Skip to main content
Search

University of Warwick

4 related articles

Subscribe to updates
  1. arXiv

    Bacterial genome hotspots no longer rest on a few summary statistics: neural posterior estimation separates recombination from mutation rates and assigns hotspot probabilities

    Recombination and mutation rates vary along bacterial genomes, the two signals overlap, and clonal relationships confound inference, while earlier methods often assumed a constant mutation rate or relied on a small set of summary statistics. This work trains a simulator-based neural posterior estimator conditioned on a fixed clonal genealogy to infer local recombination and mutation rates for each gene segment, then applies a circular posterior-propagated Difference-of-Gaussians detector that carries posterior uncertainty into hotspot calls. On simulated data the inferred rates track the true values and credible intervals narrow as gene length grows; candidate hotspots are reported for three real datasets in the Bacillus cereus group, Staphylococcus aureus and Escherichia coli.

  2. arXiv

    Heating control stops guessing from fixed thermometers: a time-free PINN holds room temperature fields within 1.85% error

    Indoor heating usually relies on one or a few fixed-point temperature readings, which miss the spatial variation around an occupant. A physics-informed neural network drops time as an explicit input and instead infers how the temperature field evolves from spatial coordinates plus a few sensor readings and their rates of change; across nine horizontal planes in three room-scale configurations it matched a validated COMSOL multiphysics simulation with a maximum relative error below 1.85%. Wired into occupant-centric heating control, it showed an average numerical error of about 0.67% against COMSOL-based offline control, with inference fast enough for real-time use.

  3. arXiv

    OPUR unifies jailbreak objectives via a harmfulness-reweighted distribution, lifting AgentHarm attack success on Llama-3.1-8B to 75.00% under stochastic decoding

    The work reformulates LLM-agent jailbreaking probabilistically: it shows that the gradient of the log expected harmfulness with respect to the input equals the expected log-likelihood gradient under a harmfulness-reweighted output distribution, unifying expected-harmfulness maximization with target-likelihood optimization; building on this, OPUR constructs surrogate targets from the model's own stochastic rollouts to guide adversarial suffix optimization, achieving higher targeted attack success rates on AgentHarm, InjecAgent, and WebShop.

  4. arXiv

    MLCommons releases Jailbreak Benchmark v1.0: unsafe-response rate across eight open-weight models rises from 11.08% to 18.65% under jailbreak conditions, with an average Resilience Gap of 7.57%

    MLCommons Jailbreak Benchmark v1.0 introduces an end-to-end methodology that evaluates eight open-weight large language models with paired baseline and adversarial conditions using 264 seed prompts spanning eleven hazard categories and representative attacks drawn from the MLCommons Jailbreak Taxonomy, judging responses with the AILuminate Assessment Standard v1.4 and measuring robustness through the Resilience Gap; the unsafe-response rate rose from 11.08% at baseline to 18.65% under jailbreak conditions, yielding an average Resilience Gap of 7.57%, with accessible systems showing a larger mean gap and attack effectiveness varying substantially across attack categories and hazards.

You’re all caught up