Fundus2Globe: Generative AI-Driven 3D Digital Twins for Personalized Myopia Management

Myopia, projected to affect 50% population globally by 2050, is a leading cause of vision loss. Eyes with pathological myopia exhibit distinctive shape distributions, which are closely linked to the progression of vision-threatening complications. Recent understanding of eye-shape-based biomarkers requires magnetic resonance imaging (MRI), however, it is costly and unrealistic in routine ophthalmology clinics. We present Fundus2Globe, the first AI framework that synthesizes patient-specific 3D eye globes from ubiquitous 2D color fundus photographs (CFPs) and routine metadata (axial length, spherical equivalent), bypassing MRI dependency. By integrating a 3D morphable eye model (encoding biomechanical shape priors) with a latent diffusion model, our approach achieves submillimeter accuracy in reconstructing posterior ocular anatomy efficiently. Fundus2Globe uniquely quantifies how vision-threatening lesions (e.g., staphylomas) in CFPs correlate with MRI-validated 3D shape abnormalities, enabling clinicians to simulate posterior segment changes in response to refractive shifts. External validation demonstrates its robust generation performance, ensuring fairness across underrepresented groups. By transforming 2D fundus imaging into 3D digital replicas of ocular structures, Fundus2Globe is a gateway for precision ophthalmology, laying the foundation for AI-driven, personalized myopia management.

MATH-Perturb: Benchmarking LLMs' Math Reasoning Abilities against Hard Perturbations

Large language models have demonstrated impressive performance on challenging mathematical reasoning tasks, which has triggered the discussion of whether the performance is achieved by true reasoning capability or memorization. To investigate this question, prior work has constructed mathematical benchmarks when questions undergo simple perturbations -- modifications that still preserve the underlying reasoning patterns of the solutions. However, no work has explored hard perturbations, which fundamentally change the nature of the problem so that the original solution steps do not apply. To bridge the gap, we construct MATH-P-Simple and MATH-P-Hard via simple perturbation and hard perturbation, respectively. Each consists of 279 perturbed math problems derived from level-5 (hardest) problems in the MATH dataset (Hendrycksmath et. al., 2021). We observe significant performance drops on MATH-P-Hard across various models, including o1-mini (-16.49%) and gemini-2.0-flash-thinking (-12.9%). We also raise concerns about a novel form of memorization where models blindly apply learned problem-solving skills without assessing their applicability to modified contexts. This issue is amplified when using original problems for in-context learning. We call for research efforts to address this challenge, which is critical for developing more robust and reliable reasoning models.

Self-Regulation and Requesting Interventions

Human intelligence involves metacognitive abilities like self-regulation, recognizing limitations, and seeking assistance only when needed. While LLM Agents excel in many domains, they often lack this awareness. Overconfident agents risk catastrophic failures, while those that seek help excessively hinder efficiency. A key challenge is enabling agents with a limited intervention budget $C$ is to decide when to request assistance. In this paper, we propose an offline framework that trains a "helper" policy to request interventions, such as more powerful models or test-time compute, by combining LLM-based process reward models (PRMs) with tabular reinforcement learning. Using state transitions collected offline, we score optimal intervention timing with PRMs and train the helper model on these labeled trajectories. This offline approach significantly reduces costly intervention calls during training. Furthermore, the integration of PRMs with tabular RL enhances robustness to off-policy data while avoiding the inefficiencies of deep RL. We empirically find that our method delivers optimal helper behavior.

Phi-4 Technical Report

We present phi-4, a 14-billion parameter language model developed with a training recipe that is centrally focused on data quality. Unlike most language models, where pre-training is based primarily on organic data sources such as web content or code, phi-4 strategically incorporates synthetic data throughout the training process. While previous models in the Phi family largely distill the capabilities of a teacher model (specifically GPT-4), phi-4 substantially surpasses its teacher model on STEM-focused QA capabilities, giving evidence that our data-generation and post-training techniques go beyond distillation. Despite minimal changes to the phi-3 architecture, phi-4 achieves strong performance relative to its size -- especially on reasoning-focused benchmarks -- due to improved data, training curriculum, and innovations in the post-training scheme.

NODE-AdvGAN: Improving the transferability and perceptual similarity of adversarial examples by dynamic-system-driven adversarial generative model

Understanding adversarial examples is crucial for improving the model's robustness, as they introduce imperceptible perturbations that deceive models. Effective adversarial examples, therefore, offer the potential to train more robust models by removing their singularities. We propose NODE-AdvGAN, a novel approach that treats adversarial generation as a continuous process and employs a Neural Ordinary Differential Equation (NODE) for simulating the dynamics of the generator. By mimicking the iterative nature of traditional gradient-based methods, NODE-AdvGAN generates smoother and more precise perturbations that preserve high perceptual similarity when added to benign images. We also propose a new training strategy, NODE-AdvGAN-T, which enhances transferability in black-box attacks by effectively tuning noise parameters during training. Experiments demonstrate that NODE-AdvGAN and NODE-AdvGAN-T generate more effective adversarial examples that achieve higher attack success rates while preserving better perceptual quality than traditional GAN-based methods.

BackdoorMBTI: A Backdoor Learning Multimodal Benchmark Tool Kit for Backdoor Defense Evaluation

We introduce BackdoorMBTI, the first backdoor learning toolkit and benchmark designed for multimodal evaluation across three representative modalities from eleven commonly used datasets. BackdoorMBTI provides a systematic backdoor learning pipeline, encompassing data processing, data poisoning, backdoor training, and evaluation. The generated poison datasets and backdoor models enable detailed evaluation of backdoor defense methods. Given the diversity of modalities, BackdoorMBTI facilitates systematic evaluation across different data types. Furthermore, BackdoorMBTI offers a standardized approach to handling practical factors in backdoor learning, such as issues related to data quality and erroneous labels. We anticipate that BackdoorMBTI will expedite future research in backdoor defense methods within a multimodal context. Code is available at https://anonymous.4open.science/r/BackdoorMBTI-D6A1/README.md.

Hybrid Precoding with Per-Beam Timing Advance for Asynchronous Cell-free mmWave Massive MIMO-OFDM Systems

Cell-free massive multiple-input-multiple-output (CF-mMIMO) is regarded as one of the promising technologies for next-generation wireless networks. However, due to its distributed architecture, geographically separated access points (APs) jointly serve a large number of user-equipments (UEs), there will inevitably be a discrepancies in the arrival time of transmitted signals. In this paper, we investigate millimeter-wave (mmWave) CF-mMIMO orthogonal frequency division multiplexing (OFDM) systems with asynchronous reception in a wide area coverage scenario, where asynchronous timing offsets may extend far beyond the cyclic prefix (CP) range. A comprehensive asynchronous beam-domain signal transmission model is presented for mmWave CF-mMIMO-OFDM systems in both downlink and uplink, incorporating phase offset, inter-carrier interference (ICI) and inter-symbol interference (ISI). To address the issue of asynchronous reception, we propose a novel per-beam timing advance (PBTA) hybrid precoding architecture and analyze the spectral efficiency (SE) in the beam domain for downlink and uplink asynchronous receptions. Both scalable centralized and distributed implementations are taken into account, and the asynchronous delay phase is utilized to design precoding/combining vectors. Furthermore, we formulate the sum rate maximization problem and develop two low-complexity joint beam selection and UE association algorithms considering the impact of asynchronous timing offset exceeding the CP range. Simulation results demonstrate that the performance will be severely limited by ICI and ISI, and our proposed PBTA hybrid precoding architecture effectively mitigates asynchronous interference compared to the nearest AAU/UE-based timing-advance scheme. Additionally, numerical results show that our proposed low-complexity joint beam selection and UE association algorithms achieve superior SE performance.

Delving into the Reversal Curse: How Far Can Large Language Models Generalize?

While large language models (LLMs) showcase unprecedented capabilities, they also exhibit certain inherent limitations when facing seemingly trivial tasks. A prime example is the recently debated "reversal curse", which surfaces when models, having been trained on the fact "A is B", struggle to generalize this knowledge to infer that "B is A". In this paper, we examine the manifestation of the reversal curse across various tasks and delve into both the generalization abilities and the problem-solving mechanisms of LLMs. This investigation leads to a series of significant insights: (1) LLMs are able to generalize to "B is A" when both A and B are presented in the context as in the case of a multiple-choice question. (2) This generalization ability is highly correlated to the structure of the fact "A is B" in the training documents. For example, this generalization only applies to biographies structured in "[Name] is [Description]" but not to "[Description] is [Name]". (3) We propose and verify the hypothesis that LLMs possess an inherent bias in fact recalling during knowledge application, which explains and underscores the importance of the document structure to successful learning. (4) The negative impact of this bias on the downstream performance of LLMs can hardly be mitigated through training alone. Based on these intriguing findings, our work not only presents a novel perspective for interpreting LLMs' generalization abilities from their intrinsic working mechanism but also provides new insights for the development of more effective learning methods for LLMs.

TreeBoN: Enhancing Inference-Time Alignment with Speculative Tree-Search and Best-of-N Sampling

Inference-time alignment enhances the performance of large language models without requiring additional training or fine-tuning but presents challenges due to balancing computational efficiency with high-quality output. Best-of-N (BoN) sampling, as a simple yet powerful approach, generates multiple responses and selects the best one, achieving improved performance but with a high computational cost. We propose TreeBoN, a novel framework that integrates a speculative tree-search strategy into Best-of-N (BoN) Sampling. TreeBoN maintains a set of parent nodes, iteratively branching and pruning low-quality responses, thereby reducing computational overhead while maintaining high output quality. Our approach also leverages token-level rewards from Direct Preference Optimization (DPO) to guide tree expansion and prune low-quality paths. We evaluate TreeBoN using AlpacaFarm, UltraFeedback, GSM8K, HH-RLHF, and TutorEval datasets, demonstrating consistent improvements. Specifically, TreeBoN achieves a 65% win rate at maximum lengths of 192 and 384 tokens, outperforming standard BoN with the same computational cost. Furthermore, TreeBoN achieves around a 60% win rate across longer responses, showcasing its scalability and alignment efficacy.

A Common Pitfall of Margin-based Language Model Alignment: Gradient Entanglement

Reinforcement Learning from Human Feedback (RLHF) has become the predominant approach for language model (LM) alignment. At its core, RLHF uses a margin-based loss for preference optimization, specifying ideal LM behavior only by the difference between preferred and dispreferred responses. In this paper, we identify a common pitfall of margin-based methods -- the under-specification of ideal LM behavior on preferred and dispreferred responses individually, which leads to two unintended consequences as the margin increases: (1) The probability of dispreferred (e.g., unsafe) responses may increase, resulting in potential safety alignment failures. (2) The probability of preferred responses may decrease, even when those responses are ideal. We demystify the reasons behind these problematic behaviors: margin-based losses couple the change in the preferred probability to the gradient of the dispreferred one, and vice versa, often preventing the preferred probability from increasing while the dispreferred one decreases, and thus causing a synchronized increase or decrease in both probabilities. We term this effect, inherent in margin-based objectives, gradient entanglement. Formally, we derive conditions for general margin-based alignment objectives under which gradient entanglement becomes concerning: the inner product of the gradients of preferred and dispreferred log-probabilities is large relative to the individual gradient norms. We theoretically investigate why such inner products can be large when aligning language models and empirically validate our findings. Empirical implications of our framework extend to explaining important differences in the training dynamics of various preference optimization algorithms, and suggesting potential algorithm designs to mitigate the under-specification issue of margin-based methods and thereby improving language model alignment.