From 0a80a28cc5522931021e58f25d4840ef366be6d8 Mon Sep 17 00:00:00 2001
From: LaoWang <257199637@qq.com>
Date: Tue, 2 Jun 2026 12:53:44 +0000
Subject: [PATCH] =?UTF-8?q?feat:=20=E6=96=B0=E5=A2=9E=209=20=E7=AF=87?=
 =?UTF-8?q?=E5=85=B3=E9=94=AE=E7=BC=BA=E4=BD=8D=E8=AE=BA=E6=96=87=20(189?=
 =?UTF-8?q?=E7=AF=87)?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

架构 5 篇:
- MiniMax-01: Lightning Attention + MoE (2501.08313)
- Byte Latent Transformer: tokenization-free (2412.09871)
- Large Concept Models: 句子级推理 (2412.08821)
- xLSTM: LSTM复兴, matrix memory (2405.04517)
- LLaDA: Diffusion LLM, 非自回归 (2502.09992)

后训练 2 篇:
- MiniMax-M1: CISPO RL 推理涌现 (2506.13585)
- Coconut: 潜在空间连续推理 (2412.06769)

Agent 1 篇:
- Generative Agents: Stanford AI Town (2304.03442)

多模态 1 篇:
- GPT-4V/4o/5.4: 视觉-语音原生多模态
---
 data/papers.json | 90 ++++++++++++++++++++++++++++++++++++++++++++++++
 1 file changed, 90 insertions(+)

diff --git a/data/papers.json b/data/papers.json
index 03524a1..f79a4dc 100644
--- a/data/papers.json
+++ b/data/papers.json
@@ -188,6 +188,16 @@
               "前沿"
             ],
             "arxiv": "2407.08608"
+          },
+          {
+            "title": "MiniMax-01: Lightning Attention + MoE (456B, 45.9B active)",
+            "authors": "MiniMax",
+            "year": 2025,
+            "venue": "arXiv",
+            "tags": [
+              "关键节点"
+            ],
+            "arxiv": "2501.08313"
           }
         ],
         "branches": [
@@ -526,6 +536,16 @@
               "支线"
             ],
             "arxiv": "2411.13676"
+          },
+          {
+            "title": "xLSTM: Extended Long Short-Term Memory (sLSTM + mLSTM, matrix memory, linear attention revival)",
+            "authors": "Beck et al.",
+            "year": 2024,
+            "venue": "arXiv",
+            "tags": [
+              "支线"
+            ],
+            "arxiv": "2405.04517"
           }
         ],
         "forward": [
@@ -548,6 +568,36 @@
               "前瞻"
             ],
             "arxiv": "2406.02528"
+          },
+          {
+            "title": "Byte Latent Transformer: Patches Scale Better Than Tokens (tokenization-free, entropy-based patching)",
+            "authors": "Pagnoni et al. / Meta FAIR",
+            "year": 2024,
+            "venue": "arXiv",
+            "tags": [
+              "前瞻"
+            ],
+            "arxiv": "2412.09871"
+          },
+          {
+            "title": "Large Concept Models: Language Modeling in a Sentence Representation Space (SONAR, 200 languages)",
+            "authors": "Baranchuk et al. / Meta FAIR",
+            "year": 2024,
+            "venue": "arXiv",
+            "tags": [
+              "前瞻"
+            ],
+            "arxiv": "2412.08821"
+          },
+          {
+            "title": "LLaDA: Large Language Diffusion with mAsking (diffusion-based LLM, 8B, non-autoregressive generation)",
+            "authors": "Nie et al.",
+            "year": 2025,
+            "venue": "arXiv",
+            "tags": [
+              "前瞻"
+            ],
+            "arxiv": "2502.09992"
           }
         ]
       }
@@ -613,6 +663,16 @@
               "前沿"
             ],
             "arxiv": "2602.02276"
+          },
+          {
+            "title": "GPT-4V / GPT-4o / GPT-5.4: Multimodal Frontier (vision + speech + text native, real-time voice)",
+            "authors": "OpenAI",
+            "year": 2026,
+            "venue": "OpenAI",
+            "tags": [
+              "关键节点"
+            ],
+            "arxiv": ""
           }
         ],
         "branches": [
@@ -1336,6 +1396,16 @@
               "前沿"
             ],
             "arxiv": "2507.20534"
+          },
+          {
+            "title": "MiniMax-M1: CISPO (Curriculum-Informed Synthetic Planning Optimization) RL for reasoning emergence",
+            "authors": "MiniMax",
+            "year": 2025,
+            "venue": "arXiv",
+            "tags": [
+              "关键节点"
+            ],
+            "arxiv": "2506.13585"
           }
         ],
         "branches": [
@@ -1370,6 +1440,16 @@
               "前瞻"
             ],
             "arxiv": "2403.09629"
+          },
+          {
+            "title": "Coconut: Training LLMs to Reason in a Continuous Latent Space (摆脱文字链, 潜在空间连续推理)",
+            "authors": "Hao et al. / Meta FAIR",
+            "year": 2024,
+            "venue": "NeurIPS 2024",
+            "tags": [
+              "前瞻"
+            ],
+            "arxiv": "2412.06769"
           }
         ]
       }
@@ -1795,6 +1875,16 @@
               "前沿"
             ],
             "arxiv": "2508.04700"
+          },
+          {
+            "title": "Generative Agents: Interactive Simulacra of Human Behavior (Stanford AI Town, memory-stream architecture, 25 agents)",
+            "authors": "Park et al. / Stanford",
+            "year": 2023,
+            "venue": "UIST 2023",
+            "tags": [
+              "起点"
+            ],
+            "arxiv": "2304.03442"
           }
         ],
         "branches": [