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 However, they are trained on static corpora and their knowledge can become outdated quickly in the fast-changing world.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. 

Abstract The response of sediment routing to climatic changes across icehouse-to-greenhouse turnovers is not well documented in Earth's pre-Cenozoic sedimentary record. Southwest Gondwana hosts one of the thickest and most laterally extensive records of Earth's penultimate icehouse, the late Paleozoic ice age. We present the first high-resolution U-Pb zircon chemical abrasion–isotope dilution–thermal ionization mass spectrometry (CA-ID-TIMS) analysis of late Paleozoic ice age deposits in the Kalahari Basin of southern Africa, which, coupled with existing CA-ID-TIMS zircon records from the Paraná and Karoo Basins, we used to refine the late Paleozoic ice age glacial history of SW Gondwana. Key findings from this work suggest that subglacial evidence in the Kalahari region is restricted to the Carboniferous (older than 300 Ma), with glacially influenced deposits culminating in this region by the earliest Permian (296 Ma). The U-Pb detrital zircon geochronologic records from the Paraná Basin of South America, which was located downstream of the Kalahari Basin in the latest Carboniferous and Permian, indicate that large-scale changes in sediment supplied to the Paraná were contemporaneous with shifts in the SW Gondwana ice record. Gondwanan deglaciation events were associated with the delivery of far-field, African-sourced sediments into the Paraná Basin. In contrast, Gondwanan glacial periods were associated with the restriction of African-sourced sediments into the basin. We interpret the influx of far-field sediments into the Paraná Basin as an expansion of the catchment area for the Paraná Basin during the deglaciation events, which occurred in the latest Carboniferous (300–299 Ma), early Permian (296 Ma), and late early Permian (<284 Ma). The coupled ice and detrital zircon records for this region of Gondwana present opportunities to investigate climate feedbacks associated with changes in freshwater and nutrient delivery to late Paleozoic ocean basins across the turnover from icehouse to greenhouse conditions. 

The location, longevity, and geographic extent of late Paleozoic ice centers in west-central Gondwana remain ambiguous. Paleovalleys on the Rio Grande do Sul Shield of southernmost Brazil have previously been interpreted as fjords carved by outlet glaciers that originated in Africa and emptied into the Paraná Basin (Brazil). In this study, the sedimentology, stratigraphy, and provenance of sediments infilling two such paleovalleys (the Mariana Pimentel and Leão paleovalleys) were examined in order to test the hypothesis that an ice center over present day Namibia drained across southernmost Brazil during the Carboniferous and Permian. Contrary to previous findings, the facies assemblage from within the paleovalleys is inconsistent with a fjord setting and no clear evidence for glaciation was observed. The facies show a transition from a non-glacial lacustrine/estuarine environment, to a fluvial-dominated setting, and finally to a restricted marine/estuarine environment. Detrital zircon results present a single population of Neoproterozoic ages (c. 800–550 Ma) from the paleovalley fill that matches the ages of underlying igneous and metamorphic basement (Dom Feliciano Belt) and is incongruent with African sources that contain abundant older (Mesoproterozoic, Paleoproterozoic, and Archean) zircons. Furthermore, results suggest that the formation of the paleovalleys and the deposition of their fill were controlled by the reactivation of Neoproterozoic basement structures during the Carboniferous and Permian. The lack of evidence for glaciation in these paleovalleys highlights the need for detailed studies of supposed late Paleozoic glacial deposits. These results are supportive of the hypothesis that well-established glacial sediments on the Rio Grande do Sul Shield (southern margin of the Paraná Basin) may be the product of a separate lobe extending north across Uruguay, rather than a single, massive ice sheet draining west from Africa. 

The timing and geographic distribution of glaciers in high-latitude southern Gondwana during the Late Paleozoic Ice Age remain poorly constrained, ultimately precluding our ability to estimate ice volume and associated climate teleconnections and feedbacks during Earth's penultimate icehouse. Current aerial extents of glaciers, constrained by sedimentary flow directions, near exclusively infer paleo-glaciation to be highland-driven and may underestimate potential ice sources in continental regions from which ice sheets may have emanated. Here, we report new U-Pb ages and Hf isotope compositions of detrital zircons recovered from diamictites in two key mid- to high-latitude Gondwanan basins (Paraná, Brazil and Tepuel, Argentine Patagonia). The results indicate regional sediment sources for both basins during the early period of late Paleozoic glaciation evolving into more distal sources during the final deglaciation along southern and western Gondwana. Similar age sediment sourced from diamictites in the Congo Basin, that require an ice center in eastern Africa suggest the possibility of a large ice sheet in this area of Africa proximal to the Carboniferous-Permian boundary, which may have sourced sediments to the Paraná Basin. An inferred distal southern source of glacial sediment for the Tepuel Basin argues for the presence of an ice sheet(s) in the Ellsworth Block of Antarctica towards the end of the glaciation history in Patagonia. These findings indicate an evolution during the Late Paleozoic Ice Age from proximally to extrabasinally sourced sediment reflecting continental-scale glaciation and subsequent drainage from the Windhoek Highlands, Ellsworth Block and an east African source in west-central Gondwana. Coincidence with a long-term fall in atmospheric pCO2 during the Pennsylvanian to a minimum across the Carboniferous-Permian boundary and a subsequent rise in the early Permian suggests a primary CO2-driver for deglaciation in the Paraná Basin. Additional boundary conditions including availability of moisture and paleogeography likely further contributed to the timing of nucleation, growth and demise of these Gondwanan glaciers.