NeurIPS 2024 Papers

Despite the success of reinforcement learning from human feedback (RLHF) in aligning language models with human values, reward hacking, also termed reward overoptimization, remains a critical challenge. This issue primarily arises from reward misgeneralization, where reward models (RMs) compute reward using spurious features that are irrelevant to human preferences. In this work, we tackle this problem from an information-theoretic perspective and propose a framework for reward modeling, namely InfoRM, by introducing a variational information bottleneck objective to filter out irrelevant information. Notably, we further identify a correlation between overoptimization and outliers in the IB latent space of InfoRM, establishing it as a promising tool for detecting reward overoptimization. Inspired by this finding, we propose the Cluster Separation Index (CSI), which quantifies deviations in the IB latent space, as an indicator of reward overoptimization to facilitate the development of online mitigation strategies. Extensive experiments on a wide range of settings and RM scales (70M, 440M, 1.4B, and 7B) demonstrate the effectiveness of InfoRM. Further analyses reveal that InfoRM's overoptimization detection mechanism is not only effective but also robust across a broad range of datasets, signifying a notable advancement in the field of RLHF. The code will be released upon acceptance.

3D Gaussian Splatting (3DGS) has already become the emerging research focus in the fields of 3D scene reconstruction and novel view synthesis. Given that training a 3DGS requires a significant amount of time and computational cost, it is crucial to protect the copyright, integrity, and privacy of such 3D assets. Steganography, as a crucial technique for encrypted transmission and copyright protection, has been extensively studied. However, it still lacks profound exploration targeted at 3DGS. Unlike its predecessor NeRF, 3DGS possesses two distinct features: 1) explicit 3D representation; and 2) real-time rendering speeds. These characteristics result in the 3DGS point cloud files being public and transparent, with each Gaussian point having a clear physical significance. Therefore, ensuring the security and fidelity of the original 3D scene while embedding information into the 3DGS point cloud files is an extremely challenging task. To solve the above-mentioned issue, we first propose a steganography framework for 3DGS, dubbed GS-Hider, which can embed 3D scenes and images into original GS point clouds in an invisible manner and accurately extract the hidden messages. Specifically, we design a coupled secured feature attribute to replace the original 3DGS's spherical harmonics coefficients and then use a scene decoder and a message decoder to disentangle the original RGB scene and the hidden message. Extensive experiments demonstrated that the proposed GS-Hider can effectively conceal multimodal messages without compromising rendering quality and possesses exceptional security, robustness, capacity, and flexibility. Our project is available at: https://xuanyuzhang21.github.io/project/gshider.

Direct Preference Optimization (DPO) is a widely used offline preference optimization algorithm that reparameterizes reward functions in reinforcement learning from human feedback (RLHF) to enhance simplicity and training stability. In this work, we propose SimPO, a simpler yet more effective approach. The effectiveness of SimPO is attributed to a key design: using the _average_ log probability of a sequence as the implicit reward. This reward formulation better aligns with model generation and eliminates the need for a reference model, making it more compute and memory efficient. Additionally, we introduce a target reward margin to the Bradley-Terry objective to encourage a larger margin between the winning and losing responses, further improving the algorithm's performance. We compare SimPO to DPO and its latest variants across various state-of-the-art training setups, including both base and instruction-tuned models such as Mistral, Llama 3, and Gemma 2. We evaluate on extensive chat-based evaluation benchmarks, including AlpacaEval 2, MT-Bench, and Arena-Hard. Our results demonstrate that SimPO consistently and significantly outperforms existing approaches without substantially increasing response length. Specifically, SimPO outperforms DPO by up to 6.4 points on AlpacaEval 2 and by up to 7.5 points on Arena-Hard. Our top-performing model, built on Gemma-2-9B-it, achieves a 72.4\% length-controlled win rate on AlpacaEval 2, a 59.1\% win rate on Arena-Hard, and ranks 1st on Chatbot Arena among $<$10B models with real user votes.

Complex 3D scene understanding has gained increasing attention, with scene encoding strategies built on top of visual foundation models playing a crucial role in this success. However, the optimal scene encoding strategies for various scenarios remain unclear, particularly compared to their image-based counterparts. To address this issue, we present the first comprehensive study that probes various visual encoding models for 3D scene understanding, identifying the strengths and limitations of each model across different scenarios. Our evaluation spans seven vision foundation encoders, including image, video, and 3D foundation models. We evaluate these models in four tasks: Vision-Language Scene Reasoning, Visual Grounding, Segmentation, and Registration, each focusing on different aspects of scene understanding. Our evaluation yields key intriguing findings: Unsupervised image foundation models demonstrate superior overall performance, video models excel in object-level tasks, diffusion models benefit geometric tasks, language-pretrained models show unexpected limitations in language-related tasks, and the mixture-of-vision-expert (MoVE) strategy leads to consistent performance improvement. These insights challenge some conventional understandings, provide novel perspectives on leveraging visual foundation models, and highlight the need for more flexible encoder selection in future vision-language and scene understanding tasks.

Structured state-space models (SSMs) are gaining popularity as effective foundational architectures for sequential data, demonstrating outstanding performance across a diverse set of domains alongside desirable scalability properties. Recent developments show that if the linear recurrence powering SSMs allows for a selectivity mechanism leveraging multiplicative interactions between inputs and hidden states (e.g. Mamba, GLA, Hawk/Griffin, HGRN2), then the resulting architecture can surpass attention-powered foundation models trained on text in both accuracy and efficiency, at scales of billion parameters. In this paper, we give theoretical grounding to the selectivity mechanism, often linked to in-context learning, using tools from Rough Path Theory. We provide a framework for the theoretical analysis of generalized selective SSMs, fully characterizing their expressive power and identifying the gating mechanism as the crucial architectural choice. Our analysis provides a closed-form description of the expressive powers of modern SSMs, such as Mamba, quantifying theoretically the drastic improvement in performance from the previous generation of models, such as S4. Our theory not only motivates the success of modern selective state-space models, but also provides a solid framework to understand the expressive power of future SSM variants. In particular, it suggests cross-channel interactions could play a vital role in future improvements.

Offline model-based optimization seeks to optimize against a learned surrogate model without querying the true oracle objective function during optimization. Such tasks are commonly encountered in protein design, robotics, and clinical medicine where evaluating the oracle function is prohibitively expensive. However, inaccurate surrogate model predictions are frequently encountered along offline optimization trajectories. To address this limitation, we propose *generative adversarial model-based optimization* using **adaptive source critic regularization (aSCR)**—a task- and optimizer- agnostic framework for constraining the optimization trajectory to regions of the design space where the surrogate function is reliable. We propose a computationally tractable algorithm to dynamically adjust the strength of this constraint, and show how leveraging aSCR with standard Bayesian optimization outperforms existing methods on a suite of offline generative design tasks. Our code is available at https://github.com/michael-s-yao/gabo.

Trained models are often composed with post-hoc transforms such as temperature scaling (TS), ensembling and stochastic weight averaging (SWA) to improve performance, robustness, uncertainty estimation, etc. However, such transforms are typically applied only after the base models have already been finalized by standard means. In this paper, we challenge this practice with an extensive empirical study. In particular, we demonstrate a phenomenon that we call post-hoc reversal, where performance trends are reversed after applying post-hoc transforms. This phenomenon is especially prominent in high-noise settings. For example, while base models overfit badly early in training, both ensembling and SWA favor base models trained for more epochs. Post-hoc reversal can also prevent the appearance of double descent and mitigate mismatches between test loss and test error seen in base models. Preliminary analyses suggest that these transforms induce reversal by suppressing the influence of mislabeled examples, exploiting differences in their learning dynamics from those of clean examples. Based on our findings, we propose post-hoc selection, a simple technique whereby post-hoc metrics inform model development decisions such as early stopping, checkpointing, and broader hyperparameter choices. Our experiments span real-world vision, language, tabular and graph datasets. On an LLM instruction tuning dataset, post-hoc selection results in >1.5x MMLU improvement compared to naive selection.

Visual representations become progressively more abstract along the cortical hierarchy. These abstract representations define notions like objects and shapes, but at the cost of spatial specificity. By contrast, low-level regions represent spatially local but simple input features. How do spatially non-specific representations of abstract concepts in high-level areas flexibly modulate the low-level sensory representations in appropriate ways to guide context-driven and goal-directed behaviors across a range of tasks? We build a biologically motivated and trainable neural network model of dynamics in the visual pathway, incorporating local, lateral, and feedforward synaptic connections, excitatory and inhibitory neurons, and long-range top-down inputs conceptualized as low-rank modulations of the input-driven sensory responses by high-level areas. We study this ${\bf D}$ynamical ${\bf C}$ortical ${\bf net}$work ($DCnet$) in a visual cue-delay-search task and show that the model uses its own cue representations to adaptively modulate its perceptual responses to solve the task, outperforming state-of-the-art DNN vision and LLM models. The model's population states over time shed light on the nature of contextual modulatory dynamics, generating predictions for experiments. We fine-tune the same model on classic psychophysics attention tasks, and find that the model closely replicates known reaction time results. This work represents a promising new foundation for understanding and making predictions about perturbations to visual processing in the brain.

Large pre-trained models excel in zero/few-shot learning for language and vision tasks but face challenges in multivariate time series (TS) forecasting due to diverse data characteristics. Consequently, recent research efforts have focused on developing pre-trained TS forecasting models. These models, whether built from scratch or adapted from large language models (LLMs), excel in zero/few-shot forecasting tasks. However, they are limited by slow performance, high computational demands, and neglect of cross-channel and exogenous correlations. To address this, we introduce Tiny Time Mixers (TTM), a compact model (starting from 1M parameters) with effective transfer learning capabilities, trained exclusively on public TS datasets. TTM, based on the light-weight TSMixer architecture, incorporates innovations like adaptive patching, diverse resolution sampling, and resolution prefix tuning to handle pre-training on varied dataset resolutions with minimal model capacity. Additionally, it employs multi-level modeling to capture channel correlations and infuse exogenous signals during fine-tuning. TTM outperforms existing popular benchmarks in zero/few-shot forecasting by (4-40\%), while reducing computational requirements significantly. Moreover, TTMs are lightweight and can be executed even on CPU-only machines, enhancing usability and fostering wider adoption in resource-constrained environments. The model weights for reproducibility and research use are available at https://huggingface.co/ibm/ttm-research-r2/, while enterprise-use weights under the Apache license can be accessed as follows: the initial TTM-Q variant at https://huggingface.co/ibm-granite/granite-timeseries-ttm-r1, and the latest variants (TTM-B, TTM-E, TTM-A) weights are available at https://huggingface.co/ibm-granite/granite-timeseries-ttm-r2. The source code for the TTM model along with the usage scripts are available at https://github.com/ibm-granite/granite-tsfm/tree/main/tsfm_public/models/tinytimemixer

Many training data attribution (TDA) methods aim to estimate how a model's behavior would change if one or more data points were removed from the training set. Methods based on implicit differentiation, such as influence functions, can be made computationally efficient, but fail to account for underspecification, the implicit bias of the optimization algorithm, or multi-stage training pipelines. By contrast, methods based on unrolling address these issues but face scalability challenges. In this work, we connect the implicit-differentiation-based and unrolling-based approaches and combine their benefits by introducing Source, an approximate unrolling-based TDA method that is computed using an influence-function-like formula. While being computationally efficient compared to unrolling-based approaches, Source is suitable in cases where implicit-differentiation-based approaches struggle, such as in non-converged models and multi-stage training pipelines. Empirically, Source outperforms existing TDA techniques in counterfactual prediction, especially in settings where implicit-differentiation-based approaches fall short.

This paper introduces OpenGaussian, a method based on 3D Gaussian Splatting (3DGS) that possesses the capability for 3D point-level open vocabulary understanding. Our primary motivation stems from observing that existing 3DGS-based open vocabulary methods mainly focus on 2D pixel-level parsing. These methods struggle with 3D point-level tasks due to weak feature expressiveness and inaccurate 2D-3D feature associations. To ensure robust feature presentation and 3D point-level understanding, we first employ SAM masks without cross-frame associations to train instance features with 3D consistency. These features exhibit both intra-object consistency and inter-object distinction. Then, we propose a two-stage codebook to discretize these features from coarse to fine levels. At the coarse level, we consider the positional information of 3D points to achieve location-based clustering, which is then refined at the fine level. Finally, we introduce an instance-level 3D-2D feature association method that links 3D points to 2D masks, which are further associated with 2D CLIP features. Extensive experiments, including open vocabulary-based 3D object selection, 3D point cloud understanding, click-based 3D object selection, and ablation studies, demonstrate the effectiveness of our proposed method. The source code is available at our project page https://3d-aigc.github.io/OpenGaussian.

This paper considers the distributed learning problem where a group of agents cooperatively minimizes the summation of their local cost functions based on peer-to-peer communication. Particularly, we propose a highly efficient algorithm, termed ``B-ary Tree Push-Pull'' (BTPP), that employs two B-ary spanning trees for distributing the information related to the parameters and stochastic gradients across the network. The simple method is efficient in communication since each agent interacts with at most $(B+1)$ neighbors per iteration. More importantly, BTPP achieves linear speedup for smooth nonconvex objective functions with only $\tilde{O}(n)$ transient iterations, significantly outperforming the state-of-the-art results to the best of our knowledge.

Recently, prompt learning has emerged as the state-of-the-art (SOTA) for fair text-to-image (T2I) generation. Specifically, this approach leverages readily available reference images to learn inclusive prompts for each target Sensitive Attribute (tSA), allowing for fair image generation. In this work, we first reveal that this prompt learning-based approach results in degraded sample quality. Our analysis shows that the approach's training objective--which aims to align the embedding differences of learned prompts and reference images-- could be sub-optimal, resulting in distortion of the learned prompts and degraded generated images. To further substantiate this claim, **as our major contribution**, we deep dive into the denoising subnetwork of the T2I model to track down the effect of these learned prompts by analyzing the cross-attention maps. In our analysis, we propose a novel prompt switching analysis: I2H and H2I. Furthermore, we propose new quantitative characterization of cross-attention maps. Our analysis reveals abnormalities in the early denoising steps, perpetuating improper global structure that results in degradation in the generated samples. Building on insights from our analysis, we propose two ideas: (i) *Prompt Queuing* and (ii) *Attention Amplification* to address the quality issue. Extensive experimental results on a wide range of tSAs show that our proposed method outperforms SOTA approach's image generation quality, while achieving competitive fairness. More resources at FairQueue Project site: https://sutd-visual-computing-group.github.io/FairQueue

Heterophily, or the tendency of connected nodes in networks to have different class labels or dissimilar features, has been identified as challenging for many Graph Neural Network (GNN) models. While the challenges of applying GNNs for node classification when class labels display strong heterophily are well understood, it is unclear how heterophily affects GNN performance in other important graph learning tasks where class labels are not available. In this work, we focus on the link prediction task and systematically analyze the impact of heterophily in node features on GNN performance. We first introduce formal definitions of homophilic and heterophilic link prediction tasks, and present a theoretical framework that highlights the different optimizations needed for the respective tasks. We then analyze how different link prediction encoders and decoders adapt to varying levels of feature homophily and introduce designs for improved performance. Based on our definitions, we identify and analyze six real-world benchmarks spanning from homophilic to heterophilic link prediction settings, with graphs containing up to 30M edges. Our empirical analysis on a variety of synthetic and real-world datasets confirms our theoretical insights and highlights the importance of adopting learnable decoders and GNN encoders with ego- and neighbor-embedding separation in message passing for link prediction tasks beyond homophily.

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 $256 \times 256$.

Performative learning addresses the increasingly pervasive situations in which algorithmic decisions may induce changes in the data distribution as a consequence of their public deployment. We propose a novel view in which these performative effects are modelled as push forward measures. This general framework encompasses existing models and enables novel performative gradient estimation methods, leading to more efficient and scalable learning strategies. For distribution shifts, unlike previous models which require full specification of the data distribution, we only assume knowledge of the shift operator that represents the performative changes. This approach can also be integrated into various change-of-variable-based models, such as VAEs or normalizing flows. Focusing on classification with a linear-in-parameters performative effect, we prove the convexity of the performative risk under a new set of assumptions. Notably, we do not limit the strength of performative effects but rather their direction, requiring only that classification becomes harder when deploying more accurate models. In this case, we also establish a connection with adversarially robust classification by reformulating the performative risk as a min-max variational problem. Finally, we illustrate our approach on synthetic and real datasets.

Offline reinforcement learning (RL) has progressed with return-conditioned supervised learning (RCSL), but its lack of stitching ability remains a limitation. We introduce $Q$-Aided Conditional Supervised Learning (QCS), which effectively combines the stability of RCSL with the stitching capability of $Q$-functions. By analyzing $Q$-function over-generalization, which impairs stable stitching, QCS adaptively integrates $Q$-aid into RCSL's loss function based on trajectory return. Empirical results show that QCS significantly outperforms RCSL and value-based methods, consistently achieving or exceeding the highest trajectory returns across diverse offline RL benchmarks. QCS represents a breakthrough in offline RL, pushing the limits of what can be achieved and fostering further innovations.

Many algorithms for aligning LLMs with human preferences assume that human preferences are binary and deterministic. However, human preferences can vary across individuals, and therefore should be represented distributionally. In this work, we introduce the distributional soft preference labels and improve Direct Preference Optimization (DPO) with a weighted geometric average of the LLM output likelihood in the loss function. This approach adjusts the scale of learning loss based on the soft labels such that the loss would approach zero when the responses are closer to equally preferred. This simple modification can be easily applied to any DPO-based methods and mitigate over-optimization and objective mismatch, which prior works suffer from. Our experiments simulate the soft preference labels with AI feedback from LLMs and demonstrate that geometric averaging consistently improves performance on standard benchmarks for alignment research. In particular, we observe more preferable responses than binary labels and significant improvements where modestly-confident labels are in the majority.

Tokenizers serve as crucial interfaces between models and linguistic data, substantially influencing the efficacy and precision of large language models (LLMs). Traditional tokenization methods often rely on static frequency-based statistics and are not inherently synchronized with LLM architectures, which may limit model performance. In this study, we propose a simple but effective method to learn tokenizers specifically engineered for seamless integration with LLMs. Initiating with a broad initial vocabulary, we refine our tokenizer by monitoring changes in the model’s perplexity during training, allowing for the selection of a tokenizer that is closely aligned with the model’s evolving dynamics. Through iterative refinement, we develop an optimized tokenizer. Our empirical evaluations demonstrate that this adaptive approach significantly enhances accuracy compared to conventional methods, maintaining comparable vocabulary sizes and affirming its potential to improve LLM functionality.