Gradient-based methods are the primary approach for training ne ural networks. In recent years, research in learning theory has shown that neural networks can efficiently lea rn various data classes using empirical risk minimization (ERM) methods. In many real-world settings, data a rrive sequentially in a non-stationary manner, requiring the learner to maintain performance on past tas ks while acquiring new capabilities. In such cases, a learning model must be continually learnable, meaning it should retain previously acquired knowledge when trained on new tasks. On the other hand, various le arning systems, including deep learning architectures, can be prone to catastrophic forgetting, that is, updating a model on new data causes a dramatic drop in performance on previously learned tasks [ McCloskey and Cohen, 1989, Goodfellow et al., 2013 ]. The goal of continual (lifelong) learning is to develop models and methods that, even without retraining on old data, experience minimal forgetting when incorporating new inform ation. Despite deep learning's ubiquity, characterizing the power and limitat ions of neural networks is still an ongoing research direction. While several recent works aim to unde rstand the power of gradient descent (GD) for training neural networks with stylized data distributions, these works are still limited to single-task scenarios (for some examples see [ Du et al., 2019, Bartlett et al., 2021, Abbe et al., 2022 ]).

Conducting contamination-free evaluation of mathematical capabilities can be difficult for two reasons: models may memorize a test set once it is made public, and current mathematical benchmarks are prone to overfitting due to having limited diversity of symbols and rules, coupled with closed-ended answers. This paper proposes a method to leverage these shortcomings as useful features to a construct dynamic, counterfactual benchmark, which can be used to both reveal overfitting and measure true reasoning. We demonstrate this via MatheMagic, which generates math test instances with the interpretations of numbers and operators altered, yet has automatically verifiable answers. Test instances are randomly seeded and constructed at test time to evaluate a model's induction or deduction capability, offering stability, extensibility, comparability, and robustness to overfitting. Our experiments find that models solve deduction more easily than induction, but they revert to standard math. Further analysis reveals that math-adapted models fail to exhibit a general "skill" of reasoning, and fine-tuning on induction tasks generalizes poorly.

Trustworthy verifiers are essential for the success of reinforcement learning with verifiable reward (RLVR), which is the core methodology behind various large reasoning models such as DeepSeek-R1. In complex domains like mathematical reasoning, rule-based verifiers have been widely adopted in previous works to train strong reasoning models. However, the reliability of these verifiers and their impact on the RL training process remain poorly understood. In this work, we take mathematical reasoning as a case study and conduct a comprehensive analysis of various verifiers in both static evaluation and RL training scenarios. First, we find that current open-source rule-based verifiers often fail to recognize equivalent answers presented in different formats across multiple commonly used mathematical datasets, resulting in non-negligible false negative rates. This limitation adversely affects RL training performance and becomes more pronounced as the policy model gets stronger. Subsequently, we investigate model-based verifiers as a potential solution to address these limitations. While the static evaluation shows that model-based verifiers achieve significantly higher verification accuracy, further analysis and RL results imply that they are highly susceptible to hacking, where they misclassify certain patterns in responses as correct, particularly after fine-tuning. This vulnerability is exploited during policy model optimization, leading to artificially inflated rewards. Our findings underscore the unique challenges inherent to both rule-based and model-based verifiers and provide insights toward developing more accurate and robust reward systems for reinforcement learning.

Recent advances in large language models (LLMs) have enabled the creation of highly effective chatbots. However, the compute costs of widely deploying LLMs have raised questions about profitability. Companies have proposed exploring ad-based revenue streams for monetizing LLMs, which could serve as the new de facto platform for advertising. This paper investigates the implications of personalizing LLM advertisements to individual users via a between-subjects experiment with 179 participants. We developed a chatbot that embeds personalized product advertisements within LLM responses, inspired by similar forays by AI companies. The evaluation of our benchmarks showed that ad injection only slightly impacted LLM performance, particularly response desirability. Results revealed that participants struggled to detect ads, and even preferred LLM responses with hidden advertisements. Rather than clicking on our advertising disclosure, participants tried changing their advertising settings using natural language queries. We created an advertising dataset and an open-source LLM, Phi-4-Ads, fine-tuned to serve ads and flexibly adapt to user preferences.

Video reasoning, the task of enabling machines to infer from dynamic visual content through multi-step logic, is crucial for advanced AI. While the Chain-of-Thought (CoT) mechanism has enhanced reasoning in text-based tasks, its application to video understanding remains underexplored. This paper presents a systematic analysis revealing that CoT often degrades performance in video reasoning, generating verbose but misleading internal monologues, and leading to hallucinated visual details and overridden correct intuitions - a phenomenon we term "visual thinking drift". We explain this drift through a Bayesian lens, positing that CoT traces often diverge from actual visual evidence, instead amplifying internal biases or language priors, causing models to storytell rather than engage in grounded reasoning. To counteract this, we introduce Visual Evidence Reward (VER), a novel reinforcement learning framework that explicitly rewards the generation of reasoning traces that are verifiably grounded in visual evidence. Comprehensive evaluation across 10 diverse video understanding benchmarks demonstrates that our Video-VER consistently achieves top performance. Our work sheds light on the distinct challenges of video-centric reasoning and encourages the development of AI that robustly grounds its inferences in visual evidence - for large multimodal models that not only "think before answering", but also "see while thinking".

An enduring challenge in contagion theory is that the pathways contagions follow through social networks exhibit emergent complexities that are difficult to predict using network structure. Here, we address this challenge by developing a causal modeling framework that (i) simulates the possible network pathways that emerge as contagions spread and (ii) identifies which edges and nodes are most impactful on diffusion across these possible pathways. This yields a surprising discovery. If people require exposure to multiple peers to adopt a contagion (a.k.a., 'complex contagions'), the pathways that emerge often only work in one direction. In fact, the more complex a contagion is, the more asymmetric its paths become. This emergent directedness problematizes canonical theories of how networks mediate contagion. Weak ties spanning network regions - widely thought to facilitate mutual influence and integration - prove to privilege the spread contagions from one community to the other. Emergent directedness also disproportionately channels complex contagions from the network periphery to the core, inverting standard centrality models. We demonstrate two practical applications. We show that emergent directedness accounts for unexplained nonlinearity in the effects of tie strength in a recent study of job diffusion over LinkedIn. Lastly, we show that network evolution is biased toward growing directed paths, but that cultural factors (e.g., triadic closure) can curtail this bias, with strategic implications for network building and behavioral interventions.

Low-Rank Adaptation (LoRA) has emerged as a dominant method in Parameter-Efficient Fine-Tuning (PEFT) for large language models, which augments the transformer layer with one down-projection $A$ and one up-projection $B$. However, LoRA's reliance on a single down-projection matrix ($A$) creates a representational bottleneck, as this solitary feature extractor is inherently insufficient for capturing the diverse signals required by complex tasks. This motivates our architectural shift to focus on enriching the feature adaptation to improve the downstream task adaptation ability. We propose MASA (Multi-$A$ Shared Adaptation), an architecture that implements a multi-$A$, single-$B$ structure where the multi-$A$ expert ensemble is asymmetrically shared across layers to ensure parameter efficiency. In MASA, these specialized experts capture diverse features, which are then integrated by a single, layer-specific $B$-matrix. The effectiveness and versatility of our method are validated through a comprehensive suite of experiments spanning multi-domain generalization, single-domain specialization, and multi-task reasoning. For example, on the MMLU benchmark, MASA achieves an average accuracy of 59.62%, outperforming the standard LoRA by 1.08 points (a relative improvement of 1.84%) with comparable learnable parameters of 0.52%.