Researchers examined text from nine leading AI models and found they use constructions that people find hard to read more often than human writing does. These include inanimate things performing human actions, long strings of nouns, unusual word combinations and longer grammatical gaps. The patterns differed between model families. The authors suggest the models are tuned more for tasks like math and reasoning than for clear communication.

arXiv · AI research

The researchers built a large open-source AI model, IQuest-Q1, with 320 billion total parameters, of which 15 billion are active at a time. They trained it in three stages: first, stable and efficient pre-training; second, fine-grained mid-training that expanded the mix of data across domains and formats and gradually extended how much context the model could handle; and third, post-training across several different setups and versions that combined instruction tuning, domain-specific reinforcement learning, and other training signals to strengthen its ability to work as an agent. They also merged different experts and checkpoints from different stages while reducing conflicts among their updates. The model shows abilities in both coding and general tasks.

arXiv · AI research

Researchers built a way for AI agents to improve the setup that runs them, called a harness. The system uses AI-generated tests instead of official ones. Agents look at past runs, plan changes, and try them out. Instead of one setup, it offers three versions: cheap, balanced, and high performance. In tests, the best version did slightly better on tasks while cutting the average cost per task by about a quarter.

arXiv · AI research

Researchers tested a simple approach to giving chatbots long-term memory: search raw conversation history and pick the best passages, without the complex reorganizing and rewriting that other memory systems use. Across several conversation memory tests, this plain approach ranked first in six of seven settings, beating the strongest complex system by 0.13 to 5.53 points. The reason is that raw history covers more of the evidence needed to answer.

arXiv · AI research

The researchers studied how verification is organized in agent systems: when to check an agent's work, who checks it, and how checking leads to revision. They compared 17 designs across seven coding models. Initial instructions and reminders during work give small, model-dependent gains without much extra output. Completion gates give larger gains but cost more. Independent review by a separate critic improves accuracy by 5.9 to 9.5 percentage points while costing 1.13 to 1.25 times baseline.

arXiv · AI research

Researchers argue that memory for AI agents should be judged by whether it improves what the agent does next, not by what it can recall. They propose a way to organize memory by where it is stored, what form it takes, and how it is managed, with an inner loop updating memory and an outer loop refining how memory is handled. They also suggest a way to test retention, adaptation, and transfer, and a framework for how performance grows with experience, storage, and computing budget.

arXiv · AI research

The paper looks at how companies assign decisions between people and AI systems. Studying two company examples, it argues that giving a person formal authority over a decision does not keep their ability to judge well. Four forces wear that ability down: losing skill from less practice, misplaced trust, unclear accountability, and workers doubting themselves. The authors offer a four-part check for leaders.

SSRN · AI and managerial work

The researchers studied card transactions from 174 million US consumers to see how people buy paid generative AI services. As of April 2026, about 2.9 percent of consumers pay for at least one major service, mostly younger and higher-income people, and nearly all pick basic plans costing $30 or less a month. Switching between providers rises around big model releases, and yearly retention runs from 20 to 45 percent, which the authors say points to weak loyalty and little differentiation. A demand model shows buyers are very price sensitive and treat services as close substitutes.

NBER · AI and economics

The researchers compared earnings in occupations more and less exposed to generative AI, using US survey data from 2018 through June 2026. After ChatGPT's public release, the pay advantage from an extra year of schooling fell in exposed occupations, while the pay advantage from experience rose early in a career, held steady around fifteen years, and turned negative later. The results appear with the release and hold up under checks.

SSRN · AI and jobs

The paper describes a practical method for recreating software when the original source code is lost. Instead of maintaining existing code, it builds a working duplicate from the behavior of the original program. The authors show how teams can recover essential functionality without the original files, letting them run and update a system they no longer have instructions for.

AI and software development

Researchers looked at how software agents that fix bugs in large codebases are trained with reinforcement learning. Usually, when a fix fails testing, the training throws that failure away instead of using it. The new method keeps the failed fix and the tester's feedback, resets the code to its original state, and has the agent try again with that context. Grouping all attempts together lets successful recovery not erase credit for earlier failures. On 500 real tasks, the agent resolved more issues with the same model.

arXiv · AI and software development

An economics professor at Brown University made the final exam proctored after unusually high take-home midterm scores. Among 59 students, average scores fell from 95.7 to 48.8 out of 100. The released scores show the link between a student's midterm and final is almost nil, with changes ranging from a 4-point gain to a 100-point loss. The data cannot say whether unauthorized AI use, test anxiety, a harder final or other causes explain the drop.

arXiv · AI and education