arXiv subjects

David Yanni

Publications and source records attributed to David Yanni.

6 recordsLinked to original sources

arXiv2025

Hedging and Non-Affirmation: Quantifying LLM Alignment on Questions of Human Rights

Hedging and non-affirmation are behaviors exhibited by large language models (LLMs) that limit the clear endorsement of specific statements. While these behaviors are desirable in subjective contexts, they are undesirable in the context of human rights - which apply unambiguously to all groups. We present a systematic framework to measure these behaviors in unconstrained LLM responses regarding various identity groups. We evaluate six large proprietary models as well as one open-weight LLM on 4738 prompts across 205 national and stateless ethnic identities and find that 4 out of 7 display hedging and non-affirmation that is significantly dependent on the identity of the group. While factors like conflict signals, sovereignty (whether identity is stateless), or economic indicators (GDP) also influence model behavior, their effect sizes are consistently weaker than the impact of identity itself. The systematic disparity is robust to methods of rephrasing the prompts. Since group identity is the strongest predictor of these behaviors, we use open-weight models to explore whether applying steering and orthogonalization techniques to these group identities can mitigate the rates of hedging and non-affirmation behaviors. We find that group steering is the most effective debiasing approach across query types and is robust to downstream forgetting.

cs.CY↗

arXiv2021

Cellular organization in lab-evolved and extant multicellular species obeys a maximum entropy law

The prevalence of multicellular organisms is due in part to their ability to form complex structures. How cells pack in these structures is a fundamental biophysical issue, underlying their functional properties. However, much remains unknown about how cell packing geometries arise, and how they are affected by random noise during growth - especially absent developmental programs. Here, we quantify the statistics of cellular neighborhoods of two different multicellular eukaryotes: lab-evolved "snowflake" yeast and the green alga $Volvox~carteri$. We find that despite large differences in cellular organization, the free space associated with individual cells in both organisms closely fits a modified gamma distribution, consistent with maximum entropy predictions originally developed for granular materials. This 'entropic' cellular packing ensures a degree of predictability despite noise, facilitating parent-offspring fidelity even in the absence of developmental regulation. Together with simulations of diverse growth morphologies, these results suggest that gamma-distributed cell neighborhood sizes are a general feature of multicellularity, arising from conserved statistics of cellular packing.

cond-mat.soft↗

arXiv2019

Cyber-physical risks of hacked Internet-connected vehicles

The integration of automotive technology with Internet-connectivity promises to both dramatically improve transportation, while simultaneously introducing the potential for new unknown risks. Internet-connected vehicles are like digital data because they can be targeted for malicious hacking. Unlike digital data, however, Internet-connected vehicles are cyber-physical systems that physically interact with each other and their environment. As such, the extension of cybersecurity concerns into the cyber-physical domain introduces new possibilities for self-organized phenomena in traffic flow. Here, we study a scenario envisioned by cybersecurity experts leading to a large number of Internet-connected vehicles being suddenly and simultaneously disabled. We investigate post-hack traffic using agent-based simulations, and discover the critical relevance of percolation for probabilistically predicting the outcomes on a multi-lane road in the immediate aftermath of a vehicle-targeted cyber attack. We develop an analytic percolation-based model to rapidly assess road conditions given the density of disabled vehicles and apply it to study the street network of Manhattan (NY, USA) revealing the city's vulnerability to this particular cyber-physical attack.

cs.CR↗

arXiv2019

Viscosity independent diffusion mediated by death and reproduction in biofilms

Bacterial biofilms, surface-attached communities of cells, are in some respects similar to colloidal solids; both are densely packed with non-zero yield stresses. However, unlike non-living materials, bacteria reproduce and die, breaking mechanical equilibrium and inducing collective dynamic responses. We report experiments and theory investigating the motion of immotile Vibrio cholerae, which can kill each other and reproduce in biofilms. We vary viscosity by using bacterial variants that secrete different amounts of extracellular matrix polymers, but are otherwise identical. Unlike thermally-driven diffusion, in which diffusivity decreases with increased viscosity, we find that cellular motion mediated by death and reproduction is independent of viscosity over timescales relevant to bacterial reproduction. To understand this surprising result, we use two separate modeling approaches. First we perform explicitly mechanical simulations of one-dimensional chains of Voigt-Kelvin elements that can die and reproduce. Next, we perform an independent statistical approach, modeling Brownian motion with the classic Langevin equation under an effective temperature that depends on cellular division rate. The diffusion of cells in both approaches agrees quite well, supporting a kinetic interpretation for the effective temperature used here and developed in previous work. As the viscoelastic behavior of biofilms is believed to play a large role in their anomalous biological properties, such as antibiotic resistance, the independence of cellular diffusive motion --- important for biofilm growth and remodeling --- on viscoelastic properties likely holds ecological, medical, and industrial relevance.

cond-mat.soft↗

arXiv2017

Collective behavior and emergent risks in a model of human- and autonomously-driven vehicles

While much effort has been invested in studies of traffic flow as a physics problem, two emerging trends in technology have broadened the subject for new investigations. The first trend is the development of self-driving vehicles. This highly-anticipated shift from human- to autonomous-drivers is expected to offer substantial benefits for traffic throughput by streamlining large-scale collective behavior. The second trend is the widespread hacking of Internet-connected devices, which as of 2015, includes vehicles. While the first proof-of-concept automobile hack was done at the single-vehicle scale, undesirable collective effects can easily arise if this activity becomes more common. Motivated by these two trends, we explore the phenomena that arise in an active matter model with lanes and lane-changing behavior. Our model incorporates a simplified minimal description of essential differences between human- and autonomous-drivers. We study the emergent collective behavior as the population of vehicles shifts from all-human to all-autonomous. Within the context of our model, we explore a worst-case scenario where Internet-connected autonomous vehicles are disabled simultaneously and \textit{en masse}. Our approach reveals a model-independent role for percolation in interpreting the results. A broad lesson our work highlights is that seemingly minor malicious activity can ultimately have major impacts when magnified through the action of collective behavior.

physics.soc-ph↗

arXiv2017

Life in the coffee-ring: how evaporation-driven density gradients dictate the outcome of inter-bacterial competition

When a drop dries, it often leaves a ring-shaped stain through a ubiquitous phenomenon known as the coffee-ring effect. This also occurs when the liquid contains suspended microbes; evaporation leaves cells at higher concentrations in the ring than the drop interior. Using biofilm experiments and cellular automata simulations, we show that the physical structure created by the coffee-ring effect can dramatically alter the outcome of inter-bacterial competition. We experimentally study this effect using two strains of $\textit{Vibrio cholerae}$ that compete using a contact-dependent killing mechanism termed the type 6 secretion system. By creating a heterogeneous density profile, the coffee-ring effect changes the outcome of competition: the bacterial strain that wins in the low-density interior loses in the far denser ring. Through simulations parameterized with experimentally-determined density profiles, we recapitulate our experimental findings. We examine the role of a $\textit{V. cholerae}$ strain's frequency, its relative efficacy at killing competitors, and the initial concentration of cells in the droplet in determining the outcome of competition. By scaling from individual cellular interactions to overall changes in strain frequency, our work demonstrates how the coffee-ring effect plays a powerful role in structuring microbial communities, indirectly driving ecological changes in community composition.

cond-mat.soft↗