Simon Fraser University @ NeurIPS 2024

TBD

I will discuss a concept-centric paradigm for building agents that can learn continually and reason flexibly across multiple domains and input modalities. The concept-centric agent utilizes a vocabulary of neuro-symbolic concepts. These concepts, including object, relation, and action concepts, are grounded on sensory inputs and actuation outputs. They are also compositional, allowing for the creation of novel concepts through their structural combination. To facilitate learning and reasoning, the concepts are typed and represented using a combination of symbolic programs and neural network representations. Leveraging such neuro-symbolic concepts, the agent can efficiently learn and recombine them to solve various tasks across different domains and data modalities, ranging from 2D images, videos, 3D scenes, temporal data, and robotic manipulation data.

[Zoom Link] https://sfu.zoom.us/j/89273707787?pwd=btj3OeM7kDEmY561XGiVRJ5PIDneab.1

High quality data is crucial for training LLMs with superior performance. In this talk, I will present two theoretically-rigorous approaches to find smaller subsets of examples that can improve the performance and efficiency of training LLMs. First, I will present a one-shot data selection method for supervised fine-tuning of LLMs. Then, I'll talk about an iterative data selection strategy to pretrain or fine-tune LLMs on imbalanced mixtures of language data. I'll conclude by showing empirical results confirming that the above data selection strategies can effectively improve the performance of various LLMs during fine-tuning and pretraining.

Generative models have transformed content creation, and the next frontier may be simulating realistic experiences in response to actions by humans and agents. In this talk, I will talk about a line of work that involves learning a real-world simulator (i.e., a world model) to emulate interactions through generative modeling of video content. I will then talk about the applications of using this world model to train embodied agents through reinforcement learning (RL) and planning, which have demonstrated zero-shot real-world transfer. Lastly, I will talk about how to improve generative world models from real-world feedback.

We show that even the most recent safety-aligned LLMs are not robust to simple adaptive jailbreaking attacks. First, we demonstrate how to successfully leverage access to logprobs for jailbreaking: we initially design an adversarial prompt template (sometimes adapted to the target LLM), and then we apply random search on a suffix to maximize a target logprob (e.g., of the token "Sure"), potentially with multiple restarts. In this way, we achieve 100% attack success rate -- according to GPT-4 as a judge -- on Vicuna-13B, Mistral-7B, Phi-3-Mini, Nemotron-4-340B, Llama-2-Chat-7B/13B/70B, Llama-3-Instruct-8B, Gemma-7B, GPT-3.5, GPT-4o, and R2D2 from HarmBench that was adversarially trained against the GCG attack. We also show how to jailbreak all Claude models -- that do not expose logprobs -- via either a transfer or prefilling attack with a 100% success rate. In addition, we show how to use random search on a restricted set of tokens for finding trojan strings in poisoned models -- a task that shares many similarities with jailbreaking -- which is the algorithm that brought us the first place in the SaTML'24 Trojan Detection Competition. The common theme behind these attacks is that adaptivity is crucial: different models are vulnerable to different prompting templates (e.g., R2D2 is very sensitive to in-context learning prompts), some models have unique vulnerabilities based on their APIs (e.g., prefilling for Claude), and in some settings, it is crucial to restrict the token search space based on prior knowledge (e.g., for trojan detection). For reproducibility purposes, we provide the code, logs, and jailbreak artifacts in the JailbreakBench format at https://github.com/tml-epfl/llm-adaptive-attacks.

TBD

TBD

Neural fields, particularly neural signed distance functions (neural SDFs), have recently become a popular representation for reconstructing point clouds without input normals. While the smoothness of this representation gives many regularization benefits, it also causes a lot of problems for optimization, with the end reconstruction being highly biased by the initialization of the field. In this talk, I will discuss our line of work on improving optimization strategies for neural SDF optimization. I will also talk about our work on a newer representation for this task that is quickly becoming popular, the generalized winding number field. This representation alleviates many of the drawbacks of neural SDFs for this task, and our work shows that many techniques developed for neural field optimization are beneficial here.