Latest AI News & Industry Updates

TRACE: Training Reasoning Agents for Causal Exploration

TRACE demonstrates that simulation-based verification can make diagnostic reasoning tasks scalable for reinforcement learning, outperforming larger models with fewer tool calls.

#TRACE #Qwen3.5-35B-A3B #Claude Opus 5 #reasoning models

arXiv cs.LG@Rui Sun et al. (3)Source

Papers BuildersSelected73°09:29

Maverick: Private LLM Inference Protocol

Maverick enables private LLM inference with 45x speedup over local processing, addressing privacy concerns for users who must use third-party providers.

#Maverick #Qwen3-4B #LLM inference #privacy

arXiv cs.LG@Ben Merbaum et al. (6)Source

Papers Builders12:15

Chinese contentCo-Evolving Harnesses and Models: On-Policy Correction Helps Weaker Models Catch Up Where Imitation Fails

研究解决了小模型模仿专家时的性能下降问题，提出了新的在线策略修正方法，让小模型在特定任务中更接近大模型表现。

#智能体 #系统提示 #Qwen3-Coder #Gemma 4

arXiv cs.AI@Zhou Yu et al. (11)Source

Papers BuildersSelected11:20

Domain Coverage Optima in Mid-Training

The study reveals optimal data distribution for multi-domain models, helping developers balance coverage during mid-training to maximize final performance.

#Qwen3-8B-Base #KOR-Bench #mid-training #alignment

arXiv cs.AI@Yunpeng Xu, Kun ZhengSource

Models BuildersSelected73°11:15

PlannerForge Framework for Autonomous Testing

PlannerForge improves autonomous driving testing efficiency, with open-source models matching commercial performance and reducing collisions by over 50% through cost-tuning.

#PlannerForge #Qwen3.6:35B #autonomous driving #LLM agents

arXiv cs.AI@Yuan Gao et al. (8)Source

Papers BuildersSelected10:33

LLM-Generated Asset Administration Shell Quality Metrics

The study provides concrete metrics for evaluating AI-generated industrial data, crucial for manufacturers adopting LLMs in Industry 4.0 implementations.

#Asset Administration Shell #GPT-4o-mini #Qwen3 #DeepSeek-R1

arXiv: DeepSeek@Janek Groß et al. (3)Source

Papers Builders73°09:47

Chinese contentOSOL缓解多领域强化学习中的高阶干扰

清华团队提出OSOL方法，通过分析token足迹解决多领域RL性能退化问题，在Qwen3-30B-A3B上提升5.7%性能。

#OSOL #Qwen3-30B-A3B #强化学习 #多领域

arXiv cs.LG@Zihan Lin et al. (7)Source

How-to BuildersSelected73°09:33

Speculative Uncertainty Method for Agentic Coding

Developers using LLM agents for coding can reduce errors and costs by 14-19% without retraining, transferring across models and benchmarks.

#Speculative Uncertainty #Qwen3-Coder-480B #Claude 3.5 Sonnet #coding assistants

arXiv cs.LG@Konstantin Grotov, Valentin MalykhSource

Models BuildersSelected73°09:23

ACE framework for MoE-based LLMs

ACE improves efficiency for MoE models without retraining, benefiting developers working with large language models in resource-constrained environments.

#ACE #MoE #Qwen3.6-35B-A3B #expert skipping

arXiv cs.AI@Zukang Xu et al. (8)Source

Papers Builders78°11:28

Chinese contentESPO：错误结构化提示优化方法

ESPO方法让提示更短更准，在Qwen3 GSM8K上从15%提升到91.4%，比GEPA短47%还更准。

#ESPO #提示词工程 #NLP #Qwen3

arXiv cs.AI@Lihao Liu et al. (4)Source

ModelsSelected73°11:25

Terminal-Universe Framework for Agent Environments

Developers can use Terminal-Universe to create more realistic training environments for terminal-based code agents, improving their performance on real-world coding tasks.

#Terminal-Universe #Qwen3.5-27B #Terminal-Bench #code agents

arXiv cs.AI@Jie Wu et al. (14)Source

PapersMulti-source11:23

Chinese content终端智能体环境演化方法研究

研究人员提出环境演化方法，让终端智能体训练环境持续变难，在Qwen模型上提升了18%性能。

#环境演化 #终端智能体 #Qwen3.6 #Claude Opus

Event

arXiv cs.AI@Zhiyuan Fan et al. (12)4 sourcesSource

PapersSelected73°11:22

Minima Quantizes Hybrid LLM Recurrent Layers

This quantization method enables smaller, faster hybrid LLMs without significant performance loss, making large models more accessible for resource-constrained environments.

#Minima #Qwen3.8-27B #Gated DeltaNet #quantization

Event

arXiv cs.AI@Sergii Kozyrev, Davyd Maiboroda2 sourcesSource

Papers11:13

Chinese contentLLM能否从代码提交中提取架构设计决策

研究测试了四种大模型从代码提交中提取架构设计决策的能力，发现虽有一定效果但仍有改进空间。

#LLM #架构设计决策 #代码提交 #Gemini 3 Pro

Event

arXiv: DeepSeek@Amey Karan et al. (4)2 sourcesSource

PapersSelected73°12:17

SafeEvolve Framework for Agent Safety

Developers can implement SafeEvolve to improve agent safety without sacrificing utility, achieving better safety-utility tradeoffs than existing methods.

#SafeEvolve #LLM-based agents #AgentDojo #safety alignment

arXiv cs.AI@Qinghua Mao et al. (11)Source

Papers11:53

Chinese contentGraph Machine：通过边实现更优预训练

Graph Machine 用边对象替代传统状态，在 Qwen3-0.6B 上预训练时，稀疏层仅需少量检索就能保持性能。

#Graph Machine #Qwen3-0.6B #Transformer #稀疏路由

arXiv cs.LG@Lintai HouSource

PapersSelected73°09:12

Multi-Teacher Self-Distillation Policy Optimization

MT-SDPO improves multi-domain LLM performance by 14.79 points on Qwen3-8B, offering better domain balance than single-teacher approaches.

#MT-SDPO #Qwen3-8B #Multi-Teacher Self-Distillation #distillation

arXiv cs.LG@Xixiang He et al. (7)Source

PapersSelected73°10:14

Quantization Damage in LLMs

Global quantization granularity beats selective layer repair by up to 52 accuracy points, offering a better default for LLM deployment.

#quantization #LLMs #Qwen3-8B #OpenLLaMA

arXiv cs.LG@Jundong Hu, Shekar RamachandranSource

ModelsSelected73°09:36

HiveTraceGuard-Pro generative guardrail model

Released under Apache-2.0 license with lowest median latency at 14.3 ms, offering robust protection against Russian and English prompt injections.

#HiveTraceGuard-Pro #Qwen3-0.6B #guardrail models #prompt injection

arXiv cs.AI@Nikita Oblakov et al. (3)Source

Papers09:32

Chinese contentPersianAnonymizer：评估基于LLM标注的波斯语NER匿名化

波斯语数据匿名化新方案，用LLM标注训练轻量NER模型，消费级GPU上2分钟处理4万消息。

#PersianAnonymizer #NER #DeepSeek-V3-0324 #GPT-OSS-120B

arXiv: DeepSeek@Mohammad Hossein Shalchian et al. (3)Source

PapersSelected73°11:24

LLM Scale Study for Ontology Learning

The study provides concrete guidance for selecting LLMs in ontology engineering, showing model architecture matters more than size for specific tasks.

#Qwen3.5 #Qwen3.6 #GPT #ontology learning

arXiv cs.AI@Hamed Babaei Giglou et al. (3)Source

PapersSelected73°11:24

BLOOM-WILT auditing pipeline for LLM behavior

BLOOM-WILT improves automated LLM auditing efficiency, making it valuable for companies deploying AI systems in ASEAN markets.

#BLOOM-WILT #Qwen3.5-4B #LLM auditing #behavior elicitation

arXiv cs.AI@Adrians Skapars, Edoardo ManinoSource

PapersSelected73°11:21

LLM Sequence Score Recovery Method

The method improves reasoning accuracy in existing models without retraining, making it valuable for developers working with constrained resources in ASEAN.

#Qwen3.5 #OLMo-2-1B #Llama-3.1-8B #reasoning models

arXiv cs.AI@Qiyao Yan et al. (3)Source

Models78°10:14

Chinese contentTuringLLM推出20B参数专家混合模型

Turing团队发布20B参数专家混合模型，每激活2B参数，性能接近9B模型，支持超长上下文。

#TuringLLM #Turing-20B-A2B #专家混合模型 #长上下文

arXiv cs.AI@Yuheng Zhang et al. (22)Source

PapersSelected09:20

Reasoning Models Show Asymmetric Disclosure

The findings reveal vulnerabilities in AI oversight mechanisms that developers and policymakers should consider when implementing reasoning model monitoring systems.

#Qwen3-14B #Qwen3-32B #Qwen3-235B #reasoning models

arXiv: DeepSeek@Zimo Shi et al. (3)Source

PapersSelected73°09:30

Reward-Aligned On-Policy Distillation

RA-OPD improves model performance on math and code tasks, making it valuable for developers working with LLMs in technical domains.

#RA-OPD #Qwen3 #DeepSeek-R1 #on-policy distillation

arXiv: DeepSeek@Siyuan Gan et al. (8)Source

Papers73°09:05

Chinese contentVISTA：基于验证器的学生-教师自适应蒸馏方法

VISTA方法改进了OPSD蒸馏技术，通过验证器指导教师自适应，在数学推理任务上显著提升性能。

#VISTA #OPSD #Qwen3 #蒸馏

arXiv cs.LG@Zewen Ding et al. (8)Source

Papers73°09:05

Chinese contentPASK：结构化生成中的条件化KV持久化

PASK让结构化生成更高效，在保持准确性的同时大幅降低内存使用，提升处理速度。

#PASK #结构化生成 #KV压缩 #Qwen3-4B

arXiv cs.LG@Linze Wu, Xinrui ChenSource

PapersSelected73°19:25

Block Drafting Model Gaps Analysis

The findings help developers understand why block drafters underperform and guide improvements to proposal quality in text generation systems.

#block drafting #Qwen3-4B #information floor #model gap

arXiv cs.LG@Xinwei Qiang et al. (5)Source

PapersSelected73°18:45

Mixture-of-Experts Inference Optimization Study

The findings impact MoE deployment efficiency for researchers and companies using these models in production environments.

#Mixture-of-Experts #OLMoE-1B-7B #DeepSeek-V2-Lite #Qwen3-30B-A3B

arXiv: DeepSeek@Gokulakannan Sakthivel et al. (4)Source

Papers10:26

Chinese contentDualOPSD:自适应策略自蒸馏方法

DualOPSD在模型规模和性能上都有显著提升，是自蒸馏领域的一个新突破。与OPSD相比，DualOPSD在多个基准测试中取得了更好的成绩，值得关注。

#自蒸馏 #模型优化 #性能提升 #Qwen3-8B

arXiv cs.AI@Yutong Chen et al. (4)Source

Papers09:25

Chinese contentLocalLSTC：用于本地GUI代理的长期短期控制架构

LocalLSTC通过时间组织控制信息，显著提升了本地GUI代理的性能，值得研究。

#LocalLSTC #长期短期控制 #GUI代理 #Qwen3.6-27B

arXiv cs.AI@Weiming Li et al. (3)Source

ModelsMulti-sourceSelected09:21

JIT-Agent: Harness Intelligence Model

JIT-Agent improves performance across multiple model families including DeepSeek V4 and Qwen3.6, offering a scalable alternative to manual harness design.

#JIT-Agent #DeepSeek-V4-Flash #GLM-5.2 #agents

Event

arXiv: DeepSeek@Guibin Zhang et al. (16)5 sourcesSource

Papers BuildersMulti-source09:18

Chinese contentEVOMAL: 自进化的编码智能体中的自毒化问题

研究揭示了自进化的编码智能体中的自毒化问题，并提出了counter-prompt防御措施，有效降低了恶意技能的传播。对于关注AI安全和智能体研究的读者来说，这是一篇不容错过的论文。

#智能体 #AI安全 #自毒化 #防御措施

Event

arXiv cs.AI@Xiaodong Wu et al. (8)3 sourcesSource

Papers Builders11:29

Chinese contentRecursive Experiential-Working Memory Evolution for Long-Horizon Agent Harnesses

Recuris模型在长时程任务中显著提升了GPT-5.6 Sol和Claude Opus 5的表现，是长时程智能体驱动的递归记忆架构的一个突破。