Chinese content美国TwIL-LM3-Pro超越中国VibeThinker-3B
美国3.66B模型TwIL-LM3-Pro在形式逻辑测试中大幅领先中国VibeThinker-3B等模型，且支持本地运行。
Latest
Every published item, in order · 825 items
Oct 1
Sep 30
Bilibili releases Index-Translate models
This release expands open-source multilingual translation capabilities with specialized models for different use cases. Developers and content creators working with multilingual content can leverage these models for community expression translation, voice cloning, and maintaining consistency in long documents. The models offer practical alternatives to commercial translation services with more control over output style and formatting.
Qwen3.8-27B model available via Nebius
The model's compact size makes it efficient for complex multi-step workflows without requiring massive computational resources. Developers in ASEAN can leverage this for building applications that need advanced reasoning capabilities, offering a cost-effective alternative to larger models while maintaining strong performance in planning and task completion.
AA-AgentPerf-Local open-source inference tool
This tool democratizes performance testing for local AI deployments, enabling developers to make informed hardware decisions without relying solely on cloud benchmarks. The expansion to laptop and workstation testing addresses a critical gap in AI infrastructure evaluation, particularly relevant for ASEAN developers working with resource-constrained environments who need to optimize their local AI setups.
MetaCtrl: Metacognitive Controller for LRMs
MetaCtrl addresses a critical inefficiency in reasoning models by dynamically determining when additional computation is beneficial, eliminating the need for predefined budgets or retraining. This plug-and-play solution significantly improves accuracy while reducing inference costs, making it valuable for developers working with resource-constrained environments. The open-source implementation allows immediate integration into existing reasoning pipelines.
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.
EASE framework for self-evolving agents
EASE addresses a critical limitation in agent skill curation by solving cross-executor degradation through behavior adaptation. Developers working with multiple AI models can now use a single, more efficient curator that reduces inference tokens by 9.1-14.5% while maintaining performance across diverse executors, making it particularly valuable for organizations deploying multi-model systems.
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.
Sep 29
Chinese content阿里Qoder平台Qwen3.8-Flash免费活动延期
阿里Qoder平台Qwen3.8-Flash免费活动延期了，10月后还能继续免费使用这个多模态模型。
IT之家Source
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.
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.
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.
Chinese contentMeta RAM 团队提出 RL-XAR：用专家对齐评分标准做写作 RL
Meta 把 RL 那套可验证奖励搬进了写作：先让 Kimi-K2.6 自动生成能区分专家和 AI 的评分标准，再用它训练 Qwen3.5-27B，盲评 95% 胜率打赢人类专家文本。
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。
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。
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%。
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。
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。
Chinese content用 vLLM-Omni 在 SageMaker AI 上部署 Qwen3-TTS 实时语音应用（一）
AWS 手把手教程：在 SageMaker 上用 vLLM-Omni 跑 Qwen3-TTS，流式输出语音，想做实时语音应用的可以跟着搭。
Sep 28
Open-source visual decision model Valen scores actions from screenshots
The interesting design choice is putting a scoring head on top of Qwen3.5 small models, positioning Valen as a fast reflex layer under a slower planning agent rather than a standalone reasoning model. Developers building GUI, game or robotics agents can use it to offload millisecond-level perception-to-action decisions. It competes with hand-tuned heuristics or heavyweight VLM calls for the same reflex-loop role.
Chinese contentSimon Willison 主题演讲：2026 上半年 LLM 关键事件时间线
Simon Willison 把 2026 年的 agent 出逃、OpenClaw 爆红、模型王座更迭串成一条线，信息量很大，读完对全年脉络就清楚了。
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.
Chinese contentQwen3.8-Max 通过递归自我改进完成 33 轮迭代，Artificial Analysis 得分升至 45 分
阿里这次没发新模型，而是让 Qwen3.8-Max 自己训自己，33 轮迭代后成绩追上 Claude 和 GPT，还顺手把多模态价格砍了九成。
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。
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。
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.
Sep 26
Chinese content用户用 Qwen3.8-Flash-Next-Uncensored 破解 IDM 引发授权争议
有人花 5 欧元租 GPU，让无审查版 Qwen 三十分钟改一个字节就破解了 IDM，这帖子把 AI 逆向能力聊得很具体。
OpenRouter Releases Kev, a LoRA Adapter on Qwen3.5-4B-Base
The interesting bit is the architecture: a small pointer head plus a LoRA on a 4B base suggests task-specific capabilities can ship as lightweight add-ons rather than full model releases. Builders can pull the weights from Hugging Face and serve them through SiliconFlow or their own stack. It also shows a US product layer building directly on Alibaba's Qwen (Qwen3.5) base models.
SGLang adds scoring API and multi-item scoring for decision models
Classification and ranking workloads waste GPU cycles recomputing shared context per candidate, and MIS cuts that cost by an order of magnitude at higher candidate counts. Developers serving scoring endpoints for agent action selection or reranking can switch to /v1/score and measure the p95 gains directly using the benchmark commands in the blog. The LinkedIn contribution shows production demand for this pattern beyond the LMSYS ecosystem.
Chinese contentQwen3.8-Max 完成 33 轮自动迭代，Artificial Analysis 分数从 40 升到 45
阿里 Qwen3.8-Max 被曝跑了 33 轮自动训练迭代，AA 分数从 40 涨到 45，还有 2.4T 参数的说法。
Qwen3.8-Omni-Flash: Qwen's multimodal model for long-horizon agent tasks
Omni-modal models are shifting from perception demos to executing multi-step agent work on audio and video, a capability most Western agent stacks still lack. Builders can plug Qwen-MM-Plugins into existing harnesses or use Qwen-Live-Harness for real-time pipelines without building memory and tool routing from scratch. As an open release from Alibaba (Qwen), it gives developers in Southeast Asia and beyond a free alternative to closed multimodal APIs.