Posted "Rethinking Failure Studies in the Age of AI" Personally, whether it's AI4Science or AI solving open problems in mathematics, it doesn't matter to me whether a breakthrough comes from humans or AI—the value of truth doesn't change. At the same time, though, I believe the tacit knowledge and hands-on experience gained through the actual process of research and execution belong to an entirely different dimension—one that humanity still needs to absorb and learn from.

Introducing Hyra-1.0, the first version of Hunyuan Research Agent. 💡💡💡 Built to recursively improve solutions for performance-driven research and engineering tasks. Explore our demos in AI4AI, AI4Science, and AI4Fun:

🔬 Even the top AI agents leave about a third of scientific code repair tasks unsolved within a one-hour budget. Title: ScienceIDE: Turning World's Scientific Codebase into Agent Learnable Environments URL: 🧩 Overview AItonomy, Qwen, and PhAI-Labs built ScienceIDE, infrastructure that turns 27 scientific codebases, spanning astrophysics to ocean modeling, into executable environments AI agents can actually learn from. It ships 64 environments and 2,812 tasks. ⚙️ Problem it solves Scientific code encodes decades of expertise, but fragmented toolchains and undocumented numerical conventions make it hard to turn that knowledge into reliable learning experience. 🛠 Methodology "Scientific checks" comparing numerical tolerances and conserved quantities validate task correctness, across 7 task categories like repair and implementation. For RL, a "truncation masking" trick prevents unfairly penalizing trajectories that hit the budget cutoff. 📊 Results Under a one-hour budget, Fable 5.1 leads the agent leaderboard at 67.1%, with Astra at 63.1%. Models trained on ScienceIDE experience (PhAI-IDE) gain up to +33.33 points on scientific code repair, and RL training lifts reward on LAPS tasks 2.4x. #AIAgents# #AI4Science#

🔬 Can AI actually "discover" science? Tested on 90 real Nature papers, even the strongest agent beat SOTA on under 20% of tasks. Title: NatureBench: Can Coding Agents Match the Published SOTA of Nature-Family Papers? URL: 📋 Overview Using 90 tasks drawn from Nature-family journals across six domains, this work tests whether AI coding agents can reproduce and surpass published SOTA. An "information firewall" hides the original method, and NatureGym builds per-task containerized environments for strict, unified evaluation with web search disabled. 🎯 The problem it solves Prior agent-on-research benchmarks were fragmented and low in credibility. This work forces genuine discovery instead of reproduction and creates a common ground for cross-domain comparison. 📐 Methodology Tasks are scored with a SOTA-normalized relative gap (g > 0.1 to surpass, g >= 0 to match), comparing across 81 metrics. Ten agents are evaluated across three harnesses under a 4-hour budget with task-based GPU allocation. 📊 Results ・Even the top agent, Claude Opus 4.7, surpassed SOTA on only 17.8% of tasks and matched on 47.8% ・45.5% of successes were just translating a science task into supervised prediction; genuine domain reasoning was only 8.3% ・Failures were driven by wrong method choice (45.1%) and insufficient compute (24.4%), with task misunderstanding at only 3.1% ・Interdisciplinary tasks consistently scored worse A sober snapshot: today's agents are good at "translation" but weak at real "discovery." #AI4Science# #CodingAgents#