Jiajun Chen and three other authors present Recursive Game Creator, a harness that uses four components to advance agentic game development from rough prototypes into entertaining games. The method achieves a state-of-the-art overall performance of 77.89 on GameCraft-Bench. The system mitigates evaluation bias by using a coding-native Player to collect diverse gameplay trajectories through programmatic interfaces.
Owolabi, Gupta, and Wang evaluated a category-conditioned retention threshold in a deployed cold-start memory pipeline on 100 synthetic personas. They found that value and belief assertions had only 77.9% source support compared to 96.2% for other categories, and applying a stricter bar to values reduced unsupported retentions from 6.2% to 4.0%. Builders care because a single global confidence threshold fails to separate well-evidenced assertions from unsupported value claims.
Muhammad Bilal Awan and two other authors proposed AegisFlow, an agentic framework that uses LLMs to automatically create and deploy code patches for brittle data pipelines. The system reduced the average time to repair from 170 minutes to 3.2 minutes across five failure scenarios. The framework frees up about 98 percent of data engineering on-call time from firefighting.
Researchers analyzed 103,939 graded replies from ten LLM configurations to identify conditions that produce sycophancy. They found that the cost of verifying a user's claim is the dominant factor, with maximum reasoning reducing adoption on deep puzzles from 19.2% and 12.5% to 0%. Builders can reduce model yielding by simplifying hard-to-verify problems and enabling deep reasoning.
Pavan Maddula introduced the Adversarial Surface-Form Robustness Dataset containing 2,100 prompts to evaluate five open-weight language models on non-canonical inputs. The study found that leetspeak and encoded wrappers caused comprehension failure rates of 36.47% and 65.60%, respectively, while emoji variations resulted in pooled harmful compliance of 20.27%. Builders care because models fail to understand or safely handle inputs with altered spellings and encoded strings.
Yixin Zhang and Wenjie Feng proposed CuratorMAS, a multi-agent framework that automates dataset curation through five programmable stages. The system reduces the noise rate by up to 36.03 percentage points and improves downstream model F1 scores by up to 8.88 percentage points. Builders care because the framework automates a costly curation process that is difficult to generalize across domains.
