PAPER PLAINE

When multiple AI agents work on the same problem in parallel, combining their edits often destroys the improvements each one made individually. Researchers developed a method that selects which edits to merge based on graph theory, guaranteeing that at least a provable fraction of all verified improvements survive the combination. Experiments on code repair and route optimization show the approach keeps 15% more of the value that independent agents produce, compared to naive merging.

A new system inspects precision components by combining two types of AI models—one that spots defects in specific regions and another that flags overall anomalies—then filters their results through manufacturing rules to make final accept/reject decisions. On fuel injector needles, this approach caught all defective parts while cutting false rejections from 57.6% down to 10.9%.

Researchers tested whether AI models trained to predict stock market order flows also learn realistic market impact—the price movement caused by large trades. They found that impact does emerge naturally in these models, but with slightly stronger effects than observed in real markets, and the models failed to capture how markets recover after large trades, likely because they don't remember enough context.

Most coding agents are tested with single, polished requests—but real users change their minds and add requirements as they go. Researchers built SWE-DIVE, a testing system that simulates realistic multi-turn conversations between a user and a coding agent, complete with a "user simulator" that behaves like actual people do and a validator that checks whether the agent actually solved what was asked. The framework reveals how well agents handle the messiness of actual development work, not just textbook problems.

Researchers showed theoretically that a vibrating membrane can control how free electrons scatter light and absorb momentum—and in the process, detect tiny external forces better than conventional optical methods. The membrane's motion creates special light patterns that electrons can interact with, encoding force information directly into how the electrons recoil.

When generative models like GPT train on synthetic data they've created themselves, they gradually lose diversity and performance in a process called model collapse. Researchers found that this collapse appears as a mathematical narrowing in the model's internal structure—subdominant eigenvalues approaching one—which they can see directly by analyzing the model's spectral properties. A simple fix called Dreaming Learning, which trains the model alongside its own high-temperature samples, keeps this mathematical collapse from happening by preventing the model's internal values from becoming vanishingly small.

A new AI system called UniWALM teaches robots to manipulate objects by learning video, actions, and language all at once in a single model, rather than bolting separate reasoning systems together. The robot generates text descriptions of what it's doing and remembers them for the next step—no external memory bank needed. On a real-world task of finding and fetching items, adding this self-generated memory boosted success from 11% to 61%, far outpacing existing robot control systems.

Researchers proved that stochastic gradient descent—the algorithm used to train virtually all modern neural networks—reliably converges to good solutions, even in simplified deep learning models. The proof works without requiring unrealistic assumptions like perfectly clean data, and holds for both fixed and adaptive learning rates.

Cells behave differently depending on their neighbors and position in tissue, but existing methods for analyzing this spatial context miss important details or rely on arbitrary coordinates. A new framework called Tessera captures both which genes are active and how cells are organized relative to each other—encoding direction and local relationships rather than just distance—and outperforms current methods even when applied to tissues it has never seen before.

Researchers combined two techniques for speeding up large language models—extreme compression and predicting multiple words at once—which hadn't worked well together before. Their method, BitMTP, achieves 2–3x faster text generation across different GPU types while maintaining the same output quality, solving practical engineering problems that prevented these speedups from working in practice.

Researchers discovered that a single malicious prompt injection can trick AI assistants into repeating harmful instructions across multiple sessions—sometimes dozens of times—long after an attacker has gone. Testing five major AI models, they found that once malicious code gets planted in an agent's persistent memory or scheduled tasks, it re-fires reliably across follow-up sessions with almost no degradation, and defenses at the prompt level alone don't stop it.

Researchers trained artificial intelligence agents to act as market makers on prediction markets like Polymarket, where they quote buy and sell prices to earn the spread between them. The key innovation was splitting the problem into two cooperating agents—one deciding what prices to quote and another managing the risk of holding too much inventory—using a technique called cooperative multi-agent reinforcement learning. The redesigned system quoted prices more often and lost less money than baseline approaches, though none yet achieved consistent profitability on held-out test markets.