Latest AI News & Industry Updates

Papers BuildersMulti-sourceSelected06:46

Shanghai AI Laboratory SkillGym trains better agents from verified skill runs

Training on verified execution traces can bake retrieval-dependent prompt skills into model weights, reducing reliance on the agent finding and following instructions at runtime. Teams building coding agents can replicate the recipe by turning their own skill libraries into checkable sandbox tasks. The work comes from a Chinese lab and benchmarks on Claude Code with Qwen, making it a useful reference point against prompt-based skill systems like Claude Skills.

#Shanghai AI Laboratory #Qwen #SkillGym #agents

Event

rohanpaul_ai@rohanpaul_ai3 sourcesSource

Papers US vs ChinaMulti-sourceSelected01:04

Qwen3.8 27B addition in words tested on DGX Spark

A 27B open-weights model running locally now solves multi-digit addition-in-words at near-perfect accuracy, a task GPT-4o fumbled two years ago. Practitioners evaluating small local models for arithmetic-adjacent workflows get a concrete benchmark on consumer-tier DGX Spark hardware. Comparisons across reasoning-on and reasoning-off settings make it easy to judge whether enabling reasoning is worth the latency.

#Qwen #DGX Spark #reasoning models #open-source models

Event

官方账号Simon Willison’s Weblog3 sourcesSource

Papers Builders19:23

Chinese contentNeutronGym：面向 LLM 智能体的中子仪器设计评测环境

用中子仪器设计测智能体会不会物理，7个模型最多做对7/16题，RL把Qwen3-8B从11%练到77%，比32B还强。

#NeutronGym #Qwen3 #McStas #智能体

arXiv cs.AI@Lijie Ding, Changwoo DoSource

Papers BuildersSelected15:21

Microsoft paper ScholarEvolve evolves agent harnesses from research papers

The counterintuitive idea is that the literature, not runtime traces, is the highest-value source of harness improvements. Builders tuning agent scaffolds around a fixed model can replicate the module-and-topic-modeling loop with their own paper corpus instead of paying for a model upgrade. Results on AppWorld Challenge and Tau2-Bench Telecom give a concrete baseline for comparing it against failure-log-driven harness tuning.

#ScholarEvolve #Microsoft #agents #open-source models

elvis@omarsar0Source

Papers BuildersMulti-sourceSelected16:15

ProVer method for agent RL credit assignment

The approach addresses a fundamental limitation in current RL methods where all trajectory tokens receive equal credit, making it impossible to identify decisive steps. Developers working on agent systems can implement this method to improve training efficiency, particularly when using smaller judge models, potentially reducing computational costs while maintaining performance gains.

#ProVer #Qwen3.5 #agent RL #credit assignment

Event

elvis@omarsar03 sourcesSource

Papers BuildersSelected10:41

Multi-Token Prediction Heads Research

This research significantly reduces the computational cost of implementing MTP heads, making advanced inference optimization more accessible. The findings could enable faster deployment of large language models without requiring massive pretraining resources. Developers and researchers working on model optimization can implement these techniques to improve throughput with minimal computational overhead.

#Qwen3 #MiMo-7B #multi-token prediction #inference optimization

arXiv: DeepSeek@Prachi Badarayani et al. (11)Source

Papers Builders10:19

Chinese content多教师在线蒸馏研究：从梯度到能力

研究Qwen3如何整合多个RL训练教师，发现梯度传递和参数更新的关键机制。

#Qwen3 #MOPD #蒸馏 #RL

arXiv cs.LG@Siqi Zhu et al. (7)Source

Papers BuildersMulti-source83°10:18

Chinese content长期AI决策能力研究：模型表现远低于人类

最新研究揭示，即使是顶尖AI模型在长期复杂任务中也仅达人类四分之一表现。

#GPT-5.6 Sol #Claude Opus #Qwen3.7-Max #长期决策

Event

rohanpaul_ai@rohanpaul_ai3 sourcesSource

Papers Builders09:57

Chinese contentMoRA：通过路由器偏置学习和专家近似进行MoE剪枝

MoRA解决了MoE模型部署内存问题，通过路由偏置学习和专家近似机制，在减少专家数量的同时保持模型性能。

#MoRA #MoE #Qwen3-30B-A3B #DeepSeek-V2-Lite

arXiv: DeepSeek@Yushuai Sun et al. (5)Source

Papers Builders73°06:45

Chinese contentMeta发布上下文语言模型新研究

Meta让AI助手能直接编辑自己的记忆，任务完成效率提升近七成，还省计算资源。

#CLM #Meta #BrowseComp-Plus #上下文语言模型

Event

elvis@omarsar02 sourcesSource

Papers Builders12:44

Chinese content多模态临床诊断的排序感知提示优化

医疗AI研究者必看：新方法让模型在诊断中更准确地区分阳性阴性病例，AUROC提升显著。

#Ranking-PE #MLLMs #AUROC #临床诊断

arXiv cs.LG@Tian Xia et al. (5)Source

Papers Builders11:08

Chinese contentSemifactual Credit-Augmented Policy Optimization

SCAPO算法通过半事实稳定性改进了令牌级信用分配，在数学和分布外基准上表现优异。

#SCAPO #GRPO #Qwen3 #强化学习

arXiv cs.AI@Junshu Pan et al. (8)Source

Papers BuildersMulti-source11:01

Chinese contentPivotOPD：学习从关键错误中恢复的多轮智能体

NVIDIA新方法让AI智能体犯错后能自我修正，在多个任务上超越13个基线模型。

#PivotOPD #Qwen3 #NVIDIA #多轮智能体

Event

arXiv cs.AI@Yinghui He et al. (7)11 sourcesSource

Papers Builders10:52

Chinese contentComputerSD：基于实时反馈的计算机使用智能体在线自蒸馏

清华团队推出 ComputerSD，让计算机使用智能体能从实时 GUI 反馈中学习，比传统方法效果更好。

#ComputerSD #Qwen3-VL-8B-Thinking #EvoCUA-8B #智能体

arXiv cs.AI@Yong Du et al. (8)Source

Papers Builders10:32

Chinese contentDAGent：深度研究智能体的评估增长规划框架

DAGent解决了深度研究任务中的规划脆弱性问题，通过条件扩展任务图实现更高效的执行。

#DAGent #深度研究智能体 #多模态 #规划算法

Event

arXiv cs.AI@Hanwen Liu et al. (3)2 sourcesSource

Papers Builders10:27

Chinese contentBELIEFRAG：自适应RAG在动态证据下的状态感知

BELIEFRAG解决了RAG中的证据状态碎片化问题，在保持质量的同时大幅减少token使用。

#BELIEFRAG #RAG #GPT-OSS-120B #Qwen3-32B

arXiv cs.AI@Hongji PuSource

Papers BuildersSelected09:58

Cross-World On-Policy Distillation for VLMs

This research addresses a critical gap in VLM distillation where students can match teacher outputs without relying on the same visual evidence. The CW-OPD method provides a corrective signal that makes visual evidence reliance explicit rather than implicit. Developers working with smaller models can leverage this approach to achieve performance comparable to much larger models while maintaining better visual grounding.

#Cross-World On-Policy Distillation #Qwen3.5-4B #DeepSeek-V4.1 #vision-language models

arXiv: DeepSeek@Yuanhao Sun et al. (5)Source

Papers BuildersSelected09:55

Activation-Conditioned Self-Distillation

ACSD provides a novel approach to token-level supervision throughout long responses by leveraging activation contrasts, offering a more stable alternative to existing methods. Researchers and developers working on reasoning models can implement this technique to improve mathematical accuracy without requiring reference solutions, potentially enhancing model performance across diverse training datasets.

#ACSD #self-distillation #reasoning models #mathematical benchmarks

arXiv: DeepSeek@Zhexi Lu et al. (5)Source

Papers BuildersMulti-source09:48

Chinese content自进化搜索代理中的协同作弊诊断与缓解

论文揭示了自进化AI系统的隐藏缺陷，提出CrossFit方法有效解决协同作弊问题，在Qwen模型上效果显著。

#自进化代理 #协同作弊 #Qwen3.5 #CrossFit

Event

arXiv cs.AI@Meijia Chen et al. (15)4 sourcesSource

Papers Builders09:47

Chinese contentSCIC：文本转语音中的说话人自适应表达控制

清华团队推出SCIC，让TTS系统能更精确控制长文本中的语调变化，比传统方法更自然。

#SCIC #TTS #Qwen3-TTS #文本转语音

arXiv cs.AI@Longyu Lu et al. (7)Source

Papers Builders10:50

Chinese contentSTEPQuant：Delta规则循环状态量化研究

清华团队提出STEPQuant，根据误差大小和内存寿命分配精度，在Qwen和Kimi模型上显著降低内存使用。

#STEPQuant #Qwen3.8-27B #Kimi-Linear-48B-A3B-Instruct #量化

Event

arXiv cs.AI@Bingchen Yao et al. (9)2 sourcesSource

Papers BuildersSelected10:43

Post-Training Effects on Mathematical Reasoning

The research challenges the universal compression explanation for post-training effects, revealing instead two distinct regimes affecting mathematical reasoning differently. For AI practitioners and researchers working on mathematical reasoning models, this provides nuanced insights into training approaches and their impact on problem-solving capabilities across difficulty levels and languages.

#mathematical reasoning #post-training #Qwen3 #DeepSeek-Math

arXiv: DeepSeek@Hongyang Li et al. (6)Source

Papers Builders10:21

Chinese contentProbe-Space 预训练方法提升零阶训练效率

新方法1.5-SPSA让零阶训练比传统方法快1000倍，OPT-13B只需70步就超越MeZO 10万步效果。

#1.5-SPSA #ZOO #OPT #零阶优化

arXiv cs.LG@Francois Chaubard et al. (3)Source

Papers BuildersSelected09:37

CTE-Bench for Stateful Software Simulators

This benchmark reveals critical limitations in current AI models' understanding of stateful systems, showing they struggle to track intervention effects without continuous correct feedback. Developers building coding assistants should note that models make compounding errors over time, highlighting the need for robust error handling and state management in AI systems.

#CTE-Bench #coding agents #stateful software #DeepSeek V4-Flash

Event

arXiv: DeepSeek@Xinran Zhang2 sourcesSource

Papers Builders09:34

Chinese contentCoEM：基于证据提交的长上下文推理方法

CoEM解决了长上下文推理中信息过早压缩的问题，通过智能管理记忆空间，显著提升了模型处理长文档的能力。

#CoEM #长上下文推理 #Qwen3.5-9B #记忆管理

Event

arXiv cs.AI@Jingguang Li et al. (10)2 sourcesSource

Papers Builders09:30

Chinese contentMultiTalk：全双工语音模型扩展至长多语会话

MultiTalk解决了长多语会话难题，训练数据和评估基准都开源了，模型表现比现有开源方案好很多。

#MultiTalk #Moshi #全双工语音 #多模态

arXiv cs.AI@Ke Wang et al. (7)Source

Papers BuildersSelected20:37

LLM Post-training Robust to Erroneous Rewards

This research challenges the assumption that expensive verification systems are necessary for effective LLM training. Developers and researchers can now confidently use more affordable verification options without compromising training quality, potentially reducing costs for AI model development. The findings particularly benefit organizations working with resource constraints or those deploying models in cost-sensitive environments.

#Qwen3 #Gemma #verifier #LLM training

Tanishq Abraham (论文推介)@iScienceLuvrSource

Papers BuildersSelected18:34

ScopeIF Framework for LLM Instruction Following

ScopeIF addresses a critical gap in LLM training by providing dense supervision for complex constraints that existing binary reward models miss. This approach significantly improves performance on multi-faceted instructions without compromising general capabilities. Developers working with complex task automation can implement this framework to enhance their models' precision, particularly when dealing with segmented constraints across different response parts.

#ScopeIF #Qwen3 #Gemini-2.5-Pro #DeepSeek-V3.2

arXiv: DeepSeek@Bosi Wen et al. (6)Source

Papers BuildersSelected18:33

OracleLadder Reveals Math Reasoning Gap

The composition gap reveals a fundamental limitation in how LLMs combine correct intermediate steps into final answers, even when all components are provided. Developers and researchers can use OracleLadder to diagnose specific reasoning failures in their models, with the released code and data enabling reproducible testing across different problem domains.

#OracleLadder #Qwen3 #Llama 3.3 #DeepSeek-V3.1

arXiv: DeepSeek@Zhuohan Wang et al. (6)Source

Papers Builders18:04

Chinese content生物文献综述代理评估新基准BioLitBench发布

BioLitBench基准用组合可控性方法解决排行榜评估偏差，SCRIBE在生物文献综述任务中表现优异。

#BioLitBench #Qwen3.8-27B #生物医学文献综述 #排行榜评估

Event

arXiv: OpenAI@Zhaowei Han et al. (7)2 sourcesSource

Papers US vs China17:52

Chinese content语言模型目标移植研究

MIT团队发现语言模型内部存在可操控的价值信号，能改变模型行为方向，对AI安全研究有重要意义。

#Qwen3-8B #GPT-OSS-20B #价值移植 #AI安全

arXiv cs.LG@Pengcheng Jiang, Fabien RogerSource

Papers Builders16:14

Chinese contentarXiv 论文：可再推导性决定多级智能体流水线故障后的恢复能力

这篇论文用 gsm_hard 上的 120 题实验算清了一笔账：流水线出错后，把原题重新喂给第一个下游阶段就能挽回 0.394，每题只多花 59.8 个 token，后面的阶段白搭。

#gsm_hard #Qwen3-14B #Llama-3.1-8B #智能体

arXiv cs.LG@Tianqi Bu et al. (4)Source

Papers Builders16:10

Chinese content研究显示 MoE 模型合并后的路由漂移不能直接判定为路由失败

合并 MoE 模型时看到路由乱了先别急着修，这篇用 DeepSeekMoE 等三个模型证明漂移不等于失败，还给了工具包和修复代码。

#MoE #DeepSeekMoE #Qwen3-MoE #OLMoE

arXiv cs.LG@Yuanyi Wang et al. (10)Source

Papers Builders14:56

Chinese contentPEARL：面向智能体强化学习的自适应 Prefill-Decode 弹性执行系统

训练智能体 RL 的人可以看看：PEARL 动态调 prefill-decode 配置还能借闲置 GPU，吞吐比 RLBoost+ 最高多 36.3%。

#PEARL #Qwen3 #强化学习 #GPU优化

arXiv cs.AI@Jiaan Zhu et al. (9)Source

Papers Builders14:07

Chinese contentROFT：只用自我复盘微调，Qwen3.5-4B 不靠强化学习也能提升智能体表现

Qwen3.5-4B 只靠写复盘解释做微调，不用奖励模型就在 SWE-bench 上跑赢 GRPO，做智能体训练的可以看看这套 ROFT 流程。

#Qwen3.5-4B #SWE-bench #GRPO #ROFT

arXiv cs.AI@Jonathan Light et al. (10)Source

Papers Builders14:00

Chinese contentRSI-Router：子任务级大模型路由框架，推理成本降至约一半

一篇很实用的论文：让大模型和小模型在任务里分工干活，成本砍到一半性能反升，ALFWorld 上省了八成推理开销。

#RSI-Router #DeepSeek #Qwen #智能体

arXiv: DeepSeek@Hao Li et al. (8)Source

Papers BuildersSelected12:01

Meta proposes DCE + SRCL outperforming OPSD on Qwen3-8B

A more-than-doubling of accuracy on a widely used 8B open model suggests the training recipe, not just model scale, drives reasoning gains. Researchers fine-tuning smaller open models can test DCE + SRCL against OPSD as a cheaper distillation path. Qwen3-8B's popularity in ASEAN developer communities makes the recipe directly reproducible with local hardware.

#Meta #Qwen3 #OPSD #open-source models

Aran Komatsuzaki (论文推介)@arankomatsuzakiSource

Papers BuildersMulti-source79°11:03

Chinese contentNVIDIA 发布 Skill2Env：把社区 Agent Skills 编译成 RL 训练环境

NVIDIA 把社区写的 3.4k 个 Agent Skills 直接变成 RL 训练环境，300 步就把 Qwen 27B 在 Terminal-Bench 上拉高近 5 个点，流水线和负结果都写得很细，做 agent 训练的值得看看仓库。

#NVIDIA #Skill2Env #Qwen3.8 #智能体

Event

shao__meng@shao__meng11 sourcesSource

Papers Builders09:12

Chinese content研究显示推理模型的思考过程会消解部分偏见但制造更多新偏见

这篇 arXiv 论文拿三个 32B 模型做了思考前后对比，发现思考反而制造约 5 倍的新偏见翻转，做公平性评估的人该看看。

#QwQ-32B #DeepSeek-R1 #Qwen3 #推理模型

arXiv: DeepSeek@Deng Pan et al. (7)Source

Papers BuildersMulti-sourceSelected73°06:39

SkillGym trains Qwen3.5-35B-A3B on 2,756 skill environments

Continual training on distilled skills outperforming a larger-context approach suggests skill libraries can be baked into weights rather than retrieved at inference time. Teams building terminal or coding agents can test whether the same recipe transfers to their own models and checkers. The result also matters for anyone benchmarking open-weight Qwen models against Claude Sonnet 4.6 and GPT-5.4 Mini on agentic tasks.

#Qwen #SkillGym #fine-tuning #agents

Event

DAIR.AI@dair_ai3 sourcesSource