Indian Institute of Technology Gandhinagar

Vision-language models are increasingly deployed in high-stakes settings, yet their susceptibility to sycophantic manipulation remains poorly understood, particularly in relation to how these models represent visual information internally. Whether models whose visual representations more closely mirror human neural processing are also more resistant to adversarial pressure is an open question with implications for both neuroscience and AI safety. We investigate this question by evaluating 12 open-weight vision-language models spanning 6 architecture families and a 40\times parameter range (256M--10B) along two axes: brain alignment, measured by predicting fMRI responses from the Natural Scenes Dataset across 8 human subjects and 6 visual cortex regions of interest, and sycophancy, measured through 76,800 two-turn gaslighting prompts spanning 5 categories and 10 difficulty levels. Region-of-interest analysis reveals that alignment specifically in early visual cortex (V1--V3) is a reliable negative predictor of sycophancy (r = -0.441, BCa 95\% CI [-0.740, -0.031]), with all 12 leave-one-out correlations negative and the strongest effect for existence denial attacks (r = -0.597, p = 0.040). This anatomically specific relationship is absent in higher-order category-selective regions, suggesting that faithful low-level visual encoding provides a measurable anchor against adversarial linguistic override in vision-language models. We release our code on \href{this https URL}{GitHub} and dataset on \href{this https URL}{Hugging Face}

An intercomparison of climate change impacts projected by nine regional-scale hydrological models for 12 large river basins on all continents was performed, and sources of uncertainty were quantified in the framework of the ISIMIP project. The models ECOMAG, HBV, HYMOD, HYPE, mHM, SWAT, SWIM, VIC and WaterGAP3 were applied in the following basins: Rhine and Tagus in Europe, Niger and Blue Nile in Africa, Ganges, Lena, Upper Yellow and Upper Yangtze in Asia, Upper Mississippi, MacKenzie and Upper Amazon in America, and Darling in Australia. The model calibration and validation was done using WATCH climate data for the period 1971-2000. The results, evaluated with 14 criteria, are mostly satisfactory, except for the low flow. Climate change impacts were analyzed using projections from five global climate models under four representative concentration pathways. Trends in the period 2070-2099 in relation to the reference period 1975-2004 were evaluated for three variables: the long-term mean annual flow and high and low flow percentiles Q 10 and Q 90, as well as for flows in three months high- and low-flow periods denoted as HF and LF. For three river basins: the Lena, MacKenzie and Tagus strong trends in all five variables were found (except for Q 10 in the MacKenzie); trends with moderate certainty for three to five variables were confirmed for the Rhine, Ganges and Upper Mississippi; and increases in HF and LF were found for the Upper Amazon, Upper Yangtze and Upper Yellow. The analysis of projected streamflow seasonality demonstrated increasing streamflow volumes during the high-flow period in four basins influenced by monsoonal precipitation (Ganges, Upper Amazon, Upper Yangtze and Upper Yellow), an amplification of the snowmelt flood peaks in the Lena and MacKenzie, and a substantial decrease of discharge in the Tagus (all months). The overall average fractions of uncertainty for the annual mean flow projections in the multi-model ensemble applied for all basins were 57% for GCMs, 27% for RCPs, and 16% for hydrological models.