Repository for Publications and Research Data
News from the ETH Library
Recently Added
The Non-Effects of the Introduction of Innovation Tax Incentives on Research and Development in Switzerland
Item type: Working Paper
Spescha, Andrin (2026)
This study investigates the introduction of the cantonal innovation tax incentives in Switzerland in the year 2020. Because some of the 26 cantons did not introduce innovation tax incentives at all, while others introduced them at different levels, we can exploit rich quasi-experimental variation. The study runs difference-in-differences using estimators that can deal with alternative forms of treatment effect heterogeneity. It thereby relies on two distinct data sources that are complementary to each other and serve as mutual replications. Whereas the policy induced changes in the B-Index serve as treatment for the intensive margin how much to invest in R&D, the policy induced changes in the maximum R&D-related corporate tax rate serve as treatment for the extensive margin whether to invest in R&D or not. The results show a null or even negative effect of the innovation tax incentives on firms’ R&D expenditures and a positive but rather small effect on the share of firms that pursue R&D activity. The new entries into R&D can also not increase total R&D expenditures in a significant way. These particular results are likely due to a lack of immediate needs for and a sub-optimal design of the introduced innovation tax incentives.
Particle transformers for identifying Lorentz-boosted Higgs bosons decaying to a pair of W bosons
Item type: Journal Article
Hayrapetyan A.; Makarenko V.; Tumasyan A.; et al. (2026)
A novel deep neural network classifier, a “particle transformer” (ParT), is introduced for the identification of highly Lorentz-boosted resonances reconstructed as single, multipronged jets in measurements and searches performed by the CMS Collaboration at the CERN LHC. Based on a self-attention mechanism that allows the model to weigh the importance of different particles, ParT is trained on a wide variety of topologies, notably demonstrating strong performance for the first time on jets originating from boosted Higgs boson decays to W bosons. The ParT algorithm achieves a tagging efficiency of more than 50% for such jets at a background efficiency of 1%, while maintaining decorrelation from the jet mass. A calibration is performed in proton-proton collision data collected by CMS at a center-of-mass energy of 13 TeV, with a data set corresponding to a total luminosity of 138 fb−1. Data-to-simulation selection efficiency scale factors are measured to be in the 0.9–1.0 range, with relative uncertainties between 7 and 23%. The tagging capability of ParT enhances the sensitivity of standard model measurements and searches for beyond-the-standard-model resonances decaying to hadronic diboson systems.
Indoor, Perceived, and Municipal Seismic Intensity in the Greater Tokyo Metropolitan Area: Persistent Divergence and No Detectable Perceptual Calibration
Item type: Journal Article
Furuya T.; Tsuruoka H.; Schorlemmer D.; et al. (2026)
Discrepancies between announced regional seismic intensity and shaking experienced indoors may affect household mitigation and risk communication. We report a three-year field experiment (2021–2024) in the Greater Tokyo Metropolitan Area combining citizen-hosted low-cost sensors with subjective reports. Up to 92 volunteers answered questionnaires after 96 target earthquakes; sensor hosts kept plug-in micro-electromechanical systems (MEMS) seismometers at home. Delivery logs fixed the denominator: 5724 responses to 8054 attempts. First, indoor seismic intensity exceeded municipal Japan Meteorological Agency (JMA) seismic intensity by a median of +0.34 classes against unrounded instrumental values; site-year means of the excess spread over 1.4 classes against the municipal class, and the excess varied between events within a home (median standard deviation 0.41 classes). Perceived seismic intensity fell between the two, partly by definition of the JMA scale. The excess of perceived over municipal intensity was confined to weak shaking: +0.31 classes at municipal intensities of 1.0–1.5, and not resolved above 1.5. Second, local intensity was associated with feeling an event independently of magnitude and distance (odds ratio 2.42 per class), whereas floor level was not resolved. Third, in this highly prepared volunteer panel, repeated feedback over up to 96 mostly weak events, with receipt unverified, produced no detectable calibration of the individual bias.
Benchmarking large language models for HIV medical decision support
Item type: Journal Article
Cardenal-Antolin G.; Fellay J.; Jaha B.; et al. (2026)
Background: Large language models (LLMs) are emerging as tools to support clinical decision making. HIV management is a compelling use case due to its complexity and dynamic nature, involving diverse treatment options, comorbidities, and adherence challenges. However, integrating LLMs into clinical practice raises concerns about accuracy, safety, and clinician acceptance. Despite growing interest, their performance in HIV care remains poorly studied, and benchmarking is lacking. Methods: We developed HIVMedQA, a clinician-curated benchmark of HIV-related open-ended medical question-answer pairs spanning basic knowledge, clinical reasoning, complex patient vignettes, and bias-modified scenarios. We evaluated seven general-purpose and three medical LLMs. Performance was assessed using lexical similarity and an extended medical LLM-as-a-judge framework capturing key clinical dimensions, including question comprehension, reasoning, knowledge recall, bias, potential harm, and factual accuracy, to better capture nuances relevant to the medical domain, with additional evaluation by HIV-experienced physicians. Results: Performance varies substantially across models and task complexity. Gemini 2.5 Pro achieves the highest overall scores, followed by Claude 3.5 Sonnet and MedGemma-27B. Knowledge recall is generally stronger than question comprehension or clinical reasoning. Medical LLMs do not consistently outperform general-purpose models, and model size alone does not predict performance. Several models are sensitive to cognitive bias prompts. LLM-as-a-judge scoring aligns better with clinician assessment than lexical metrics. Conclusions: HIVMedQA provides a structured benchmark for evaluating LLMs in HIV clinical decision support. Current LLMs show promise, but limitations in reasoning, bias robustness, and safety indicate that careful validation, domain-specific evaluation, and clinician oversight remain essential before clinical deployment.
Global-scale variability of glacier velocity and mass loss across terminus environments
Item type: Journal Article
Mallalieu J.; Carrivick J.L.; Huss M.; et al. (2026)
The environment in which a glacier terminates can strongly influence its behaviour, stability and future evolution, yet the relationship between terminus type and glacier dynamics at the global scale remains poorly constrained. Here we present a globally-complete classification of glacier termini distinguishing between land, lake, marine and shelf environments, and integrate it with datasets of glacier geometry, dynamics and climate. We show that although lake-, marine- and shelf-terminating glaciers comprise only 3.2% of glaciers globally, they account for over half of global glacier area and mass loss. We also show that lake- and marine-terminating glaciers flow, on average, 10% and 44% faster than land-terminating glaciers, respectively. Furthermore, specific (area-normalised) mass loss rates at lake-terminating glaciers are 57% greater than those at land- and marine-terminating glaciers. Our findings underscore the need for improved representation of non-land glacier termini in models of glacier evolution and sea-level rise, and suggest that distinguishing the environmental and geometric controls of termini on glacier dynamics remains an important challenge for future research.