How does the conversational past of large language models (LLMs) influence their future performance? Recent work suggests that LLMs are affected by their conversational history in unexpected ways. For instance, hallucinations in prior interactions may influence subsequent model responses. In this work, we introduce History-Echoes, a framework that investigates how conversational history biases subsequent generations. The framework explores this bias from two perspectives: probabilistically, we model conversations as Markov chains to quantify state consistency; geometrically, we measure the consistency of consecutive hidden representations. Across three model families and six datasets spanning diverse phenomena, our analysis reveals a strong correlation between the two perspectives. By bridging these perspectives, we demonstrate that behavioral persistence manifests as a geometric trap, where gaps in the latent space confine the model's trajectory. Code available at https://github.com/technion-cs-nlp/OldHabitsDieHard.
Abstract While numerous studies have examined the initial adoption of integrated soil fertility management (ISFM) practices, little is known about the factors that sustain their continued use over time. This study addresses this knowledge gap by analyzing the determinants of sustained ISFM adoption among smallholder farmers in northern Ghana. The analysis draws on primary data from a 2023 structured survey of 412 randomly selected maize-farming households across 15 communities—formerly part of the Northern Region, now administratively designated as the Savannah and Northeast regions—and highlights the roles of insurance awareness, extension services, market accessibility, and credit availability. The results show that insurance awareness is positively associated with intercropping or relay cropping, yet negatively correlated with continued use of improved seeds, suggesting the need for insurance products that better align with specific input strategies. Market inaccessibility emerges as a major constraint, underscoring the importance of rural infrastructure development to support ISFM uptake. Credit availability exhibits a positive but statistically insignificant association with chemical fertilizer use, while it is significantly negatively associated with organic fertilizer use, indicating a potential trade-off whereby improved financial access may favor chemical inputs over organic ones. Extension services and household labor availability are both positively associated with input intensity, emphasizing the roles of information and labor resources in sustaining ISFM practices. Overall, the findings suggest the need for integrated policy measures combining tailored insurance schemes, infrastructure investment, and balanced financial support to promote long-term ISFM adoption and sustainable agricultural intensification.
History of scholarship and learning. The humanities, Social Sciences
Komala Subramanyam Cherukuri, Pranav Abishai Moses, Aisa Sakata
et al.
Oral histories are vital records of lived experience, particularly within communities affected by systemic injustice and historical erasure. Effective and efficient analysis of their oral history archives can promote access and understanding of the oral histories. However, Large-scale analysis of these archives remains limited due to their unstructured format, emotional complexity, and high annotation costs. This paper presents a scalable framework to automate semantic and sentiment annotation for Japanese American Incarceration Oral History. Using LLMs, we construct a high-quality dataset, evaluate multiple models, and test prompt engineering strategies in historically sensitive contexts. Our multiphase approach combines expert annotation, prompt design, and LLM evaluation with ChatGPT, Llama, and Qwen. We labeled 558 sentences from 15 narrators for sentiment and semantic classification, then evaluated zero-shot, few-shot, and RAG strategies. For semantic classification, ChatGPT achieved the highest F1 score (88.71%), followed by Llama (84.99%) and Qwen (83.72%). For sentiment analysis, Llama slightly outperformed Qwen (82.66%) and ChatGPT (82.29%), with all models showing comparable results. The best prompt configurations were used to annotate 92,191 sentences from 1,002 interviews in the JAIOH collection. Our findings show that LLMs can effectively perform semantic and sentiment annotation across large oral history collections when guided by well-designed prompts. This study provides a reusable annotation pipeline and practical guidance for applying LLMs in culturally sensitive archival analysis. By bridging archival ethics with scalable NLP techniques, this work lays the groundwork for responsible use of artificial intelligence in digital humanities and preservation of collective memory. GitHub: https://github.com/kc6699c/LLM4OralHistoryAnalysis.
Zefanya Bramantasaputra, Dea Daniella Wangsawijaya, Bharathram Ganapathisubramani
The present study experimentally investigates the recovery of smooth-wall turbulent boundary layers (TBLs) following non-equilibrium pressure gradients (PGs). The imposed pressure gradient history (PGH) comprises favourable-adverse pressure gradient (FAPG) sequences of varying strength, followed by recovery to zero-pressure-gradient (ZPG) conditions. Hot-wire anemometry measurements were obtained at multiple downstream stations in the recovery region, with friction Reynolds numbers $Re_τ$ ranging from 2000 to 6000 depending on downstream development. Comparative analysis at matched $Re_τ$ and Clauser pressure gradient parameter $β$ enables clear assessment of history effects on TBL behaviour. Results show that increasing PGH strength enhances the wake in mean velocity profiles and amplifies turbulence intensities across the boundary layer, including the inner peak, logarithmic region, and outer peak (a signature of APG). Downstream, the mean flow gradually recovers toward a ZPG-like state, but turbulence in the outer region retains a lasting impact of PGH. Spectral analysis indicates that PGH primarily affects outer-layer scales, introducing a distinct PG peak and modifying the VLSM peak - with energy amplification dependent on PGH strength and spatial characteristics governed by history effects. Downstream recovery involves merging of large-scale wavelengths and the reorganisation of turbulence structures toward a ZPG-like state - although the `recovered' VLSM streamwise length becomes shortened due to the mixing of lengthscales with the PG peak. These results demonstrate that even under matched local parameters, TBLs retain a clear imprint of their upstream history, consistent with the findings of Preskett et al. (2025); moreover, this study provides new insights regarding the central role of scale interactions in the recovery mechanism of TBL subjected to complex PGH.
Considering inflationary magnetogenesis induced by time-dependent kinetic and axial couplings of a massless Abelian vector boson field breaking the conformal invariance we show in this article that, surprisingly, the spectral shape of the large-scale primordial magnetic field power spectrum is insensitive to the post-inflationary history, namely the barotropic parameter ($w$) and the gauge coupling functions of the post-inflationary era.
Large Language Models (LLMs) have recently shown a promise and emergence of Theory of Mind (ToM) ability and even outperform humans in certain ToM tasks. To evaluate and extend the boundaries of the ToM reasoning ability of LLMs, we propose a novel concept, taxonomy, and framework, the ToM reasoning with Zero, Finite, and Infinite Belief History and develop a multi-round text-based game, called $\textit{Pick the Right Stuff}$, as a benchmark. We have evaluated six LLMs with this game and found their performance on Zero Belief History is consistently better than on Finite Belief History. In addition, we have found two of the models with small parameter sizes outperform all the evaluated models with large parameter sizes. We expect this work to pave the way for future ToM benchmark development and also for the promotion and development of more complex AI agents or systems which are required to be equipped with more complex ToM reasoning ability.
China, with the severe burden of hepatitis B, plays a significant role in the global efforts towards eliminating hepatitis B disease by 2030. Vaccination is recognized as the most effective measure to prevent infectious diseases. However, vaccine hesitancy remains a significant barrier to achieving herd immunity across diverse populations. To address this issue, the health ministries and public health authorities in China have implemented various measures to encourage hepatitis B vaccination. China’s National Hepatitis B Immunization Program, initiated in 1985, has been successful in controlling this vaccine-preventable disease. Given the challenges in eliminating hepatitis B, strengthening the National Hepatitis Immunization Program in China is of utmost importance. Through an analysis of policy documents, reports, and scientific papers, the history of the program was summarized, and effective approaches to address vaccine hesitancy were identified. This will help achieve universal health coverage of vaccines and effectively work towards meeting the goals set for 2030.