Researchers presented Benchproofer, a pipeline for formal verification of LLM-generated code. It turns a coding task with a known correct patch into a formally verified one, writing a specification for the new code and summarizing existing functions. This extends to 500 real issues on SWE-bench Verified, catching what tests miss and improving resolution rates from 85% to 95% for Opus 4.8. However, writing faithful specifications is a challenge.

Researchers introduced a method to detect hallucinations in LLMs by analyzing attention graph topologies and information flow patterns. Their approach provides consistent improvements over existing methods and identifies impaired context sharing as a key characteristic of hallucinations. This affects developers building and operating LLMs, as it may inform improvements to LLM architectures and training procedures.

Researchers studied the vulnerability of multi-agent trading systems built on large language models (LLMs) to black-box, input-only attacks. They found that even simple attacks can degrade risk-return profiles and reduce Sharpe ratios. However, they also discovered that suitably designed multi-agent topologies and coordinator prompts can improve average robustness under identical poisoning budgets.

A new algorithm, SemanticAdv, generates unrestricted adversarial examples by manipulating the semantic attributes of images. This affects deep neural networks, which are used in various AI applications. The algorithm can be used to fool AI models in different learning tasks, including face verification and landmark detection.