arXiv subjects

Derek Liu

Publications and source records attributed to Derek Liu.

5 recordsLinked to original sources

arXiv2025

Response-Based Knowledge Distillation for Multilingual Jailbreak Prevention Unwittingly Compromises Safety

Large language models (LLMs) are increasingly deployed worldwide, yet their safety alignment remains predominantly English-centric. This allows for vulnerabilities in non-English contexts, especially with low-resource languages. We introduce a novel application of knowledge distillation (KD) in the context of multilingual jailbreak prevention, examining its efficacy. We distill the refusal behaviors of a proprietary teacher model (OpenAI o1-mini) with Low-Rank Adaptation (LoRA) into three open-source student models: Meta-Llama-3-8B-Instruct, Gemma-2-2B-IT, and Qwen3-8B, using ~28,000 multilingual jailbreak prompts from XSafety via black-box response-based, parameter-efficient fine-tuning (PEFT). Evaluation on the MultiJail benchmark reveals a counterintuitive behavior: standard fine-tuning on the teacher's ``safe'' refusal data inadvertently increases Jailbreak Success Rate (JSR) for all student models, up to 16.6 percentage points. Our experiments reveal a divergent generalization to unseen languages during distillation, with varying outcomes depending on the base model. By removing a primary source of safety degradation, nuanced `boundary' refusals, we mitigate or even reverse safety declines in student models, although reductions in reasoning performance (GSM8K) persist. Overall, our exploratory study highlights the challenges and potential of KD as a technique for multilingual safety alignment, offering a foundation for future research in this direction.

cs.CL↗

arXiv2025

Cube: A Roblox View of 3D Intelligence

Foundation models trained on vast amounts of data have demonstrated remarkable reasoning and generation capabilities in the domains of text, images, audio and video. Our goal at Roblox is to build such a foundation model for 3D intelligence, a model that can support developers in producing all aspects of a Roblox experience, from generating 3D objects and scenes to rigging characters for animation to producing programmatic scripts describing object behaviors. We discuss three key design requirements for such a 3D foundation model and then present our first step towards building such a model. We expect that 3D geometric shapes will be a core data type and describe our solution for 3D shape tokenizer. We show how our tokenization scheme can be used in applications for text-to-shape generation, shape-to-text generation and text-to-scene generation. We demonstrate how these applications can collaborate with existing large language models (LLMs) to perform scene analysis and reasoning. We conclude with a discussion outlining our path to building a fully unified foundation model for 3D intelligence.

cs.CV↗

arXiv2025

Homology in Combinatorial Refraction Billiards

Given a graph $G$ with vertex set $\{1,\ldots,n\}$, we can project the graphical arrangement of $G$ to an $(n-1)$-dimensional torus to obtain a toric hyperplane arrangement. Adams, Defant, and Striker constructed a toric combinatorial refraction billiard system in which beams of light travel in the torus, refracting (with refraction coefficient $-1$) whenever they hit one of the toric hyperplanes in this toric arrangement. Each billiard trajectory in this system is periodic. We adopt a topological perspective and view the billiard trajectories as closed loops in the torus. We say $G$ is ensnaring if all of the billiard trajectories are contractible, and we say $G$ is expelling if none of the billiard trajectories is contractible. Our first main result states that a graph is expelling if and only if it is bipartite. We then provide several necessary conditions and several sufficient conditions for a graph to be ensnaring. For example, we show that the complement of an ensnaring graph cannot have a clique as a connected component. We also discuss ways to construct ensnaring graphs from other ensnaring graphs. For example, gluing two ensnaring graphs at a single vertex always yields another ensnaring graph.

math.CO↗

arXiv2023

KEPLET: Knowledge-Enhanced Pretrained Language Model with Topic Entity Awareness

In recent years, Pre-trained Language Models (PLMs) have shown their superiority by pre-training on unstructured text corpus and then fine-tuning on downstream tasks. On entity-rich textual resources like Wikipedia, Knowledge-Enhanced PLMs (KEPLMs) incorporate the interactions between tokens and mentioned entities in pre-training, and are thus more effective on entity-centric tasks such as entity linking and relation classification. Although exploiting Wikipedia's rich structures to some extent, conventional KEPLMs still neglect a unique layout of the corpus where each Wikipedia page is around a topic entity (identified by the page URL and shown in the page title). In this paper, we demonstrate that KEPLMs without incorporating the topic entities will lead to insufficient entity interaction and biased (relation) word semantics. We thus propose KEPLET, a novel Knowledge-Enhanced Pre-trained LanguagE model with Topic entity awareness. In an end-to-end manner, KEPLET identifies where to add the topic entity's information in a Wikipedia sentence, fuses such information into token and mentioned entities representations, and supervises the network learning, through which it takes topic entities back into consideration. Experiments demonstrated the generality and superiority of KEPLET which was applied to two representative KEPLMs, achieving significant improvements on four entity-centric tasks.

cs.CL↗

arXiv2023

NFT Wash Trading Detection

Wash trading is a form of market manipulation where the same entity sells an asset to themselves to drive up market prices, launder money under the cover of a legitimate transaction, or claim a tax loss without losing ownership of an asset. Although the practice is illegal with traditional assets, lack of supervision in the non-fungible token market enables criminals to wash trade and scam unsuspecting buyers while operating under regulators radar. AnChain.AI designed an algorithm that flags transactions within an NFT collection history as wash trades when a wallet repurchases a token within 30 days of previously selling it. The algorithm also identifies intermediate transactions within a wash trade cycle. Testing on 7 popular NFT collections reveals that on average, 0.14% of transactions, 0.11% of wallets, and 0.16% of tokens in each collection are involved in wash trading. These wash trades generate an overall total price manipulation, sales, and repurchase profit of \$900K, \$1.1M, and negative \$1.6M respectively. The results draw attention to the prevalent market manipulation taking place and inform unsuspecting buyers which tokens and sellers may be involved in criminal activity.

q-fin.GN↗