NeurIPS 2024 Papers

Large language models (LLMs) demonstrate surprising capabilities, but we do not understand how they are implemented. One hypothesis suggests that these capabilities are primarily executed by small subnetworks within the LLM, known as circuits. But how can we evaluate this hypothesis? In this paper, we formalize a set of criteria that a circuit is hypothesized to meet and develop a suite of hypothesis tests to evaluate how well circuits satisfy them. The criteria focus on the extent to which the LLM's behavior is preserved, the degree of localization of this behavior, and whether the circuit is minimal. We apply these tests to six circuits described in the research literature. We find that synthetic circuits -- circuits that are hard-coded in the model -- align with the idealized properties. Circuits discovered in Transformer models satisfy the criteria to varying degrees. To facilitate future empirical studies of circuits, we created the \textit{circuitry} package, a wrapper around the \textit{TransformerLens} library, which abstracts away lower-level manipulations of hooks and activations. The software is available at \url{https://github.com/blei-lab/circuitry}.

Large Language Models (LLMs) struggle with reliably generating highly structured outputs, such as program code, mathematical formulas, or well-formed markup. Constrained decoding approaches mitigate this problem by greedily restricting what tokens an LLM can output at each step to guarantee that the output matches a given constraint. Specifically, in grammar-constrained decoding (GCD), the LLM's output must follow a given grammar. In this paper we demonstrate that GCD techniques (and in general constrained decoding techniques) can distort the LLM's distribution, leading to outputs that are grammatical but appear with likelihoods that are not proportional to the ones given by the LLM, and so ultimately are low-quality. We call the problem of aligning sampling with a grammar constraint, grammar-aligned decoding (GAD), and propose adaptive sampling with approximate expected futures (ASAp), a decoding algorithm that guarantees the output to be grammatical while provably producing outputs that match the conditional probability of the LLM's distribution conditioned on the given grammar constraint. Our algorithm uses prior sample outputs to soundly overapproximate the future grammaticality of different output prefixes. Our evaluation on code generation and structured NLP tasks shows how ASAp often produces outputs with higher likelihood (according to the LLM's distribution) than existing GCD techniques, while still enforcing the desired grammatical constraints.

Large Language Models (LLMs) are trained on vast amounts of data, most of which is automatically scraped from the internet. This data includes encyclopedic documents that harbor a vast amount of general knowledge (*e.g.*, Wikipedia) but also potentially overlap with benchmark datasets used for evaluating LLMs. Consequently, evaluating models on test splits that might have leaked into the training set is prone to misleading conclusions. To foster sound evaluation of language models, we introduce a new test dataset named RepLiQA, suited for question-answering and topic retrieval tasks. RepLiQA is a collection of five splits of test sets, four of which have not been released to the internet or exposed to LLM APIs prior to this publication. Each sample in RepLiQA comprises (1) a reference document crafted by a human annotator and depicting an imaginary scenario (*e.g.*, a news article) absent from the internet; (2) a question about the document’s topic; (3) a ground-truth answer derived directly from the information in the document; and (4) the paragraph extracted from the reference document containing the answer. As such, accurate answers can only be generated if a model can find relevant content within the provided document. We run a large-scale benchmark comprising several state-of-the-art LLMs to uncover differences in performance across models of various types and sizes in a context-conditional language modeling setting. Released splits of RepLiQA can be found here: https://huggingface.co/datasets/ServiceNow/repliqa.

While 2D diffusion models generate realistic, high-detail images, 3D shape generation methods like Score Distillation Sampling (SDS) built on these 2D diffusion models produce cartoon-like, over-smoothed shapes. To help explain this discrepancy, we show that the image guidance used in Score Distillation can be understood as the velocity field of a 2D denoising generative process, up to the choice of a noise term. In particular, after a change of variables, SDS resembles a high-variance version of Denoising Diffusion Implicit Models (DDIM) with a differently-sampled noise term: SDS introduces noise i.i.d. randomly at each step, while DDIM infers it from the previous noise predictions. This excessive variance can lead to over-smoothing and unrealistic outputs. We show that a better noise approximation can be recovered by inverting DDIM in each SDS update step. This modification makes SDS's generative process for 2D images almost identical to DDIM. In 3D, it removes over-smoothing, preserves higher-frequency detail, and brings the generation quality closer to that of 2D samplers. Experimentally, our method achieves better or similar 3D generation quality compared to other state-of-the-art Score Distillation methods, all without training additional neural networks or multi-view supervision, and providing useful insights into relationship between 2D and 3D asset generation with diffusion models.

We introduce Equivariant Neural Diffusion (END), a novel diffusion model for molecule generation in 3D that is equivariant to Euclidean transformations. Compared to current state-of-the-art equivariant diffusion models, the key innovation in END lies in its learnable forward process for enhanced generative modelling. Rather than pre-specified, the forward process is parameterized through a time- and data-dependent transformation that is equivariant to rigid transformations. Through a series of experiments on standard molecule generation benchmarks, we demonstrate the competitive performance of END compared to several strong baselines for both unconditional and conditional generation.

This paper presents innovative enhancements to diffusion models by integrating a novel multi-resolution network and time-dependent layer normalization. Diffusion models have gained prominence for their effectiveness in high-fidelity image generation. While conventional approaches rely on convolutional U-Net architectures, recent Transformer-based designs have demonstrated superior performance and scalability. However, Transformer architectures, which tokenize input data (via "patchification"), face a trade-off between visual fidelity and computational complexity due to the quadratic nature of self-attention operations concerning token length. While larger patch sizes enable attention computation efficiency, they struggle to capture fine-grained visual details, leading to image distortions. To address this challenge, we propose augmenting the **Di**ffusion model with the **M**ulti-**R**esolution network (DiMR), a framework that refines features across multiple resolutions, progressively enhancing detail from low to high resolution. Additionally, we introduce Time-Dependent Layer Normalization (TD-LN), a parameter-efficient approach that incorporates time-dependent parameters into layer normalization to inject time information and achieve superior performance. Our method's efficacy is demonstrated on the class-conditional ImageNet generation benchmark, where DiMR-XL variants surpass previous diffusion models, achieving FID scores of 1.70 on ImageNet $256 \times 256$ and 2.89 on ImageNet $512 \times 512$. Our best variant, DiMR-G, further establishes a state-of-the-art 1.63 FID on ImageNet $256 \times 256$.

Code generation models have increasingly become integral to aiding software development. Although current research has thoroughly examined the correctness of the code produced by code generation models, a vital aspect that plays a pivotal role in green computing and sustainability efforts — the efficiency of the generated code — has often been neglected. This paper presents Effibench, a benchmark with 1,000 efficiency-critical coding problems to assess the efficiency of code generated by code generation models. EffiBench contains a diverse set of LeetCode coding problems. Each problem is paired with an executable human-written canonical solution, which obtains the SOTA efficiency on the LeetCode solution leaderboard. With EffiBench, we empirically examine the ability of 42 large language models (35 open-source and 7 closed-source) to generate efficient code. Our evaluation results demonstrate that the efficiency of the code generated by LLMs is generally worse than the efficiency of human-written canonical solutions. For example, GPT-4 generated code has an average \textbf{3.12} times execution time that of the human-written canonical solutions. In the most extreme cases, the execution time and total memory usage of GPT-4 code are \textbf{13.89} and \textbf{43.92} times that of the canonical solutions. The source code of EffiBench is released on https://github.com/huangd1999/EffiBench. We also provide the LeaderBoard in https://huggingface.co/spaces/EffiBench/effibench-leaderboard.

Text-to-image diffusion models benefit artists with high-quality image generation. Yet their stochastic nature hinders artists from creating consistent images of the same subject. Existing methods try to tackle this challenge and generate consistent content in various ways. However, they either depend on external restricted data or require expensive tuning of the diffusion model. For this issue, we propose a novel one-shot tuning paradigm, termed OneActor. It efficiently performs consistent subject generation solely driven by prompts via a learned semantic guidance to bypass the laborious backbone tuning. We lead the way to formalize the objective of consistent subject generation from a clustering perspective, and thus design a cluster-conditioned model. To mitigate the overfitting challenge shared by one-shot tuning pipelines, we augment the tuning with auxiliary samples and devise two inference strategies: semantic interpolation and cluster guidance. These techniques are later verified to significantly improve the generation quality. Comprehensive experiments show that our method outperforms a variety of baselines with satisfactory subject consistency, superior prompt conformity as well as high image quality. Our method is capable of multi-subject generation and compatible with popular diffusion extensions. Besides, we achieve a $4\times$ faster tuning speed than tuning-based baselines and, if desired, avoid increasing the inference time. Furthermore, our method can be naturally utilized to pre-train a consistent subject generation network from scratch, which will implement this research task into more practical applications. (Project page: https://johnneywang.github.io/OneActor-webpage/)

Realistic trajectory generation with natural language control is pivotal for advancing autonomous vehicle technology. However, previous methods focus on individual traffic participant trajectory generation, thus failing to account for the complexity of interactive traffic dynamics. In this work, we propose InteractTraj, the first language-driven traffic trajectory generator that can generate interactive traffic trajectories. InteractTraj interprets abstract trajectory descriptions into concrete formatted interaction-aware numerical codes and learns a mapping between these formatted codes and the final interactive trajectories. To interpret language descriptions, we propose a language-to-code encoder with a novel interaction-aware encoding strategy. To produce interactive traffic trajectories, we propose a code-to-trajectory decoder with interaction-aware feature aggregation that synergizes vehicle interactions with the environmental map and the vehicle moves. Extensive experiments show our method demonstrates superior performance over previous SoTA methods, offering a more realistic generation of interactive traffic trajectories with high controllability via diverse natural language commands.

Building on the success of diffusion models in visual generation, flow-based models reemerge as another prominent family of generative models that have achieved competitive or better performance in terms of both visual quality and inference speed. By learning the velocity field through flow-matching, flow-based models tend to produce a straighter sampling trajectory, which is advantageous during the sampling process. However, unlike diffusion models for which fast samplers are well-developed, efficient sampling of flow-based generative models has been rarely explored. In this paper, we propose a framework called FlowTurbo to accelerate the sampling of flow-based models while still enhancing the sampling quality. Our primary observation is that the velocity predictor's outputs in the flow-based models will become stable during the sampling, enabling the estimation of velocity via a lightweight velocity refiner. Additionally, we introduce several techniques including a pseudo corrector and sample-aware compilation to further reduce inference time. Since FlowTurbo does not change the multi-step sampling paradigm, it can be effectively applied for various tasks such as image editing, inpainting, etc. By integrating FlowTurbo into different flow-based models, we obtain an acceleration ratio of 53.1\%$\sim$58.3\% on class-conditional generation and 29.8\%$\sim$38.5\% on text-to-image generation. Notably, FlowTurbo reaches an FID of 2.12 on ImageNet with 100 (ms / img) and FID of 3.93 with 38 (ms / img), achieving the real-time image generation and establishing the new state-of-the-art. Code is available at https://github.com/shiml20/FlowTurbo.

Despite their widespread use, the mechanisms by which large language models (LLMs) represent and regulate uncertainty in next-token predictions remain largely unexplored. This study investigates two critical components believed to influence this uncertainty: the recently discovered entropy neurons and a new set of components that we term token frequency neurons. Entropy neurons are characterized by an unusually high weight norm and influence the final layer normalization (LayerNorm) scale to effectively scale down the logits. Our work shows that entropy neurons operate by writing onto an \textit{unembedding null space}, allowing them to impact the residual stream norm with minimal direct effect on the logits themselves. We observe the presence of entropy neurons across a range of models, up to 7 billion parameters. On the other hand, token frequency neurons, which we discover and describe here for the first time, boost or suppress each token’s logit proportionally to its log frequency, thereby shifting the output distribution towards or away from the unigram distribution. Finally, we present a detailed case study where entropy neurons actively manage confidence: the setting of induction, i.e. detecting and continuing repeated subsequences.

Generative AI has revolutionised visual content editing, empowering users to effortlessly modify images and videos. However, not all edits are equal. To perform realistic edits in domains such as natural image or medical imaging, modifications must respect causal relationships inherent to the data generation process. Such image editing falls into the counterfactual image generation regime. Evaluating counterfactual image generation is substantially complex: not only it lacks observable ground truths, but also requires adherence to causal constraints. Although several counterfactual image generation methods and evaluation metrics exist a comprehensive comparison within a unified setting is lacking. We present a comparison framework to thoroughly benchmark counterfactual image generation methods. We evaluate the performance of three conditional image generation model families developed within the Structural Causal Model (SCM) framework. We incorporate several metrics that assess diverse aspects of counterfactuals, such as composition, effectiveness, minimality of interventions, and image realism. We integrate all models that have been used for the task at hand and expand them to novel datasets and causal graphs, demonstrating the superiority of Hierarchical VAEs across most datasets and metrics. Our framework is implemented in a user-friendly Python package that can be extended to incorporate additional SCMs, causal methods, generative models, and datasets for the community to build on. Code: https://github.com/gulnazaki/counterfactual-benchmark.

Graph Neural Networks (GNNs) have shown great performance in various tasks, with the core idea of learning from data labels and aggregating messages within the neighborhood of nodes. However, the common challenges in graphs are twofold: insufficient accurate (high-quality) labels and limited neighbors for nodes, resulting in weak GNNs. Existing graph augmentation methods typically address only one of these challenges, often adding training costs or relying on oversimplified or knowledge-intensive strategies, limiting their generalization. To simultaneously address both challenges faced by graphs in a generalized way, we propose an elegant method called IntraMix. Considering the incompatibility of vanilla Mixup with the complex topology of graphs, IntraMix innovatively employs Mixup among inaccurate labeled data of the same class, generating high-quality labeled data at minimal cost. Additionally, it finds data with high confidence of being clustered into the same group as the generated data to serve as their neighbors, thereby enriching the neighborhoods of graphs. IntraMix efficiently tackles both issues faced by graphs and challenges the prior notion of the limited effectiveness of Mixup in node classification. IntraMix is a theoretically grounded plug-in-play method that can be readily applied to all GNNs. Extensive experiments demonstrate the effectiveness of IntraMix across various GNNs and datasets. Our code is available at: [https://github.com/Zhengsh123/IntraMix](https://github.com/Zhengsh123/IntraMix).

Recent advancements in 3D generation have leveraged synthetic datasets with ground truth 3D assets and predefined camera trajectories. However, the potential of adopting real-world datasets, which can produce significantly more realistic 3D scenes, remains largely unexplored. In this work, we delve into the key challenge of the complex and scene-specific camera trajectories found in real-world captures. We introduce Director3D, a robust open-world text-to-3D generation framework, designed to generate both real-world 3D scenes and adaptive camera trajectories. To achieve this, (1) we first utilize a Trajectory Diffusion Transformer, acting as the \emph{Cinematographer}, to model the distribution of camera trajectories based on textual descriptions. Next, a Gaussian-driven Multi-view Latent Diffusion Model serves as the \emph{Decorator}, modeling the image sequence distribution given the camera trajectories and texts. This model, fine-tuned from a 2D diffusion model, directly generates pixel-aligned 3D Gaussians as an immediate 3D scene representation for consistent denoising. Lastly, the 3D Gaussians are further refined by a novel SDS++ loss as the \emph{Detailer}, which incorporates the prior of the 2D diffusion model. Extensive experiments demonstrate that Director3D outperforms existing methods, offering superior performance in real-world 3D generation.