State Space Models (SSMs) like Mamba2 are a promising alternative to Transformers, with faster theoretical training and inference times -- especially for long context lengths. Recent work on Matryoshka Representation Learning -- and its application to Transformer backbones in works like MatFormer -- showed how to introduce nested granularities of smaller submodels in one universal elastic model. In this work, we present MatMamba: a state space model which combines Matryoshka-style learning with Mamba2, by modifying the block to contain nested dimensions to enable joint training and adaptive inference. MatMamba allows for efficient and adaptive deployment across various model sizes. We train a single large MatMamba model and are able to get a number of smaller nested models for free -- while maintaining or improving upon the performance of a baseline smaller model trained from scratch. We train language and image models at a variety of parameter sizes from 35M to 1.4B. Our results on ImageNet and FineWeb show that MatMamba models scale comparably to Transformers, while having more efficient inference characteristics. This makes MatMamba a practically viable option for deploying large-scale models in an elastic way based on the available inference compute. Code and models are open sourced at \url{https://github.com/ScaledFoundations/MatMamba}

As large language models (LLMs) are increasingly used across various applications, there is a growing need to control text generation to satisfy specific constraints or requirements. This raises a crucial question: Is it possible to guarantee strict constraint satisfaction in generated outputs while preserving the distribution of the original model as much as possible? We first define the ideal distribution - the one closest to the original model, which also always satisfies the expressed constraint - as the ultimate goal of guaranteed generation. We then state a fundamental limitation, namely that it is impossible to reach that goal through autoregressive training alone. This motivates the necessity of combining training-time and inference-time methods to enforce such guarantees. Based on this insight, we propose GUARD, a simple yet effective approach that combines an autoregressive proposal distribution with rejection sampling. Through GUARD's theoretical properties, we show how controlling the KL divergence between a specific proposal and the target ideal distribution simultaneously optimizes inference speed and distributional closeness. To validate these theoretical concepts, we conduct extensive experiments on two text generation settings with hard-to-satisfy constraints: a lexical constraint scenario and a sentiment reversal scenario. These experiments show that GUARD achieves perfect constraint satisfaction while almost preserving the ideal distribution with highly improved inference efficiency. GUARD provides a principled approach to enforcing strict guarantees for LLMs without compromising their generative capabilities.

Calibrating verbalized probabilities presents a novel approach for reliably assessing and leveraging outputs from black-box Large Language Models (LLMs). Recent methods have demonstrated improved calibration by applying techniques like Platt scaling or temperature scaling to the confidence scores generated by LLMs. In this paper, we explore the calibration of verbalized probability distributions for discriminative tasks. First, we investigate the capability of LLMs to generate probability distributions over categorical labels. We theoretically and empirically identify the issue of re-softmax arising from the scaling of verbalized probabilities, and propose using the invert softmax trick to approximate the "logit" by inverting verbalized probabilities. Through extensive evaluation on three public datasets, we demonstrate: (1) the robust capability of LLMs in generating class distributions, and (2) the effectiveness of the invert softmax trick in estimating logits, which, in turn, facilitates post-calibration adjustments.

Large vision-language models (LVLMs) integrate visual information into large language models, showcasing remarkable multi-modal conversational capabilities. However, the visual modules introduces new challenges in terms of robustness for LVLMs, as attackers can craft adversarial images that are visually clean but may mislead the model to generate incorrect answers. In general, LVLMs rely on vision encoders to transform images into visual tokens, which are crucial for the language models to perceive image contents effectively. Therefore, we are curious about one question: Can LVLMs still generate correct responses when the encoded visual tokens are attacked and disrupting the visual information? To this end, we propose a non-targeted attack method referred to as VT-Attack (Visual Tokens Attack), which constructs adversarial examples from multiple perspectives, with the goal of comprehensively disrupting feature representations and inherent relationships as well as the semantic properties of visual tokens output by image encoders. Using only access to the image encoder in the proposed attack, the generated adversarial examples exhibit transferability across diverse LVLMs utilizing the same image encoder and generality across different tasks. Extensive experiments validate the superior attack performance of the VT-Attack over baseline methods, demonstrating its effectiveness in attacking LVLMs with image encoders, which in turn can provide guidance on the robustness of LVLMs, particularly in terms of the stability of the visual feature space.

We live in a world that is experiencing an unprecedented boom of AI applications that increasingly penetrate and enhance all sectors of private and public life, from education, media, medicine, and mobility to the industrial and professional workspace, and -- potentially particularly consequentially -- robotics. As this world is simultaneously grappling with climate change, the climate and environmental implications of the development and use of AI have become an important subject of public and academic debate. In this paper, we aim to provide guidance on the climate-related regulation for data centers and AI specifically, and discuss how to operationalize these requirements. We also highlight challenges and room for improvement, and make a number of policy proposals to this end. In particular, we propose a specific interpretation of the AI Act to bring reporting on the previously unadressed energy consumption from AI inferences back into the scope. We also find that the AI Act fails to address indirect greenhouse gas emissions from AI applications. Furthermore, for the purpose of energy consumption reporting, we compare levels of measurement within data centers and recommend measurement at the cumulative server level. We also argue for an interpretation of the AI Act that includes environmental concerns in the mandatory risk assessment (sustainability risk assessment, SIA), and provide guidance on its operationalization. The EU data center regulation proves to be a good first step but requires further development by including binding renewable energy and efficiency targets for data centers. Overall, we make twelve concrete policy proposals, in four main areas: Energy and Environmental Reporting Obligations; Legal and Regulatory Clarifications; Transparency and Accountability Mechanisms; and Future Far-Reaching Measures beyond Transparency.

Large Language Models (LLMs) have exhibited strong mathematical reasoning and computational prowess, tackling tasks ranging from basic arithmetic to advanced competition-level problems. However, frequently occurring subtle errors, such as miscalculations or incorrect substitutions, limit the models' full mathematical potential. Existing studies to improve mathematical ability typically involve distilling reasoning skills from stronger LLMs or applying preference learning to step-wise response pairs. Although these methods leverage samples of varying granularity to mitigate reasoning errors, they overlook the frequently occurring subtle errors. A major reason is that sampled preference pairs involve differences unrelated to the errors, which may distract the model from focusing on subtle errors. In this work, we propose a novel preference learning framework called eRror-Injected Self-Editing (RISE), which injects predefined subtle errors into partial tokens of correct solutions to construct hard pairs for error mitigation. In detail, RISE uses the model itself to edit a small number of tokens in the solution, injecting designed subtle errors. Then, pairs composed of self-edited solutions and their corresponding correct ones, along with pairs of correct and incorrect solutions obtained through sampling, are used together for subtle error-aware DPO training. Compared with other preference learning methods, RISE further refines the training objective to focus on predefined errors and their tokens, without requiring fine-grained sampling or preference annotation. Extensive experiments validate the effectiveness of RISE, with preference learning on Qwen2-7B-Instruct yielding notable improvements of 3.0% on GSM8K and 7.9% on MATH.

Large Language Models (LLMs) have demonstrated remarkable performance across multiple tasks through in-context learning. For complex reasoning tasks that require step-by-step thinking, Chain-of-Thought (CoT) prompting has given impressive results, especially when combined with self-consistency. Nonetheless, some tasks remain particularly difficult for LLMs to solve. Tree of Thoughts (ToT) and Graph of Thoughts (GoT) emerged as alternatives, dividing the complex problem into paths of subproblems. In this paper, we propose Tree of Problems (ToP), a simpler version of ToT, which we hypothesise can work better for complex tasks that can be divided into identical subtasks. Our empirical results show that our approach outperforms ToT and GoT, and in addition performs better than CoT on complex reasoning tasks. All code for this paper is publicly available here: https://github.com/ArmelRandy/tree-of-problems.

To use generative question-and-answering (QA) systems for decision-making and in any critical application, these systems need to provide well-calibrated confidence scores that reflect the correctness of their answers. Existing calibration methods aim to ensure that the confidence score is on average indicative of the likelihood that the answer is correct. We argue, however, that this standard (average-case) notion of calibration is difficult to interpret for decision-making in generative QA. To address this, we generalize the standard notion of average calibration and introduce $\beta$-calibration, which ensures calibration holds across different question-and-answer groups. We then propose discretized posthoc calibration schemes for achieving $\beta$-calibration.

Recent advancements in Transformer-based large language models (LLMs) have set new standards in natural language processing. However, the classical softmax attention incurs significant computational costs, leading to a $O(T)$ complexity for per-token generation, where $T$ represents the context length. This work explores reducing LLMs' complexity while maintaining performance by introducing Rodimus and its enhanced version, Rodimus$+$. Rodimus employs an innovative data-dependent tempered selection (DDTS) mechanism within a linear attention-based, purely recurrent framework, achieving significant accuracy while drastically reducing the memory usage typically associated with recurrent models. This method exemplifies semantic compression by maintaining essential input information with fixed-size hidden states. Building on this, Rodimus$+$ combines Rodimus with the innovative Sliding Window Shared-Key Attention (SW-SKA) in a hybrid approach, effectively leveraging the complementary semantic, token, and head compression techniques. Our experiments demonstrate that Rodimus$+$-1.6B, trained on 1 trillion tokens, achieves superior downstream performance against models trained on more tokens, including Qwen2-1.5B and RWKV6-1.6B, underscoring its potential to redefine the accuracy-efficiency balance in LLMs. Model code and pre-trained checkpoints will be available soon.

Biased AI-generated medical advice and misdiagnoses can jeopardize patient safety, making the integrity of AI in healthcare more critical than ever. As Large Language Models (LLMs) take on a growing role in medical decision-making, addressing their biases and enhancing their accuracy is key to delivering safe, reliable care. This study addresses these challenges head-on by introducing new resources designed to promote ethical and precise AI in healthcare. We present two datasets: BiasMD, featuring 6,007 question-answer pairs crafted to evaluate and mitigate biases in health-related LLM outputs, and DiseaseMatcher, with 32,000 clinical question-answer pairs spanning 700 diseases, aimed at assessing symptom-based diagnostic accuracy. Using these datasets, we developed the EthiClinician, a fine-tuned model built on the ChatDoctor framework, which outperforms GPT-4 in both ethical reasoning and clinical judgment. By exposing and correcting hidden biases in existing models for healthcare, our work sets a new benchmark for safer, more reliable patient outcomes.