Papers by Guy Van Den Broeck
The Pitfalls of KV Cache Compression (2026.acl-long)
Copied to clipboard
| Challenge: | Recent literature has shown minimal degradation of KV cache in multi-instruction prompts . authors show that certain instructions degrade much more rapidly with compression . |
| Approach: | They propose to change KV cache eviction policies to reduce the impact of KV evict bias . they propose to use a 'simple' evviction policy to reduce ejection bias if the LLM is a multi-instruction model . |
| Outcome: | The proposed methods show that certain instructions degrade much faster with compression, causing them to be ignored by the LLM. |
Adversarial Tokenization (2025.acl-long)
Copied to clipboard
| Challenge: | Current LLM pipelines account for only one possible tokenization for a given string . authors: noncanonical tokenizations can evade LLM safety while still generating meaningful responses. |
| Approach: | They show that LLM pipelines account for only one possible tokenization for a given string . they show that tokenizers retain semantic understanding of other tokenizations . authors propose an exploit that can be exploited to evade safety and alignment restrictions . |
| Outcome: | The proposed exploit exploits a previously unknown vulnerability in subword models. |
Where is the signal in tokenization space? (2024.emnlp-main)
Copied to clipboard
| Challenge: | Autoregressive large language models (LLMs) generate text by predicting the next word sequentially, but tokenization is a significant challenge. |
| Approach: | They propose to use a deterministic, rule-based mapping from text to token sequences to compute the marginal probability over all possible tokenizations. |
| Outcome: | The proposed method improves on a range of LLM evaluation benchmarks for a variety of architectures, including transformers and state space models. |