Error Propagation and Model Collapse in Diffusion Models: A Theoretical Study

Machine learning models are increasingly trained or fine-tuned on synthetic data. Recursively training on such data has been observed to significantly degrade performance in a wide range of tasks, often characterized by a progressive drift away from the target distribution. In this work, we theoretically analyze this phenomenon in the setting of score-based diffusion models. For a realistic pipeline where each training round uses a combination of synthetic data and fresh samples from the target distribution, we obtain upper and lower bounds on the accumulated divergence between the generated and target distributions. This allows us to characterize different regimes of drift, depending on the score estimation error and the proportion of fresh data used in each generation. We also provide empirical results on synthetic data and images to illustrate the theory.

Probabilistic Modelling is Sufficient for Causal Inference

Causal inference is a key research area in machine learning, yet confusion reigns over the tools needed to tackle it. There are prevalent claims in the machine learning literature that you need a bespoke causal framework or notation to answer causal questions. In this paper, we want to make it clear that you \emph{can} answer any causal inference question within the realm of probabilistic modelling and inference, without causal-specific tools or notation. Through concrete examples, we demonstrate how causal questions can be tackled by writing down the probability of everything. Lastly, we reinterpret causal tools as emerging from standard probabilistic modelling and inference, elucidating their necessity and utility.

Wavelet-Induced Rotary Encodings: RoPE Meets Graphs

We introduce WIRE: Wavelet-Induced Rotary Encodings. WIRE extends Rotary Position Encodings (RoPE), a popular algorithm in LLMs and ViTs, to graph-structured data. We demonstrate that WIRE is more general than RoPE, recovering the latter in the special case of grid graphs. WIRE also enjoys a host of desirable theoretical properties, including equivariance under node ordering permutation, compatibility with linear attention, and (under select assumptions) asymptotic dependence on graph resistive distance. We test WIRE on a range of synthetic and real-world tasks, including identifying monochromatic subgraphs, semantic segmentation of point clouds, and more standard graph benchmarks. We find it to be effective in settings where the underlying graph structure is important.

Language models' activations linearly encode training-order recency

We show that language models' activations linearly encode when information was learned during training. Our setup involves creating a model with a known training order by sequentially fine-tuning Llama-3.2-1B on six disjoint but otherwise similar datasets about named entities. We find that the average activations of test samples for the six training datasets encode the training order: when projected into a 2D subspace, these centroids are arranged exactly in the order of training and lie on a straight line. Further, we show that linear probes can accurately (~90%) distinguish "early" vs. "late" entities, generalizing to entities unseen during the probes' own training. The model can also be fine-tuned to explicitly report an unseen entity's training stage (~80% accuracy). Interestingly, this temporal signal does not seem attributable to simple differences in activation magnitudes, losses, or model confidence. Our paper demonstrates that models are capable of differentiating information by its acquisition time, and carries significant implications for how they might manage conflicting data and respond to knowledge modifications.

Graph Random Features for Scalable Gaussian Processes

We study the application of graph random features (GRFs) - a recently introduced stochastic estimator of graph node kernels - to scalable Gaussian processes on discrete input spaces. We prove that (under mild assumptions) Bayesian inference with GRFs enjoys $O(N^{3/2})$ time complexity with respect to the number of nodes $N$, compared to $O(N^3)$ for exact kernels. Substantial wall-clock speedups and memory savings unlock Bayesian optimisation on graphs with over $10^6$ nodes on a single computer chip, whilst preserving competitive performance.

Distributional Training Data Attribution

Randomness is an unavoidable part of training deep learning models, yet something that traditional training data attribution algorithms fail to rigorously account for. They ignore the fact that, due to stochasticity in the initialisation and batching, training on the same dataset can yield different models. In this paper, we address this shortcoming through introducing distributional training data attribution (d-TDA), the goal of which is to predict how the distribution of model outputs (over training runs) depends upon the dataset. We demonstrate the practical significance of d-TDA in experiments, e.g. by identifying training examples that drastically change the distribution of some target measurement without necessarily changing the mean. Intriguingly, we also find that influence functions (IFs), a popular but poorly-understood data attribution tool, emerge naturally from our distributional framework as the limit to unrolled differentiation; without requiring restrictive convexity assumptions. This provides a new mathematical motivation for their efficacy in deep learning, and helps to characterise their limitations.

JoLT: Joint Probabilistic Predictions on Tabular Data Using LLMs

We introduce a simple method for probabilistic predictions on tabular data based on Large Language Models (LLMs) called JoLT (Joint LLM Process for Tabular data). JoLT uses the in-context learning capabilities of LLMs to define joint distributions over tabular data conditioned on user-specified side information about the problem, exploiting the vast repository of latent problem-relevant knowledge encoded in LLMs. JoLT defines joint distributions for multiple target variables with potentially heterogeneous data types without any data conversion, data preprocessing, special handling of missing data, or model training, making it accessible and efficient for practitioners. Our experiments show that JoLT outperforms competitive methods on low-shot single-target and multi-target tabular classification and regression tasks. Furthermore, we show that JoLT can automatically handle missing data and perform data imputation by leveraging textual side information. We argue that due to its simplicity and generality, JoLT is an effective approach for a wide variety of real prediction problems.

Spectral-factorized Positive-definite Curvature Learning for NN Training

Many training methods, such as Adam(W) and Shampoo, learn a positive-definite curvature matrix and apply an inverse root before preconditioning. Recently, non-diagonal training methods, such as Shampoo, have gained significant attention; however, they remain computationally inefficient and are limited to specific types of curvature information due to the costly matrix root computation via matrix decomposition. To address this, we propose a Riemannian optimization approach that dynamically adapts spectral-factorized positive-definite curvature estimates, enabling the efficient application of arbitrary matrix roots and generic curvature learning. We demonstrate the efficacy and versatility of our approach in positive-definite matrix optimization and covariance adaptation for gradient-free optimization, as well as its efficiency in curvature learning for neural net training.

Tighter sparse variational Gaussian processes

Sparse variational Gaussian process (GP) approximations based on inducing points have become the de facto standard for scaling GPs to large datasets, owing to their theoretical elegance, computational efficiency, and ease of implementation. This paper introduces a provably tighter variational approximation by relaxing the standard assumption that the conditional approximate posterior given the inducing points must match that in the prior. The key innovation is to modify the conditional posterior to have smaller variances than that of the prior at the training points. We derive the collapsed bound for the regression case, describe how to use the proposed approximation in large data settings, and discuss its application to handle orthogonally structured inducing points and GP latent variable models. Extensive experiments on regression benchmarks, classification, and latent variable models demonstrate that the proposed approximation consistently matches or outperforms standard sparse variational GPs while maintaining the same computational cost. An implementation will be made available in all popular GP packages.

Efficient Few-Shot Continual Learning in Vision-Language Models

Vision-language models (VLMs) excel in tasks such as visual question answering and image captioning. However, VLMs are often limited by their use of pretrained image encoders, like CLIP, leading to image understanding errors that hinder overall performance. On top of that, real-world applications often require the model to be continuously adapted as new and often limited data continuously arrive. To address this, we propose LoRSU (Low-Rank Adaptation with Structured Updates), a robust and computationally efficient method for selectively updating image encoders within VLMs. LoRSU introduces structured and localized parameter updates, effectively correcting performance on previously error-prone data while preserving the model's general robustness. Our approach leverages theoretical insights to identify and update only the most critical parameters, achieving significant resource efficiency. Specifically, we demonstrate that LoRSU reduces computational overhead by over 25x compared to full VLM updates, without sacrificing performance. Experimental results on VQA tasks in the few-shot continual learning setting, validate LoRSU's scalability, efficiency, and effectiveness, making it a compelling solution for image encoder adaptation in resource-constrained environments.