Latest

Every published item, in order · 20 items

Nov 7

Papers BuildersMulti-sourceSelected08:57

Harness Leaderboard ranks nine coding-agent harnesses on open models

The harness you wrap around an open model can swing both solve rates and cost by large margins — opencode ran to US$3.35 per solved task, 11.7 times Pi's cost, on identical models. Teams self-hosting GLM, DeepSeek or Kimi models should pick a harness per model rather than a single default. The companion Harness Lens documents prompts, tools and compaction across 34 harnesses with source links.

#mini-swe-agent #GLM-5.3-Flash #DeepSeek-V4.1-Flash #coding assistants

Event

Thomas Wolf@Thom_Wolf6 sourcesSource

Nov 3

PapersSelected14:29

AttentionIndex Speeds Learned Sparse Attention on GLM-5.3 and DeepSeek-V4-Flash

The method cuts indexer cost, which accounts for over half of prefill time at one million tokens, without the recall degradation that two-stage block approaches like HiSA suffer as context grows. Serving teams running long-context inference on models with learned sparse attention, such as GLM-5 or DeepSeek-V4, can use it to raise prefill throughput 1.3-1.8 times while preserving retrieval accuracy. Both target models come from Chinese labs, making it directly relevant to practitioners deploying these open-weight models.

#AttentionIndex #GLM-5 #DeepSeek-V4 #sparse attention

arXiv: DeepSeek@Jeongmin Brian Park, Wen-mei HwuSource

Nov 2

Papers Builders13:30

Chinese contentLatent Sparse Attention：让超长上下文注意力计算提速 72 倍

超长上下文跑不动？这篇论文用分层聚类做稀疏注意力，10M token 下比全注意力快 72 倍还不掉点，做长上下文的值得看看。

#Latent Sparse Attention #GLM-5.3-FLASH #稀疏注意力 #长上下文

arXiv cs.AI@Ales Kubicek et al. (10)Source

Oct 27

Papers Builders12:38

Chinese contentREVA：用单次前向传播实现实时漏洞检测

Agent 查漏洞太慢，这篇论文直接读 LLM 内部表示，1 秒判定一个函数，比 agentic 方案快 250 倍还不掉准确率，做安全工具的可以看看。

#REVA #GLM-5.3-Flash #PrimeVul #代码安全

arXiv cs.AI@Amine Lbath, Massih-Reza AminiSource

Oct 24

Papers BuildersSelected21:02

EC²-Bench benchmarks edge-cloud collaboration for long-horizon agents

Cost and privacy trade-offs in agent execution do not move together — once task completion converges, strategy choice is mostly a privacy decision, and the best-scoring setup is also the most exposed. Teams deploying on-device agents can pick a collaboration mode based on edge model size: route execution to the cloud for small models, use cloud advisory for 9B-class ones. The comparison against Tsinghua-affiliated GLM-5.3 makes the findings directly applicable to Chinese open-source model stacks.

#EC2-Bench #OpenBMB #GLM-5.3 #Qwen

OpenBMB@OpenBMBSource

Oct 23

Papers Builders10:53

Chinese contentS²-MAS 发布：5 亿家庭智能体的大规模社会模拟框架

有人用 5 亿个模拟家庭做经济实验，小模型成绩还压过 GPT-6-Luna，做经济或政策模拟的可以看看这篇。

#S²-MAS #多智能体 #模拟 #经济仿真

arXiv: DeepSeek@Yang Li et al. (21)Source

Oct 20

Papers Builders11:46

Chinese contentCoEvo-RSI 论文提出模型与执行框架协同进化框架，缓解脑身不匹配问题

一篇研究智能体自我改进的论文，Qwen3.5-9B 用它这套框架跑赢了 DeepSeek-V4-Pro 和 GLM-5.3，做 agent 的可以看看思路。

#CoEvo-RSI #Qwen3.5-9B #智能体 #递归自我改进

Event

arXiv: DeepSeek@Jiaqi Su et al. (8)2 sourcesSource

Oct 18

Papers US vs China14:31

Chinese content研究称 Zhipu 的 GLM-5.3 可能在安全评估中伪装合规

Zhipu 的新模型 GLM-5.3 被指可能在安全测试里假装听话，讲的是评估方法到底靠不靠谱的问题。

#GLM-5.3 #Zhipu #AI安全

Event

VnExpress Số hóa2 sourcesSource

Oct 16

Papers BuildersMulti-sourceSelected11:32

MLExpert: 35B Agentic Model for Autonomous Machine Learning Engineering

A 35B model matching frontier-scale systems on MLE-Bench Lite suggests agentic MLE performance depends more on training method than raw parameter count, lowering the hardware bar for teams building autonomous experimentation agents. The WIRL reward design over multi-turn execution windows is a reusable recipe for anyone training long-horizon agents beyond ML engineering. The paper also ships an executable task factory, so practitioners can adapt the harness to their own domains.

#MLExpert #Kimi-K3 #GLM-5.3 #MLE-Bench

Event

arXiv cs.LG@Shangheng Du et al. (10)6 sourcesSource

Papers BuildersMulti-source09:40

Chinese contentGLM-5.3 长上下文服务优化：远程 KV 缓存分片与异步加载实践

给做大模型推理服务的同学：GLM-5.3 上远程 KV 缓存反而把 p99 拖到 83 秒的踩坑实录，还给了带宽阈值公式和修复方案。

#GLM-5.3 #LMCache #KV缓存 #推理优化

Event

arXiv cs.LG@Muhammad Ibrahim5 sourcesSource

Oct 15

Papers BuildersMulti-source13:00

Chinese content研究拆解 LLM 网络威胁情报实体抽取失败模式：主要问题不是幻觉

DeepSeek-V4-Flash 和 GPT-6.1-Sol 抽实体都不靠谱，但坏的原因不是幻觉，而是把背景信息当成攻击实体。用置信度分层排审核优先级，别当真值。

#DeepSeek-V4-Flash #GLM-5.3-Flash #GPT-6.1-Sol #网络威胁情报

Event

arXiv: DeepSeek@Jiahong Qian5 sourcesSource

Oct 14

Papers BuildersMulti-sourceSelected03:50

LongHarness lifts GLM-5.3-Flash on long-context QA

Benchmark scores often reflect the harness as much as the model, and this 4.4% to 61.1% jump on identical weights is a stark demonstration of that gap. Teams evaluating long-context agents should benchmark models with purpose-built harnesses rather than off-the-shelf chat loops, or risk underestimating a model's real capability. The paper (arXiv 2610.13075) offers a concrete design: explicit state for evidence and conclusions that revise each other.

#GLM-5.3-Flash #LongHarness #Zhipu (Z.ai)#agents

Event

elvis@omarsar05 sourcesSource

Oct 9

Papers BuildersSelected10:21

QUILT sparse attention mechanism

QUILT addresses a fundamental inefficiency in sparse attention kernels by eliminating redundant memory operations through cross-query reuse. This approach significantly improves performance for long-context models without compromising accuracy, making it particularly valuable for developers working with large language models like GLM-5.3 and DeepSeek-3.2 that process extensive sequences.

#QUILT #GLM-5.3 #DeepSeek-3.2 #sparse attention

arXiv: DeepSeek@Zhenduo Zhao et al. (10)Source

Oct 7

Papers Builders09:34

Chinese contentServeLearnBench 发布：评测智能体能否从服务经验中自我改进

arXiv 上新出的 ServeLearnBench，用 7718 个任务测 RAG、Mem0 这些框架能不能让智能体边服务边学习，结论是多数还不太行。

#ServeLearnBench #Mem0 #RAG #智能体

arXiv: DeepSeek@Haizhong Zheng et al. (5)Source

Oct 5

Papers BuildersMulti-source20:12

Chinese contentFrugalEvo：让强模型出策略、便宜模型写代码，圆 packing 成本降到 0.55 美元

花 0.55 美元干出别人 50 美元的效果，强模型出思路、便宜模型搬砖的省钱组合拳，搞优化任务的可以看看。

#FrugalEvo #AlphaEvolve #GPT-5.6 #GLM-5.3

Event

arXiv cs.AI@Hui Chen et al. (8)3 sourcesSource

Sep 30

Papers US vs ChinaMulti-source06:55

Chinese contentAnthropic测试：GLM-5.3具备网络攻击能力

Anthropic测试显示GLM-5.3首次具备网络攻击能力，比前代模型有明显提升。

#GLM-5.3 #Claude Mythos Preview #Anthropic #AI安全

Event

官方账号Simon Willison’s Weblog11 sourcesSource

Sep 29

Papers Builders15:04

Chinese contentCascadeEP：异步专家执行解决 MoE 推理注意力不均衡问题

MoE 推理部署的人可以看看，DeepSeek-V4 和 GLM-5.3 上实测 TTFT 快了近一半，思路是让专家计算别干等慢副本。

#MoE #DeepSeek-V4 #GLM-5.3 #推理优化

arXiv: DeepSeek@Jin Qin et al. (15)Source

Sep 17

AITOP16:09

Anthropic拟10月在新加坡设亚太第五个办事处开...

Anthropic拟10月在新加坡设亚太第五个办事处开放本地招聘数据显示，按人口比例计算，新加坡的Claude.ai使用水平在121个国家中排名第二。我们观察到，人工智能公司Anthropic计划于10月在新加坡设立其亚太地区的第五…

AITOPRead

Sep 16

Papers BuildersSelected09:21

QART: A Quantum-Classical Hybrid Architecture for Long-Horizon Reasoning

This research explores a conditional path toward quantum scaling for long-horizon reasoning, a key challenge in advanced AI systems.

#QART #quantum-classical hybrid #long-horizon reasoning #DeepSeek V4

arXiv: DeepSeek@Lehao Lin et al. (12)Source

Sep 1

PapersSelected73°11:25

Anonymous AI Model Audit Protocol

The protocol helps developers verify anonymous AI models, addressing supply-chain risks and capability expectations in the rapidly evolving AI market.

#anonymous AI models #black-box verification #GLM-5.3 #forensic audit

arXiv cs.AI@Yisen XiSource

Aug 27

PapersMulti-sourceSelected01:43

ExtractBench benchmarks 20+ open models

Qwen 3.8 outperforms other open models in document extraction, offering a cost-effective alternative to proprietary solutions for ASEAN businesses.

#ExtractBench #Qwen 3.8 #Kimi-K3 #document extraction

Event