publications

2026

1. ICML

   Temperature Scaling in Discrete Sequence (Language) Models

   Hannah Scheufele^*, Peter Blohm^*, and Vikas K. Garg

   In International Conference on Machine Learning (ICML), 2026

   ICML 2026 Poster

   Abs PDF Blog Code Poster

   Temperature scaling is widely used to control generation behavior and improve reasoning performance in discrete sequence models. However, sequence-level temperature scaling is challenging because the partition function is typically intractable and token-wise temperature scaling can bias maximum a posteriori estimates. This work introduces a unified formalism for autoregressive and diffusion-based language models, partition-free metrics for quantifying temperature scaling, and fine-tuning objectives that reliably achieve the desired change in model temperature. Experiments with GPT2 and BD3LM show more consistent generation with lower perplexity, with additional evidence for improved reasoning performance.

2. ICLR

   

   Composition of Pretrained Diffusion Models: A Logic-Based Calculus

   Peter Blohm and Vikas K. Garg

   In International Conference on Learning Representations (ICLR), 2026

   ICLR 2026 Poster

   Abs PDF Blog Code

   Composing pretrained diffusion models provides a cost-effective mechanism to encode constraints and unlock complex generative capabilities. Prior work relies on set-theoretic composition operators but suffers from limited mode coverage, biased sampling, instability under negation queries, and violations of core compositional laws. We introduce a principled calculus grounded in fuzzy logic, defining conjunction, disjunction, and negation operators that generalize classical mixtures and avoid these pathologies. Beyond existing methods, Dombi operators enable richer compositions such as Exclusive-OR, and we derive Feynman-Kac correctors to mitigate sampling bias. Empirical results on image generation with Stable Diffusion and multi-objective molecular generation support the conceptual, theoretical, and methodological benefits.

2025

1. NeurIPS

   

   Discovering Opinion Intervals from Conflicts in Signed Graphs

   Peter Blohm^*, Florian Chen^*, Aristides Gionis, and 1 more author

   In Advances in Neural Information Processing Systems (NeurIPS), 2025

   NeurIPS 2025 Oral

   Top 1.6% PDF News Code

   Oral Presentation

2. ICML

   Probably Approximately Global Robustness Certification

   Peter Blohm, Patrick Indri, Thomas Gärtner, and 1 more author

   In International Conference on Machine Learning (ICML), 2025

   ICML 2025 Poster

   Abs PDF Code

   We propose probabilistic guarantees for adversarial robustness of classification algorithms. Traditional formal verification is often intractable for modern models, while sampling-only approaches lack formal guarantees. Our method samples an epsilon-net and evaluates a local robustness oracle on the sample, yielding probably approximately global robustness guarantees. The required sample size is independent of input dimensionality, number of classes, and learning algorithm, which enables application to larger neural networks than typical formal global verification methods. Experiments show that the approach better characterizes robustness than prior sampling-based baselines and scales favorably compared with formal methods.