
本期从 arXiv 8月3日 recent 分组精选11篇 AI 论文,拆解世界模型 critic、动态推理、精确分词、KV 重建、Agent 安全与多模态评测如何把中间过程变成可复核接口。



| 设置 | IND avg | OOD avg |
|---|---|---|
| π0 + WCM | 84.4±1.2 | 51.5±1.5 |
| π0.5 + WCM | 91.9±0.4 | 64.4±1.4 |
| OpenVLA-OFT + WCM | 99.0±0.4 | 77.9±0.8 |













| 指标 | TokTier 结果 | 论文给出的边界 |
|---|---|---|
| correctness campaigns divergence | 0 | 17 个 tokenizer families 的测试版本 |
| 1M 字符 GPU encode | 0.87 ms | exact reference-equivalent IDs |
| 1M 字符增量修复 | 约 0.5–1.1 ms(随长度区间) | 受 append/window 影响 |
| 50 ms P99 容量 | 1,821 req/s | 四核 repair pool + 一张 GPU |














| 模型 | In-domain quality | tokens | OOD quality | tokens |
|---|---|---|---|---|
| TwT-Qwen2.5-7B | 66.27 | 274 | 66.25 | 269 |
| TwT-Qwen2.5-14B | 66.62 | 291 | 66.46 | 288 |
| DeepSeek-R1 | 65.40 | 586 | 65.62 | 561 |
| GPT-5 | 66.05 | 660 | 66.29 | 645 |






| 设置 | baseline | +ResKV |
|---|---|---|
| LongBench,LLaMA,10% retained | 45.64 | 47.39 |
| LongBench,LLaMA,20% retained | 47.25 | 48.45 |
| RULER,LLaMA,4K,10% retained | 58.94 | 70.51 |
| RULER,LLaMA,4K,20% retained | 75.44 | 83.34 |




| reader | LoCoMo F1 | memory-operation tokens |
|---|---|---|
| Zero-Mem + GPT-4o-mini | 59.15 | 0 |
| Zero-Mem + Qwen2.5-14B | 57.57 | 0 |
| GAM + GPT-4o-mini | 53.75 | 28,570,674 |






| Benchmark | MoRoute AVG | fixed multi-layer AVG |
|---|---|---|
| IntelligentVBench | 3.98 | 4.00(TIV2V 子集) |
| OpenVE-Bench | 3.83 | 3.63 |
| RefVIE-Bench | 4.11 | — |


| 方法 | Default success | Variant success | Drop |
|---|---|---|---|
| RGB | 42.5 | 24.0 | 18.5 |
| RayViT-cls | 43.2 | 37.0 | 6.2 |
| RayViT | 39.8 | 37.3 | 2.5 |








| 任务 | MOT-SR NMSE | 对照 |
|---|---|---|
| Oscillation 1 | 1.27e-15 | LLM-SR 2.55e-5 |
| LSR-Synth–Chemistry | 3.85e-7 | LLM-SR 8.01e-6 |
| EMRI held-out | 1.17e-3 | LLM-SR 3.13e-2 |










| 方法 | AgentHarm refusal | 平均 accuracy | InjecAgent ASR |
|---|---|---|---|
| Base | 23.8 | 60.9 | 25.6 |
| SafeKeep | 70.6 | 79.6 | 2.5 |







| 系统 | Value F1 | 成本/页 | word grounding F1 | page grounding F1 |
|---|---|---|---|---|
| LlamaExtract Agentic Plus | 95.6 | 8.1¢ | 46.4 | 84.9 |
| Codex GPT-5.5 | 93.6 | 27.8¢ | — | — |
| Reducto Deep | 90.4 | — | 43.3 | 71.7 |
| LlamaExtract Cost-Effective | 86.8 | 1.0¢ | — | — |




| 系统 | Judge@1 | ID@1 | Judge@3 | ID@3 |
|---|---|---|---|---|
| Gemini-3.1-Pro-Preview | 64.7 | 61.4 | 73.4 | 64.0 |
| Claude Opus 4.8 | 64.4 | 55.0 | 72.0 | 59.9 |
| Qwen3.5-122B-A10B | 5.5 | 4.8 | 5.9 | 5.9 |
| Qwen3.5-122B-A10B + SFT | 52.9 | 49.8 | 67.5 | 55.4 |
This story was produced automatically by a channel. One sentence is all it takes for Neodrop to keep producing for you.