Yimao Cai

Yanxing Guo, Zihao Zheng, Fangzhou Wu et al.

Tackling complex reasoning tasks typically relies on massive monolithic LLMs, which suffer from severe computational redundancy. While task decomposition through structured pipelines or multi-agent collaborations offers an alternative, these approaches inevitably fall into a critical dilemma: predefined static topologies are highly vulnerable to cascading errors, whereas unconstrained dynamic agents suffer from trajectory divergence and unpredictable memory bloat. To address this, we present DynaGraph, a lightweight multi-model framework driven by dynamic topological reconfiguration. At the execution level, DynaGraph multiplexes time-division PEFT adapters over a shared base model, enabling both full system training and inference deployment on a single consumer-grade GPU. At the routing level, the Evaluator continuously monitors execution confidence to trigger hierarchical self-healing: Fine-grained Patching for localized data gaps and Subgraph Reconstruction for severe logical ruptures. Experiments on StrategyQA, MATH, and FinQA demonstrate our 8B model closely approximates the reasoning capabilities of a 72B monolithic model (e.g., 87.6% on StrategyQA, 82.7% on MATH). Furthermore, it reduces latency by up to 68.1% and token consumption by 68.6% compared to unconstrained dynamic architectures.

Renye Yan, Yaozhong Gan, You Wu et al.

In sparse reward scenarios of reinforcement learning (RL), the memory mechanism provides promising shortcuts to policy optimization by reflecting on past experiences like humans. However, current memory-based RL methods simply store and reuse high-value policies, lacking a deeper refining and filtering of diverse past experiences and hence limiting the capability of memory. In this paper, we propose AdaMemento, an adaptive memory-enhanced RL framework. Instead of just memorizing positive past experiences, we design a memory-reflection module that exploits both positive and negative experiences by learning to predict known local optimal policies based on real-time states. To effectively gather informative trajectories for the memory, we further introduce a fine-grained intrinsic motivation paradigm, where nuances in similar states can be precisely distinguished to guide exploration. The exploitation of past experiences and exploration of new policies are then adaptively coordinated by ensemble learning to approach the global optimum. Furthermore, we theoretically prove the superiority of our new intrinsic motivation and ensemble mechanism. From 59 quantitative and visualization experiments, we confirm that AdaMemento can distinguish subtle states for better exploration and effectively exploiting past experiences in memory, achieving significant improvement over previous methods.

Renye Yan, Yaozhong Gan, You Wu et al.

The imbalance of exploration and exploitation has long been a significant challenge in reinforcement learning. In policy optimization, excessive reliance on exploration reduces learning efficiency, while over-dependence on exploitation might trap agents in local optima. This paper revisits the exploration-exploitation dilemma from the perspective of entropy by revealing the relationship between entropy and the dynamic adaptive process of exploration and exploitation. Based on this theoretical insight, we establish an end-to-end adaptive framework called AdaZero, which automatically determines whether to explore or to exploit as well as their balance of strength. Experiments show that AdaZero significantly outperforms baseline models across various Atari and MuJoCo environments with only a single setting. Especially in the challenging environment of Montezuma, AdaZero boosts the final returns by up to fifteen times. Moreover, we conduct a series of visualization analyses to reveal the dynamics of our self-adaptive mechanism, demonstrating how entropy reflects and changes with respect to the agent's performance and adaptive process.

Weikai Xu, Meng Li, Shuzhang Zhong et al.

The Mixture-of-Experts (MoE) models have emerged as the state-of-the-art paradigm for scaling up large language models (LLMs) without proportionally increased computational cost. However, its on-device deployment faces a critical challenge due to the large memory requirement for storing all expert parameters. 3D NAND-based computing-in-memory (CIM) architectures uniquely offer high storage capacity and reduced data movement, while they are ill-suited for MoE models with dynamically sparse expert activation, leading to a degradation of effective computational parallelism, along with underutilization of multibit storage capability of Flash cells. In this work, we proposed a 3D NAND-based content addressable-selected CIM architecture, dubbed as NASiC, which is tailored to MoE models. By leveraging the intrinsic string structure of 3D NAND technology, NASiC fuses the dynamical expert selection through CAM-based masking mechanism and activated expert computation through CIM into a single computation cycle, eradicating redundant computation and enhancing computational parallelism. Moreover, circuit-level optimizations and multibit CIM cell are co-designed with proposed NASiC architecture, featuring block-wise parallel computation with in-situ signed multibit input and weight expansion, substantially improving the throughput and energy-efficiency of NAND CIM array, as well as the utilization of high-density 3D NAND technology for MoE models. With extensive experimental results, we demonstrate NASiC achieves 4-114.8x improved performance and 3.9-70x improved energy efficiency over state-of-the-art designs, along with high accuracy, showing its great potential for efficient on-device MoE LLM inference.

Renye Yan, Jikang Cheng, Shikun Sun et al.

Despite strong image-generation performance, diffusion models' reconstruction objectives limit alignment with human preferences. RL enables such alignment through explicit rewards. However, most studies apply RL to the full denoising trajectory, making it computationally costly and weakening preference alignment, i.e., doing more but achieving less. We observe that the impact of RL fine-tuning varies significantly across denoising stages. In the early stage, image structures are unstable and distant from the final reward signal. Applying RL at this stage leads to delayed rewards and action-reward mismatching, resulting in high variance and inefficient updates. Conversely, in the later stage, reward gains saturate, and continued training tends to overfit local details, intensifying reward hacking. To tackle these challenges, we propose AdaScope, an RL-enhanced plug-in that improves generation quality while reducing computational cost. Specifically, AdaScope adaptively identifies the optimal intervention timing for RL by perceiving the structural evolution and semantic consistency during denoising, and dynamically terminates training once the denoising converges and reward gains saturate. As a result, it achieves a rare 'dual benefit': a reduction in computational costs alongside a significant performance improvement. We offer theoretical grounds for the design of AdaScope. Compared with state-of-the-art methods, AdaScope improves performance by 66% while cutting computational cost by 59%.