Archives AI News

arXiv:2508.19414v1 Announce Type: new Abstract: We present a mechanistic case study of a format-dependent reasoning failure in Llama-3.1-8B-Instruct, where the model incorrectly judges "9.11" as larger than "9.8" in chat or Q&A formats, but answers correctly in simple format. Through systematic intervention, we discover transformers implement even/odd attention head specialization: even indexed heads handle numerical comparison, while odd heads serve incompatible functions. The bug requires exactly 8 even heads at Layer 10 for perfect repair. Any combination of 8+ even heads succeeds, while 7 or fewer completely fails, revealing sharp computational thresholds with perfect redundancy among the 16 even heads. SAE analysis reveals the mechanism: format representations separate (10% feature overlap at Layer 7), then re-entangle with different weightings (80% feature overlap at Layer 10), with specific features showing 1.5x amplification in failing formats. We achieve perfect repair using only 25% of attention heads and identify a 60% pattern replacement threshold, demonstrating that apparent full-module requirements hide sophisticated substructure with implications for interpretability and efficiency. All of our code is available at https://github.com/gussand/surgeon.

arXiv:2407.01991v4 Announce Type: replace Abstract: To find the shortest paths for all pairs on manifolds with infinitesimally defined metrics, we introduce a framework to generate them by predicting midpoints recursively. To learn midpoint prediction, we propose an actor-critic approach. We prove the soundness of our approach and show experimentally that the proposed method outperforms existing methods on several planning tasks, including path planning for agents with complex kinematics and motion planning for multi-degree-of-freedom robot arms.

arXiv:2508.19419v1 Announce Type: new Abstract: Accurate subsurface reservoir pressure control is extremely challenging due to geological heterogeneity and multiphase fluid-flow dynamics. Predicting behavior in this setting relies on high-fidelity physics-based simulations that are computationally expensive. Yet, the uncertain, heterogeneous properties that control these flows make it necessary to perform many of these expensive simulations, which is often prohibitive. To address these challenges, we introduce a physics-informed machine learning workflow that couples a fully differentiable multiphase flow simulator, which is implemented in the DPFEHM framework with a convolutional neural network (CNN). The CNN learns to predict fluid extraction rates from heterogeneous permeability fields to enforce pressure limits at critical reservoir locations. By incorporating transient multiphase flow physics into the training process, our method enables more practical and accurate predictions for realistic injection-extraction scenarios compare to previous works. To speed up training, we pretrain the model on single-phase, steady-state simulations and then fine-tune it on full multiphase scenarios, which dramatically reduces the computational cost. We demonstrate that high-accuracy training can be achieved with fewer than three thousand full-physics multiphase flow simulations -- compared to previous estimates requiring up to ten million. This drastic reduction in the number of simulations is achieved by leveraging transfer learning from much less expensive single-phase simulations.

arXiv:2412.18164v3 Announce Type: replace Abstract: Diffusion models have emerged as powerful tools for generative modeling, demonstrating exceptional capability in capturing target data distributions from large datasets. However, fine-tuning these massive models for specific downstream tasks, constraints, and human preferences remains a critical challenge. While recent advances have leveraged reinforcement learning algorithms to tackle this problem, much of the progress has been empirical, with limited theoretical understanding. To bridge this gap, we propose a stochastic control framework for fine-tuning diffusion models. Building on denoising diffusion probabilistic models as the pre-trained reference dynamics, our approach integrates linear dynamics control with Kullback-Leibler regularization. We establish the well-posedness and regularity of the stochastic control problem and develop a policy iteration algorithm (PI-FT) for numerical solution. We show that PI-FT achieves global convergence at a linear rate. Unlike existing work that assumes regularities throughout training, we prove that the control and value sequences generated by the algorithm maintain the regularity. Additionally, we explore extensions of our framework to parametric settings and continuous-time formulations, and demonstrate the practical effectiveness of the proposed PI-FT algorithm through numerical experiments. Our code is available at https://github.com/yinbinhan/fine-tuning-of-diffusion-models.

arXiv:2508.19424v1 Announce Type: new Abstract: Motivation. Understanding the pan-cancer mutational landscape offers critical insights into the molecular mechanisms underlying tumorigenesis. While patient-level machine learning techniques have been widely employed to identify tumor subtypes, cohort-level clustering, where entire cancer types are grouped based on shared molecular features, has largely relied on classical statistical methods. Results. In this study, we introduce a novel unsupervised contrastive learning framework to cluster 43 cancer types based on coding mutation data derived from the COSMIC database. For each cancer type, we construct two complementary mutation signatures: a gene-level profile capturing nucleotide substitution patterns across the most frequently mutated genes, and a chromosome-level profile representing normalized substitution frequencies across chromosomes. These dual views are encoded using TabNet encoders and optimized via a multi-scale contrastive learning objective (NT-Xent loss) to learn unified cancer-type embeddings. We demonstrate that the resulting latent representations yield biologically meaningful clusters of cancer types, aligning with known mutational processes and tissue origins. Our work represents the first application of contrastive learning to cohort-level cancer clustering, offering a scalable and interpretable framework for mutation-driven cancer subtyping.

arXiv:2504.11195v2 Announce Type: replace Abstract: Vision-language models (VLMs), such as CLIP, have gained significant popularity as foundation models, with numerous fine-tuning methods developed to enhance performance on downstream tasks. However, due to their inherent vulnerability and the common practice of selecting from a limited set of open-source models, VLMs suffer from a higher risk of adversarial attacks than traditional vision models. Existing defense techniques typically rely on adversarial fine-tuning during training, which requires labeled data and lacks of flexibility for downstream tasks. To address these limitations, we propose robust test-time prompt tuning (R-TPT), which mitigates the impact of adversarial attacks during the inference stage. We first reformulate the classic marginal entropy objective by eliminating the term that introduces conflicts under adversarial conditions, retaining only the pointwise entropy minimization. Furthermore, we introduce a plug-and-play reliability-based weighted ensembling strategy, which aggregates useful information from reliable augmented views to strengthen the defense. R-TPT enhances defense against adversarial attacks without requiring labeled training data while offering high flexibility for inference tasks. Extensive experiments on widely used benchmarks with various attacks demonstrate the effectiveness of R-TPT. The code is available in https://github.com/TomSheng21/R-TPT.

arXiv:2508.19441v1 Announce Type: new Abstract: Partial differential equations (PDEs) underpin the modeling of many natural and engineered systems. It can be convenient to express such models as neural PDEs rather than using traditional numerical PDE solvers by replacing part or all of the PDE's governing equations with a neural network representation. Neural PDEs are often easier to differentiate, linearize, reduce, or use for uncertainty quantification than the original numerical solver. They are usually trained on solution trajectories obtained by long time integration of the PDE solver. Here we propose a more sample-efficient data-augmentation strategy for generating neural PDE training data from a computer model by space-filling sampling of local "stencil" states. This approach removes a large degree of spatiotemporal redundancy present in trajectory data and oversamples states that may be rarely visited but help the neural PDE generalize across the state space. We demonstrate that accurate neural PDE stencil operators can be learned from synthetic training data generated by the computational equivalent of 10 timesteps' worth of numerical simulation. Accuracy is further improved if we assume access to a single full-trajectory simulation from the computer model, which is typically available in practice. Across several PDE systems, we show that our data-augmented synthetic stencil data yield better trained neural stencil operators, with clear performance gains compared with naively sampled stencil data from simulation trajectories.

arXiv:2506.07666v3 Announce Type: replace Abstract: Adversarial Robustness Distillation (ARD) has emerged as an effective method to enhance the robustness of lightweight deep neural networks against adversarial attacks. Current ARD approaches have leveraged a large robust teacher network to train one robust lightweight student. However, due to the diverse range of edge devices and resource constraints, current approaches require training a new student network from scratch to meet specific constraints, leading to substantial computational costs and increased CO2 emissions. This paper proposes Progressive Adversarial Robustness Distillation (ProARD), enabling the efficient one-time training of a dynamic network that supports a diverse range of accurate and robust student networks without requiring retraining. We first make a dynamic deep neural network based on dynamic layers by encompassing variations in width, depth, and expansion in each design stage to support a wide range of architectures. Then, we consider the student network with the largest size as the dynamic teacher network. ProARD trains this dynamic network using a weight-sharing mechanism to jointly optimize the dynamic teacher network and its internal student networks. However, due to the high computational cost of calculating exact gradients for all the students within the dynamic network, a sampling mechanism is required to select a subset of students. We show that random student sampling in each iteration fails to produce accurate and robust students.

arXiv:2508.19443v1 Announce Type: new Abstract: Producing large complex simulation datasets can often be a time and resource consuming task. Especially when these experiments are very expensive, it is becoming more reasonable to generate synthetic data for downstream tasks. Recently, these methods may include using generative machine learning models such as Generative Adversarial Networks or diffusion models. As these generative models improve efficiency in producing useful data, we introduce an internal tensor decomposition to these generative models to even further reduce costs. More specifically, for multidimensional data, or tensors, we generate the smaller tensor factors instead of the full tensor, in order to significantly reduce the model's output and overall parameters. This reduces the costs of generating complex simulation data, and our experiments show the generated data remains useful. As a result, tensor decomposition has the potential to improve efficiency in generative models, especially when generating multidimensional data, or tensors.

arXiv:2508.03772v3 Announce Type: replace Abstract: Policy-based optimizations are widely adopted today for the training and alignment of language models, where one of the most recent and effective approaches is Group-relative Policy Optimization (GRPO). In this paper, we reveals and analyze two major limitations of GRPO: (i) tokens frequently appear in completions with both positive and negative rewards, leading to conflicting gradient updates that can reduce their output probability, even though can be essential for maintaining proper structure; (ii) negatively rewarded completions may penalize confident responses and shift model decisions toward unlikely tokens, progressively flattening the output distribution and degrading learning. To address these issues and provide a more stable and effective policy optimization strategy, we introduce GTPO (Group-relative Trajectory-based Policy Optimization), which identifies conflict tokens, tokens appearing in the same position across completions with opposite rewards, protects them by skipping negative updates, while amplifying positive ones. To further prevent policy collapse, GTPO filters out completions whose entropy exceeds a provable threshold. Unlike GRPO, GTPO does not rely on KL-divergence regularization, eliminating the need for a reference model during training, while still ensuring greater training stability and improved performance, validated through multiple experiments on GSM8K, MATH and AIME 2024 benchmarks.