Search for: All records

Large language models (LLMs) have achieved remarkable performance on various natural language tasks. However, they are trained on static corpora and their knowledge can become outdated quickly in the fast-changing world. This motivates the development of knowledge editing (KE) to update specific knowledge in LLMs without changing unrelated others or compromising their pre-trained capabilities. Previous efforts sought to update a small amount of parameters of a LLM and proved effective for making selective updates. Nonetheless, the edited LLM often exhibits degraded ability to reason about the new knowledge. In this work, we identify a key issue: \textit{heterogeneous token overfitting} (HTO), where the LLM overfits different tokens in the provided knowledge at varying rates. To tackle this, we propose {\NAME}, a token-level smoothing method that mitigates HTO by adaptively refining the target distribution. Theoretically, {\NAME} offers better parameter updates with negligible computation overhead. It also induces an implicit DPO but does not require preference data pairs. Extensive experiments across four editing methods, two LLMs, and diverse scenarios demonstrate the effectiveness and versatility of our method. 

International Ocean Discovery Program (IODP) Expedition 399 collected new cores from the Atlantis Massif (30°N; Mid-Atlantic Ridge), an oceanic core complex that hosts the Lost City hydrothermal field (LCHF). Studies of the Atlantis Massif and the LCHF have transformed our understanding of tectonic, magmatic, hydrothermal, and microbial processes at slow-spreading ridges. The Atlantis Massif was the site of four previous expeditions (Integrated Ocean Drilling Program Expeditions 304, 305, and 340T and IODP Expedition 357) and numerous dredging and submersible expeditions. The deepest IODP hole in young (<2 My) oceanic lithosphere, Hole U1309D, was drilled ~5 km north of the LCHF and reached 1415 meters below seafloor (mbsf) through a series of primitive gabbroic rocks. A series of 17 shallow (<16.4 mbsf) holes were also drilled at 9 sites across the south wall of the massif during Expedition 357, recovering heterogeneous rock types including hydrothermally altered peridotites, gabbroic, and basaltic rocks. The hydrologic regime differs between the two locations, with a low-permeability conductive regime in Hole U1309D and a high- (and possibly deep-reaching) permeability regime along the southern wall. Expedition 399 targeted Hole U1309D and the southern wall area to collect new data on ancient processes during deformation and alteration of detachment fault rocks. The recovered rocks and fluids are providing new insights into past and ongoing water-rock interactions, processes of mantle partial melting and gabbro emplacement, deformation over a range of temperatures, abiotic organic synthesis reactions, and the extent and diversity of life in the subseafloor in an actively serpentinizing system. We sampled fluids and measured temperature in Hole U1309D before deepening it to 1498 mbsf. The thermal structure was very similar to that measured during Expedition 340T, and lithologies were comparable to those found previously in Hole U1309D. A significant zone of cataclasis and alteration was found at 1451–1474 mbsf. A new Hole U1601C (proposed Site AMDH-02A) was drilled on the southern ridge close to Expedition 357 Hole M0069A, where both deformed and undeformed serpentinites had previously been recovered. Rapid drilling rates achieved a total depth of 1267.8 mbsf through predominantly ultramafic (68%) and gabbroic (32%) rocks, far surpassing the previous drilling record in a peridotite-dominated system of 201 m. Recovery was excellent overall (71%) but particularly high in peridotite-dominated sections where recovery regularly exceeded 90%. The recovery of sizable sections of largely intact material will provide robust constraints on the architecture and composition of the oceanic mantle lithosphere. The deepest portions of the newly drilled borehole may be beyond the known limits of life, providing the means to assess the role of biological activity across the transition from a biotic to an abiotic regime. Borehole fluids from both holes were collected using both the Kuster Flow-Through Sampler and the new Multi-Temperature Fluid Sampler. Wireline logging in Hole U1601C provided information on downhole density and resistivity, imaged structural features, and documented fracture orientations. A reentry system was installed in Hole U1601C, and both it and Hole U1309D were left open for future deep drilling, fluid sampling, and potential borehole observatories. 

Social media discourse involves people from different backgrounds, beliefs, and motives. Thus, often such discourse can devolve into toxic interactions. Generative Models, such as Llama and ChatGPT, have recently exploded in popularity due to their capabilities in zero-shot question-answering. Because these models are increasingly being used to ask questions of social significance, a crucial research question is whether they can understand social media dynamics. This work provides a critical analysis regarding generative LLM’s ability to understand language and dynamics in social contexts, particularly considering cyberbullying and anti-cyberbullying (posts aimed at reducing cyberbullying) interactions. Specifically, we compare and contrast the capabilities of different large language models (LLMs) to understand three key aspects of social dynamics: language, directionality, and the occurrence of bullying/anti-bullying messages. We found that while fine-tuned LLMs exhibit promising results in some social media understanding tasks (understanding directionality), they presented mixed results in others (proper paraphrasing and bullying/anti-bullying detection). We also found that fine-tuning and prompt engineering mechanisms can have positive effects in some tasks. We believe that a understanding of LLM’s capabilities is crucial to design future models that can be effectively used in social applications.