搜索“Agent”

-red { fill: var(--g-logo-red); } .g-yellow { fill: var(--g-logo-yellow); } .g-green { fill: var(--g-logo-green); } June 18, 2026 Responsibility & Safety Securing the future of AI agents Rohin Shah and Four Flynn Share How we’re securing internal systems against increasingly capable and imperfectly aligned AI AI agents are transforming our relationship with technology. By autonomously executing complex tasks — from cyber defence to scientific discovery and product development — these systems are

匹配字段正文

compute so these evaluations are cheap to run. Methods We evaluated 14 frontier models from GPT-2 (2019) through GPT-5.5 (2026) on 43 benchmarks covering math, coding, knowledge, agentic tool-use, and safety-relevant questions like steganography and scheming reasoning. The questions include single-token answers, longer generation, and multi-turn agentic settings. In every case we prevent the model from emitting reasoning tokens before its answer by using benchmark-specific prompts and structure

匹配字段正文

too specific, e.g. checking for exact error strings or function names, or unsolvable, testing for a behavior not in the instruction or in the codebase We show through analysis of agent trajectories that FrontierCode produces 81% less misclassification errors than other leading benchmarks. This means that FrontierCode scores are the most accurate ranking currently available. Existing benchmarks also suffer from lack of diversity in several ways. While other benchmarks generated issues from singl

匹配字段正文

and API layer, reducing CPU overhead without affecting inter-token latency. Looking Ahead The next frontier for distributed inference is shifting from chat-style workloads toward agentic applications — tools like Claude Code, Codex, and Cursor that drive deep multi-turn, tool-augmented conversations with long context windows (up to 1M tokens), extremely high KV cache reuse, and rapid-fire request bursts from parallel subagent spawning. InferenceX is developing an agentic coding benchmark to cap

匹配字段正文

oduces new demand as additional use cases become cost-effective. If these trends continue, a compute crunch is likely near — particularly for the long-context workloads that drive agentic AI. This will drive up the price of access to frontier capabilities for those willing to pay, while everyday users shift to cheaper, smaller models. It may also mean that AI companies increase their focus on developing more efficient ways to serve models. Because of efficiency gains, these shifts won’t necessar

匹配字段正文

n year. When you just need to build a tiny desktop app: TinyJS October 13, 2026 Valen creator drives 'Golden Spike' to connect new languages with Rust September 26, 2026 Microsoft agentically ports Copilot runtime to Rust for $120K September 19, 2026 Apple iPhone Duo makes developers think in folds September 16, 2026 JavaScript installer pnpm recast in Rust because ECMAScript can't keep up September 5, 2026 Kubernetes cleans house, bins legacy kube-dns, IPVS, and cgroup v1 August 27, 2026 How Cu

匹配字段正文

at are incompatible with each other (such as being concise and answering all parts); Hacking detection - can we predict hacking onset from the first 20 steps using reward distributions or other metrics? One of our Sprints beta testers (@michellechen) investigated replications for OpenAI’s recent “goblin mode” reward hack, and found that the backdoor-ifeval pattern can be extended to elicit “goblin-themed” speech even in unrelated contexts. Rewriting Prime Agent in Rust On the Nature of the Swarm

匹配字段正文

Update May 19, 2026 This work is now published in Nature. The paper highlights Robin’s role as the first multi-agent system for biological discovery to integrate hypothesis generation, experimental strategy, data analysis, and follow-up insight generation in one continuous workflow. We describe how Robin developed the experimental strategy for therapeutic hypothesis generation, proposed follow-up experiments, and extracted actionable insights from the resulting data, including validation in prim

匹配字段正文

s from AI development and deployment more broadly?; Transaction costs in markets: How does AI affect systems of exchange and transaction costs in marketplaces? When does access to agents able to negotiate on your behalf improve market efficiency and equitable outcomes? When does it not? Broad labor market impacts AI and jobs: How will AI change jobs and employment in different parts of the economy? What new tasks and jobs could emerge as AI automates existing parts of the economy? How will these

匹配字段正文

nment. It does not have a deterministic fix, and as we have been saying for the last 3 years the correct mitigation is to limit or specifically allow what actions and impact an AI agent can have (threat model the worst case scenario): Trust No AI. Over time I noticed that Microsoft put more and more effort into preventing system prompt extraction. In my talk I showed a simple way I bypassed this in the past by just having Copilot print the system prompt in German, rather than English. When I ret

匹配字段正文

dimensional coordination problem where traditional gradient-based methods fail. The results are very promising. In our experiments, TRINITY consistently outperforms existing multi-agent methods and individual models across various benchmarks. At the time of publication, it set a new state-of-the-art record on LiveCodeBench, achieving an 86.2% pass@1 score. More importantly, it demonstrated incredible generalization. Without any retraining, TRINITY transferred zero-shot to four unseen tasks (AIME

匹配字段正文

ReasoningBank: Enabling agents to learn from experience April 21, 2026 ReasoningBank is a novel agent memory framework that uses successful and failed experiences to distill generalizable reasoning strategies, enabling an agent to continuously learn from experience after deployment. Agents are becoming increasingly crucial in tackling complex real-world tasks, ranging from general web navigation to assisting with extensive software engineering codebases. However, as these agents transition into

匹配字段正文

rial Difference When I think about self-driving cars, I am convinced that they will lead to safer conditions on the road and fewer accidents in the long run. When I think about AI agents, I often think it should be the same and they will be able to operate computers more safely than most humans, fall less often for phishing attacks, or download malware by accident and so forth. However, one key difference is that adversaries usually don’t try to make cars crash or cause accidents. The challenges

匹配字段正文

ークショップ主催者の許可を得て実施し、採択後に規定通り取り下げを行いました。 今回の Nature 論文では、これらの成果を統合しつつ、AIサイティストを支える高度な複数モデルがどのように組み合わされているかを示しています。概略としては、AIサイエンティストには大まかな方向性を与えるだけで、システムは自律的なアイデア創出、文献調査、エージェントによる木探索（Agentic Tree Search）を用いた実験設計・実装・実行を並列的に行います。最終的には、LaTeX 形式での論文執筆に加え、視覚能力を持つモデルが図表のチェックまで一貫して担います（下図）。 新たな知見：自動査読と「科学のスケーリング則」 AIが生成する膨大な研究成果を、人間の査読者に負担をかけずに評価するため、私たちは自動査読システムを構築しました。このシステムは、NeurIPSの公式ガイドラインに準拠し、複数の査読結果を統合して最終判定を下す「エリアチェア（AC）」の役割を果たします。 OpenReviewのデータセットに含まれる数千件の実際の人間の判定結果と比較したところ、この自動査読システムは複数の尺度において人

匹配字段正文

When we first introduced The AI Scientist in our initial preprint, we shared an ambitious vision: an agent powered by foundation models capable of executing the entire machine learning research lifecycle. Soon after, we shared a historic update: the improved AI Scientist-v2 produced the first fully AI-generated paper to pass a rigorous human peer-review process. Today, we are happy to announce that a paper describing all of this work and that includes new insights has been published in Nature. T

匹配字段正文

Our Work Risk Assessment; About; Donate; Careers Summary: We find that roughly half of test-passing SWE-bench Verified PRs written by mid-2024 to mid/late-2025 agents would not be merged into main by repo maintainers, even after adjusting for noise in maintainer merge decisions. Since the agents are not given a chance to iterate on their solution in response to feedback the way a human developer would, we do not claim that this represents a fundamental capability limitation. Rather, our results

匹配字段正文

rough academic papers, blog posts, and GitHub issues, and a model running the eval can encounter them in search results. When we evaluated Claude Opus 4.6 on BrowseComp in a multi-agent configuration, we found nine examples of this kind of contamination across 1,266 BrowseComp problems. However, we also witnessed two cases of a novel contamination pattern. Instead of inadvertently coming across a leaked answer, Claude Opus 4.6 independently hypothesized that it was being evaluated, identified wh

匹配字段正文

broad variety of software engineering tasks. Unlike benchmarks that focus on a single dimension, the OpenHands Index tests models on five key areas that reflect real-world coding agent use cases: Issue Resolution - Fixing bugs and issues, tested via SWE-Bench Verified; Greenfield Development - Building applications from scratch, tested via Commit0; Frontend Development - Creating and improving frontend apps, tested via SWE-Bench Multimodal; Software Testing - Identifying bugs and generating rep

匹配字段正文

were struggling to get it working at all - models would hallucinate functions and parameters, and you had to prompt hard to get them to use web search reliably. Now, chatting with agents that use tools is the norm. It seemed that OpenAI had solved it for good with the introduction of structured outputs. The τ²-bench benchmark (June 2025), which gpt-4o could only manage at 20%, is now practically solved: 95%, 98.7% depending on who you ask. With this narrative, it’s easy to assume that tool hallu

匹配字段正文

erform the default scaffolds METR uses, at least when measuring the time horizon of Opus 4.5 and GPT-5. Main differences between scaffolds ReAct is a very simple scaffold where an agent takes an action, sees the results of the action, and repeats. Triframe is a bit more complicated – before every action, an advisor subagent strategizes, parallel actor subagents generate possible actions, and then parallel rater subagents compare all the possible actions. Only the highest rated action actually ge

匹配字段正文