arXiv subjects

Hoang Phan

Publications and source records attributed to Hoang Phan.

At least 19 recordsLinked to original sources

arXiv2026

Beyond Reasoning Gains: Mitigating General-Capability Forgetting in Large Reasoning Models

Reinforcement learning with verifiable rewards (RLVR) has delivered impressive gains in mathematical and multimodal reasoning and has become a standard post-training paradigm for contemporary language and vision-language models. However, the RLVR recipe introduces a significant risk of capability regression, in which models forget foundational skills after prolonged training without employing regularization strategies. We empirically confirm this concern, observing that open-source reasoning models suffer performance degradation on core capabilities such as perception and faithfulness. While imposing regularization terms like KL divergence can help prevent deviation from the base model, these terms are computed on the current task and therefore do not guarantee preservation of broader knowledge. Meanwhile, commonly used experience replay across heterogeneous domains makes it nontrivial to decide how much training emphasis each objective should receive. To address this, we propose RECAP-a replay strategy with dynamic objective reweighting for general knowledge preservation. Our reweighting mechanism adapts online using short-horizon signals of convergence and instability, shifting the post-training focus away from saturated objectives and toward underperforming or volatile ones. Our method is end-to-end and readily applicable to existing RLVR pipelines without training additional models or heavy tuning. Extensive experiments on benchmarks using Qwen2.5-VL-3B and Qwen2.5-VL-7B demonstrate the effectiveness of our method, which not only preserves general capabilities but also improves reasoning by enabling more flexible trade-offs among in-task rewards.

cs.LG↗

arXiv2026

Route-and-Generate: KL-Regularized Router Improvement for Reasoning Routing in LLMs

Large language models often apply complex reasoning indiscriminately, leading to overthinking on simple tasks and underthinking on harder ones. We present Route-and-Generate (Rouge), a hierarchical training framework that enables LLMs to adaptively decide when and how much reasoning is required for each input. Our approach decouples reasoning effort allocation from answer generation by introducing a routing policy that allocates a fixed sampling budget across multiple reasoning formats, while the base model is optimized with GRPO using per-sample rewards. To promote efficient reasoning, we employ a rank-based reward that favors shorter correct solutions, and update the routing policy via a closed-form KL-regularized policy improvement step distilled into a neural router. This design encourages concise answers on easy problems while reserving longer reasoning for harder ones, achieving improved accuracy-efficiency trade-offs across multiple reasoning benchmarks.

cs.AI↗

arXiv2026

MIMESIS: Learning User Simulators as Training Environments for Interactive Agents

Training and evaluating interactive language agents typically requires rich user interactions, yet collecting human feedback is expensive and difficult to scale. Simulated users offer a scalable alternative, but they must both resemble real user behavior and provide useful learning experiences for agents. In contrast, most agent-training frameworks rely on off-the-shelf assistant LLMs, whose helpfulness can make them overly cooperative, explicit, and behaviorally homogeneous compared with real users. We introduce MIMESIS, a purpose-built user simulator trained on human conversations with explicit reasoning supervision and 13 realistic behavioral patterns derived from real user interactions. Empirically, our 9B model achieves a SOUL-Index of 65.7, surpassing the strongest frontier model. Compared with Claude-Opus-5, the strongest baseline on RealUserSim and SimulatorArena, MIMESIS improves behavioral fidelity by 13.4 points and reduces Turing distance by 3.6 points, respectively. We then freeze the simulator and train an agent by interacting with the frozen simulator using multi-turn reinforcement learning. Across eight environments, training with MIMESIS yields better agent performance than training with GPT-5.5 under all nine unseen user simulators, demonstrating stronger generalization to new user simulators. Moreover, we propose Coached On-Policy Self-Distillation (CSD), which leverages simulator-generated private reasoning traces and subsequent utterances as feedback on how well the agent addresses user needs. A coach converts this information into concise coaching notes that describe how the agent can better anticipate user needs and adapt its behavior over the course of an interaction. CSD turns this feedback into dense, token-level supervision beyond sparse task rewards, yielding further gains across all nine evaluation user models.

cs.AI↗

arXiv2026

Rationale-Guided Policy Optimization: Learning to Reason with Adaptive Rationale Scaffolding

On-policy reinforcement learning has become a central paradigm for improving the reasoning abilities of large language models. However, its effectiveness is often limited by reward sparsity: when a model fails to discover correct trajectories for difficult problems, the optimization process receives little useful signal and may stagnate. Existing approaches mitigate this issue by incorporating off-policy demonstrations, expert traces, or model-generated solutions, but they typically require the auxiliary data to match the format of the reinforcement-learning task, often relying on rejection sampling from stronger models to obtain suitable training trajectories. We introduce Rationale-Guided Policy Optimization (RGPO), a framework that adaptively leverages ground-truth rationale information according to the model's current capability while preserving its freedom to explore. Rather than treating reference solutions as fixed imitation targets, RGPO uses them as temporary scaffolds: rationales help the model generate improved responses, after which only higher-reward, model-generated solutions are transferred back to the original unguided setting. This design allows training to exploit available ground-truth information without requiring off-policy data to follow the same format as the RL task. Across both language-only and vision-language reasoning settings, RGPO consistently improves performance over RLVR baselines, and ablation studies show that adaptive rationale guidance is a key contributor to these gains. These results suggest that RGPO offers a practical and general approach for reducing reward sparsity, stabilizing reinforcement learning, and improving reasoning performance in both text-only and multimodal models.

cs.AI↗

arXiv2026

Sharpening Tax in Post-Training

An emerging hypothesis about reinforcement learning (RL) post-training of large language models (LLMs) is that it merely sharpens existing behaviors of a base model, improving single-shot accuracy at the cost of solution coverage. Although this trade-off has been observed in math and coding tasks, it need not extend to agentic tasks, where multi-turn tool use and interaction may require capabilities newly acquired during post-training. Our surprising finding is that pre-trained LLMs, equipped with a light inference harness, can serve as capable agents. Despite far lower accuracy (pass@1), they often surpass their post-trained counterparts in solution coverage (pass@K) given a sufficient test-time budget. We further analyze the underlying mechanism and show that post-training pushes tasks toward two extremes, always solved or never solved, and thereby improves sampling efficiency and consistency at the cost of solution coverage. To measure this cost, we propose Sharpening Tax, a diagnostic metric that quantifies the loss in test-time scalability after post-training. Across 14 base/post-trained model pairs from four families and three agentic benchmarks (42 cases in total), the tax is prevalent in most settings, can be estimated from a few rollouts, and correlates well with other metrics. Finally, we present posterior-tempered group sampling (PTGS), a simple plug-and-play Bayesian sampler that adapts the sampling temperature per prompt to its estimated difficulty. Applied during RL training in two agentic environments, PTGS pays a smaller tax than the fixed-temperature baseline, solving more tasks under repeated sampling while also improving single-shot accuracy.

cs.AI↗

arXiv2026

Decoding the Critique Mechanism in Large Reasoning Models

Large Reasoning Models (LRMs) exhibit backtracking and self-verification mechanisms that enable them to revise intermediate steps and reach correct solutions, yielding strong performance on complex logical benchmarks. We hypothesize that such behaviors are beneficial only when the model has sufficiently strong ``critique'' ability to detect its own mistakes. This work systematically investigates how current LRMs recover from errors by inserting arithmetic mistakes in their intermediate reasoning steps. Notably, we discover a peculiar yet important phenomenon: despite the error propagating throughout the entire chain-of-thought (CoT) without any verbalized correction, the model still reaches the correct final answer after the thinking process finishes. This recovery implies the existence of an internal mechanism helping the model to detect errors and trigger self-correction, which we refer to as the \textit{hidden critique ability}. Building on feature space analysis, we identify a highly interpretable \textit{critique vector} representing this behavior. Extensive experiments across multiple model scales and families demonstrate that steering latent representations with this vector improves the model's error detection capability and enhances the performance of test-time scaling at no extra training cost. Our findings provide a valuable understanding of LRMs' critique behavior, suggesting a promising direction to control and improve their self-verification mechanism. Our code is available at: https://github.com/mail-research/lrm-critique-vectors.

cs.LG↗

arXiv2026

An Optimal Transport-driven Approach for Cultivating Latent Space in Online Incremental Learning

In online incremental learning, data continuously arrives with substantial distributional shifts, creating a significant challenge because previous samples have limited replay value when learning a new task. Prior research has typically relied on either a single adaptive centroid or multiple fixed centroids to represent each class in the latent space. However, such methods struggle when class data streams are inherently multimodal and require continual centroid updates. To overcome this, we introduce an online Mixture Model learning framework grounded in Optimal Transport theory (MMOT), where centroids evolve incrementally with new data. This approach offers two main advantages: (i) it provides a more precise characterization of complex data streams, and (ii) it enables improved class similarity estimation for unseen samples during inference through MMOT-derived centroids. Furthermore, to strengthen representation learning and mitigate catastrophic forgetting, we design a Dynamic Preservation strategy that regulates the latent space and maintains class separability over time. Experimental evaluations on benchmark datasets confirm the superior effectiveness of our proposed method.

cs.LG↗

arXiv2026

Toward a Holistic Approach to Continual Model Merging

We present a holistic framework for Continual Model Merging (CMM) that intervenes at three critical stages: pre-merging, during merging, and post-merging-to address two fundamental challenges in continual learning. In particular, conventional approaches either maintain a growing list of per-domain task vectors, leading to scalability issues or rely solely on weight-space merging when old data is inaccessible, thereby losing crucial functional information. Our method overcomes these limitations by first fine-tuning the main model within its tangent space on domain-specific data; this linearization amplifies per-task weight disentanglement, effectively mitigating across-task interference. During merging, we leverage functional information from available optimizer states beyond mere parameter averages to avoid the need to revisit old data. Finally, a post-merging correction aligns the representation discrepancy between pre- and post-merged models, reducing bias and enhancing overall performance-all while operating under constant memory constraints without accessing historical data. Extensive experiments on standard class-incremental and domain-incremental benchmarks demonstrate that our approach not only achieves competitive performance but also provides a scalable and efficient solution to the catastrophic forgetting problem.

cs.LG↗

arXiv2026

Hyperparameter Transfer Enables Consistent Gains of Matrix-Preconditioned Optimizers Across Scales

Several recently introduced deep learning optimizers utilizing matrix-level preconditioning have shown promising speedups relative to the current dominant optimizer AdamW, particularly in relatively small-scale experiments. However, efforts to validate and replicate their successes have reported mixed results. To better understand the effectiveness of these optimizers at scale, in this work we investigate how to scale preconditioned optimizers via hyperparameter transfer, building on prior works such as $μ$P. We study how the optimal learning rate and weight decay should scale with model width and depth for a wide range of optimizers, including Shampoo, SOAP, and Muon, accounting for the impact of commonly used techniques such as blocking and grafting. We find that scaling the learning rate according to $μ$P improves transfer, but can still suffer from significant finite-width deviations that cause drifting optimal learning rates, which we show can be mitigated by blocking and explicit spectral normalization. For compute-optimal scaling, we find scaling independent weight decay as $1/\mathrm{width}$ is nearly optimal across optimizers. Applying these scaling rules, we show Muon, SOAP and Shampoo consistently achieve near $1.4\times$ speedup over AdamW for training Llama-architecture language models of sizes ranging from $190$M to $1.4$B, whereas the speedup vanishes rapidly with scale under incorrect scaling. Based on these results and further ablations, we argue that studying optimal hyperparameter transfer is essential for reliably comparing optimizers at scale given a realistic tuning budget.

cs.LG↗

arXiv2025

Unveiling Concept Attribution in Diffusion Models

Diffusion models have shown remarkable abilities in generating realistic and high-quality images from text prompts. However, a trained model remains largely black-box; little do we know about the roles of its components in exhibiting a concept such as objects or styles. Recent works employ causal tracing to localize knowledge-storing layers in generative models without showing how other layers contribute to the target concept. In this work, we approach diffusion models' interpretability problem from a more general perspective and pose a question: \textit{``How do model components work jointly to demonstrate knowledge?''}. To answer this question, we decompose diffusion models using component attribution, systematically unveiling the importance of each component (specifically the model parameter) in generating a concept. The proposed framework, called \textbf{C}omponent \textbf{A}ttribution for \textbf{D}iffusion Model (CAD), discovers the localization of concept-inducing (positive) components, while interestingly uncovers another type of components that contribute negatively to generating a concept, which is missing in the previous knowledge localization work. Based on this holistic understanding of diffusion models, we introduce two fast, inference-time model editing algorithms, CAD-Erase and CAD-Amplify; in particular, CAD-Erase enables erasure and CAD-Amplify allows amplification of a generated concept by ablating the positive and negative components, respectively, while retaining knowledge of other concepts. Extensive experimental results validate the significance of both positive and negative components pinpointed by our framework, demonstrating the potential of providing a complete view of interpreting generative models. Our code is available \href{https://github.com/mail-research/CAD-attribution4diffusion}{here}.

cs.CV↗

arXiv2025

Generalization Bounds for Robust Contrastive Learning: From Theory to Practice

Contrastive Learning first extracts features from unlabeled data, followed by linear probing with labeled data. Adversarial Contrastive Learning (ACL) integrates Adversarial Training into the first phase to enhance feature robustness against attacks in the probing phase. While ACL has shown strong empirical results, its theoretical understanding remains limited. Furthermore, while a fair amount of theoretical works analyze how the unsupervised loss can support the supervised loss in the probing phase, none has examined its role to the robust supervised loss. To fill this gap, our work develops rigorous theories to identify which components in the unsupervised training can help improve the robust supervised loss. Specifically, besides the adversarial contrastive loss, we reveal that the benign one, along with a global divergence between benign and adversarial examples can also improve robustness. Proper experiments are conducted to justify our findings.

cs.LG↗

arXiv2025

Beyond Losses Reweighting: Empowering Multi-Task Learning via the Generalization Perspective

Multi-task learning (MTL) trains deep neural networks to optimize several objectives simultaneously using a shared backbone, which leads to reduced computational costs, improved data efficiency, and enhanced performance through cross-task knowledge sharing. Although recent gradient manipulation techniques aim to find a common descent direction that benefits all tasks, conventional empirical loss minimization still leaves models vulnerable to overfitting and gradient conflicts. To address this, we introduce a novel MTL framework that leverages weight perturbation to regulate gradient norms, thus improving generalization. By adaptively modulating weight perturbations, our approach harmonizes task-specific gradients, reducing conflicts and encouraging more robust learning across tasks. Theoretical insights reveal that controlling the gradient norm through weight perturbation directly contributes to better generalization. Extensive experiments across diverse applications demonstrate that our method significantly outperforms existing gradient-based MTL techniques in terms of task performance and overall model robustness.

cs.LG↗

arXiv2025

Think Twice, Generate Once: Safeguarding by Progressive Self-Reflection

Large language models (LLMs) have revolutionized natural language processing with their ability to generate coherent and contextually relevant text. However, their deployment raises significant concerns about the potential for generating harmful or inappropriate content. In this paper, we introduce Progressive Self-Reflection (PSR), a novel inference-time technique that empowers LLMs to self-monitor and correct their outputs dynamically. Experimental results demonstrate that applying our proposed method to Llama-3.1-8B-Instruct reduces the attack success rate from 77.5\% to 5.9\%, to Llama-3.1-8B base from 89.7\% to 5.6\%, and to Qwen2.5-7B-Instruct from 44.4\% to 3.8\%, without additional training, while maintaining their original performance on benign tasks. Our approach acts as a test-time scaling method, where additional self-reflection rounds enhance safety at the cost of inference overhead. To balance safety with computational efficiency, we introduce a lightweight self-reflection predictor that estimates the optimal number of reflection rounds based on input complexity. This adaptive mechanism prevents unnecessary self-assessment on benign inputs while ensuring thorough evaluation when encountering potentially harmful content. Our findings suggest that Progressive Self-Reflection serves as a scalable test-time approach, enhancing LLM safety by dynamically allocating computational resources in proportion to the input's risk profile.

cs.CL↗

arXiv2025

Preserving Clusters in Prompt Learning for Unsupervised Domain Adaptation

Recent approaches leveraging multi-modal pre-trained models like CLIP for Unsupervised Domain Adaptation (UDA) have shown significant promise in bridging domain gaps and improving generalization by utilizing rich semantic knowledge and robust visual representations learned through extensive pre-training on diverse image-text datasets. While these methods achieve state-of-the-art performance across benchmarks, much of the improvement stems from base pseudo-labels (CLIP zero-shot predictions) and self-training mechanisms. Thus, the training mechanism exhibits a key limitation wherein the visual embedding distribution in target domains can deviate from the visual embedding distribution in the pre-trained model, leading to misguided signals from class descriptions. This work introduces a fresh solution to reinforce these pseudo-labels and facilitate target-prompt learning, by exploiting the geometry of visual and text embeddings - an aspect that is overlooked by existing methods. We first propose to directly leverage the reference predictions (from source prompts) based on the relationship between source and target visual embeddings. We later show that there is a strong clustering behavior observed between visual and text embeddings in pre-trained multi-modal models. Building on optimal transport theory, we transform this insight into a novel strategy to enforce the clustering property in text embeddings, further enhancing the alignment in the target domain. Our experiments and ablation studies validate the effectiveness of the proposed approach, demonstrating superior performance and improved quality of target prompts in terms of representation.

cs.CV↗

arXiv2025

DiMSUM: Diffusion Mamba -- A Scalable and Unified Spatial-Frequency Method for Image Generation

We introduce a novel state-space architecture for diffusion models, effectively harnessing spatial and frequency information to enhance the inductive bias towards local features in input images for image generation tasks. While state-space networks, including Mamba, a revolutionary advancement in recurrent neural networks, typically scan input sequences from left to right, they face difficulties in designing effective scanning strategies, especially in the processing of image data. Our method demonstrates that integrating wavelet transformation into Mamba enhances the local structure awareness of visual inputs and better captures long-range relations of frequencies by disentangling them into wavelet subbands, representing both low- and high-frequency components. These wavelet-based outputs are then processed and seamlessly fused with the original Mamba outputs through a cross-attention fusion layer, combining both spatial and frequency information to optimize the order awareness of state-space models which is essential for the details and overall quality of image generation. Besides, we introduce a globally-shared transformer to supercharge the performance of Mamba, harnessing its exceptional power to capture global relationships. Through extensive experiments on standard benchmarks, our method demonstrates superior results compared to DiT and DIFFUSSM, achieving faster training convergence and delivering high-quality outputs. The codes and pretrained models are released at https://github.com/VinAIResearch/DiMSUM.git.

cs.CV↗

arXiv2025

Enhancing Domain Adaptation through Prompt Gradient Alignment

Prior Unsupervised Domain Adaptation (UDA) methods often aim to train a domain-invariant feature extractor, which may hinder the model from learning sufficiently discriminative features. To tackle this, a line of works based on prompt learning leverages the power of large-scale pre-trained vision-language models to learn both domain-invariant and specific features through a set of domain-agnostic and domain-specific learnable prompts. Those studies typically enforce invariant constraints on representation, output, or prompt space to learn such prompts. In contrast, we cast UDA as a multiple-objective optimization problem in which each objective is represented by a domain loss. Under this new framework, we propose to align per-objective gradients to foster consensus between them. Additionally, to prevent potential overfitting when fine-tuning this deep learning architecture, we penalize the norm of these gradients. To achieve these goals, we devise a practical gradient update procedure that can work under both single-source and multi-source UDA. Empirically, our method consistently outperforms other vision-language model adaptation methods. The implementation is available at https://github.com/VietHoang1512/PGA.

cs.LG↗

arXiv2025

Leveraging Hierarchical Taxonomies in Prompt-based Continual Learning

Humans perceive the world as a series of sequential events, which can be hierarchically organized with different levels of abstraction based on conceptual knowledge. Drawing inspiration from human learning behaviors, this work proposes a novel approach to mitigate catastrophic forgetting in Prompt-based Continual Learning models by exploiting the relationships between continuously emerging class data. We find that applying human habits of organizing and connecting information can serve as an efficient strategy when training deep learning models. Specifically, by building a hierarchical tree structure based on the expanding set of labels, we gain fresh insights into the data, identifying groups of similar classes could easily cause confusion. Additionally, we delve deeper into the hidden connections between classes by exploring the original pretrained model's behavior through an optimal transport-based approach. From these insights, we propose a novel regularization loss function that encourages models to focus more on challenging knowledge areas, thereby enhancing overall performance. Experimentally, our method demonstrated significant superiority over the most robust state-of-the-art models on various benchmarks.

cs.LG↗

arXiv2024

Personalized Large Vision-Language Models

The personalization model has gained significant attention in image generation yet remains underexplored for large vision-language models (LVLMs). Beyond generic ones, with personalization, LVLMs handle interactive dialogues using referential concepts (e.g., ``Mike and Susan are talking.'') instead of the generic form (e.g., ``a boy and a girl are talking.''), making the conversation more customizable and referentially friendly. In addition, PLVM is equipped to continuously add new concepts during a dialogue without incurring additional costs, which significantly enhances the practicality. PLVM proposes Aligner, a pre-trained visual encoder to align referential concepts with the queried images. During the dialogues, it extracts features of reference images with these corresponding concepts and recognizes them in the queried image, enabling personalization. We note that the computational cost and parameter count of the Aligner are negligible within the entire framework. With comprehensive qualitative and quantitative analyses, we reveal the effectiveness and superiority of PLVM.

cs.CV↗