PAPER PLAINE

Current AI systems can answer cybersecurity questions but struggle to translate those answers into working commands for real tools—even tiny syntax errors cause failures. Researchers created KaliBench, a test set of 8,500 real commands across 1,600 cybersecurity tools, and found that no open-source AI model gets more than 42% of commands exactly right without help. Training models on these exact-command tests improved performance dramatically, bringing small models up to the level of much larger ones.

AI has made it cheap to produce customized weather forecasts that could help hundreds of millions of small farmers in poor countries. But without clear standards to measure forecast quality, low-cost but unreliable predictions could dominate the market, crowding out genuinely useful ones. The authors propose evaluation principles and protocols that would let farmers and others judge which forecasts are actually worth using.

Two widely used sampling methods—Wasserstein gradient flows and forward-only diffusion processes—come with theoretical guarantees of fast convergence, but researchers show these promises don't hold for distributions with multiple distinct peaks. The problem is fundamental: moving probability mass between well-separated modes requires exponentially long times, no matter how you adjust the algorithm's schedule.

A new tool called MatLoom uses AI language models to generate material textures from text descriptions—but instead of producing just images, it outputs compact, readable code that explains how each texture layer is built. In human preference tests, materials created this way won 59% of votes against leading competitor methods, and the code stays short enough (around 21 lines) that artists can edit it by hand afterward.

A step in preparing medical images for quality control—identifying which parts are actual tissue versus background—quietly determines which real samples get flagged as defects. Researchers found that switching to a better tissue-detection method dropped false alarms from 10% to 1.6%, simply by changing what the system learned to recognize as normal.

Two neural networks can solve the same task using completely different internal representations—or use similar representations while computing things in fundamentally different ways. The researchers found that representational similarity (shared internal codes) and functional similarity (solving the same problem) are independent properties, and that networks built to withstand weight changes are actually forced to use task-specific, less flexible representations.

Researchers created the first major benchmark for editing text that appears on real-world surfaces in videos—like storefront signs or whiteboards—while keeping everything else unchanged. Testing eight existing video editors revealed a fundamental trade-off: systems can either get the text right, maintain smooth motion, or preserve the background, but not all three together.

Researchers developed a new mathematical method that treats wind speed as something to be optimized rather than simply calculated, allowing them to embed the complex physics of how turbines interfere with each other directly into a computer optimization program. This approach uses linear equations instead of the usual complicated nonlinear ones, making it possible to find genuinely optimal turbine layouts much faster than existing methods.

When LLM-powered AI agents make financial decisions, they often trigger collective failures even when no single agent is trying to break the system—bank runs failed 77% of the time and debt crises 83% of the time in controlled tests. Three structural fixes (advance commitments, centralized agreements, and agent-led coalitions) all reduced failures, but only if they locked in cooperation early, before panic spread.

When language models reconstruct graphs, researchers usually report one summary number measuring how wrong the reconstruction is. This paper proves that number conceals crucial details: two mathematical bounds reveal whether the model only added edges, only deleted them, or did both—and when it does both, that signals the model hallucinated connections while also losing real ones. Testing three AI models on 45 graphs, the researchers found that 29 reconstructions invented and deleted edges simultaneously, even in cases where the total edge count stayed the same.

Researchers created Imagine3D-LLM, a visual AI system that builds a rough 3D mental model of a scene from multiple camera angles before answering questions about it — similar to how humans visualize spaces. This approach outperformed existing systems that try to track detailed geometric correspondences between views, suggesting that coarse 3D reasoning works better than pixel-level precision for spatial understanding tasks.

Time series foundation models—AI systems that predict future values in data streams—produce forecasts that vary in quality depending on how you set them up, but no single configuration works best for all situations. Researchers developed a method that automatically combines multiple forecasts from the same model by learning in latent space which configurations work well for different parts of the time series, achieving results competitive with traditional ensemble methods while keeping the models off-the-shelf.