Neuron-level Balance between Stability and Plasticity in Deep Reinforcement Learning

In contrast to the human ability to continuously acquire knowledge, agents struggle with the stability-plasticity dilemma in deep reinforcement learning (DRL), which refers to the trade-off between retaining existing skills (stability) and learning new knowledge (plasticity). Current methods focus on balancing these two aspects at the network level, lacking sufficient differentiation and fine-grained control of individual neurons. To overcome this limitation, we propose Neuron-level Balance between Stability and Plasticity (NBSP) method, by taking inspiration from the observation that specific neurons are strongly relevant to task-relevant skills. Specifically, NBSP first (1) defines and identifies RL skill neurons that are crucial for knowledge retention through a goal-oriented method, and then (2) introduces a framework by employing gradient masking and experience replay techniques targeting these neurons to preserve the encoded existing skills while enabling adaptation to new tasks. Numerous experimental results on the Meta-World and Atari benchmarks demonstrate that NBSP significantly outperforms existing approaches in balancing stability and plasticity.

FADConv: A Frequency-Aware Dynamic Convolution for Farmland Non-agriculturalization Identification and Segmentation

Cropland non-agriculturalization refers to the conversion of arable land into non-agricultural uses such as forests, residential areas, and construction sites. This phenomenon not only directly leads to the loss of cropland resources but also poses systemic threats to food security and agricultural sustainability. Accurate identification of cropland and non-cropland areas is crucial for detecting and addressing this issue. Traditional CNNs employ static convolution layers, while dynamic convolution studies demonstrate that adaptively weighting multiple convolutional kernels through attention mechanisms can enhance accuracy. However, existing dynamic convolution methods relying on Global Average Pooling (GAP) for attention weight allocation suffer from information loss, limiting segmentation precision. This paper proposes Frequency-Aware Dynamic Convolution (FADConv) and a Frequency Attention (FAT) module to address these limitations. Building upon the foundational structure of dynamic convolution, we designed FADConv by integrating 2D Discrete Cosine Transform (2D DCT) to capture frequency domain features and fuse them. FAT module generates high-quality attention weights that replace the traditional GAP method,making the combination between dynamic convolution kernels more reasonable.Experiments on the GID and Hi-CNA datasets demonstrate that FADConv significantly improves segmentation accuracy with minimal computational overhead. For instance, ResNet18 with FADConv achieves 1.9% and 2.7% increases in F1-score and IoU for cropland segmentation on GID, with only 58.87M additional MAdds. Compared to other dynamic convolution approaches, FADConv exhibits superior performance in cropland segmentation tasks.

Retrieval-Augmented Perception: High-Resolution Image Perception Meets Visual RAG

High-resolution (HR) image perception remains a key challenge in multimodal large language models (MLLMs). To overcome the limitations of existing methods, this paper shifts away from prior dedicated heuristic approaches and revisits the most fundamental idea to HR perception by enhancing the long-context capability of MLLMs, driven by recent advances in long-context techniques like retrieval-augmented generation (RAG) for general LLMs. Towards this end, this paper presents the first study exploring the use of RAG to address HR perception challenges. Specifically, we propose Retrieval-Augmented Perception (RAP), a training-free framework that retrieves and fuses relevant image crops while preserving spatial context using the proposed Spatial-Awareness Layout. To accommodate different tasks, the proposed Retrieved-Exploration Search (RE-Search) dynamically selects the optimal number of crops based on model confidence and retrieval scores. Experimental results on HR benchmarks demonstrate the significant effectiveness of RAP, with LLaVA-v1.5-13B achieving a 43% improvement on $V^*$ Bench and 19% on HR-Bench.

Stable-SPAM: How to Train in 4-Bit More Stably than 16-Bit Adam

This paper comprehensively evaluates several recently proposed optimizers for 4-bit training, revealing that low-bit precision amplifies sensitivity to learning rates and often causes unstable gradient norms, leading to divergence at higher learning rates. Among these, SPAM, a recent optimizer featuring momentum reset and spike-aware gradient clipping, achieves the best performance across various bit levels, but struggles to stabilize gradient norms, requiring careful learning rate tuning. To address these limitations, we propose Stable-SPAM, which incorporates enhanced gradient normalization and clipping techniques. In particular, Stable-SPAM (1) adaptively updates the clipping threshold for spiked gradients by tracking their historical maxima; (2) normalizes the entire gradient matrix based on its historical $l_2$-norm statistics; and $(3)$ inherits momentum reset from SPAM to periodically reset the first and second moments of Adam, mitigating the accumulation of spiked gradients. Extensive experiments show that Stable-SPAM effectively stabilizes gradient norms in 4-bit LLM training, delivering superior performance compared to Adam and SPAM. Notably, our 4-bit LLaMA-1B model trained with Stable-SPAM outperforms the BF16 LLaMA-1B trained with Adam by up to $2$ perplexity. Furthermore, when both models are trained in 4-bit, Stable-SPAM achieves the same loss as Adam while requiring only about half the training steps. Code is available at https://github.com/TianjinYellow/StableSPAM.git.

Parameter Efficient Merging for Multimodal Large Language Models with Complementary Parameter Adaptation

Fine-tuning pre-trained models with custom data leads to numerous expert models on specific tasks. Merging models into one universal model to empower multi-task ability refraining from data leakage has gained popularity. With the expansion in data and model size, parameter efficient tuning becomes the common practice for obtaining task-specific models efficiently. However, we observe that existing methods designed for full fine-tuning merging fail under efficient tuning. To address the issues, we analyze from low-rank decomposition and reveal that maintaining direction and compensating for gap between singular values are crucial for efficient model merging. Consequently, we propose CoPA-Merging, a training-free parameter efficient merging method with complementary parameter adaptation. Specifically, we (1) prune parameters and construct scaling coefficients from inter-parameter relation to compensate for performance drop from task interference and (2) perform cross-task normalization to enhance unseen task generalization. We establish a benchmark consisting of diverse multimodal tasks, on which we conduct experiments to certificate the outstanding performance and generalizability of our method. Additional study and extensive analyses further showcase the effectiveness.

A Survey on Mechanistic Interpretability for Multi-Modal Foundation Models

The rise of foundation models has transformed machine learning research, prompting efforts to uncover their inner workings and develop more efficient and reliable applications for better control. While significant progress has been made in interpreting Large Language Models (LLMs), multimodal foundation models (MMFMs) - such as contrastive vision-language models, generative vision-language models, and text-to-image models - pose unique interpretability challenges beyond unimodal frameworks. Despite initial studies, a substantial gap remains between the interpretability of LLMs and MMFMs. This survey explores two key aspects: (1) the adaptation of LLM interpretability methods to multimodal models and (2) understanding the mechanistic differences between unimodal language models and crossmodal systems. By systematically reviewing current MMFM analysis techniques, we propose a structured taxonomy of interpretability methods, compare insights across unimodal and multimodal architectures, and highlight critical research gaps.

PEARL: Towards Permutation-Resilient LLMs

The in-context learning (ICL) capability of large language models (LLMs) enables them to perform challenging tasks using provided demonstrations. However, ICL is highly sensitive to the ordering of demonstrations, leading to instability in predictions. This paper shows that this vulnerability can be exploited to design a natural attack - difficult for model providers to detect - that achieves nearly 80% success rate on LLaMA-3 by simply permuting the demonstrations. Existing mitigation methods primarily rely on post-processing and fail to enhance the model's inherent robustness to input permutations, raising concerns about safety and reliability of LLMs. To address this issue, we propose Permutation-resilient learning (PEARL), a novel framework based on distributionally robust optimization (DRO), which optimizes model performance against the worst-case input permutation. Specifically, PEARL consists of a permutation-proposal network (P-Net) and the LLM. The P-Net generates the most challenging permutations by treating it as an optimal transport problem, which is solved using an entropy-constrained Sinkhorn algorithm. Through minimax optimization, the P-Net and the LLM iteratively optimize against each other, progressively improving the LLM's robustness. Experiments on synthetic pre-training and real-world instruction tuning tasks demonstrate that PEARL effectively mitigates permutation attacks and enhances performance. Notably, despite being trained on fewer shots and shorter contexts, PEARL achieves performance gains of up to 40% when scaled to many-shot and long-context scenarios, highlighting its efficiency and generalization capabilities.

On Theoretical Limits of Learning with Label Differential Privacy

Label differential privacy (DP) is designed for learning problems involving private labels and public features. While various methods have been proposed for learning under label DP, the theoretical limits remain largely unexplored. In this paper, we investigate the fundamental limits of learning with label DP in both local and central models for both classification and regression tasks, characterized by minimax convergence rates. We establish lower bounds by converting each task into a multiple hypothesis testing problem and bounding the test error. Additionally, we develop algorithms that yield matching upper bounds. Our results demonstrate that under label local DP (LDP), the risk has a significantly faster convergence rate than that under full LDP, i.e. protecting both features and labels, indicating the advantages of relaxing the DP definition to focus solely on labels. In contrast, under the label central DP (CDP), the risk is only reduced by a constant factor compared to full DP, indicating that the relaxation of CDP only has limited benefits on the performance.

Edit Once, Update Everywhere: A Simple Framework for Cross-Lingual Knowledge Synchronization in LLMs

Knowledge editing allows for efficient adaptation of large language models (LLMs) to new information or corrections without requiring full retraining. However, prior methods typically focus on either single-language editing or basic multilingual editing, failing to achieve true cross-linguistic knowledge synchronization. To address this, we present a simple and practical state-of-the-art (SOTA) recipe Cross-Lingual Knowledge Democracy Edit (X-KDE), designed to propagate knowledge from a dominant language to other languages effectively. Our X-KDE comprises two stages: (i) Cross-lingual Edition Instruction Tuning (XE-IT), which fine-tunes the model on a curated parallel dataset to modify in-scope knowledge while preserving unrelated information, and (ii) Target-language Preference Optimization (TL-PO), which applies advanced optimization techniques to ensure consistency across languages, fostering the transfer of updates. Additionally, we contribute a high-quality, cross-lingual dataset, specifically designed to enhance knowledge transfer across languages. Extensive experiments on the Bi-ZsRE and MzsRE benchmarks show that X-KDE significantly enhances cross-lingual performance, achieving an average improvement of +8.19%, while maintaining high accuracy in monolingual settings.

Zero Token-Driven Deep Thinking in LLMs: Unlocking the Full Potential of Existing Parameters via Cyclic Refinement

Resource limitations often constrain the parameter counts of Large Language Models (LLMs), hindering their performance. While existing methods employ parameter sharing to reuse the same parameter set under fixed budgets, such approaches typically force each layer to assume multiple roles with a predetermined number of iterations, restricting efficiency and adaptability. In this work, we propose the Zero Token Transformer (ZTT), which features a head-tail decoupled parameter cycling method. We disentangle the first (head) and last (tail) layers from parameter cycling and iteratively refine only the intermediate layers. Furthermore, we introduce a Zero-Token Mechanism, an internal architectural component rather than an input token, to guide layer-specific computation. At each cycle, the model retrieves a zero token (with trainable key values) from a Zero-Token Pool, integrating it alongside regular tokens in the attention mechanism. The corresponding attention scores not only reflect each layer's computational importance but also enable dynamic early exits without sacrificing overall model accuracy. Our approach achieves superior performance under tight parameter budgets, effectively reduces computational overhead via early exits, and can be readily applied to fine-tune existing pre-trained models for enhanced efficiency and adaptability.