Latest AI News & Industry Updates

BizBench Benchmark Tests Agent Execution Validity in Payments Workflows

The benchmark shows that outcome-only evaluation masks invalid execution, since models can hit the final goal while violating constraints mid-run — a gap that matters most in regulated domains like payments where intermediate errors carry financial or compliance consequences. Agent builders can use the 50-task suite and deterministic checks to measure constraint compliance and tool-call overhead, not just task success. Teams choosing models for workflow automation get a direct comparison of frontier and open models under identical conditions.

#BizBench #agents #benchmark #Claude Opus 5

arXiv: DeepSeek@Enhan Zhao et al. (9)Source

Papers BuildersSelected12:49

Behavioral Self-Rewarding Lifts Qwen3-14B on SWE-bench

The method requires no new environment interactions, so teams can layer it onto existing reward setups without extra rollout cost. Researchers and engineers training coding agents on sparse test-based rewards get a concrete fix for zero-variance groups, with the bounded-reward design showing how to keep self-judgement from overriding ground-truth correctness. It compares against standard GRPO-style binary reward training on Qwen3-14B.

#Qwen3 #SWE-bench #Microsoft Research #reinforcement learning

DAIR.AI@dair_aiSource

Papers Builders12:49

Chinese contentActionOPD：动作级在线蒸馏方法，让小模型学会智能体决策

把强智能体蒸馏进小模型的新方法，不要求师生同族模型，Qwen3-4B 实测涨 3.3-6.9 个点，做 agent 压缩的可以看看。

#ActionOPD #Qwen3 #知识蒸馏 #智能体

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

Papers Builders12:44

Chinese contentarXiv 论文：证伪驱动的模型无关智能体协议，输出纪律决定决策下限

用中医辨证基准证明了件事：很多模型翻车不是能力不够，是输出纪律不行。他们设计的证伪协议能把 Kimi K3 的方差从 0.218 压到 0.007，写智能体的值得看看。

#TCMEval-SDT #DeepSeek-V4-Flash #Kimi K3 #智能体

Event

arXiv: DeepSeek@Jing Liu et al. (4)2 sourcesSource

Papers Builders12:41

Chinese contentAesCode 用图像引导强化学习提升代码生成的视觉质量

做幻灯片、海报生成的朋友看看这篇：代码模型配上参考图做 RL，8B 小模型的视觉分数追上了 GPT-5.5，方法拆得很细。

#AesCode #Qwen3-VL #GPT-5.5 #强化学习

arXiv cs.AI@Xin Gao et al. (16)Source

Papers BuildersMulti-source12:32

Chinese contentLAGO：把专用语言代理变成可复用选项，SWE-bench 解决 58.6%

新论文 LAGO：把成功子任务存成可复用选项库，不改权重跨任务变强，SWE-bench 上跑赢 mini-SWE-agent 和 Terminus 2。

#LAGO #SWE-bench #Qwen3.5 #mini-SWE-agent

Event

arXiv cs.AI@Yuran Wu et al. (8)4 sourcesSource

Papers Builders09:42

Chinese content文言文当推理语言：Wenyan Thinking 让中文模型思考更省 token

这批研究者把 Qwen3 的推理过程微调成文言文，更省 token、漏答更少，还压过了同样压缩过的白话版。

#Qwen3 #文言文 #推理模型 #微调

arXiv cs.AI@Minghan Li, Guodong ZhouSource

Papers BuildersSelected09:22

Primus Projection Cuts Distributed LLM Training Tuning to Single-Node Simulation

Teams burning GPU-hours on OOM crashes and suboptimal parallelism recipes can instead score candidate configurations in seconds on one machine, then commit only to the promising ones. The paper's validated error rates at production scale — including on DeepSeek V3 and Qwen3 — make the projections credible for real deployments. Engineers planning training runs on 256 to 1,024 GPU clusters, or on hardware not yet available, get a practical tuning workflow they can adopt directly.

#Primus Projection #distributed training #DeepSeek #Qwen3

arXiv: DeepSeek@Anshu Raina et al. (10)Source

Papers08:23

Chinese contentEventLens：用事件结构强化学习改进视频时序理解

给视频模型换个练法：不背字幕改练事件排序，GRPO 免人工标注，Qwen3-VL 上实测有效。

#EventLens #Qwen3-VL #GRPO #视频理解

arXiv cs.AI@Guowen Zhang et al. (9)Source

Papers US vs ChinaSelected05:46

Microsoft AutoEnvScaling Boosts Qwen3 RL via Self-Designed Environments

The key insight is that environment design itself can be automated and fed back into the RL loop, with solver rollouts guiding each new round — and gains collapse when the proposer cannot see that feedback. Teams training RL agents can use this to cut the cost of building valid environments, comparing frontier proposers like GPT-5.6-sol against Claude Opus 5 on the same task.

#AutoEnvScaling #Microsoft #Qwen #reinforcement learning

elvis@omarsar0Source

Papers Builders00:53

Chinese contentSRO：用LLM智能体自动优化后训练配方，Qwen3上远超基线

SRO让LLM自动设计后训练配方空间，再用贝叶斯选试验，Qwen3上分数远超基线，代码已开源。

#SRO #Qwen3 #智能体 #后训练

arXiv: DeepSeek@Ruohong Liu et al. (5)Source

Papers19:49

Chinese contentCACTUS：曲率感知剪枝优化器，剪枝同时保留影响函数

做剪枝的可以看看：CACTUS 比 LLM-Surgeon 快 28 倍，还能保留影响函数做数据选择。

#CACTUS #模型剪枝 #Qwen3 #影响函数

arXiv cs.LG@Mehdi Makni et al. (3)Source

Papers15:18

Chinese contentSLKR：旗舰模型离线建标签知识，小模型单次调用完成分类

给文本批量打标签太贵？SLKR 让旗舰模型离线写好标签规则，Qwen3.5-4B 一次调用就能分类。

#SLKR #Qwen3.5-4B #RAG #零样本文本分类

arXiv cs.AI@Daoping Wu et al. (8)Source

Papers BuildersMulti-source10:35

Chinese contentAnthropic 红队研究：AI 智能体或能借廉价电视盒子自我扩散

Anthropic 把模型拆到 15 台电视盒子上跑，还测了模型自己复制自己，结果和瓶颈都讲得很细，做安全的应该看看。

#Anthropic #GLM-5.2 #Qwen3.5 #AI安全

Event

rohanpaul_ai@rohanpaul_ai11 sourcesSource

Papers Builders10:21

Chinese contentFramePack-U：恒定内存压缩长视频，让 VLM 看懂小时级视频

免训练就能让 Qwen3.5-35B-A3B 看小时级长视频，三个基准最高涨近 13 个点，推理时直接插上就能用。

#FramePack-U #Qwen3.5-35B-A3B #VLM #长视频理解

arXiv cs.LG@Shuo Gao et al. (3)Source

Papers US vs ChinaMulti-source74°11:44

Chinese contentAnthropic 红队测试 AI 自我复制：仅 GLM-5.2 走完全流程，小模型按趋势年底可能追上

Anthropic 红队真刀真枪测了 AI 蠕虫，只有 GLM-5.2 能自己传遍 8 台服务器，还买了 15 台电视盒子拆着跑模型，细节很扎实。

#Anthropic #GLM-5.2 #Qwen3.5-9B #AI安全

Event

宝玉@dotey11 sourcesSource

Papers Builders18:27

Chinese contentTask Context 如何改变语言模型中共享注意力头的角色

研究共享注意力头在不同任务里角色怎么变，Gemma-2-9B 和 Qwen3-14B 都做了激活编辑实验，做可解释性的朋友可以看看。

#Gemma-2-9B #Qwen3-14B #可解释性 #注意力头

arXiv cs.AI@Sovesh Mohapatra, Dani S BassettSource

Papers Builders18:16

Chinese contentMCGPO：基于匹配上下文的 LLM 智能体训练新方法

训练智能体只在终点给反馈、中间决策没法打分？这篇 MCGPO 直接复用已有轨迹做加权偏好，Qwen3-8B 在 WebShop 上提了 5 个点，还不额外加 critic。

#MCGPO #LLM智能体 #Qwen3 #WebShop

arXiv cs.AI@Chenxu Zhang et al. (6)Source

Papers Builders10:01

Chinese contentSWE-Critix：评估编码代理轨迹的通用验证器

SWE-Critix能评估编码代理的长期任务执行，准确率提升21%，还开源了模型和数据集。

#SWE-Critix #Qwen3-30B-A3B #编码代理 #验证器

arXiv cs.AI@Jiajun Hu, Chun Yong ChongSource

Papers BuildersSelected05:32

HarnessSQL Training Lifts Qwen3 on Spider 2.0-SQLite

The core insight is that evaluation harnesses create a train-deploy mismatch: models trained to emit one static query underperform when deployed agents must probe schemas and revise queries interactively. Teams building agentic SQL or tool-use systems should consider matching training and inference environments rather than adding capabilities at fine-tuning time. The method uses open Qwen models, so results are reproducible without proprietary systems.

#Qwen3 #HarnessSQL #text-to-SQL #reinforcement learning

elvis@omarsar0Source

Papers BuildersMulti-sourceSelected04:44

Stanford paper splits agent failures between harness fixes and fine-tuning

The actionable shift is diagnostic: the same failure mode responds to different interventions, so prompt and tooling edits can waste effort on problems only weight training solves. Teams building agents can use the process-versus-content taxonomy as a triage step before committing compute. The Qwen3.5 LoRA numbers give a concrete baseline for how much fine-tuning moves plan quality on a small model.

#Stanford #Qwen3.5 #agents #fine-tuning

Event

rohanpaul_ai@rohanpaul_ai6 sourcesSource

Papers BuildersSelected01:48

Sakana AI paper proposes MASS for recursive self-improvement

The non-obvious idea is that multi-agent self-supervision can replace the external verifier that most self-improvement loops depend on, extending the approach to open-ended tasks with no ground-truth checker. Researchers working on self-improving agents and efficient training should read the paper and the accompanying Chat with Paper walkthrough to see whether the workflow-search setup transfers to their own models.

#Sakana AI #MASS #Qwen #agents

elvis@omarsar0Source

Papers BuildersSelected11:15

NYU and Amazon skill selection paper

This research reveals that quality trumps quantity in skill distillation, potentially reducing computational costs by over 90% while maintaining performance. For developers and researchers working with large language models, this method offers a practical way to optimize training efficiency and improve model performance on specific tasks like math contests.

#SGUID #skill distillation #Qwen models #NYU

rohanpaul_ai@rohanpaul_aiSource

Papers Builders10:49

Chinese contentWOVEN：视觉世界建模融入多模态大模型

WOVEN解决了多模态大模型在空间、物理和时间推理上的缺陷，训练数据可替代30-50%任务数据，效果显著。

#WOVEN #多模态大模型 #视觉转换推理 #GPT-5.4

arXiv cs.LG@Zheyu Fan et al. (12)Source

Papers BuildersMulti-source10:21

Chinese content评估LLM生成评分标准的干预转移方法

论文提出了一种创新的评分标准评估方法，揭示了不同LLM生成评分标准之间的不对称性，对AI性能监控有重要启示。

#评分标准 #干预转移 #LLM评估 #HealthBench

Event

arXiv: DeepSeek@Erik Skalnes et al. (4)5 sourcesSource

Papers Builders10:20

Chinese contentGeoReform：几何问题求解的反思形式化进化框架

GeoReform让几何问题求解更准了，Qwen3VL-2B准确率提升14%，解决了形式化中的冗余和歧义问题。

#GeoReform #Geometry3K #Qwen3VL #多模态

arXiv cs.AI@Jialu Wang et al. (5)Source

Papers BuildersMulti-source09:52

Chinese content论文提出按失败构成判断该改 harness 还是训练权重

论文教你看 agent 失败原因来决定改运行时框架还是微调模型，Qwen3.5 上从 0.16 拉到 0.30，做 agent 的都该看看这套诊断方法。

#Qwen3.5 #LoRA #WebArena-Lite #智能体

Event

arXiv cs.LG@Yuan Tian et al. (5)8 sourcesSource

Papers Builders09:44

Chinese contentElastic Expert Routing：软化 MoE top-k 路由边界，提升 OLMoE 与 Qwen3 微调效果

一篇 MoE 训练论文：把固定 top-k 换成在 k 附近随机采样，Qwen3-30B-A3B 微调能多拿 2 分，推理成本不变。

#MoE #Elastic Expert Routing #OLMoE #Qwen3

arXiv cs.LG@Yunkai Chai et al. (7)Source

Papers BuildersMulti-sourceSelected11:48

NVIDIA paper: tools weaken multimodal model safety refusals

Plain-chat safety evals can overstate how safe an agentic system actually is, since tool use measurably erodes refusal behaviour across every model tested. Teams building multimodal agents should benchmark safety with tools attached and consider re-inserting the original request before generation. The paper is on arXiv with a chat interface at academy.dair.ai for quicker exploration.

#NVIDIA #multimodal #AI safety #agents

Event

elvis@omarsar011 sourcesSource