搜索“RAG”

entic coding RL for two weeks. With these improvements, Inkling-Small surpassed Inkling on reasoning and agentic coding benchmarks. Inkling maintains an advantage on knowledge coverage and factuality. Reasoning and Agentic Tasks Inkling-Small matches or exceeds Inkling on reasoning and agentic tasks. On Humanity’s Last Exam it scores 31.6%, ahead of Inkling’s 29.7%, and the advantage holds at every thinking budget: Inkling-Small’s test-time compute curves sit above Inkling’s throughout. On SWEBe

匹配字段正文

g based on the provided visual information. It demonstrates strong performance on charts, diagrams, and mathematical visual reasoning tasks. During inference, Inkling can also leverage a Python tool to support image understanding through operations such as zooming and cropping, while seamlessly integrating visual reasoning with code-based reasoning. As our first release, Inkling establishes a robust multimodal foundation for future work. We expect its multimodal capabilities to continue improvin

匹配字段正文

action. A typical response has the following form: Coordinates are integer-normalized to [0, 999]. The action vocabulary is platform aware: Mobile: Click, DoubleClick, LongPress, Drag, Swipe, Type, LaunchApp, Wait, CallUser, GetScreenshot, device navigation, Answer, and Finished; Desktop: pointing actions, RightClick, Drag, Swipe, Type, Hotkey, Wait, CallUser, and Finished; Web: pointing actions, Drag, directional Scroll, Hover, Type, URL Launch, Hotkey, browser navigation, CallUser, and Finishe

匹配字段正文

ource-reported action scaffolds may still differ. MobileWorld uses GUI-only success rate on 117 tasks with 50 steps, WebVoyager uses the refreshed 595-task split, Odysseys uses average rubric score over 200 tasks, VenusBench-CAPTCHA uses micro Pass@1 over all 219 examples, and VenusBench-GD uses English-instruction micro-average accuracy. “ ” denotes results reproduced by us. [2026/09] We release VenusBench-CAPTCHA, a comprehensive benchmark for evaluating GUI agents on diverse CAPTCHA tasks. Se

匹配字段正文

to introduce RoboBrain-X0, a groundbreaking cross-ontology foundation model designed to overcome the limitations of single-robot systems in heterogeneous ontology transfer. By leveraging End-Effector pose representation in SE(3) task space, coupled with a Unified Action Vocabulary (UAV) and action tokenizer, RoboBrain-X0 achieves efficient zero-shot generalization and complex task decomposition. Its Grouped Residual Quantizer (GRVQ) maps continuous control sequences from diverse degrees of freed

匹配字段正文

first-party and third-party use (API and Enterprise) unless customers meet strict criteria for removal. To prevent CSAM generation, we have built a dedicated safety stack that leverages system mitigations used across our other products as well as additional safeguards developed specifically for Sora. For full details on our specific CSAM safety stack can be found on our Sora System Card1. OpenAI Sora System Card: https://openai.com/index/sora-system-card/↩︎; Teen Safety: Sora 2 has a number of a

匹配字段正文

mentally trained from GPT-5.5 with training data related to beneficial life sciences capability. Unlike GPT-5.5, it is trained not to refuse sophisticated biology queries, and leverages a trusted access and responsible deployment structure as the primary safeguard. 3. Beneficial Capability Evaluations The Beneficial Capability Evaluations below measure how well GPT-Rosalind-5.5 performs on the intended research-support tasks it was designed to improve. Preparedness evaluations, in the following

匹配字段正文

fine-tuned the gpt-oss-120b model for these two categories. OpenAI’s Safety Advisory Group (“SAG”) reviewed this testing and concluded that, even with robust fine-tuning that leveraged OpenAI’s field-leading training stack, gpt-oss-120b did not reach High capability in Biological and Chemical Risk or Cyber risk; Would releasing gpt-oss-120b significantly advance the frontier of biological capabilities in open foundation models? We found that the answer is no: For most of the evaluations, the de

匹配字段正文

They were built around cases in which our existing models were not yet giving ideal responses, and this is reflected in the scores below. Error rates are not representative of average production traffic. The primary metric is safe completion rate, checking that the model’s response and actions are not disallowed according to our safety policies. Our evaluations are run on the model without system-level safeguards to ensure the model’s underlying behavior meets our safety bar. We continue monito

匹配字段正文

They were built around cases in which our existing models were not yet giving ideal responses, and this is reflected in the scores below. Error rates are not representative of average production traffic. The primary metric is safe completion rate, checking that the model’s response and actions are not disallowed according to our safety policies. Our evaluations are run on the model without system-level safeguards, to ensure the model’s underlying behavior meets our safety bar. We continue monit

匹配字段正文

rovements to prompt injection and instruction hierarchy robustness compared to their predecessors. GPT-6.1 Sol is also highly robust. The plot below shows defender success rate averaged per defender query on indirect prompt injection attacks (higher is better). Figure 3 Table 6: Defender success rate averaged per defender query on instruction hierarchy attacks (higher is better). 5. Health For a full description of these evaluations, please see the Health section in the GPT-6 Astra card. 5.1 Hea

匹配字段正文

They were built around cases in which our existing models were not yet giving ideal responses, and this is reflected in the scores below. Error rates are not representative of average production traffic. The primary metric is not unsafe, checking that the model did not produce output that is disallowed under the relevant OpenAI policy. Our evaluations are run on the model without system-level safeguards, to ensure the model’s underlying behavior meets our safety bar. We continue monitoring thes

匹配字段正文

They were built around cases in which our existing models were not yet giving ideal responses, and this is reflected in the scores below. Error rates are not representative of average production traffic. The metric is not unsafe, checking that the model did not produce output that is disallowed under the relevant OpenAI policy. Values from previously launched models are from the latest versions of those models, and are subject to some variation. Values may vary slightly from values published at

匹配字段正文

g based on the provided visual information. It demonstrates strong performance on charts, diagrams, and mathematical visual reasoning tasks. During inference, Inkling can also leverage a Python tool to support image understanding through operations such as zooming and cropping, while seamlessly integrating visual reasoning with code-based reasoning. As our first release, Inkling establishes a robust multimodal foundation for future work. We expect its multimodal capabilities to continue improvin

匹配字段正文

entic coding RL for two weeks. With these improvements, Inkling-Small surpassed Inkling on reasoning and agentic coding benchmarks. Inkling maintains an advantage on knowledge coverage and factuality. Reasoning and Agentic Tasks Inkling-Small matches or exceeds Inkling on reasoning and agentic tasks. On Humanity’s Last Exam it scores 31.6%, ahead of Inkling’s 29.7%, and the advantage holds at every thinking budget: Inkling-Small’s test-time compute curves sit above Inkling’s throughout. On SWEBe

匹配字段正文

rom interactions generated more than a day earlier. Learning to reason efficiently We trained Beam with a controllable length penalty that rewards successful solutions while discouraging unnecessary tokens. Early in RL, performance improved even as completion lengths fell: the model learned to solve tasks more effectively with less reasoning. Later, as Beam developed stronger agentic capabilities, completion lengths grew again, but those additional tokens supported further gains in performance.

匹配字段正文

testing approach. Microsoft tested Decision-1 with eight kinds of perturbations, including reordered options and paraphrased descriptions, which changed 1.3% of its answers on average. JevOut didn’t test Decision-1, so it’s unclear how the model would hold up against similar attacks. Microsoft also hasn’t confirmed whether Decision-1 fully supports the System One API competitors adopted, leaving questions about interoperability and how much trust to place in its confidence scores. The post Micr

匹配字段正文

8-token context window; Runs on: Hosted API only (Microsoft Foundry, OpenRouter via Azure). No open weights, no quantized variants, hardware not disclosed; Performance: Highest average accuracy in Microsoft’s 36-benchmark comparison, at 85 ms p50 latency; Best: 83.5% average accuracy across 36 benchmarks, ahead of Quyet-1.0-Large at 81.9%; Worst: Calibration of 92.2, second to Quyet-1.0-Large at 93.1. Text-only, no explanations; Bottom line: Best thing: fast, cheap, calibrated decisions at $0.04

匹配字段正文

long engineering sessions, navigate massive repositories, and orchestrate terminal tools. Kimi K3 also excels in tasks blending software engineering with visual reasoning — it leverages screenshots and visuals to optimize game dev, frontend, and CAD. The case studies below show how Kimi K3's coding capability translates into open-ended software creation and scientific research. Kernel Optimization We tested the models' capability to optimize GPU kernels. Each model works independently in an iden

匹配字段正文

ruction adherence. It follows prompts with higher precision, addressing each requirement clearly and systematically—often expanding on every mentioned point to ensure thorough coverage. In academic, research, and long-form analytical writing, it excels at producing rigorous, logically coherent, and substantively rich content, making it particularly effective in scholarly and professional contexts. Personal & Emotional: When addressing personal or emotional questions, K2 Thinking responds with mo

匹配字段正文