We use cookies to analyze the browsing and usage of our website and to personalize your experience. You can disable these technologies at any time, but this may limit certain functionalities of the site. Read our Privacy Policy for more information.
Setting cookies
You can enable and disable the types of cookies you wish to accept. However certain choices you make could affect the services offered on our sites (e.g. suggestions, personalised ads, etc.).
Essential cookies
These cookies are necessary for the operation of the site and cannot be deactivated. (Still active)
Analytics cookies
Do you accept the use of cookies to measure the audience of our sites?
Multimedia Player
Do you accept the use of cookies to display and allow you to watch the video content hosted by our partners (YouTube, etc.)?
Scaling language models unlocks impressive capabilities, but the accompanying computational and memory demands make both training and deploy… (see more)ment expensive. Existing efficiency efforts typically target either parameter sharing or adaptive computation, leaving open the question of how to attain both simultaneously. We introduce Mixture-of-Recursions (MoR), a unified framework that combines the two axes of efficiency inside a single Recursive Transformer. MoR reuses a shared stack of layers across recursion steps to achieve parameter efficiency, while lightweight routers enable adaptive token-level thinking by dynamically assign recursion depth to tokens, thereby focusing quadratic attention computation only where it is most useful. Further enhancing its efficiency, MoR incorporates a recursion-wise key-value caching mechanism that eliminates redundant memory access across recursion steps by selectively storing only the key-value caches for designated tokens. Across pretraining runs at model scales ranging from 135M to 1.7B parameters, MoR forms a new Pareto frontier: at equal training FLOPs and smaller model sizes, it significantly lowers validation perplexity and improves few-shot accuracy, while delivering higher throughput compared with vanilla and existing recursive baselines. These gains demonstrate that MoR is an effective path towards large-model quality without incurring large-model cost.
A key challenge in AI alignment is guiding large language models (LLMs) to follow desired behaviors at test time. Activation steering, which… (see more) modifies internal model activations during inference, offers a potential solution. However, prior work in dense activation spaces struggles with superposition, wherein multiple features become entangled, limiting interpretability and precise control. In contrast, sparse representations provide an untapped opportunity for more interpretable behavior modulation. In this work, we introduce sparse activation steering (SAS), a method that leverages sparse autoencoders (SAEs) to steer LLM behavior in sparse spaces. By isolating behavior-specific features through a contrastive prompt-pairing approach, we define a set of features that can selectively reinforce or suppress behaviors. Experiments on Gemma 2 LLMs show that SAS vectors enable nuanced behavioral modulation and finer-grained control. Furthermore, scaling SAEs improves monosemanticity of SAS vectors, suggesting more reliable and interpretable interventions.
A key challenge in AI alignment is guiding large language models (LLMs) to follow desired behaviors at test time. Activation steering, which… (see more) modifies internal model activations during inference, offers a potential solution. However, prior work in dense activation spaces struggles with superposition, wherein multiple features become entangled, limiting interpretability and precise control. In contrast, sparse representations provide an untapped opportunity for more interpretable behavior modulation. In this work, we introduce sparse activation steering (SAS), a method that leverages sparse autoencoders (SAEs) to steer LLM behavior in sparse spaces. By isolating behavior-specific features through a contrastive prompt-pairing approach, we define a set of features that can selectively reinforce or suppress behaviors. Experiments on Gemma 2 LLMs show that SAS vectors enable nuanced behavioral modulation and finer-grained control. Furthermore, scaling SAEs improves monosemanticity of SAS vectors, suggesting more reliable and interpretable interventions.
Neural networks often learn simple explanations that fit the majority of the data while memorizing exceptions that deviate from these explan… (see more)ations.This behavior leads to poor generalization when the learned explanations rely on spurious correlations. In this work, we formalize the interplay between memorization and generalization, showing that spurious correlations would particularly lead to poor generalization when are combined with memorization. Memorization can reduce training loss to zero, leaving no incentive to learn robust, generalizable patterns. To address this, we propose memorization-aware training (MAT), which uses held-out predictions as a signal of memorization to shift a model's logits. MAT encourages learning robust patterns invariant across distributions, improving generalization under distribution shifts.
Neural networks often learn simple explanations that fit the majority of the data while memorizing exceptions that deviate from these explan… (see more)ations. This leads to poor generalization when the learned explanations are spurious. In this work, we formalize
Neural networks often learn simple explanations that fit the majority of the data while memorizing exceptions that deviate from these explan… (see more)ations. This leads to poor generalization when the learned explanations are spurious. In this work, we formalize
Adversarial Training (AT) is a well-known framework designed to mitigate adversarial vulnerabilities in neural networks. Recent research ind… (see more)icates that incorporating adversarial examples (AEs) in training can enhance models' generalization capabilities. To understand the impact of AEs on learning dynamics, we study AT through the lens of sample difficulty methodologies. Our findings show that AT leads to more stable learning dynamics compared to Natural Training (NT), resulting in gradual performance improvements and less overconfident predictions. This suggests that AT steers training away from learning easy, perturbable spurious features toward more resilient and generalizable ones. However, a trade-off exists between adversarial robustness and generalization gains, due to robust overfitting, limiting practical deployment. To address this, we propose using synthesized data to bridge this gap. Our results demonstrate that AT benefits significantly from synthesized data, whereas NT does not, enhancing generalization without compromising robustness and offering new avenues for developing robust and generalizable models.
Vision-Language Models (VLMs) have witnessed a surge in both research and real-world applications. However, as they becoming increasingly pr… (see more)evalent, ensuring their robustness against adversarial attacks is paramount. This work systematically investigates the impact of model design choices on the adversarial robustness of VLMs against image-based attacks. Additionally, we introduce novel, cost-effective approaches to enhance robustness through prompt formatting. By rephrasing questions and suggesting potential adversarial perturbations, we demonstrate substantial improvements in model robustness against strong image-based attacks such as Auto-PGD. Our findings provide important guidelines for developing more robust VLMs, particularly for deployment in safety-critical environments.