作三种能力", "because": ["作者Limitation章节原文明确承认这一权衡在小规模模型上更明显", "当前实验规模(4B/0.8B)远小于当前主流大模型规模"], "failure_mode": ["模型容量不足时,提升某一种读出能力的训练信号,可能挤占/削弱另一种读出能力的表现"], "operational_cost": ["要真正验证这套世界潜空间范式的上限,需要投入远超4B的模型规模和更大比例的库存数据,这部分工程/算力成本论文本身尚未投入"], "original_type": "boundary"}}
{"card_id": "new_boundary_121", "paper_id": "orca_2026_07", "type": "negative_boundary", "claim": "用冻结预训练视觉编码器做监督目标是设计妥协,作者自己承认非理想方案", "evidence_ids": ["Orca.ev_0067"], "interpretation": "Orca当前用预训练VLM+冻结视觉编码器的输出作为状态转移的监督目标(ground truth)。作者明确承认:这种做法简化了训练流程,但也把学到的状态空间和某个预训练模型的语义空间对齐了——一个真正通用的世界基础模型,应该直接从多源信号学习统一的世界空间,这些信号应共同定义和约束状态,而不是依赖任何单一预训练模态空间作为监督目标。", "confidence": "high", "tags": ["embodied_ai", "world_model", "vla", "action_generation", "robotics", "latent_state", "state_prediction", "具身智能", "世界模型"], "meta": {"paper": "Orca", "type": "boundary", "title": "用冻结预训练视觉编码器做监督目标是设计妥协,作者自己承认非理想方案", "content": "Orca当前用预训练VLM+冻结视觉编码器的输出作为状态转移的监督目标(ground truth)。作者明确承认:这种做法简化了训练流程,但也把学到的状态空间和某个预训练模型的语义空间对齐了——一个真正通用的世界基础模型,应该直接从多源信号学习统一的世界空间,这些信号应共同定义和约束状态,而不是依赖任何单一预训练模态空间作为监督目标。", "source": "new_paper", "lifecycle": "active", "prior_confidence": 0.75, "review_due": "2026-12-31", "posterior_log": [], "heuristic": "作者自己区分'当前简化方案'与'理想方案',这层区分要如实保留", "falsification": "若后续版本改用多源信号联合监督(不依赖单一预训练视觉模型)训练出的Orca,下游各项读出能力相比当前版本没有明显提升,则'依赖单一预训练模态空间是值得担忧的限制'这一判断的实际影响程度需要下调。", "connections": [], "isolated": {"searched_n": 20, "method": "bge-m3+deepseek", "votes": 3, "date": "2026-07-05"}, "not_claim": "Orca当前的监督方式就是构建世界基础模型的最终/理想方案", "because": ["作者原文明确指出这是'naive setup'(朴素设置),并明确提出了'理想方案应该是什么'的对照标准"], "failure_mode": ["依赖单一预训练视觉模型的语义空间做监督,可能让学到的'世界状态'带有该预训练模型自身的偏差/局限性,而非真正独立的物理世界表征"], "operational_cost": ["要摆脱对单一预训练模态空间的依赖、改为直接从多源信号学习,需要重新设计监督目标和训练范式,不是当前架构的简单扩展"], "original_type": "boundary"}}
{"card_id": "new_boundary_122", "paper_id": "orca_2026_07", "type": "negative_boundary", "claim": "PRICE-V0.1图像预测基准规模有限,作者自称只是初步评测", "evidence_ids": ["Orca.ev_0069"], "interpretation": "论文提出的PRICE-V0.1图像预测评测基准,虽然覆盖多个真实世界数据源,但作者自己承认其规模、多样性、交互丰富度仍然有限,希望它能作为迈向更全面真实世界状态预测评测的第一步,而非最终评测标准。", "confidence": "medium", "tags": ["embodied_ai", "world_model", "vla", "action_generation", "robotics", "latent_state", "state_prediction", "具身智能", "世界模型"], "meta": {"paper": "Orca", "type": "boundary", "title": "PRICE-V0.1图像预测基准规模有限,作者自称只是初步评测", "content": "论文提出的PRICE-V0.1图像预测评测基准,虽然覆盖多个真实世界数据源,但作者自己承认其规模、多样性、交互丰富度仍然有限,希望它能作为迈向更全面真实世界状态预测评测的第一步,而非最终评测标准。", "source": "new_paper", "lifecycle": "active", "prior_confidence": 0.6, "review_due": "2026-12-31", "posterior_log": [], "heuristic": "新提出的评测基准,规模/覆盖面局限性优先记录", "falsification": "若后续独立团队在更大规模、更多样化的真实世界数据上复现出与PRICE-V0.1一致的模型排名结果,则该基准的代表性可以上调评级。", "connections": [], "isolated": {"searched_n": 20, "method": "bge-m3+deepseek", "votes": 3, "date": "2026-07-05"}, "not_claim": "PRICE-V0.1已经是评估真实世界状态预测能力的完备、权威基准", "because": ["作者原文明确用'初步'(initial step)定性这个基准,而非'完整'或'最终'"], "failure_mode": ["基准规模/多样性不足时,模型在该基准上的表现可能无法充分代表其在更广泛真实世界场景下的状态预测能力"], "operational_cost": ["扩大基准的规模和多样性需要额外的数据采集与标注投入,论文本身没有投入这部分成本"], "original_type": "boundary"}}
{"card_id": "new_fact_123", "paper_id": "orca_2026_07", "type": "experiment_result", "claim": "抓取失败恢复:Orca能纠偏继续任务,π0.5反复失败后仍不稳定", "evidence_ids": ["Orca.ev_0058", "Orca.ev_0105"], "interpretation": "在抓勺子等真机操作任务中,Orca表现出更强的错误恢复能力:FNS(失败阶段数)更高,说明即便轨迹最终失败,Orca也能推进到更晚的任务阶段才终止;DRR(回撤恢复比率)更高,说明Orca更擅长在进度下降后纠偏、继续推进任务,而不是卡在原地反复失败。对比之下,π0.5在同样任务中抓取失败后表现不稳定,反复尝试无明显进展。", "confidence": "high", "tags": ["embodied_ai", "world_model", "vla", "action_generation", "robotics", "latent_state", "state_prediction", "具身智能", "世界模型"], "meta": {"paper": "Orca", "type": "fact", "title": "抓取失败恢复:Orca能纠偏继续任务,π0.5反复失败后仍不稳定", "content": "在抓勺子等真机操作任务中,Orca表现出更强的错误恢复能力:FNS(失败阶段数)更高,说明即便轨迹最终失败,Orca也能推进到更晚的任务阶段才终止;DRR(回撤恢复比率)更高,说明Orca更擅长在进度下降后纠偏、继续推进任务,而不是卡在原地反复失败。对比之下,π0.5在同样任务中抓取失败后表现不稳定,反复尝试无明显进展。", "source": "new_paper", "lifecycle": "active", "prior_confidence": 0.7, "review_due": "2026-12-31", "posterior_log": [], "heuristic": "失败恢复能力比单次成功率更能反映真实鲁棒性", "falsification": "若独立第三方复现同类失败恢复实验(用FNS/DRR等指标度量),发现Orca与π0.5在恢复能力上的差距明显小于本文报述,则这一优势的实际幅度需要重新评估。", "connections": [], "isolated": {"searched_n": 20, "method": "bge-m3+deepseek", "votes": 3, "date": "2026-07-05"}, "original_type": "fact"}}
{"card_id": "new_fact_124", "paper_id": "orca_2026_07", "type": "method_summary", "claim": "消融实验:三个训练目标缺一不可,单独用观察目标最差(29.3),三者全用最好(48.0)", "evidence_ids": ["Orca.ev_0117", "Orca.ev_0061"], "interpretation": "消融实验对比不同训练目标组合的下游综合表现(文本/图像/动作三项平均):仅用观察-only状态转移(λobs)得29.3分,加上事件条件状态转移(λobs+λevt)提升到44.6分,换成观察+VQA(λobs+λvqa)得41.6分,事件条件+VQA(λevt+λvqa)得42.6分,三者全用(λobs+λevt+λvqa)达到最高48.0分。去掉λobs对动作生成影响最大(该行动作分骤降到10.2),去掉λevt则图像预测直接失效(该行图像分为'-')。", "confidence": "high", "tags": ["embodied_ai", "world_model", "vla", "action_generation", "robotics", "latent_state", "state_prediction", "具身智能", "世界模型"], "meta": {"paper": "Orca", "type": "fact", "title": "消融实验:三个训练目标缺一不可,单独用观察目标最差(29.3),三者全用最好(48.0)", "content": "消融实验对比不同训练目标组合的下游综合表现(文本/图像/动作三项平均):仅用观察-only状态转移(λobs)得29.3分,加上事件条件状态转移(λobs+λevt)提升到44.6分,换成观察+VQA(λobs+λvqa)得41.6分,事件条件+VQA(λevt+λvqa)得42.6分,三者全用(λobs+λevt+λvqa)达到最高48.0分。去掉λobs对动作生成影响最大(该行动作分骤降到10.2),去掉λevt则图像预测直接失效(该行图像分为'-')。", "source": "new_paper", "lifecycle": "active", "prior_confidence": 0.75, "review_due": "2026-12-31", "posterior_log": [], "heuristic": "消融实验的具体数字比'三个目标都重要'这种笼统结论更有判断价值", "falsification": "若独立复现该消融实验,发现某个训练目标组合的表现明显偏离本文报告(如λobs+λevt的44.6分复现出显著更高或更低的分数),则该消融数字的可信度应下调。", "connections": [{"type": "validates", "target": "new_fact_032", "votes": "2/3", "auto": true}], "original_type": "fact"}}