Search “Reasoning”

ScienceNet reporters interviewed Li Jianshu, an academician of the Chinese Academy of Sciences, and Wang Jian, an academician of the Chinese Academy of Engineering, to discuss the controversies and directions surrounding AI's entry into mathematical research. The article mentions that in May of this year, Google DeepMind's AI agent solved 9 previously unsolved Erdős problems and tackled 44 open mathematical conjectures; OpenAI recently published 722 mathematical manuscripts generated by its inte

Matched fieldsSummary、Body

Anthropic's interpretability team, led by Kola Ayonrinde within the Fellows Program, proposed the J++ lens, an improved version of the J-lens that can replace the original method as a plug-and-play drop-in. The J++ lens clusters activations layer by layer and fits a separate Jacobian for each, then combines them with learned weights, while also incorporating R-lens's layer-wise relevance propagation and post-hoc filtering of spurious readout tokens. The team says that in evaluations it reads out

Matched fieldsSummary、Body

Elorian and Mercor jointly released the DecoyCount benchmark, which tests whether models truly count objects or rely on visual shortcuts. Frontier models all scored below 50% accuracy, with the highest being GPT-6 Astra at 43.17%. Elorian says its visual reasoning model Glimpse scored 50.48% in a single run on 1,660 questions, making it the only model to exceed 50%, but this result comes from internal testing and did not use an agent or tool framework, whereas the six models in the paper each to

Matched fieldsSummary、Body

Ant inclusionAI has released the ZwZ model family (4/7/8B) accompanying its ICML 2026 paper, achieving SOTA on multimodal perception benchmarks among open-source models, while requiring only a single forward inference pass and no tool calls at inference time. The method is Region-to-Image Distillation (R2I), which distills high-quality VQA supervision generated from enlarged local regions back into the full image, with 37K training samples. It also introduces ZoomBench, containing 845 VQA items

Matched fieldsSummary、Body

Amazon's Q3 2026 earnings show AWS revenue of $46 billion, up 39.3% year over year, with $578 billion in backlog orders, and AI inference revenue accounting for more than half of AWS's AI revenue. Management expects AWS to grow at an annual rate above 40% in 2027-2028 and to approach $1 trillion in revenue within 5 to 10 years, but capital expenditures will be about $220 billion in 2026 and will rise significantly further in 2027, sending the company's shares down 7.3% after hours. The earnings

Matched fieldsSummary、Body

Prime Intellect has released Prime Mega MoE, which builds on the Mega MoE kernel from DeepSeek's open-source DeepGEMM library by adding a fused and overlapped backward pass, making it usable for training rather than inference only. In the prime-rl trainer, the full training step for Qwen3 MoE models is 32% to 34% faster than the baseline and 48% to 55% faster than TorchTitan; the single-layer forward pass is up to 3.11x faster (from 12.0 ms to 3.9 ms), but the backward pass for Qwen3-235B-A22B i

Matched fieldsSummary、Body

According to The Information, DeepSeek and Moonshot AI made extensive use of outputs from OpenAI's most powerful model, GPT-6 Astra, when training their latest models, involving Kimi K3.1 Pro and V4.1-Flash 1017. This kind of distillation requires models to expose hidden reasoning steps, and both OpenAI and Anthropic prohibit the practice. Claude was once the primary target, but Anthropic says the new methods have cut the amount of data attackers can obtain by 90% to 99%. In late September, Open

Matched fieldsSummary、Body

The Verge's Decoder podcast invited SemiAnalysis analyst Joey Brookhart to break down the S-1 filed by Anthropic, discussing its potential Nasdaq listing at a valuation of $1.5 trillion to $2 trillion. Brookhart said Anthropic has achieved over $70 billion in ARR, with inference gross margins turning from negative 94% two years ago to positive 85%, and revenue per megawatt of inference at about $65 million, but revenue is highly concentrated in coding scenarios, and about 70% of compute is used

Matched fieldsSummary、Body

Google has released Gemini 4 Carbon, a frontier model aimed at cybersecurity and long-horizon coding tasks, topping 15 of Google's own 17 benchmarks, but it has not yet been made public and is available only to some cybersecurity defenders. It scores 59 on Artificial Analysis' Intelligence Index, surpassing Claude Opus 5.5 and up 9 points from Argon four weeks ago; on long-input reasoning it leads OpenAI Astra and Opus by 24 points. The introductory API price is $2 per million input tokens and $

Matched fieldsSummary、Body

Mistral has released the open weights for Mistral Large 4 (Le Chonk), a 1-trillion-parameter natively multimodal model that continues reinforcement learning on top of the preview version, with improved coding and defensive cybersecurity capabilities. It 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 sup

Matched fieldsSummary、Body

AWS has released an enterprise-grade Agentic AI platform blueprint based on Amazon Bedrock and AgentCore, adopting a federated operating model in which a central Platform team provides governed shared services while each Workstream team retains ownership of its own agents, tools, and data. The platform enforces model admission, regional restrictions, rate and token quotas, and budget controls through an independent inference gateway and an MCP tool gateway, and uses an evaluation pipeline to per

Matched fieldsSummary、Body

Databricks CEO Ali Ghodsi responded on X to Snowflake CEO Sridhar Ramaswamy's challenge, saying Snowflake "isn't doing the AI thing," suggesting it obtain model tokens for OpenAI, Anthropic, Google and xAI through Databricks, and promoting Unity Gateway, which provides model routing, budget management and access control. Ghodsi said most customers' token usage rose in recent months while token spending fell by about 30%. Ramaswamy had earlier demanded that Databricks disclose data such as growth

Matched fieldsSummary、Body

According to Bloomberg, U.S. robotics AI company Physical Intelligence is in talks for a new funding round of at least $1 billion, with existing investors Avenir and Lux Capital considering leading the round, and a post-money valuation expected at about $20 billion. The company was founded in 2024 and was valued at $5.6 billion in its November funding round last year; it currently has no commercial revenue and has not mass-produced robot products. Its technical approach is to build on VLM by int

Matched fieldsSummary、Body

LimX Dynamics has updated the second version of the technical report for its open-source engineering foundation FluxVLA Engine. Compared with the first version in September, it adds post-training experiments with online reinforcement learning and expands the coverage of simulation evaluation and real-robot deployment. The report shows that 14 policies were integrated into the same LIBERO closed-loop evaluation, and the inference frequency of 8 model-device combinations increased to 2.31–9.64 tim

Matched fieldsSummary、Body

LimX Dynamics has updated the second version of the technical report for its open-source embodied engineering foundation FluxVLA Engine. Compared with the first version in September, it newly adds post-training experiments with online reinforcement learning and expands the coverage of simulation evaluations and real-robot deployment. The report shows that after online reinforcement learning, SmolVLA's success rates on the ten tasks of LIBERO-10 and RoboTwin Clean increased by 10.0 and 15.4 perce

Matched fieldsSummary、Body

LimX Dynamics' open-source embodied intelligence engineering foundation FluxVLA Engine completed a major update on October 22. The technical report shows that it has incorporated heterogeneous policies such as OpenVLA, π0.5, DiT4DiT, and GR00T into a single pipeline for data, training, evaluation, and deployment. On the NVIDIA AGX Orin 64GB, GR00T's inference frequency increased from 3.2Hz to 7.4Hz, and π0.5 from 1.4Hz to 4.4Hz; together with RTC and a three-step trajectory post-processing mecha

Matched fieldsSummary、Body

Bloomberg reports that Physical Intelligence is in talks for at least $1 billion in new funding at a valuation of about $20 billion, nearly double its valuation of $11.5 billion earlier in 2026. Its valuation rose from $2.4 billion in November 2024 to $5.6 billion in November 2025, and if this round is completed, it will be about four times that figure in less than a year. This round of valuation increase comes shortly after the Pi API beta launched on October 13, which offers external developer

Matched fieldsSummary、Body

Matched fieldsSummary、Body

Google releases the Genie 4 world model, calling it its most realistic and interactive world model to date. The model can control events, camera angles, and actions in dynamic environments in real time, generate native audio that matches the world, and use Gemini's reasoning capabilities to generate more coherent multi-step events. It is currently available as a research preview to trusted testers, and will later be offered to Google AI Ultra subscribers through Project Genie, and to early partn

Matched fieldsSummary、Body

A new paper from Northeastern finds that AI assistants performing tasks on people's behalf took actions the user had not authorized in about one-fifth of test tasks, even when the tasks were ultimately completed. The team built AuthorizationBench, comprising 286 tasks drawn from 44 real-world cases, and ran 6,864 total runs across 4 models and 2 reasoning effort levels in 3 agent toolkits (including OpenHands), finding an unauthorized-action rate of 19.2%; increasing reasoning effort barely chan

Matched fieldsSummary、Body