Berea College

Many existing benchmarks, such as MMLU, are limited to measuring large language models’ (LLM) true task understanding due to their reliance on statistical patterns in the training data. We suggest new approaches to improve how benchmarks can capture task-specific understanding in LLMs, revealing insights into their reasoning ability.

Dog trainers make widespread claims that training dogs to improve self-control is essential to solving behavior problems in domesticated dogs. We investigated whether one aspect of self-control, motor inhibition, covaries in dogs with varying levels and contexts of formal training. We tested pet dogs with little formal training and dogs with high levels of training who compete in scent work, barn hunt, and agility. All dogs were tested for motor inhibition on the cylinder task, a validated self-control task. First, in familiarization trials, dogs watched a piece of food placed inside an opaque cylinder and were subsequently allowed to retrieve it. Then, in test trials, dogs watched a piece of food placed inside a transparent cylinder. This transparent variation of the task requires dogs to inhibit the response to directly approach the now-visible food and instead detour to an open side. In contrast to the belief that highly trained dogs would uniformly demonstrate better self-control, trained dogs did not perform better than untrained dogs. In fact, dogs who competed in barn hunt performed significantly worse compared to untrained pet dogs. These findings suggest that self-control may involve a complex set of mental processes that do not uniformly generalize to multiple contexts.

Artificial light at night (ALAN) changes animal behavior in multiple invertebrates and vertebrates and can result in decreased fitness. However, ALAN effects have not been studied in European honey bees (Apis mellifera), an important pollinator in which foragers show strong circadian rhythmicity. Colonies can be exposed to ALAN in swarm clusters, when bees cluster outside the nest on hot days and evenings, and, in limited cases, when they build nests in the open. We captured and maintained foragers in incubated cages and subjected them to constant light (LL), constant dark (DD), or 12 h light:12 h dark (LD) cycle, and observed them with infrared cameras. After 79 h, there was a significant interaction of treatment and time because LL bees slept less. In detail, the bees maintained a regular sleep pattern for three days but LL bees showed a shift on the fourth day. LL bees had the largest sleep differences from LD controls, with trends of lengthened periods and increased phase misalignment from both LD and DD bees. LL bees also experienced significantly more disturbances from their nestmates and preferred to sleep in the lower portion of the cages, which had significantly lower light intensity. These findings suggest that ALAN can disrupt the sleep of honey bee foragers, which has implications for their behavior and overall colony health.

Climate-smart forestry is an increasingly important topic in forest policy and for practices. However, what does and does not constitute climate-smart forestry is subject of debate. At stake are billions of dollars of investment aimed at encouraging climate-smart forestry practices in the United States. As a leading voice for ecologically, economically, and socially responsible forestry, The Forest Stewards Guild (FSG) has produced a position statement based on the organization's vision, mission, and principles to guide conversations around climate-smart forestry for all interested stakeholders. This forest perspective presents and expands on the findings of the FSG position on climate-smart forestry. There are three common aspects in the multiple co-existing definitions of climate-smart forestry: 1) adapting forests to expected future climate conditions, 2) mitigating climate change by leveraging carbon sequestration and storage functions of forests, and 3) improving social outcomes. There are potential trade-offs with other benefits forests provide if climate-smart forestry is pursued without holistic consideration of forest ecosystems. We suggest that such trade-offs can be minimized if the goals of climate-smart forestry projects are communicated transparently, system boundaries are made as comprehensive as possible, potential trade-offs are assessed along with climate benefits, climate-smart practices are tailored to the social-ecological contexts, and uncertainty is recognized.

The role of parasites as prey for nonhost species has received increasing attention over the last 3 decades as (1) such a reduction of infectious agents decreases disease transmission by reducing parasite numbers and (2) the importance of parasites in energy flow within aquatic food webs is acknowledged. Previous predation studies of digenetic trematode cercariae have assessed species in which large numbers of cercariae are released daily by individual snails. By contrast, the cercaria of the azygiid trematode Proterometra macrostoma averages less than 1 cercaria released per infected snail per day. The objective of this study was to evaluate the preference of crayfish (i.e., Procambarus spp. and Faxonius juvenilis) as predators of 3 different prey items of similar size and swimming ability including the comparatively large cercaria (i.e., 3–9 mm) of P. macrostoma. Four experiments were conducted, 2 with F. juvenilis and 2 with Procambarus spp. Each experiment consisted of 25 crayfish, which were individually exposed to 1 of 5 treatment levels consisting of a control (i.e., no prey), individual prey (i.e., cercariae, Daphnia magna, or Xenopus laevis tadpoles only), or a mixture of all 3 prey over a 7-hr exposure at 20°C and continuous light. In both the single-prey and mixed-prey experiments, cercariae were consumed at a similar or faster rate than tadpoles and D. magna, and these differences were most notable within the first hour postexposure. These results suggest that P. macrostoma cercariae are a legitimate component of the food web in freshwater habitats, and may play an important role in energy flow between trophic levels in such systems.

This chapter explores ChatGPT's potential for assisting in a culture-focused flipped curriculum to facilitate the development of intercultural communicative competence, a core value of foreign language education in the 21st century. Three experiments assess ChatGPT's performance in generating scenarios and performance scripts for practicing intercultural communication in Chinese-speaking contexts. While ChatGPT demonstrates remarkable linguistic accuracy and comprehension abilities, it struggles to generate scripts that reflect communicative strategies specific to Chinese-speaking contexts, especially when the prompt lacks explicit instructions about Chinese cultural expectations. The limitation can be rooted in ChatGPT's training and the user's ineffective prompting. The findings suggest that ChatGPT is better suited as a reference tool than a primary learning resource in the curriculum. The implications for foreign language education and the integration of AI are also discussed.

A new Cotyledon species, C. mckayi is described from cliffs of the central Tugela River basin in KwaZulu-Natal province, South Africa. Although the flowers of this new species are superficially similar in terms of shape and size to those of C. nielsii, also from KwaZulu-Natal, C. mckayi is readily separable on both vegetative and reproductive characters, from this and all other taxa of Cotyledon that occur naturally in the province. A key to allow identification of fertile material of the Cotyledon species of KwaZulu-Natal is provided.