中文Sign inStart reading

Open claims

Judgments waiting to be verified

29 · Review window: 90 days from verification.

Summary you can quote

DisconfirmAI is tracking 29 self-reported claims. As of 2026-09-19, none has been independently reproduced. Each carries a review date and the evidence that would overturn it; results are written back when the date arrives, and revisions stay public.

ClaimReported figureOriginal materialReview dateStatus
前OpenAI研究者创办的TypeSafe发布非LLM模型Jev，输出校准概率而非文本自测 · 未经独立验证—techcrunch.com2026-12-17Pending
单作者预印本自报：Fathom按查询自适应读取K缓存位数，百万token解码步较136位稀疏扫描快1.67倍自报 · 未经独立验证1.67倍arxiv.org2026-12-17Pending
预印本自报：SmolVLA转ONNX延迟减半但Spatial成功率跌至41%，语言宽度24可恢复至75%自报41% / 75%arxiv.org2026-12-17Pending
预印本自报：M2Tok多头多码本动作分词降低重建损失、提升VLA成功率自报—arxiv.org2026-12-17Pending
预印本自报：UTTO以不确定性引导测试时优化，提升深度-only开放词汇3D分割自报 · 未经独立验证—arxiv.org2026-12-17Pending
预印本自报：TeleAntiFraud 2.0上，近域负样本使诈骗检测分类器Macro-F1从满分降至0.65–0.68自报—arxiv.org2026-12-17Pending
WoundAIssist远程伤口护理论文标注ACM期刊发表，可用性结论系作者自报自报—arxiv.org2026-12-17Pending
预印本自报：CSWAM将RoboTwin 2.0 Clean-to-Randomized成功率从10.16%提至45.18%，仍未过半自报 · 自测10.16% / 45.18%arxiv.org2026-12-17Pending
预印本自报：LCS单遍引导称VLM驾驶推理延迟较两遍CFG降约50%，数字待独立复现自报 · 待独立复现50%arxiv.org2026-12-17Pending
预印本自报：FRAUDSkill冻结权重、只在外部优化技能，TeleAntiFraud上Macro-F1报73.50%自报 · 未经独立验证73.50%arxiv.org2026-12-17Pending
预印本自报：四款医用深度传感器对比，Zivid 2M+ 60全项最优，ZED 2i体模垫底自报 · 第一方—arxiv.org2026-12-17Pending
预印本自报：RECAL称在DARPA E3入侵检测数据集上将平均误报率较最低FPR基线降低4至105倍自报105倍arxiv.org2026-12-17Pending
预印本自报：单作者以统计力学解释双下降，称增参等效增强权重正则化自报—arxiv.org2026-12-17Pending
预印本自报：VT-MUSE统一视触觉时序表征，仿真基准领先所评最强基线11个百分点自报 · 未经独立验证11个百分点arxiv.org2026-12-17Pending
预印本自报：VLA适配成本可先诊断后分配，xArm-7上0.04%参数匹配全量微调自报0.04%arxiv.org2026-12-17Pending
预印本自报：以生成接触音的响度塑造力曲线，Franka机器人零样本完成擦拭、削皮自报 · 未经独立验证—arxiv.org2026-12-17Pending
Anthropic自报：Claude已主导公司26%的AI研发，3月不足1%自报 · 第一方26% / 1%businessinsider.com2026-12-17Pending
EfficientTDMPC预印本自报样本效率新最优，未经独立复现自报—arxiv.org2026-12-17Pending
DeepMind研究所发文：可见思维链是安全优势，OpenAI系统卡已报告可监测性下降厂商自述—the-decoder.com2026-12-17Pending
苹果发表DSAS：引导强度不固定，按输入和层动态计算自报 · 尚无独立验证—machinelearning.apple.com2026-12-17Pending
Anthropic首次披露研发自动化指标：“AI主导”占26%，打分的也是Claude尚无独立验证26%the-decoder.com2026-12-17Pending
三人团队在OpenAI漏洞赏金计划下用Claude Opus 5接管员工账号，OpenAI称已修复第一方—techcrunch.com2026-12-17Pending
42位英国皇家学会数学院士联名警告AI存在性风险，敦促学会提醒政府与媒体未经独立验证—the-decoder.com2026-12-17Pending
OpenAI推出法律AI产品Astra for Law，自测基准200题通过率54%自测 · 第一方54%the-decoder.com2026-12-17Pending
Napster与GEMS Education签约，在迪拜开展AI教师数字人概念验证未经独立验证—wired.com2026-12-17Pending
AWS发布SageMaker HyperPod Inference Gateway，TTFT最多降98%出自自家基准未经独立验证 · 第一方98%aws.amazon.com2026-12-17Pending
预印本自报：自进化智能体技能池过临界规模后污染难回滚，VaG门控在Terminal-Bench 2达72% pass@1自报 · 自测72%arxiv.org2026-12-17Pending
MoRE预印本自报：相邻层组共享专家池，三规模同等预算下困惑度低于标准MoE自报 · 未经独立验证—arxiv.org2026-12-17Pending
G2G预印本：冻结基座仅训约32M参数做跨组位姿估计，自报四数据集两项任务SOTA自报—arxiv.org2026-12-17Pending

Claims appear in their original language (zh) and are not rewritten before independent verification. English summaries are added by the site desk, one at a time.

Back to overview

Judgments waiting to be verified | DisconfirmAI