Search “Reasoning”

OpenClaw launches Particle-1

On November 3, 2026, OpenClaw launched Particle-1, and the model is now available in OpenClaw. According to @particle_api, compared with DeepSeek V4.1 Flash, Particle-1 reduces output tokens by about 45%, achieves comparable benchmark accuracy, and retains no data during standard inference. Users can try it by upgrading to OpenClaw v2026.10.10. There are currently no further developments.

LG AI Research proposes DistPFN to mitigate label shift in tabular models

In an ICML 2026 blog post, LG AI Research pointed out that tabular foundation models such as TabPFN suffer from label shift: among 253 OpenML datasets, about 85% (214) are class-imbalanced, and in extreme cases 98.3% of test samples were predicted as the majority class, with minority-class recall nearly collapsing. To address this, LG AI Research proposed DistPFN, which at inference time multiplies predicted probabilities by a correction factor based on relative training class proportions, requiring no architecture changes, retraining, or estimation of the test label distribution. The work is currently published as an official blog post, with no further progress disclosed.

Report: DeepSeek V5 plans pretraining on Ascend chips

On November 3, 2026, according to a report by Star Market Daily, multiple industry chain sources said that DeepSeek's next-generation large model, DeepSeek V5, plans to conduct pretraining based on Huawei Ascend chips; previously, Ascend chips were mainly used for inference tasks of DeepSeek large models. East Money reported the same day that after the release of DeepSeek-V4 in April 2026, Huawei said that its full series of Ascend SuperNode products could fully support the V4 series models; on September 30, DeepSeek open-sourced infrastructure components for the Ascend computing platform, covering TileLang compilation tools, DeepGEMM, FlashMLA, DeepEP, and others. The report also said that domestic computing power deployment is currently still concentrated on the inference side, and that the full-parameter post-training of DeepSeek-V4-Pro previously completed by an Ascend 910C cluster did not involve basic pretraining. There is currently no official response from DeepSeek or Huawei.

ARC Prize publishes evaluation results for DeepSeek V4.1 Flash 1017

On November 3, 2026, ARC Prize published the verified ARC-AGI results for DeepSeek V4.1 Flash 1017: 78.1% on ARC-AGI-2 at $0.16 per task, and 94.5% on ARC-AGI-1 at $0.06 per task; on ARC-AGI-3 it scored 5.8% at a cost of $1.2K under the provider adapter harness, and 0.2% at $1.0K under the standard harness. ARC Prize stated that the model leads open-weight models on ARC-AGI-3 and ARC-AGI-2, and matches the best open-weight model results on ARC-AGI-1. Compared with the original V4.1 Flash of September 10, 2026, 1017's ARC-AGI-3 score rose from 2.68% to 5.81% on the semi-private set while falling from 11.05% to 8.97% on the public set; at maximum reasoning effort, the reasoning tokens consumed by its first action decreased by an average of 36% across 25 public games and by 48.6% across 55 semi-private games.

AWS releases a tutorial on distilling a shopping agent with SeqKD on SageMaker AI

On November 3, 2026, the official AWS Machine Learning blog published a tutorial on how to build an efficient shopping agent on Amazon SageMaker AI using sequence-level knowledge distillation (SeqKD). The tutorial uses GLM-5.3 on Amazon Bedrock as the teacher model to generate trajectories, uses the ShoppingBench scorer to retain only trajectories that solve the task, and then performs serverless SFT on the Qwen3.5-4B student model. On 324 held-out test questions, the average absolute success rate of the distilled 4B student rose from 0.297 to 0.541, essentially on par with the teacher model and mid-sized frontier models; when the teacher's reasoning process was removed and only tool calls were learned, the average success rate dropped to 0.415. The tutorial is now publicly available alongside the official blog post.

NVIDIA Expands Cosmos Platform Training Tools and Data Blueprints

On October 22, 2026, NVIDIA announced the expansion of the Cosmos platform, adding training tools, simulation checkpoints, and Cosmos Data Factory blueprints to help developers build world models. According to information released by the NVIDIA Newsroom, Cosmos Data Factory provides five cloud-native blueprints, now available through GitHub, Microsoft Azure, Nebius, and CoreWeave; Porsche is using it to build a shared autonomous driving data flywheel, which is expected to reduce repetitive real-vehicle testing by up to 90% and shorten development cycles by up to 85%. Cosmos Framework's training, optimization, and deployment tools are being extended in the form of agent skills to open-source models such as GR00T, Alpamayo, QWEN, and Pi0.5, and the Cosmos 3 NIM microservice delivers up to 2x faster inference under validated configurations. There are currently no further developments.

Karpathy builds HackerTube with Claude Opus 5.5 in two hours

Over the weekend, Andrej Karpathy used Claude Code together with Opus 5.5 to build the video site HackerTube in about two hours. According to a report by A Yi AI Notes, he filtered roughly 43 million Hacker News comments from the past twenty years down to 6,454 frequently recommended long videos; the site requires no login, uses a middle frame of each video as its thumbnail, and stores browsing state in the local browser session. He himself only manually applied for a Google Cloud YouTube API key, rented a VPS and bought a domain, then handed the IP over to Claude, which completed the deployment, security hardening and code optimization. A follow-up report by 36Kr · AI added that he used the desktop version of Claude Code with Opus 5.5 at Medium reasoning effort, that the page layout replicates YouTube, and that the underlying data can be exported and handed to one's own AI agent to rebuild a version. There is currently no further public progress on this matter.

Browser Use opens BU Ultrafast to all US cloud plans

On October 31, 2026, Browser Use opened its fastest browser agent, BU Ultrafast, from the waitlist to all US Cloud plans, including the free tier. Official data shows that BU Ultrafast has a median task time of 22 seconds and a cost of $0.0064 per run; by comparison, the default agent on GPT-6 Luna takes 79 seconds and costs $0.0098. To use it, simply set model to bu-ultrafast; for more step-by-step reasoning, you can switch to bu-fast from the same agent, which is billed at standard rates. There are currently no further developments.

Krea's Head of Research @sleenyre will speak at the SGLang Summit

Krea announced that its Head of Research @sleenyre will give a talk at the SGLang Summit on how to scale reinforcement learning and inference for Krea 3. The summit is scheduled to take place from November 12 to 13 at Fort Mason in San Francisco. Krea released this news on October 31, 2026, and no further details about the talk's content or preparations have been disclosed yet.

Krea Head of Research to Speak at SGLang Summit

Krea announced that its Head of Research @sleenyre will give a talk at the SGLang Summit on how to scale reinforcement learning and inference for Krea 3. The summit is scheduled to take place from November 12 to 13 at Fort Mason in San Francisco. Krea released this news on October 31, 2026, and no further details about the talk's content or preparations have been disclosed yet.

View story:Krea's Head of Research @sleenyre will speak at the SGLang Summit

OpenAI Releases GPT-Rosalind-6 System Card Addendum

On October 31, 2026, OpenAI released the GPT-Rosalind-6 system card addendum. The model uses the same set of weights as GPT-6 Astra, with no new training data or post-training added, and is provided only through trusted access to vetted scientific research institutions, enterprises, and government partners. At the highest reasoning budget, MedChemBench v7 averaged a score of 49.6% (GPT-Rosalind-5.5: 28.9%), GeneBench Pro v13 achieved a single-attempt success rate of 37.7% (previous generation: 16.7%), and LifeSciBench Gold averaged 61.3% (previous generation: 44.0%). Under the Preparedness Framework, this release is classified as High capability in the biological and chemical domains, below the Critical threshold, as Critical in the cybersecurity domain, and below High for AI self-improvement. There are currently no subsequent developments.

X.PIN Announces Benchmark Test Results for Iluvatar CoreX Tiange 300

On October 30, 2026, X.PIN relayed and published benchmark test data for the Iluvatar CoreX Tiange 300. Its report claimed that in standard inference tests, the Tiange 300's throughput was 1.05 times that of the Nvidia H200; under MoE workloads, performance improved by 15% and first-token latency decreased by 30%; and when fine-tuning Qwen3-32B, its speed was 1.33 times that of the H100. There are currently no follow-up developments or official responses.

Ibiden decides to invest 1.28 trillion won to expand CPU package substrate production

Japanese semiconductor package substrate manufacturer Ibiden has decided to invest 1.28 trillion won to expand production capacity for high-performance CPU package substrates used in AI servers, with new equipment to be phased into operation starting in fiscal 2029. Ibiden expects that demand for CPU substrates will continue to rise as AI inference workloads grow and agentic AI becomes more widespread. Another report said Ibiden plans to invest 151 billion yen (approximately 6.4 billion yuan) to expand production capacity for high-performance IC package substrates for CPUs at Building 6 of its Gama plant in Gifu Prefecture, also phasing it into operation starting in fiscal 2029, but it did not disclose the scale of the added capacity. In the first half of fiscal 2026, Ibiden posted revenue of 262.546 billion yen and operating profit of 60.116 billion yen, up 34.3% and 84.6% year on year, respectively.

Mistral releases Large 4 weights on Hugging Face

On October 29, 2026, Mistral released the open weights of Mistral Large 4 (Le Chonk), a 1-trillion-parameter native multimodal model that continues reinforcement learning on the basis of the preview version, with improved coding and defensive cybersecurity capabilities. According to data published by Mistral AI, the model scores 19.7% on FrontierCode 2.0 and 48% on the Artificial Analysis Coding Agent Index, roughly on par with GLM-5.3, but uses an average of 47% fewer tokens per task; on Terminal-Bench 4.0 it improves by 13 percentage points to 41%. The model supports three reasoning modes, uses a modified MIT Mistral Ecosystem Open-Weight license, and is priced at $0.70 per million input tokens, $0.07 for cached input, and $2.10 for output via API. On the same day, Unsloth released a 1-bit GGUF quantized version of the model, stating that it can run locally on a Mac Studio, retains 91% top-1 accuracy after quantization, and provides a download link and usage guide.

Hainan Computing Power and Tencent Cloud Sign Strategic Cooperation Agreement

On October 30, 2026, Hainan Computing Power Technology Co., Ltd. and Tencent Cloud signed a strategic cooperation agreement to jointly create an integrated "AI + computing power" solution for the Hainan government and enterprise market. According to reports by Yicai and Securities Times that day, Tencent Cloud will provide Hainan Computing Power and its customers with resources such as high-performance computing, intelligent computing power, storage and networking, to support the large-scale training and inference needs of the provincial-level artificial intelligence public service platform. The two parties will also carry out business cooperation around large-model Token invocation, and, leveraging the institutional advantages of the free trade port's cross-border data flow regime, explore commercialization paths such as Token going global and cross-border computing power export. No subsequent progress has been disclosed so far.

ARC Prize says scores on three ARC-AGI benchmarks are strongly correlated

On October 29, 2026, ARC Prize proposed a judgment that there is a strong correlation among model scores on the three ARC-AGI benchmarks. After plotting ARC-AGI model scores, the organization found that the power curve v2 = v1^4.53 could explain about 98% of the variation in ARC-AGI-2 scores; models that were flat on ARC-AGI-1 could still differ by more than 20 percentage points on ARC-AGI-2. A similar trend also appeared on ARC-AGI-3, but the low-reasoning tier of Gemini 4 Argon was a clear exception: its ARC-AGI-2 score was only 42.1%, while its ARC-AGI-3 score reached 97.5%. There have been no subsequent developments so far.

ARC Prize says models scoring high on ARC-AGI-2 also tend to score high on ARC-AGI-3

ARC Prize proposes a judgment: models that score higher on ARC-AGI-2 tend to also score higher on ARC-AGI-3. ARC Prize also points out an obvious exception—Gemini 4 Argon at the low reasoning tier scores only 42.1% on ARC-AGI-2, but reaches 97.5% on ARC-AGI-3 under the provider adapter harness. Currently this judgment is released by ARC Prize alone, and there are no other developments or facts yet.

ARC Prize announces the evaluation results of Muse Spark Open on ARC-AGI

On October 29, 2026, ARC Prize announced the verified results of Meta's Muse Spark Open on ARC-AGI: 89.0% on ARC-AGI-1 at $0.29 per task, 57.5% on ARC-AGI-2 at $0.73 per task, and 0.1% on ARC-AGI-3 under the standard harness at $3.9K per task. ARC Prize stated that its v1/v2 scores are close to those of Haiku 5.5 and GPT-6 Luna, but at roughly 9 times the cost of the latter; the model is priced on OpenRouter at $1.25 per million input tokens and $4.25 per million output tokens. As the reasoning tier increases, ARC-AGI-2 rises from 29.4% at medium (an average of 55k tokens per task) to 57.2% at xhigh (180k tokens per task).

Oracle announces the introduction of Cerebras AI chips into its cloud services

Oracle announced that it will introduce Cerebras Systems' AI chips into its cloud infrastructure, giving customers more computing power options beyond Nvidia GPUs. According to Data Center Knowledge, Oracle also plans to deploy Google TPUs and inference chips from OpenAI, Cerebras, SambaNova and Positron AI on OCI, with the TPU and Cerebras wafer-scale engine scheduled to come online in 2027, and Positron Atlas already available. CNBC Technology reported the same day that Alphabet announced Oracle will offer Google's TPUs, marking the first time customers can use TPUs outside of Google's infrastructure. No further developments have been disclosed at this time.

Google releases world model Genie 4

On October 27, 2026, Google DeepMind released the world model Genie 4, focused on real-time interactive simulation and introducing native audio to the series for the first time. According to the official introduction, the model is built on capabilities from the Gemini family: Gemini Omni is responsible for making the simulation match the appearance and behavior of the real world, while other Gemini models provide the complex reasoning needed to maintain the simulation's coherence; users can control scenes in real time using natural language, and each world also instantly generates ambient sound, music, and interactive dialogue. The next day, Google DeepMind further explained that Gemini 3.8 Flash breaks down goals in a loop and checks frames to maintain causal coherence across multi-step events, and that Gemini 4 Argon has also been used to jointly simulate tasks with Gemini Robotics; the company said it remains an early experimental model, that object permanence and complex physical interactions still drift during long runs, that all generated content carries a SynthID watermark, and that it is currently being previewed to trusted testers first. On October 28, Ethan Mollick posted that Genie 4 can run DOOM, and tested it generating a game-like environment as it goes, saying the experience is not perfect. On October 29, Peter Yang used Genie 4 to build 5 worlds in real time, arguing that world models are the next big direction for AI after LLMs.