arXiv subjects

Florian Klein

Publications and source records attributed to Florian Klein.

4 recordsLinked to original sources

arXiv2026

Predicting Future Behaviors in Reasoning Models Enables Better Steering

Deployed large reasoning models (LRMs) often behave unexpectedly. Test-time steering controls LRM outputs by intervening on their hidden representations, but it can degrade output quality. We argue that prior steering work implicitly relies on internal features that detect behavior in already generated text. We show that these detection features are poor predictors of future behavioral outcomes, and thus not the natural intervention target. Instead, we train activation probes to predict future behavior likelihoods from intermediate reasoning steps. These probes predict the most likely behavior with 64%-91% accuracy, revealing a separate type of internal prediction features. Building on these prediction features, we introduce a text-level steering method, Future Probe Controlled Generation. FPCG samples multiple candidate sentences and chooses the best one according to a probe predicting the future behavior likelihood. This enables steering with almost no output quality degradation. FPCG also enables steering in several evaluations where activation steering fails. These results show that distinguishing detection and prediction features enables a more nuanced approach to controlling LRM behaviors.

cs.LG↗

arXiv2022

Stochasticity of infectious outbreaks and consequences for optimal interventions

Global strategies to contain a pandemic, such as social distancing and protective measures, are designed to reduce the overall transmission rate between individuals. Despite such measures, essential institutions, including hospitals, schools, and food producing plants, remain focal points of local outbreaks. Here we develop a model for the stochastic outbreak dynamics in such local communities. We derive analytical expressions for the probability of containment of the outbreak, which is complementary to the probability of seeding a deterministically growing epidemic. This probability depends on the statistics of the intra-community contact network and the initial conditions, in particular, on the contact degree of patient zero. Based on this model, we suggest surveillance protocols by which individuals are tested proportionally to their degree in the contact network. We characterize the efficacy of contact-based protocols as a function of the epidemiological and the contact network parameters, and show numerically that such protocols outperform random testing.

q-bio.PE↗

arXiv2021

Design of an optimal combination therapy with broadly neutralizing antibodies to suppress HIV-1

Broadly neutralizing antibodies (bNAbs) are promising targets for vaccination and therapy against HIV. Passive infusions of bNAbs have shown promise in clinical trials as a potential alternative for anti-retroviral therapy. A key challenge for the potential clinical application of bnAbs is the suppression of viral escape, which is more effectively achieved with a combination of bNAbs. However, identifying an optimal bNAb cocktail is combinatorially complex. Here, we propose a computational approach to predict the efficacy of a bNAb therapy trial based on the population genetics of HIV escape, which we parametrize using high-throughput HIV sequence data from a cohort of untreated bNAb-naive patients. By quantifying the mutational target size and the fitness cost of HIV-1 escape from bNAbs, we reliably predict the distribution of rebound times in three clinical trials. Importantly, we show that early rebounds are dominated by the pre-treatment standing variation of HIV-1 populations, rather than spontaneous mutations during treatment. Lastly, we show that a cocktail of three bNAbs is necessary to suppress the chances of viral escape below 1%, and we predict the optimal composition of such a bNAb cocktail. Our results offer a rational design for bNAb therapy against HIV-1, and more generally show how genetic data could be used to predict treatment outcomes and design new approaches to pathogenic control.

q-bio.PE↗

arXiv2020

Predicting in vivo escape dynamics of HIV-1 from a broadly neutralizing antibody

Broadly neutralizing antibodies are promising candidates for treatment and prevention of HIV-1 infections. Such antibodies can temporarily suppress viral load in infected individuals; however, the virus often rebounds by escape mutants that have evolved resistance. In this paper, we map an in vivo fitness landscape of HIV-1 interacting with broadly neutralizing antibodies, using data from a recent clinical trial. We identify two fitness factors, antibody dosage and viral load, that determine viral reproduction rates reproducibly across different hosts. The model successfully predicts the escape dynamics of HIV-1 in the course of an antibody treatment, including a characteristic frequency turnover between sensitive and resistant strains. This turnover is governed by a dosage-dependent fitness ranking, resulting from an evolutionary tradeoff between antibody resistance and its collateral cost in drug-free growth. Our analysis suggests resistance-cost tradeoff curves as a measure of antibody performance in the presence of resistance evolution.

q-bio.PE↗