Small Vision-Language Models (VLMs) provide a computationally efficient alternative to larger models, at the cost of weaker generalization abilities and downstream task performance. These shortcomings could be addressed by test-time scaling techniques, but existing methods are typically computationally demanding, contradicting the resource-efficient design goals of small models. To address these limitations, we propose two novel and efficient test-time scaling strategies that leverage the model-internal features rather than external supervision: (i) Test-Time Augmentation (TTAug), which generates multiple augmented inputs and aggregates outputs at the token level without parameter updates, and (ii) Test-Time Adaptation (TTAdapt), which adapts model parameters during inference using consensus-based pseudolabels from TTAug. Through extensive experiments across nine benchmarks, we demonstrate consistent performance improvements while maintaining computational efficiency suitable for resource-constrained environments. The generality of our approach is demonstrated both within models at different scales and across different VLMs without additional tuning.

In recent years, the input context sizes of large language models (LLMs) have increased dramatically. However, existing evaluation methods have not kept pace, failing to comprehensively assess the efficiency of models in handling long contexts. To bridge this gap, we introduce the BABILong benchmark, designed to test language models' ability to reason across facts distributed in extremely long documents. BABILong includes a diverse set of 20 reasoning tasks, including fact chaining, simple induction, deduction, counting, and handling lists/sets. These tasks are challenging on their own, and even more demanding when the required facts are scattered across long natural text.

In this paper, we explore the limits of Microsoft Floating Point (MSFP), a new class of datatypes developed for production cloud-scale inferencing on custom hardware. Through the co-evolution of hardware design and algorithms, MSFP achieves accuracy comparable to or better than industry standards Bfloat16 and INT8 at 3x and 4x lower cost, respectively. MSFP incurs negligible impact to accuracy ( 1%), requires no changes to the model topology, and is integrated with a mature cloud production pipeline. MSFP supports various classes of deep learning models including CNNs, RNNs, and Transformers without modification. Finally, we characterize the accuracy and implementation of MSFP and demonstrate its efficacy on a number of production scenarios, including models that power major online scenarios such as web search, question-answering, and image classification.

Diffusion models (DMs) have been significantly developed and widely used in various applications due to their excellent generative qualities. However, the expensive computation and massive parameters of DMs hinder their practical use in resource-constrained scenarios. As one of the effective compression approaches, quantization allows DMs to achieve storage saving and inference acceleration by reducing bit-width while maintaining generation performance. However, as the most extreme quantization form, 1-bit binarization causes the generation performance of DMs to face severe degradation or even collapse. This paper proposes a novel method, namely BiDM, for fully binarizing weights and activations of DMs, pushing quantization to the 1-bit limit. From a temporal perspective, we introduce the Timestep-friendly Binary Structure (TBS), which uses learnable activation binarizers and cross-timestep feature connections to address the highly timestep-correlated activation features of DMs.

We analyze the capabilities of Transformer language models in learning compositional discrete tasks. To this end, we evaluate training LLaMA models and prompting GPT-4 and Gemini on four tasks demanding to learn a composition of several discrete sub-tasks. In particular, we measure how well these models can reuse primitives observable in the sub-tasks to learn the composition task. Our results indicate that compositional learning in state-of-the-art Transformer language models is highly sample inefficient: LLaMA requires more data samples than relearning all sub-tasks from scratch to learn the compositional task; in-context prompting with few samples is unreliable and fails at executing the sub-tasks or correcting the errors in multi-round code generation. Further, by leveraging complexity theory, we support these findings with a theoretical analysis focused on the sample inefficiency of gradient descent in memorizing feedforward models.

If asked to come up with a quintessentially "human" work of art, one could do worse than to name a film by Studio Ghibli. The Japanese animation studio, founded by the legendary eighty-four-year-old director Hayao Miyazaki, is known for its hand-drawn imagery, lushly organic color palettes, epic narratives, and evocation of both the emotional ambiguities of childhood and the twisting path to becoming an adult. We American millennials were blessed to have the films translated and distributed in English just as we were growing up, and so movies including "My Neighbor Totoro," "Princess Mononoke," and "Spirited Away" are nigh-universally recognizable touchstones of our youth. Any Ghibli imagery is primed to make us feel a combination of pleasurable nostalgia and mournful shivers, evoking the doomed forest creatures, greedy bathhouse ghosts, and missed connections featured in Miyazaki's cinematic story lines. Unfortunately, that sense of poignancy quickly erodes when you are bombarded with thousands of Ghibli-esque copycat images, as we all were online last week, thanks to OpenAI's latest version of its ChatGPT tool.

Causal graphs are commonly used to understand and model complex systems. Researchers often construct these graphs from different perspectives, leading to significant variations for the same problem. Comparing causal graphs is, therefore, essential for evaluating assumptions, integrating insights, and resolving disagreements. The rise of AI tools has further amplified this need, as they are increasingly used to generate hypothesized causal graphs by synthesizing information from various sources such as prior research and community inputs, providing the potential for automating and scaling causal modeling for complex systems. Similar to humans, these tools also produce inconsistent results across platforms, versions, and iterations. Despite its importance, research on causal graph comparison remains scarce. Existing methods often focus solely on structural similarities, assuming identical variable names, and fail to capture nuanced semantic relationships, which is essential for causal graph comparison. We address these gaps by investigating methods for comparing causal graphs from both semantic and structural perspectives. First, we reviewed over 40 existing metrics and, based on predefined criteria, selected nine for evaluation from two threads of machine learning: four semantic similarity metrics and five learning graph kernels. We discuss the usability of these metrics in simple examples to illustrate their strengths and limitations. We then generated a synthetic dataset of 2,000 causal graphs using generative AI based on a reference diagram. Our findings reveal that each metric captures a different aspect of similarity, highlighting the need to use multiple metrics.

Keeping large language models factually up-to-date is crucial for deployment, yet costly retraining remains a challenge. Knowledge editing offers a promising alternative, but methods are only tested on small-scale or synthetic edit benchmarks. In this work, we aim to bridge research into lifelong knowledge editing to real-world edits at practically relevant scale. We first introduce WikiBigEdit; a large-scale benchmark of real-world Wikidata edits, built to automatically extend lifelong for future-proof benchmarking. In its first instance, it includes over 500K question-answer pairs for knowledge editing alongside a comprehensive evaluation pipeline. Finally, we use WikiBigEdit to study existing knowledge editing techniques' ability to incorporate large volumes of real-world facts and contrast their capabilities to generic modification techniques such as retrieval augmentation and continual finetuning to acquire a complete picture of the practical extent of current lifelong knowledge editing.

The accelerating development and deployment of AI technologies depend on the continued ability to scale their infrastructure. This has implied increasing amounts of monetary investment and natural resources. Frontier AI applications have thus resulted in rising financial, environmental, and social costs. While the factors that AI scaling depends on reach its limits, the push for its accelerated advancement and entrenchment continues. In this paper, we provide a holistic review of AI scaling using four lenses (technical, economic, ecological, and social) and review the relationships between these lenses to explore the dynamics of AI growth. We do so by drawing on system dynamics concepts including archetypes such as "limits to growth" to model the dynamic complexity of AI scaling and synthesize several perspectives. Our work maps out the entangled relationships between the technical, economic, ecological and social perspectives and the apparent limits to growth. The analysis explains how industry's responses to external limits enables continued (but temporary) scaling and how this benefits Big Tech while externalizing social and environmental damages. To avoid an "overshoot and collapse" trajectory, we advocate for realigning priorities and norms around scaling to prioritize sustainable and mindful advancements.

We propose a framework for robust evaluation of reasoning capabilities of language models, using functional variants of benchmarks. Models that solve a reasoning test should exhibit no difference in performance over the static version of a problem compared to a snapshot of the functional variant. We have rewritten the relevant fragment of the MATH benchmark into its functional variant MATH(), with functionalization of other benchmarks to follow. When evaluating current state-of-the-art models over snapshots of MATH(), we find a reasoning gap -- the percentage difference between the static and functional accuracies. We find reasoning gaps from 58.35% to 80.31% among the state-of-the-art closed and open weights models that perform well on static benchmarks, with the caveat that the gaps are likely to be smaller with more sophisticated prompting strategies. Here we show that models which anecdotally have good reasoning performance over real-world tasks, have quantifiable lower gaps, motivating the open problem of building "gap 0" models. Code for evaluation and new evaluation datasets, three MATH() snapshots, are publicly available at https://github.com/consequentai/fneval/.