Latest AI News & Industry Updates

Meta and Michigan State paper finds LLM-built reasoning graphs unreliable

The non-obvious point is that most reasoning-graph pipelines assume the LLM's decomposition is correct without ever verifying it, so confidence scores inherit hidden graph errors. Builders of agent workflows and RAG-style decomposition systems should add graph-level validation before trusting computed confidence. The Qwen3-32B versus Qwen3-8B result is a concrete warning that scaling a model does not guarantee better sub-question answering.

#Qwen3 #Meta #reasoning models #AI safety

rohanpaul_ai@rohanpaul_aiSource

Papers Builders11:03

Chinese contentEnvironment Duels：评估 LLM 出题与解题双重能力的动态基准

让 9 个模型互相出题、互相解题，结果最强的解题者反而不是最好的出题者，小模型出题更狠，挺有意思的视角。

#LLM #benchmark #智能体 #Qwen3.8-27B

Event

arXiv cs.AI@Shibo Jin et al. (5)2 sourcesSource

Papers Builders11:00

Chinese contentBehavioral Localization 方法让工具调用智能体更高效学习软技能

教智能体学技能不用整条轨迹都学，只挑 10% 关键位置就够，成功率还能涨 6 个多点，做 Agent 训练的可以看看。

#智能体 #软提示 #Qwen3.6 #SpreadsheetBench

arXiv cs.AI@Zijian Jiang, Xilu WangSource

Papers Builders10:45

Chinese contentSPARE：用路由熵决定 MoE 各层剪枝比例的免训练方法

MoE 模型太大跑不动？这篇论文的 SPARE 方法不用训练，按每层路由熵决定剪多少专家，Qwen3 剪一半还涨了 6 个点。

#SPARE #MoE #Qwen3 #DeepSeek

arXiv: DeepSeek@Yifan Yang et al. (7)Source

Papers BuildersMulti-source10:45

Chinese contentGUI 智能体自报成功可能是假阳性，报告-轨迹审计可拦截

GUI 智能体嘴上说完成但轨迹不支持，这论文教你怎么用独立审计拦住这类假成功，还给了具体拦截率数据。

#GUI Agent #Qwen3.5 #DeepSeek #AI安全

Event

arXiv: DeepSeek@Yunfeng Zhu et al. (4)4 sourcesSource

Papers Builders09:58

Chinese contentOmniSimulator：用结构化行为推理框架训练小模型模拟短视频用户行为

有人把 Qwen3-8B 调教成短视频用户行为模拟器，提分 34% 后追平 Claude-Sonnet-4.5，8B 就能自己部署，做用户模拟的可以看看这篇。

#OmniSimulator #Qwen3-8B #Claude-Sonnet-4.5 #用户模拟

arXiv cs.AI@Jiawei Chen et al. (15)Source

Papers BuildersSelected13:22

Meta ProPerAgent Harness Lifts Proactive Agent Accuracy to 38%

The core insight is that timing, not memory volume, is the bottleneck for proactive assistants: a dedicated filtering call before the model sees a memory does most of the work. Builders of always-on or wearable agents can apply this retrieve-then-filter pattern without fine-tuning, and RL on top adds a further gain. It is a useful comparison point against standard RAG pipelines that flood the model with context.

#ProPerAgent #Meta #agents #RAG

DAIR.AI@dair_aiSource

Papers BuildersMulti-sourceSelected10:17

Microsoft Paper Co-Evolves RL Harness and Qwen3.5 Model

Co-evolving the harness and the model appears to unlock gains that optimizing either alone cannot reach, suggesting the harness itself is a first-class lever in RL training. Teams training coding agents can copy the loop to squeeze more performance from mid-size open models like Qwen3.5-35B-A3B. The distillation trick of using Opus 4.8 as the harness proposer also offers a cheaper route to borrow frontier-model judgment.

#Qwen3.5 #reinforcement learning #coding assistants #agents

Event

DAIR.AI@dair_ai4 sourcesSource

Papers Builders10:13

Chinese contentAMTP：自适应多 token 预测让 Llama-3.1 解码提速至 4.1 倍

arXiv 上这篇论文给 MTP 解码加了个轻量控制器，不用重训练就能自动选草稿长度，Llama-3.1 上最高跑出 4.1 倍解码速度，做推理优化的可以看看。

#AMTP #Llama-3.1 #Qwen3 #多 token 预测

arXiv cs.AI@Emile Anand et al. (3)Source

Papers BuildersMulti-sourceSelected74°10:11

Agent S4 Evolves Its Own Computer-Use Interfaces

Interface design itself becomes a learnable, evolving capability, meaning smaller open models can beat frontier systems on computer-use tasks at a fraction of the cost. Builders working on agents can reuse the open-source repository and apply the evolution loop to their own workflows. The results set a concrete cost-performance benchmark against GPT-5.5 and Claude Opus 4.8 that teams choosing an agent stack can weigh directly.

#Agent S4 #Qwen3.8-27B #computer-use agents #open-source models

Event

arXiv cs.AI@Saaket Agashe et al. (8)3 sourcesSource

Papers Builders09:49

Chinese contentNear-Duplicate 图片选择可向视觉语言模型传递隐含偏好

论文挺有意思：策展模型只是挑了挑几乎一样的图片，就把偏好悄悄传给了学生模型，LLaVA、Qwen 这些都中招，做数据清洗的值得看看。

#LLaVA-1.5 #Qwen2.5-VL #Qwen3-VL #AI安全

arXiv cs.LG@Jaeyoung Kim et al. (4)Source

Papers Builders09:47

Chinese contentAMD 提出 SliMOO：用多目标进化搜索做 LLM 深度剪枝

AMD 把剪枝当成多目标优化来做，Llama-3.1-70B 砍掉 25% 层后数学分还能涨 10 个点，代码也会开源。

#SliMOO #Llama-3.1 #Qwen3 #模型压缩

arXiv cs.LG@Guanchen Li et al. (5)Source

Papers Builders09:38

Chinese contentAgenticVulHunter：用本地开源模型在终端里挖漏洞

研究者做了个能在终端本地跑的漏洞挖掘智能体，用 DeepSeek V3.1 和 Qwen3-Coder 测出了比 Claude Sonnet 4 方案还高的准确率，搞安全审计的可以看看。

#AgenticVulHunter #Qwen3-Coder #DeepSeek V3.1 #代码安全

arXiv: DeepSeek@Nishant Abhishek et al. (3)Source

Papers Builders12:48

Chinese contentBrainLoRA 用参数高效微调刷新脑编码模型性能

把 LoRA 微调用在脑编码上，连 104B 的 Qwen3-Omni 都能高效适配，还拿了 Algonauts 2025 第一，做法挺巧的。

#LoRA #脑编码 #Qwen3-Omni #Algonauts

arXiv cs.LG@Zhuofan Josh Ying et al. (9)Source

Papers BuildersMulti-source12:36

Chinese contentOpenVLHarness:带多模态记忆的视觉智能体工具调用框架

一个开源的视觉 Agent 框架，给多模态模型外挂工具和记忆，22 个基准平均涨 10 分左右，成本还只要竞品的三成多。

#OpenVLHarness #Qwen3-VL #GPT-6 #多模态

Event

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

Papers Builders12:35

Chinese contentWeAct：让智能体通过权重切换获得能力

这篇论文把"换适配器"变成智能体的一种动作，让模型在执行中自己选能力，Qwen3-32B 上比静态 LoRA 合并高 27.5%，做 agent 的可以看看思路。

#WeAct #Qwen3-32B #LoRA #智能体

arXiv cs.AI@Rui Shi et al. (5)Source

Papers Builders12:32

Chinese contentTokenGate：可学习的 token 剪枝让长上下文预填充更快更省显存

一篇很实用的推理优化论文：给 Qwen3 加个小模块做 token 剪枝，TTFT 最快快 11 倍，代码开源可以直接试。

#TokenGate #Qwen3 #token剪枝 #推理优化

arXiv cs.AI@Liana Patel et al. (5)Source

Papers Builders11:49

Chinese content固定推理预算下的答案交付研究：停止策略与终结提示的组合实验

论文测了在固定 token 预算下怎么让模型真的把答案交出来，Qwen3-4B 和 DeepSeek 蒸馏版各涨了 8 和 7 个点，做推理部署的可以看看。

#GSM8K #Qwen3 #DeepSeek-R1 #推理模型

arXiv: DeepSeek@Yupeng Chang et al. (5)Source

Papers Builders11:48

Chinese contentNash-MOPD：用局部极小极大博弈优化多教师在线蒸馏的任务分配

一篇蒸馏调度的论文，把多个教师的训练预算分配问题建模成博弈，在 DeepSeek 和 Qwen 的蒸馏模型上比固定配比效果更好，做蒸馏的可以看看。

#Nash-MOPD #知识蒸馏 #DeepSeek-R1-Distill-Qwen-1.5B #Qwen3-8B-Base

arXiv: DeepSeek@Siyuan Gan et al. (7)Source

Papers Builders11:48

Chinese content研究：GEPA 与 SkillOpt 在 DeepSeek V4 Flash 和 Qwen3.8-27B 本地部署上的效果差异

GEPA 在 MATH-500 上能从 0.828 拉到 0.904，但在 HotpotQA 上完全没用，SkillOpt 还会翻车——如果你在本地跑推理模型做提示词优化，这篇能帮你避坑。

#GEPA #SkillOpt #DeepSeek #Qwen

Event

arXiv: DeepSeek@Andrea Giummarra et al. (6)2 sourcesSource

Papers BuildersSelected11:47

Addressed-Memory Writes Fail at Reasoning in Qwen3 and DeepSeek Beds

Memory write-injection looks like a cheap way to add knowledge to a deployed model, but these results show the added facts cannot be chained into multi-hop reasoning by the model itself — orchestration must live in the application layer. Builders wiring RAG-style memory into agents should chain written atoms explicitly in prompts or system logic, and benchmark vendors should adopt coverage, synergy and distortion reporting. The tested beds include Chinese frontier models Qwen3 and DeepSeek, making the findings broadly relevant across model families.

#memory #DeepSeek #Qwen #RAG

arXiv: DeepSeek@Wei Gao et al. (4)Source

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

Papers BuildersMulti-source09:30

Chinese contentBellman Language Models：把 LLM Agent 决策当 MDP 一次性求解

把 Agent 反复遇到的决策当成 MDP 求解一次然后复用，Qwen3.8-27B 在 τ²-bench 上从 29% 冲到 96%，成本降 98%，思路挺新的。

#Bellman Language Models #τ²-bench #Claude Opus 5.5 #Qwen3.8-27B

Event

arXiv cs.AI@Adarsh Vatsa, William Eiers3 sourcesSource

Papers Builders09:27

Chinese contentIF-MAP：用影响函数思路给多智能体系统排序优化优先级

多智能体系统预算有限时，这篇论文教你怎么算出该先优化哪个 agent，六个工作流实测收益翻倍。

#IF-MAP #多智能体 #Qwen3-8B #DeepSeek-V3.2

arXiv: DeepSeek@Jie Mao et al. (9)Source

Papers BuildersSelected21:31

OpenBMB's Trident fixes multimodal retrieval's text bias

Even small 2B-scale retrievers can double mixed-corpus retrieval quality once modality bias is addressed in training, which matters for anyone building RAG over documents with screenshots, tables or charts. The paper, code and project page are open, so teams can adopt the Multi-Positive View InfoNCE recipe or swap Trident-Qwen3VL-2B into their existing embedding stack.

#OpenBMB #Trident #Qwen3VL #RAG

OpenBMB@OpenBMBSource

Papers Builders10:04

Chinese contentLM2LM：用 RL 训练 LLM 提出语言模型改进方案

用世界模型当奖励模型训练 AI 自己改进模型，Qwen3.8-27B 训完居然超过 GPT-6-sol，成本还降了 33%。

#LM2LM #Qwen3.8-27B #GPT-6-sol #强化学习

Event

arXiv cs.AI@Minhua Lin et al. (10)2 sourcesSource

Papers Builders10:04

Chinese contentSchemaShield-Lite：无需重训即可修复 LLM 智能体的接口漂移问题

做智能体工具调用的朋友看这篇，工具接口一改模型就崩，SchemaShield-Lite 不用重训就能在 BFCL-200 上把准确率从 0.58 拉到 0.805。

#Qwen3.5-9B #Function Calling #智能体 #BFCL

Event

arXiv cs.AI@Aniruddha Chattopadhyay et al. (3)2 sourcesSource

Papers10:00

Chinese contentMGE-PO：用熵感知信用分配改进数学推理训练

这篇论文针对 GRPO 对同奖励响应一视同仁的问题，用生成熵给对错响应加不同权重的修正项，Qwen2.5-7B 上 avg@32 从 41.5% 提到 45.4%。

#MGE-PO #GRPO #Qwen3 #Qwen2.5

arXiv cs.AI@Jihao Huang et al. (12)Source

Papers Builders09:24

Chinese contentBeacon 框架改进 GRPO 安全对齐，安全错误率降低 32.83 个百分点

这篇论文解释了为什么 GRPO 做安全对齐会失效，还给出了 Beacon 框架，在 Qwen3 和 DeepSeek-R1-Distill 上把安全错误最多降了 32.83 个百分点。

#GRPO #Beacon #Qwen3 #DeepSeek-R1

arXiv: DeepSeek@Jiawei Chen et al. (12)Source

Papers BuildersSelected06:20

Data World Engine Generates RL Training Environments for Data Analysis

Verified-by-construction environments sidestep the usual reward-hacking problem in RL for open-ended analysis tasks, a reusable recipe for teams training agents beyond fixed-answer benchmarks. The diagnostic finding that frontier models compute correctly but conclude wrongly matters for anyone building data-analysis agents who should audit the reasoning step, not just the arithmetic. The paper is available on arXiv (2610.21656) with a Chat with Paper interface at academy.dair.ai.

#Data World #Qwen3.6 #reinforcement learning #data analysis

elvis@omarsar0Source

Papers BuildersSelected00:20

Tencent paper details recursive self-improvement for terminal agents

The reward-design audit is the most transferable part: dense progress rewards let failed runs harvest half the reward, so teams training RL agents should measure how much reward failure earns before trusting their curves. RL and agent-training engineers can apply the strict-success-versus-partial-credit split to their own reward pipelines. Compared with generic terminal-agent RL reports, this one pairs benchmark gains with a concrete failure-mode analysis.

#Tencent #Qwen3 #Terminal-Bench #PPO

elvis@omarsar0Source

Papers BuildersMulti-sourceSelected18:19

Nick Cammarata essay revisits neuron interpretability using Qwen3-32B

Neuron-level interpretability has quietly moved from a 1.7% legibility rate to about 90% on a mainstream open model, which changes how much trust practitioners can place in internal audits. Safety and interpretability teams should read the essay before planning monitoring for the coming Loop Transformer architectures, where readable chain-of-thought may disappear. Qwen3-32B being an open-weights model also means anyone can reproduce the audit rather than waiting on lab disclosures.

#Qwen3-32B #Goodfire #interpretability #AI safety

Event