#评测/基准

Today we are launching FrontierCS 2, the most frontier benchmark for AI agents in computer science R&D. We built 48 tasks from published papers across 13 areas of CS, refining each one with experts in its field. Excited to co-launch FrontierCS 2 with @kaggle @GoogleDeepMind

We found DeepSeek V4.1 Flash 1017 beats GPT 6 Sol on our multi-step offensive cybersecurity task. Sol was tested via the Daybreak trusted access program. We have been building a cybersecurity benchmark based on simulated corporate networks. This is a valuable indicator for the cyber-risk a model poses if in the hands of a bad actor. We plan to make the full evaluation public in the coming weeks but wanted to share our first immediate finding here.

⚠️ Frontier AI models still fall significantly short of human designers on design creativity. ⚠️ Introducing Design Creativity Bench, a new benchmark evaluating AI models’ creativity on UI design tasks. Despite rapid advances in their ability to generate functional websites and product UIs, leading AI models still struggle to make creative design decisions. We evaluated 16 frontier models across 80 open-ended UI tasks, each adapted to two distinct product domains alongside human designer references. Existing UI benchmarks often measure a single output quality metric. Instead, we measure three dimensions of design creativity: - Originality: How distinct a model's designs are from other models' outputs for the same brief. - Creative Range: How much a model adapts its design decisions when the product domain changes. - Appropriateness: How well the generated UI satisfies the functional requirements of the brief. The results reveal a substantial gap in overall creativity: The best model, @OpenAI GPT-6 Astra, scores 65.5 against a human reference of 92.3. We're publishing the full leaderboard, methodology, and reproducible evaluation framework to make design creativity a measurable dimension of frontier model capability. 📊 Benchmark Leaderboard: https://kombai.com/benchmark/ 📄 Paper: https://arxiv.org/abs/2610.11539 🤗 HuggingFace Datasets: https://huggingface.co/datasets/Kombai/design-creativity-bench 🔗 Github: https://github.com/kombai-io/design-creativity-bench/

New Research: We are releasing ExplorationBench, a benchmark for measuring how AI systems explore. Scientific discovery begins where known problems end: a system has to frame hypotheses, design experiments, and learn from the results. Evaluating this is hard. Genuinely new answers cannot be checked quickly, and in familiar domains a model can simply recall what it has seen. Addressing this challenge, researchers from Tencent Hy, Fudan University, and Tsinghua University built verifiable Alien Worlds. Their rules are executable, so every answer is checked exactly, and they conflict with familiar knowledge, so recall alone cannot solve the tasks. 🔹 Two sandboxes: AlienCode (31 hidden rule changes, 70 tasks) and AlienLogic (24 patched inference rules, 70 theorems) 🔹 A flawed manual, four rounds of self-designed probes, and closed-book tests after every round 🔹 Every answer graded by an interpreter or a proof checker, with no LLM judge What we found across 10 frontier AI systems: 1️⃣ Getting feedback is more effective than thinking alone. No AlienCode run starts above 15.7%; after four rounds the best reaches 89.0%, while the same turns without feedback stay at 0.5–11.0%. 2️⃣ Designing the experiments matters. Replaying a system's own best probes gives it exactly the same evidence, yet in AlienCode 9 of 10 systems do worse than when they chose the probes themselves. 3️⃣ Knowing a rule is not using it. Even when every required rule is stated correctly, tasks are solved only 73.4% of the time. 4️⃣ One score hides a lot. The same system under the same budget ended anywhere from 5.7% to 79.0%, and rankings barely transfer between the two worlds. CL-bench asked whether models can learn from context. ExplorationBench asks whether they can discover the rules themselves. 📄 Paper: https://arxiv.org/abs/2609.30199 🌐 Website & leaderboard: https://explorationbench.com 📝 Blog: https://explorationbench.com/blog/ 💻 Code (coming soon): https://github.com/Tencent-Hunyuan/ExplorationBench

Today we are announcing Harvey LAB-AA v1.1 in collaboration with Harvey. This updates our scoring methodology for the Legal Agent Benchmark (LAB) to add a hallucination check and require correct responses to not include material misstatements. LAB-AA v1.1's new headline metric, Hallucination-Gated All-Pass Rate, only credits a task when the deliverables satisfy every rubric criterion and contain no material hallucinations. We define a material hallucination as one that would mislead a reader on a substantive point, such as a wrong contractually required date, while a minor hallucination is a real error that is unlikely to meaningfully affect the legal interpretation of a deliverable. Minor hallucinations are reported separately and do not affect the headline score. Grok 4.7 (xhigh) leads at 9.4% Hallucination-Gated All-Pass Rate, and >60% of otherwise passing results across the models tested at launch contain a material hallucination. This is the first step in enhancing the methodology for Harvey LAB-AA. In future updates, we’re working with Harvey to better account for the full set of factors lawyers value, including usability features like style and tone. Key takeaways: ➤ Top of the leaderboard: Grok 4.7 (xhigh) from @SpaceXAI leads at 9.4% Hallucination-Gated All-Pass Rate, narrowly ahead of Muse Spark 1.3 (max) from @AIatMeta at 8.9% and GPT-6 Astra (max) from @OpenAI at 8.6% ➤ Hallucinations reshape the leaderboard: without the hallucination gate, Muse Spark 1.3 (max) would lead clearly with a 26.7% all-pass rate, but two thirds of those passes contain one or more material hallucinations, resulting in a Hallucination-Gated All-Pass Rate of 8.9%. GPT-6 Astra (max) retains almost all of its passes after the hallucination gate (8.9% to 8.6%) and moves from joint 10th to 3rd ➤ The GPT-6 model family is the most grounded: GPT-6 Astra (max) averages 0.03 material hallucinations per task (4 across all 120 tasks) and GPT-6 Sol (max) averages 0.07. When comparing six checker models on a 20-task subset, GPT-6 Astra had zero material hallucinations under every checker, including Claude Opus 5.5 (high). By comparison, Gemini 3.8 Flash (high) averages 13.96 material hallucinations per task. ➤ Top-scoring models aren’t the most expensive: Grok 4.7 (xhigh) takes the top spot on the leaderboard at a cost of ~$9.50 per task, under half the cost of Claude Fable 5.1 (max with fallback) at ~$21.70, the most expensive model. Muse Spark 1.3 (max) represents strong performance for its cost, landing in second for ~$4.20 per task Thank you to @nikogrupen and @ItsJulioPereyra from @harvey for their work on Harvey LAB and collaboration.