Repository for Publications and Research Data
News from the ETH Library
Recently Added
How Important is ‘Perfect’ English for Machine Translation Prompts?
Item type: Conference Paper
Schmidtová, Patrícia; Bafna, Niyati; Aycock, Seth; et al. (2026)
Large language models (LLMs) show state-of-the-art performance in machine translation, but are also known to be sensitive to errors in user prompts. Given that these models are largely trained on and respond best to prompts in standard English, this may affect the quality of LLM outputs for second language English speakers as well as real-world lay users, with potentially disproportionate effects on the former. We explore this effect by modeling a range of error types exhibited by such users, motivated by studies of L2 English, and quantifying their impact on LLM performance. We work with two related tasks: machine translation and machine translation evaluation. We find that LLMs-as-MT are brittle to natural spelling errors but not to phrasal simplifications. However, the quality drop caused by these errors is lower than the variance over the initial prompt choice, suggesting that “perfect English” for a given prompt is less important than choosing a good prompt. Since lay users and L2 speakers may use non-optimal prompts as well as display imperfect language skills, our work calls for increasing the resilience of model performance to both these phenomena, in order to best serve a diverse user base, both from a robustness and fairness perspective.
PATS: Personality-Aware Teaching Strategies with Large Language Model Tutors
Item type: Conference Paper
Rooein, Donya; Chowdhury, Sankalan Pal; Eremeeva, Mariia; et al. (2026)
Recent advances in large language models (LLMs) demonstrate their potential as educational tutors. However, different tutoring strategies benefit different student personalities, and mismatches can be counterproductive to student outcomes. Despite this, current LLM tutoring systems do not take into account student personality traits. To address this problem, we first construct a taxonomy that links pedagogical methods to personality profiles, based on pedagogical literature. We simulate student-teacher conversations and use our framework to let the LLM tutor adjust its strategy to the simulated student personality. We evaluate the scenario with human teachers and find that they consistently prefer our approach over two baselines. Our method also increases the use of less common, high-impact strategies such as role-playing, which human and LLM annotators prefer significantly. Our findings pave the way for developing more personalized and effective LLM use in educational applications.
Of Caves and Currents: Tracing North Atlantic Meltwater in Speleothems of the Last Deglaciation
Item type: Doctoral Thesis
Endres, Laura (2026)
The Atlantic Meridional Overturning Circulation (AMOC) is a central component of the global climate system, redistributing heat and nutrients across the Atlantic Ocean. Its projected weakening due to enhanced polar ice melt represents one of the major uncertainties in future climate projections. Past episodes of reduced AMOC strength, such as during the Last Deglaciation, provide natural laboratories for analysing the interactions between meltwater input, ocean circulation, and climate change. This thesis examines how meltwater from different sectors of the Northern Hemisphere ice sheets affected the AMOC during the Last Deglaciation, and how these dynamics are recorded in terrestrial archives, in particular in speleothems from the Northwest Iberian Speleothem Archive (NISA).
A new, high-resolution NISA speleothem record spanning 24–12 thousand years before present at decadal resolution reveals a sequence of meltwater events and circulation changes. Using negative δ18O excursions as a tracer for subpolar North Atlantic meltwater and δ13Cinit as a relative temperature proxy, the record indicates gradual early-deglacial meltwater inflow followed by abrupt meltwater pulses during Heinrich Stadial 1, a well-known multi-millennial cold phase. A pronounced regional cooling, lagging the first major meltwater pulse by roughly 850 years, points to a temporary AMOC weakening, whereas later pulses show only minor temperature effects—suggesting a changing AMOC sensitivity under evolving boundary conditions.
To interpret these speleothem proxy patterns, we conducted general circulation model (GCM) experiments using passive dye tracers to simulate the dispersion of meltwater from various source regions under contrasting AMOC states. The resulting surface ocean anomaly fields reveal that spatial gradients of the intensity persist for centuries and depend strongly on both the meltwater source and the circulation state. Meltwater from the European Ice Sheet Complex produces dynamic regional anomalies over the eastern subpolar North Atlantic, whereas meltwater from the Laurentide Ice Sheet, discharged via the Mississippi system into the Gulf of Mexico, becomes largely diluted within the subpolar gyre. Meltwater entering the Arctic tends to remain trapped there, while discharge into the Labrador Sea or along the
East Greenland Current disperses efficiently across the subpolar North Atlantic, particularly within the “Ruddiman Belt.” In most cases, a weaker AMOC prolongs the surface residence time of anomalies, thereby amplifying local signals.
Building on these simulations, we assessed the meltwater imprint on the NISA speleothem record using two complementary approaches and developed a new forward-modeling framework that translates arbitrary meltwater scenarios into expected δ18O anomalies. To estimate the meltwater imprint, the modeled dye anomalies were first scaled to δ18O values based on ice-sheet endmember assumptions. We then applied atmospheric back-trajectory and moisture-uptake analyses to evaluate how atmospheric transport under different AMOC states affects the recorded signal. Our results confirm that meltwater-induced surface anomalies can be transferred to terrestrial archives, producing spatially variable, source-dependent signals in speleothems and ice cores. The new forward-modeling approach generalizes the estimation of such anomalies through tabulated impulse responses, enabling quantitative reconstructions of circulation-modulated meltwater signals and paving the way for future data-constrained reconstructions.
Methodologically, this thesis advances speleothem-based reconstructions by proposing an improved micro-drilling protocol that provides tighter age–depth control, particularly for slow-growing stalagmites. Investigations of fluorescent layer formation through dripwater analysis reveal the complexity of the underlying processes and clarify the limitations of using such layers for seasonal counting. Together, these developments enhance the reliability of speleothems for detecting meltwater inputs and climate perturbations at decadal to centennial scales. By combining tracer-based experiments, forward modeling, and high-resolution proxy analysis, this work demonstrates that NISA speleothems can capture meltwater-driven ocean anomalies and provide improved temporal constraints on AMOC variability with a high temporal resolution. The results underscore the importance of integrating terrestrial and marine archives with dynamic climate models to reconstruct the spatio-temporal evolution of deglacial meltwater events—ultimately contributing to more robust predictions of future AMOC responses under continued polar melting.
A Cyber-Physical System for Freeway Ramp Meter Signal Control Using Deep Reinforcement Learning in a Connected Environment
Item type: Bachelor Thesis
Isser, Julian (2026)
The goal of this replication study is to verify the claim by Hou et al. that, for freeway ramp metering signal control, the reinforcement learning algorithms Ape-X DQN, PPO and A3C outperform the baseline controllers no ramp metering, fixed-time control and ALINEA. The comparison is conducted under mixed near-congested/congested and extremely congested traffic conditions, using traffic throughput, average vehicle speed and per-vehicle averages for stops, fuel efficiency, CO2 emissions and NOx emissions as evaluation metrics.
Convex optimization with p-norm oracles
Item type: Conference Paper
Adil, Deeksha; Bullins, Brian; Jambulapati, Arun; et al. (2026)