搜索“RAG”

head of GPT-6.1 Sol at 65 and Gemini 4 Argon at 55. Claude Opus 5.5 (49) and Claude Fable 5.1 (46) trail ➤ Today’s models are incredibly slow at our tasks, taking 30 minutes on average to complete each tabletop task ➤ Cost varies widely: GPT-6.1 Sol is the cheapest at $1.93 per task on average; GPT-6 Astra, the most expensive, costs $16.36, more than eight times as much ➤ Compared with a human, today's best-performing frontier language models are 100x slower and 10x more expensive More results,

匹配字段正文

and gave them no human hints. GPT-6-Astra, running in OpenAI's Codex coding environment, came out on top. It found the right answer for 49 of 50 problems, or 98 percent, at an average cost of under $3 per problem and just under $116 for the whole benchmark. The benchmark tests routine research work, not grand challenges The collection deliberately leaves out major open questions and focuses on what the authors call "turn-the-crank" problems, where researchers apply known methods to a slightly m

匹配字段正文

pausing after each reply for a tool-call time drawn from real tool-call durations. Requests carry up to 132K input tokens while replies are often just a short tool call, and on average 97% of a request’s input tokens repeat the previous request. If a serving stack keeps this repeated prefix in its KV cache, it only has to prefill the tokens each turn adds, rather than reprocessing the full context every time. In the example below, one of the longer trajectories grows from 10K to 116K tokens of c

匹配字段正文

lean, because a leaky one climbs just as convincingly. What matters is performance on held-out tasks, reported below. Diversity over task count Task count is easy to scale and coverage is not. Renaming a customer or swapping a ticket number produces a new task that exercises the same skill, and a thousand such variants can leave whole capabilities untouched. The variation that matters changes the structure of the work (Table 1). StructureExampleWhat the agent has to learn Look-alike recordsA tic

匹配字段正文

ations and climatology to capture recent weather trends and day-of-year effects. The two corrections are projected onto the space of valid cumulative distribution functions and averaged into the final PBC prediction. The team evaluated forecasts using the conventions of ECMWF’s AI Weather Quest competition: average 2-meter temperature and total precipitation over land plus average mean sea-level pressure on a global 1.5-degree grid with five quintile bins, at week-3 (days 19–25) and week-4 (days

匹配字段正文

hkeiten erhalten Sie in unserer Datenschutzerklärung. Sicherheit und Geräteverwaltung gehören zusammen; Wie UX-Design Hardware und Software verbindet; So werden Data-Compliance-Anfragen überschaubar; Digitale Transformation ohne Abhängigkeiten; Warum Firmen die KI-Infrastruktur anpassen sollten; Der stille Stress im digitalen Alltag; Die Arbeitsplatzarchitektur der Zukunft; Sensible Daten schützen; Cyber-Hygiene statt Cyber-Panik; Wissen statt suchen: KI bändigt Informationsflut Immer informiert

匹配字段正文

HiROv2 realistically generates strong winds flowing down mountain slopes. We also trained an updated version of ACE for this system. Over a 10-year simulation, its hurricanes, average climate patterns, and overall distributions of rain, wind, and pressure all line up well with X-SHiELD. That tells us ACE's coarse simulations are a solid foundation for producing long-term high-resolution climate data. Putting HiROv2 and ACE together, we generated a full year of global 3 km data. Its statistics c

匹配字段正文

alls only when a user question needs them. Before running complex operations, the agent builds a step-by-step analytical plan, selects appropriate fallback datasets when local coverage is sparse, and processes large map layers in a dedicated workspace. This approach avoids overwhelming the agent context with massive amounts of spatial data, while preserving accuracy and tracking data attribution. Every response shows the agent’s plan and the data sources used so experts can check assumptions or

匹配字段正文

cientists free access to Gemini 4 Argon in Antigravity at its highest usage tier. DeepMind describes Antigravity as an agentic science workbench that compresses what used to be a fragmented search across multiple databases into a single session, and said researchers can take the demonstrated conversation in new directions, such as layering in new filtering criteria, exploring different sequence lengths, or targeting other priority pathogens, broadening the search from human commensals to environ

匹配字段正文

进度对齐、混合不同来源的数据块。 这样既让训练过程中数据集逐步走向更高质量的区间，又保持局部训练区间的数据集配比近似稳定，避免质量变化同时造成领域数据比例的突变。 但数据排好了，并不等于模型能充分学到。Puro团队此前的研究发现：当训练后期学习率过低时，最后出现的优质数据难以推动参数更新。为此，他们采用了课程模型平均（Curriculum Model Averaging，CMA）在训练后期保持恒定学习率，使模型继续学习优质数据，再对末端的检查点做权重平均，来缓解较高学习率带来的训练噪声。 论文Figure 7｜课程数据构造与模型平均 进一步的消融也显示，直接在学习率衰减的训练轨迹上做模型平均，并不会稳定获益。真正需要配合的是数据顺序、后期学习率与模型平均策略。 Figure 13b｜对数据顺序、后期学习率与模型平均策略的消融实验 按Puro Cost Scaling Law这一经验曲线估算换算，相比于随机数据排布（Uniform, UD），数据课程加学习率衰减会相当于把成本效率提高到约1.65倍，完整CMA方案则达到约2.40倍。 数据评测 数据顺序解决后，还要决定选哪些数据。 他们建

匹配字段正文

will still be trusted many decades from now. Magnetar makes major progress on all three challenges. It proves programs written in Rust, Go, or C, runs 10 to 39 times faster on average than the prior fastest comparable systems, and rests its security only on standard hash functions. It is the latest result of more than a decade of research at Microsoft Research on making zero-knowledge proofs practical. A new form of computing Every result a computer produces comes with an implicit question: did

匹配字段正文

aces first in AA-Robotics v0.1 with a score of 69, followed by GPT-6.1 Sol, Gemini 4 Argon, Claude Opus 5.5 and Claude Fable 5.1. GPT-6 Astra took 30 minutes and cost $16.36 on average per task. AA-Robotics Score Each attempt is scored out of 100: completing the task scores 100, and a partial result earns points for the progress made. The eight tasks AA-Robotics v0.1 initially includes eight tabletop tasks, modeled on everyday activities, to evaluate a range of robotics capabilities: control and

匹配字段正文

fsanfänger ist laut IAB-Analyse klar die trübe Konjunktur. „Vor allem die Wirtschaftsschwäche der letzten Jahre hat zu einer sinkenden Beschäftigung von jungen Beschäftigten beigetragen, während der Einfluss von KI unterm Strich eine eher untergeordnete Rolle gespielt haben dürfte“, erklärte IAB-Forscher Oliver Schlenker. Generell seien die Stellen junger Beschäftigter besonders konjunkturanfällig. Videos by heise c't 3003; heise & ct; Peertube Ein mögliches Warnsignal haben die Forscher im vera

匹配字段正文

rnational forecasting competition. In a new paper in Nature, the researchers describe their approach and show how it enables forecasts that for the first time beat climatologic averages between two and six weeks. “Not only does this do better on time averaged forecasts for this range, which is already very useful for farmers and members of the energy industry trying to match supply with demand, but it also improves predictions of extreme weather events, whether it’s flooding or cold spells or so

匹配字段正文

behavior, then write a new implementation that behaves the same way. We score each task by the share of a hidden suite of behavioral tests that the rebuilt program passes, and average this score over tasks to get the reward we report throughout.2 The tasks range from small command-line utilities like jq and ripgrep to large systems like SQLite, FFmpeg, and the PHP interpreter. The most challenging remain far out of reach. On FFmpeg, the best entry on the public leaderboard passes about 10% of t

匹配字段正文

ersal variants I reported. Here a quick demo to show what I mean: As you can see in this mini demo, the call removed the dot. Conclusion Hope this was an insightful post that encourages folks to look more into Windows filesystem and canonicalization issues. Especially with the advent of AI, and the speed in which such bugs and variants can be found, we will likely see many cases where things go wrong because of it. Security controls around agents must hold even when the agent itself is compromis

匹配字段正文

tcuts. Instead, the share of shortcuts the model took silently — without mentioning them in its thoughts — more than tripled (from 11.2% to 36.4%). Avoiding thought penalties encourages the model’s reasoning to stay transparent and monitorable. Prompt injection resistance To trust an agent to do the job right, we need to ensure that malicious content the agent comes across cannot trick it to take a harmful action. When agents browse webpages or read code, they can encounter hidden, malicious ins

匹配字段正文

samplers’: the two runs’ pass@4 stays close throughout training. Their training reward and gradient norm also follow similar paths, despite a 4.7× larger sampler–trainer KL on average. Entropy matches until about step 160; afterwards, it rises on the homogeneous samplers and stays flat on the mixed ones. Importantly, nothing in how the correction works is specific to this model or these accelerators, so we expect this pattern to hold more broadly. Decoupling the samplers from the trainer also m

匹配字段正文

high reasoning), surpassing Claude-Sonnet-4. This phenomenon is consistent with GPT-5’s results for other benchmarks. Furthermore, we run GPT-4.1 on this subset six times. The “Average” bar represents the mean performance, achieving an average success rate of 8.7 ± 0.7%. The “Union” bar shows the overall success rate when at least one of the six runs succeeds, reaching 18.0%. These results demonstrate the significant improvement achieved through test-time scaling and highlight the strong potenti

匹配字段正文

ool Ecosystems LLM agents increasingly delegate execution authority to external tools. That delegation is powerful — but it assumes trust. Our findings suggest that this trust is fragile. Malicious tool attacks are scalable, automatable, and difficult to detect with current alignment and scanning techniques. As tool ecosystems continue to expand and agents become more autonomous, the consequences of malicious tools will grow accordingly. This raises urgent design questions. How should tool platf

匹配字段正文