Test-time alignment (TTA) aims to adapt models to specific rewards during inference. However, existing methods tend to either under-optimise or over-optimise (reward hack) the target reward function. We propose Null-Text Test-Time Alignment (Null-TTA), which aligns diffusion models by optimising the unconditional embedding in classifier-free guidance, rather than manipulating latent or noise variables. Due to the structured semantic nature of the text embedding space, this ensures alignment occurs on a semantically coherent manifold and prevents reward hacking (exploiting non-semantic noise patterns to improve the reward). Since the unconditional embedding in classifier-free guidance serves as the anchor for the model's generative distribution, Null-TTA directly steers model's generative distribution towards the target reward rather than just adjusting the samples, even without updating model parameters. Thanks to these desirable properties, we show that Null-TTA achieves state-of-the-art target test-time alignment while maintaining strong cross-reward generalisation. This establishes semantic-space optimisation as an effective and principled novel paradigm for TTA.

Neural Information Processing Systems http://nips.cc/

The safety alignment of current Large Language Models (LLMs) is vulnerable. Relatively simple attacks, or even benign fine-tuning, can jailbreak aligned models. We argue that many of these vulnerabilities are related to a shared underlying issue: safety alignment can take shortcuts, wherein the alignment adapts a model's generative distribution primarily over only its very first few output tokens. We refer to this issue as shallow safety alignment. In this paper, we present case studies to explain why shallow safety alignment can exist and provide evidence that current aligned LLMs are subject to this issue. We also show how these findings help explain multiple recently discovered vulnerabilities in LLMs, including the susceptibility to adversarial suffix attacks, prefilling attacks, decoding parameter attacks, and fine-tuning attacks. Importantly, we discuss how this consolidated notion of shallow safety alignment sheds light on promising research directions for mitigating these vulnerabilities. For instance, we show that deepening the safety alignment beyond just the first few tokens can often meaningfully improve robustness against some common exploits. Finally, we design a regularized finetuning objective that makes the safety alignment more persistent against fine-tuning attacks by constraining updates on initial tokens. Overall, we advocate that future safety alignment should be made more than just a few tokens deep.

Generative Vision-Language Models (VLMs) are prone to generate plausible-sounding textual answers that, however, are not always grounded in the input image. We investigate this phenomenon, usually referred to as "hallucination" and show that it stems from an excessive reliance on the language prior. In particular, we show that as more tokens are generated, the reliance on the visual prompt decreases, and this behavior strongly correlates with the emergence of hallucinations. To reduce hallucinations, we introduce Multi-Modal Mutual-Information Decoding (M3ID), a new sampling method for prompt amplification. M3ID amplifies the influence of the reference image over the language prior, hence favoring the generation of tokens with higher mutual information with the visual prompt. M3ID can be applied to any pre-trained autoregressive VLM at inference time without necessitating further training and with minimal computational overhead. If training is an option, we show that M3ID can be paired with Direct Preference Optimization (DPO) to improve the model's reliance on the prompt image without requiring any labels. Our empirical findings show that our algorithms maintain the fluency and linguistic capabilities of pre-trained VLMs while reducing hallucinations by mitigating visually ungrounded answers. Specifically, for the LLaVA 13B model, M3ID and M3ID+DPO reduce the percentage of hallucinated objects in captioning tasks by 25% and 28%, respectively, and improve the accuracy on VQA benchmarks such as POPE by 21% and 24%.

Variational autoencoders are powerful models for unsupervised learning. However deep models with several layers of dependent stochastic variables are difficult to train which limits the improvements obtained using these highly expressive models. We propose a new inference model, the Ladder Variational Autoencoder, that recursively corrects the generative distribution by a data dependent approximate likelihood in a process resembling the recently proposed Ladder Network. We show that this model provides state of the art predictive log-likelihood and tighter log-likelihood lower bound compared to the purely bottom-up inference in layered Variational Autoencoders and other generative models. We provide a detailed analysis of the learned hierarchical latent representation and show that our new inference model is qualitatively different and utilizes a deeper more distributed hierarchy of latent variables. Finally, we observe that batch-normalization and deterministic warm-up (gradually turning on the KL-term) are crucial for training variational models with many stochastic layers.

Some popular Machine Learning Uncertainty Quantification (ML-UQ) calibration statistics do not have predefined reference values and are mostly used in comparative studies. In consequence, calibration is almost never validated and the diagnostic is left to the appreciation of the reader. Simulated reference values, based on synthetic calibrated datasets derived from actual uncertainties, have been proposed to palliate this problem. As the generative probability distribution for the simulation of synthetic errors is often not constrained, the sensitivity of simulated reference values to the choice of generative distribution might be problematic, shedding a doubt on the calibration diagnostic. This study explores various facets of this problem, and shows that some statistics are excessively sensitive to the choice of generative distribution to be used for validation when the generative distribution is unknown. This is the case, for instance, of the correlation coefficient between absolute errors and uncertainties (CC) and of the expected normalized calibration error (ENCE). A robust validation workflow to deal with simulated reference values is proposed.

The practice of uncertainty quantification (UQ) validation, notably in machine learning for the physico-chemical sciences, rests on several graphical methods (scattering plots, calibration curves, reliability diagrams and confidence curves) which explore complementary aspects of calibration, without covering all the desirable ones. For instance, none of these methods deals with the reliability of UQ metrics across the range of input features (adaptivity). Based on the complementary concepts of consistency and adaptivity, the toolbox of common validation methods for variance- and intervals- based UQ metrics is revisited with the aim to provide a better grasp on their capabilities. This study is conceived as an introduction to UQ validation, and all methods are derived from a few basic rules. The methods are illustrated and tested on synthetic datasets and representative examples extracted from the recent physico-chemical machine learning UQ literature.

With the research directions described in this thesis, we seek to address the critical challenges in designing recommender systems that can understand the dynamics of continuous-time event sequences. We follow a ground-up approach, i.e., first, we address the problems that may arise due to the poor quality of CTES data being fed into a recommender system. Later, we handle the task of designing accurate recommender systems. To improve the quality of the CTES data, we address a fundamental problem of overcoming missing events in temporal sequences. Moreover, to provide accurate sequence modeling frameworks, we design solutions for points-of-interest recommendation, i.e., models that can handle spatial mobility data of users to various POI check-ins and recommend candidate locations for the next check-in. Lastly, we highlight that the capabilities of the proposed models can have applications beyond recommender systems, and we extend their abilities to design solutions for large-scale CTES retrieval and human activity prediction. A significant part of this thesis uses the idea of modeling the underlying distribution of CTES via neural marked temporal point processes (MTPP). Traditional MTPP models are stochastic processes that utilize a fixed formulation to capture the generative mechanism of a sequence of discrete events localized in continuous time. In contrast, neural MTPP combine the underlying ideas from the point process literature with modern deep learning architectures. The ability of deep-learning models as accurate function approximators has led to a significant gain in the predictive prowess of neural MTPP models. In this thesis, we utilize and present several neural network-based enhancements for the current MTPP frameworks for the aforementioned real-world applications.