搜索“Agent”

e org and let them figure it out. And enterprise AI is harder to begin with, navigating security, large data complexity, and privacy. Uber's answer was the scalpel. They call them Agentic Pods, and each is only two people: one AI-proficient engineer and one domain expert from Finance, Marketing, Ops, etc., running a surprisingly short 10-day sprint. Shadow the expert, build alongside them, ship. The result: a financial pacing report that took a key player two full days now runs in 10 minutes. Th

匹配字段正文

el, not inside it "Memory" here doesn't mean replaying a past transcript. It means a guideline set — strategies that worked, mistakes to avoid, and edge cases — distilled from the agent's own prior trajectories. The loop is straightforward: The agent attempts tasks and produces trajectories. ALTK-Evolve extracts behavioral guidelines from both its successful and unsuccessful runs. It consolidates those guidelines into a reusable set. At inference time, the agent receives either the full guidelin

匹配字段正文

uating AI research capabilities. We were especially uncertain about what to expect from newer models such as Claude Fable 5, Kimi K3, and GPT-5.6 Sol. In our previous experiments, agents struggled to come up with new ideas. One potential reason is that they over-focused on existing PRs. This time, we didn't give them access to the internet at all. The most striking result is the gap between models. It appears at every stage of the research process: which experiments they choose, how carefully th

匹配字段正文

e reasons to be skeptical of the magnitude. Early Work on Monitorability Evaluations We show preliminary results on a prototype evaluation that tests monitors' ability to catch AI agents doing side tasks, and AI agents' ability to bypass this monitoring. How Does Time Horizon Vary Across Domains? We build on our time-horizon work and analyze 9 benchmarks for scientific reasoning, math, robotics, computer use, and self-driving in terms of time-horizon trends; we observe generally similar rates of

匹配字段正文

的 token 上似乎相当。 这两种方法在很大程度上是互补的：它们之间只有 12 个共同漏洞。协调型集群能够将注意力集中在它认为最容易挖掘漏洞的地方，而独立智能体则被预先分配了搜索位置。集群中的智能体为自己构建了工具，并学会了专精于特定类型的漏洞发现。 我们预测，未来这种专精化与协调将胜过无协调的暴力搜索。 In the experiment above, agents in the agent swarm don’t directly rely on one-another’s work: if one misses a bug, it won’t directly undermine the work of another. But when agents do depend on one-another, coordination gets much more difficult. Larger software engineering projects are one place this matters: they typically develop rich—and dyn

匹配字段正文

to patients at natural conversational speed while simultaneously performing careful clinical reasoning and continuously processing visual and auditory streams. Currently, a single agent cannot satisfy all these requirements. Deep reasoning takes time, but conversational pauses erode patient trust and rapport. To address this challenge, AMIE (Video) uses an asynchronous multi-agent architecture that divides labor across three specialized agents working continuously in parallel: This decoupled des

匹配字段正文

of rules. A lesson five different tasks discovered is a different object from one that appeared once, and both are worth keeping. So on the core question — should you compress an agent's hard-won lessons into a tidy summary? — ACE and ALTK-Evolve give the same answer: no. Count them, don't collapse them. ACE's per-bullet counters and our support counts are two spellings of the same idea. Where we differ Two places: how the memory is built , and how it's delivered — and it's the delivery differe

匹配字段正文

prevent stale requests that finish after a close or reopen from restoring released references. We evaluate DeepSeek-V4-Pro and Qwen3.5-397B-A17B with TP8 and HiCache on SWE-bench agent trajectories. The baseline uses ordinary HiRadixCache with LRU. The comparison enables Unified Radix Cache and --enable-session-radix-cache. Because this changes both the cache implementation and the eviction policy, the observed differences should not be interpreted as an isolated ablation of session awareness.

匹配字段正文

s callback and hook system is a supported extension point for modifying responses at the forwarding layer. Injecting Tool Calls Even more interesting though, if the clients are AI agents with tool access, an injected response can carry a tool-call. Because the output is changed after inference, this bypasses prompt-level defenses. It does not by itself bypass client-side tool authorization (unless you run in yolo mode). Pretty scary stuff. Demo Walkthrough with Screenshots The manual sequence be

匹配字段正文

n automated protocol to evaluate the integrity of AI-generated papers. Large language models (LLMs) are increasingly being deployed not just as coding assistants but as autonomous agents capable of conducting end-to-end scientific research workflows. Recent systems (e.g., Sakana’s AI-Scientist , AutoResearchClaw , DeepScientist , AI-Researcher ) can review literature, formulate hypotheses, execute experiments and write complete manuscripts that are comparable to human-authored papers. However, a

匹配字段正文

ful even before the training path is fully accelerated. Beyond this ablation, humans& uses the same recipe family and components in production for large-scale, long-context, multi-agent asynchronous RL research. Online NVFP4 Model Serving The same per-token NVFP4 recipe can serve FP8/BF16 model checkpoints with online post-training quantization (PTQ), delivering good model quality without calibration artifacts, while keeping compute performance on par with traditional NVFP4 serving. Online per-t

匹配字段正文

easons making the scheme (as it currently stands) an attractive PQC signature candidate. Discovery process To find the attack, Claude Mythos Preview worked semi-autonomously in an agentic harness, with occasional human guidance and nontechnical direction. Mythos found the attack after an extensive literature review to understand the state of the art, and substantial mathematical reasoning and computational experiments. After finding the attack, Mythos implemented an end-to-end verification pipel

匹配字段正文

ject Fetch: Phase two, we’re already seeing improvements in model capability such that their ability to use off-the-shelf robots is on track to approach the ease with which coding agents use software tools. Working again with our partners at Andon Labs, we developed a new series of demonstrations and evaluations that assess AI models’ ability to use a flying drone to autonomously perform a simple locate-and-follow task of the kind used in aerial surveillance, culminating in a new benchmark: Dron

匹配字段正文

SymptomAI: Towards a conversational AI agent for everyday symptom assessment July 22, 2026 We present a first-of-its-kind research of AI for differential diagnosis and symptom checking through a national-scale study. A large proportion of clinical diagnoses can be derived from language-based interviews alone. These diagnostic interviews are typically conducted by clinicians through doctor-patient interactions during in-person or remote visits. While these interactions are the gold standard for s

匹配字段正文

other qubits. This can lead to errors, and error-prone quantum systems are neither practical nor useful. AI For QEC A lot of work is being done to see how AI – both generative and agentic – can help address error correction in quantum systems. For example, scientists from the University of Pennsylvania and Hao Tang of Peking University in China addressed the issue in a research paper early last year. Nvidia in April ran out Ising, a family of open source AI models that the vendor said can provid

匹配字段正文

Ekoparty in 2024, where he talked about injecting ANSI Escape Codes into log files. It’s a great talk, I recommend watching it. How could this be meaningfully exploited? LLMs and Agents! Rendering ANSI Escape Codes with AI Agents! Leon Derczynski discovered that LLMs can emit ANSI escape codes, e.g. the ESC (ASCII 27). When I saw that, it seemed like a great way to hijack AI agents and LLM CLI tools running on the command line and to exfiltrate data. Proof-of-Concept Exploit Here is a demo that

匹配字段正文

variation of any category we evaluated. Some systems recognized emotion exceptionally well but struggled to respond naturally. We found that access to audio did not guarantee that agents used the paralinguistic information it contained. Some systems remained largely transcript-driven, relying on the words being spoken while overlooking cues such as tone, pacing, hesitation, emphasis, and volume. Humans naturally use these cues to infer confidence, uncertainty, frustration, sarcasm, and empathy.

匹配字段正文

both models on FrontierCode 1.1 using our new Fusion architecture, Fable cost less. Unsurprisingly, it also scored higher. This post explains why, and what that means for pricing agentic work. Introduction Everyone who runs coding agents knows that stronger models give you better results, but you eat the cost. When we introduced Devin Fusion, we showed a way out: keep a frontier model in charge, let it delegate to a cheaper and faster sidekick, and you get frontier-level performance at 60% lowe

匹配字段正文

environments. To address this, we retain the inverse pendulum and hopper tasks as control tasks but also introduce a new task based on pinball arcade machines. In TwinFlipper, the agent controls a set of flippers and seeks to maximize the ball’s total airtime—the cumulative amount of time the ball remains above a specified height threshold while not touching anything—before it drops below the flippers. Although a naive solution is to just slam the ball upwards, much more airtime can be gained by

匹配字段正文

Harness Design Patterns Pattern 1: Workflow Automation; Pattern 2: File System as Persistent Memory; Pattern 3: Sub-agent and Backend Jobs; Case study: Coding Agent Harness; Harness Layer vs Core Intelligence?; Harness Optimization The concept of recursive self-improvement (RSI) dates back to I. J. Good (1965), where he defined an “ultraintelligent machine” as a system that can surpass humans in all intellectual activities and design better machines to improve itself. Yudkowsky (2008) used the p

匹配字段正文