Unverified50% confidenceBenchmarkExact time
In synthetic and real indoor scene experiments, compared with the strongest baseline, WorldRecon reduced Chamfer distance by 14.9% and 43.1% on the two test sets, respectively, while improving F-score by 17.3% and 12.0%.
1
Sources
50%
Confidence
Long-term
Relevance
10/14/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedIn real-world RAG pipelines, switching between the top 15 MTEB-ranked embedding models produces performance fluctuations of 15 to 25 percentage points.66% similarUnverifiedAQuA混合模型相比GRU基线的约0.023 IC差距对应约37%的相对提升61% similarUnverifiedAfter matched supervised training, model performance improved to 99.8%, while removing the relational links caused accuracy to plummet to 23.4%, even below the best shortcut's 27.0%.59% similarUnverified在HotpotQA等多跳推理基准上，ReAct相比Act-only方案的任务完成率提升约10-15个百分点，同时幻觉率显著降低59% similarUnverified探索性能上界实验发现现有验证粒度并非最优，理想情况下计算量（FLOPs）可减少75倍，同时准确率提升1.1%59% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1026707API
curl https://kongchang.com/api/v1/knowledge/claims/1026707MCP
get_claim(id=1026707)