As language models (LMs) are increasingly deployed as autonomous agents, their robust adherence to human-assigned objectives becomes crucial for safe operation. When these agents operate independently for extended periods without human oversight, even initially well-specified goals may gradually shift. Detecting and measuring goal drift - an agent's tendency to deviate from its original objective over time - presents significant challenges, as goals can shift gradually, causing only subtle behavioral changes. This paper proposes a novel approach to analyzing goal drift in LM agents. In our experiments, agents are first explicitly given a goal through their system prompt, then exposed to competing objectives through environmental pressures. We demonstrate that while the best-performing agent (a scaffolded version of Claude 3.5 Sonnet) maintains nearly perfect goal adherence for more than 100,000 tokens in our most difficult evaluation setting, all evaluated models exhibit some degree of goal drift. We also find that goal drift correlates with models' increasing susceptibility to pattern-matching behaviors as the context length grows.

I reiterated Bill Joy's 2000 question: Does the future need us? Little did I know then that a revolution was already brewing. By 2011, GPUs had accelerated considerably the training of deep neural networks, finally making a technology whose roots go back to the early 1940sb competitive. By 2011–2012, AlexNet, a deep neural network, won several international competitions, launching the deep-learning revolution. A decade later, generative AI, which refers to AI that can generate novel content rather than simply analyze or act on existing data, has become all the rage.

In multi-cycle assignment problems with rotational diversity, a set of tasks has to be repeatedly assigned to a set of agents. Over multiple cycles, the goal is to achieve a high diversity of assignments from tasks to agents. At the same time, the assignments' profit has to be maximized in each cycle. Due to changing availability of tasks and agents, planning ahead is infeasible and each cycle is an independent assignment problem but influenced by previous choices. We approach the multi-cycle assignment problem as a two-part problem: Profit maximization and rotation are combined into one objective value, and then solved as a General Assignment Problem. Rotational diversity is maintained with a single execution of the costly assignment model. Our simple, yet effective method is applicable to different domains and applications. Experiments show the applicability on a multi-cycle variant of the multiple knapsack problem and a real-world case study on the test case selection and assignment problem, an example from the software engineering domain, where test cases have to be distributed over compatible test machines.

In this paper we provide an algorithm for utility maximization of a movie supplier service, in two different settings, one with prices and the other without. This algorithm is provided along with an extensive experiment demonstrating its performance. We also uncover a phenomenon where movie consumers prefer watching and even paying for movies that they have already seen in the past than movies that are new to them.

Traditional recommender systems try to provide users with recommendations which maximize the probability that the user will accept them. Recent studies have shown that recommender systems have a positive effect on the provider’s revenue. In this paper we show that by giving a different set of recommendations, the recommendation system can further increase the business’ utility (e.g. revenue), without any significant drop in user satisfaction. Indeed, the recommendation system designer should have in mind both the user, whose taste we need to reveal, and the business, which wants to promote specific content. In order to study these questions, we performed a large body of experiments on Amazon Mechanical Turk. In each of the experiments, we compare a commercial state-of-the-art recommendation engine with a modified recommendation list, which takes into account the utility (or revenue) which the business obtains from each suggestion that is accepted by the user. We show that the modified recommendation list is more desirable for the business, as the end result gives the business a higher utility (or revenue). To study possible longterm effects of giving the user worse suggestions, we asked the users how they perceive the list of recommendation that they received. Our findings are that any difference in user satisfaction between the list is negligible, and not statistically significant. We also uncover a phenomenon where movie consumers prefer watching and even paying for movies that they have already seen in the past than movies that are new to them.