Jonas Geiping

Principal Researcher@ELLIS Institute Tübingen·德国·OpenReview

研究方向

Efficient Language Modeling · ML Safety · Privacy in ML · Data Poisoning for Neural Networks · Theory of Deep Learning · Optimization Theory · Variational Methods in Image Processing

Jonas Geiping

Pitfalls in Evaluating Language Model Forecasters

The Illusion of Diminishing Returns: Measuring Long Horizon Execution in LLMs

GPTailor: Large Language Model Pruning Through Layer Cutting and Stitching

Training Dynamics Impact Post-Training Quantization Robustness

Sample Smart, Not Hard: Correctness-First Decoding for Better Reasoning in LLMs

Adaptive Attacks on Trusted Monitors Subvert AI Control Protocols

Strategic Dishonesty Can Undermine AI Safety Evaluations of Frontier LLMs

Capability-Based Scaling Trends for LLM-Based Red-Teaming

When Fewer Layers Break More Chains: Layer Pruning Harms Test-Time Scaling in LLMs

Teaching Pretrained Language Models to Think Deeper with Retrofitted Recurrence

Answer matching outperforms multiple choice for LLM evaluations

Efficient Parallel Samplers for Recurrent-Depth Models and Their Connection to Diffusion Language Models

Scaling Open-Ended Reasoning to Predict the Future

Rewiring Experts on the Fly: Continuous Rerouting for Better Online Adaptation in Mixture-of-Expert models

Is your batch size the problem? Revisiting the Adam-SGD gap in language modeling

Scaling up Test-Time Compute with Latent Reasoning: A Recurrent Depth Approach

Can Language Models Falsify? Evaluating Algorithmic Reasoning with Counterexample Creation

When, Where and Why to Average Weights?

Great Models Think Alike and this Undermines AI Oversight

A Realistic Threat Model for Large Language Model Jailbreaks

An Interpretable N-gram Perplexity Threat Model for Large Language Model Jailbreaks

What do we learn from inverting CLIP models?

Gradient-based Optimization of Dataset Mixtures

Data Brittleness Estimation with Self-Supervised Features

Extracting Parameter Counts from Fine-Tuning APIs