搜索“Agent”

rsarial User Simulations; HallucinationsPerformance in Cases Flagged by Users; AlignmentEvaluations with Prompts Representative of External ChatGPT Usage; Expand Evaluating Coding Agents for Misalignment by Resampling Internal TrafficEvaluating Coding Agents for Misalignment by Resampling Internal TrafficMeasuring our ability to detect misalignment; Expand Chain of Thought EvaluationsChain of Thought EvaluationsCoT Monitorability; CoT Controllability; Bias EvaluationFirst Person Fairness Evaluat

匹配字段正文

otein engineering) were asked to transcribe biological protocols they have personally used in the lab. Each protocol must include precise step-by-step procedures, equipment, and reagents. If a protocol was adapted from a publication, experts were required to significantly alter at least several steps. From these protocols, they created three troubleshooting questions each, introducing subtle or realistic execution errors (e.g., improper homogenization technique) and describing the resulting fail

匹配字段正文

for prompt injection attacks against email connectors and regressed slightly for attacks into function cells. 3.5 Vision We ran the image input evaluations introduced with ChatGPT agent, that evaluate for not unsafe model output, given disallowed combined text and image input. Table 5: Image input evaluations, with metric not unsafe (higher is better) We find that GPT-5.4 Thinking performs generally on par with its predecessors. Minor regressions for harms-erotic had low statistical significance

匹配字段正文

Home; GPT-5.3-Codex System Card Published February 5, 2026 1. Introduction GPT‑5.3‑Codex is the most capable agentic coding model to date, combining the frontier coding performance of GPT-5.2-Codex with the reasoning and professional knowledge capabilities of GPT-5.2. This enables it to take on long-running tasks that involve research, tool use, and complex execution. Much like a colleague, you can steer and interact with GPT-5.3-Codex while it’s working, without losing context. Like other recen

匹配字段正文

Home; Addendum to GPT-5.2 System Card: GPT-5.2-Codex Published December 18, 2025 1. Introduction GPT-5.2-Codex is our most advanced agentic coding model yet for complex, real-world software engineering. A version of GPT-5.2 optimized for agentic coding in Codex, it includes further improvements on long-horizon work through context compaction, stronger performance on project-scale tasks like refactors and migrations, and improved performance in Windows environments — and significantly stronger cy

匹配字段正文

t-5.1-instant performs better than its predecessor, and that gpt-5.1-thinking is on par with its predecessor. 2.3 Vision We ran the image input evaluations introduced with ChatGPT agent, that evaluate for not unsafe model output, given disallowed combined text and image input. Table 3: Image input evaluations, with metric not unsafe (higher is better) We find that both the instant and thinking variations of GPT-5.1 perform generally on par with their predecessors. We are observing a regression o

匹配字段正文

Home; GPT-5.1-Codex-Max System Card Introduction; Baseline model safety evaluationsDisallowed content evaluations; Jailbreaks; Vision; Product-specific risk mitigationsAgent sandbox; Network access; Model-specific risk mitigationsExpand Harmful tasksHarmful tasksRisk description; Expand MitigationMitigationSafety training; Expand Prompt injectionPrompt injectionRisk description; Expand MitigationMitigationSafety training; Expand Avoid data-destructive actionsAvoid data-destructive actionsRisk de

匹配字段正文

hreat Creation; Radiological and Nuclear Threat Creation; Persuasion; Model Autonomy; Conclusion and Next Steps; Acknowledgments; References 1. Introduction Deep research is a new agentic capability that conducts multi-step research on the internet for complex tasks. The deep research model is powered by an early version of OpenAI o3 that is optimized for web browsing. Deep research leverages reasoning to search, interpret, and analyze massive amounts of text, images, and PDFs on the internet, p

匹配字段正文

AI agents are rapidly scaling: longer sessions, deeper reasoning chains, more autonomy. But the Transformer architecture underneath them still pays a memory tax that grows with the number of tokens it processes. As agents become more capable, that cost scales with them. Now, there’s a proliferation of hybrid architectures that use some combination of methods to achieve a better balance of memory, speed, and quality. Those methods include sparse attention, which selects a subset of past tokens to

匹配字段正文

Self-improving agents have always needed a human to write the grader. A new AWS + HSBC paper removes that step: the grader evolves alongside the agent from 10 labeled examples and trains it about as well as ground truth. Double Ratchet evolves the grader alongside the agent's skills, building it from small, mostly rule-based checks for 1 failure type each. A grader version is chosen for agreeing with 10 labeled examples, never for raising the agent's score. It kept 88% to 110% of the gain that g

匹配字段正文

新，让强化学习在 Kimi K3 的 2.8T 参数规模上变得切实可行，从而实现更快、更稳定的训练。 在本报告中，我们详细介绍了自定义内核、长上下文缓存和压缩权重传输，这些技术降低了迭代开销；此外还介绍了数值对齐和 router replay，以保持训练稳定。 这套基础设施支撑了 Harvey Tenet，其在 2026 年 8 月的评测中，在 Legal Agent Benchmark 上的全通过率达到 base K3 的 1.8 倍，并在 LAB Contracts 上取得最先进的结果。 Fireworks 无服务器训练让团队可以通过 Fireworks Training API 在 Kimi K3 上测试自己的任务和奖励函数，共享训练与 rollout 容量，无需配置专用 GPU。 对前沿模型进行专门化，意味着将其行为适配到特定领域的任务、工具和成功标准。强化学习（RL）通过一个重复循环来实现这一点：模型生成轨迹——即动作和观察的序列——评估器分配奖励，训练器更新策略，使成功行为更有可能出现。 在 Kimi K3 的规模下，实用的 RL 需要在整个训练循环中实现高效执行。Roll

匹配字段正文

AI agent teams deliver almost no better results than single agents, research finds. Evals company Vals AI tested GPT-6 Sol and Claude Opus 5.5 on the "Vibe Code Bench," both solo and as teams, at two reasoning levels: medium and maximum reasoning effort. The teams cost between 1.8x and 5.1x more than single agents. Out of four comparisons between teams and solo agents, only one showed a statistically significant improvement: GPT-6 Sol at medium reasoning, where the team scored 7.3 points higher.

匹配字段正文

in September. By March 2028, that intern is supposed to become an autonomous AI researcher under human oversight, but a key ability still seems to be missing: scientific judgment. Agents were asked to invent a new training method With its benchmark "InnovationEval," Epoch AI tested whether AI agents can conduct research on their own. The task was to invent a new method for improving language models after their initial training, then implement, test, and refine it independently. The starting poin

匹配字段正文

New Tsinghua paper finds that AI agents improving game bots from match replays can top human leaderboards, but mostly stall on games with complex rules. Getting AI to learn a winning game strategy from a limited number of matches is still hard, especially against changing rivals. They built AAArena from 12 games in Tsinghua's yearly bot-building contest, with 1,920 archived human programs as rivals. A coding agent, with its model weights unchanged, reads the rules, picks opponents, studies repla

匹配字段正文

务。 具体来说，Harness会通过任务清单记录尚未完成的事项，通过进度文件和代码记录保存已经完成的工作。 这样一来，即使模型进入新的上下文窗口，也能通过读取这些信息，了解此前做了什么、接下来该做什么。 复杂任务由此可以分步执行、检查结果，并在不同上下文窗口之间接续工作。 可以说，这套主要通过文字和代码来保存任务状态的Harness，在一定程度上推动了这一波Agent能力的爆发。 然而，在现实环境里， 光有文字记忆，显然还不够 。 因为一旦进入真实的视觉环境，模型不仅需要记住物体的位置、朝向，还要理解每次动作前后，环境究竟发生了哪些变化。 而且 模型第一次看到一张图时，未必知道哪个细节会在后面变得重要。 与此同时，这些视觉信息，也很难提前用一份文字笔记完整记录下来，更难保证模型后续执行任务时，记录的信息依然够用。 在ARC-AGI-3这样的基准中，一种已有思路是先把画面转换成由数字组成的文本网格，再让模型编写程序，模拟环境规则、验证行动方案。 （注释：ARC-AGI-3是一套交互式视觉推理基准，不预先提供游戏规则和目标，智能体需要通过观察和行动，自行摸索如何通关） 但问题在于，环境越复

匹配字段正文

New Meta paper shows that AI agents can build on each other's work, even failed attempts, when every step lives in a shared Git repo. Most ways to give agents more compute treat each run on its own. When a run ends, its partial results and failures vanish, so later runs may need to rediscover them. GitSwarm runs many identical agents on a shared repo, with no boss assigning tasks. Each agent reads past work, picks what to try, and commits its result with a list of earlier commits it used, even f

匹配字段正文

er robot doesn't always require a stronger foundation model. Sometimes, it just needs a better way to act. Recent research from SenseNova team offers an interesting example of how agent system design can improve embodied AI performance. With GPT-6 Astra's model weights unchanged, SenseNova-RoboRSI reports a RoboDojo average score of 56.83, compared with a published baseline of 28.97, alongside a 50.83% task success rate. It also reports strong results on LIBERO-PRO (94.50%) and RoboCasa (64.60).

匹配字段正文

New Stanford paper finds that harness changes fix agents that loop or stall, while agents that deliver bad plans need weight training instead. An agent can be improved by editing its harness, the prompts, tools, and checks around the model, or by fine-tuning its weights. They sorted failed runs into process failures, such as loops and used-up step budgets, and content failures, where a poor plan was delivered. On a travel-planning benchmark, an LLM-driven loop rewrote the harness, and its best r

匹配字段正文

Recent years have seen the rise of AI agents, large language models (LLMs) that take multi-step actions on behalf of users. These systems have improved rapidly, and can now send emails, browse the web, execute code, and make financial transactions, often with minimal human oversight. How are these AI agents actually being used in the real world? Despite growing adoption, little public data exists on where and how agents are being deployed. To address this gap, we analysed 177,436 AI agent tools

匹配字段正文

or that it was the only route left. GPT-6 Astra often asked for permission to perform unsanctioned actions on out-of-scope targets (Figure 5). As our evaluations are simulating an agent running autonomously, if the model attempts to ask the user a question, it always receives an automated message only: “Please proceed to the next step using your best judgement”. This is a standard approach in evaluation harnesses, including the default Inspect ReAct agent used in AISI’s cyber evaluations, and in

匹配字段正文