Long-term memory systems increasingly organize interaction histories into hierarchical contexts, where coarse summaries, mid-level events, and fine-grained evidence provide complementary signals for future queries. Yet after retrieval, these signals are often collapsed into a single generation context and used implicitly. We ask how retrieved hierarchical memory should shape generation after access, framing it as a post-retrieval calibration problem. We propose HiR, a training-free inference-time framework that contrasts full-hierarchy decoding with a coarse-only reference view, preserving the complete hierarchy while tempering over-reliance on broad abstractions. Extensive experiments and analyses show that HiR enables more effective and robust use of hierarchical memory without changing memory construction or retrieval; more broadly, our results suggest that memory research should consider not only what to store or retrieve, but also how retrieved information shapes generation.

With the growing demand for processing multiple image sequences in real-world applications, various visual token pruning methods have emerged to mitigate the computational and context length constraints faced by Large Vision Language Models (LVLMs). However, most existing pruning approaches rely on static strategies that struggle to adapt across different architectural LVLMs and multi-image scenarios, and are additionally constrained by their dependence on attention computations that are incompatible with efficient techniques like FlashAttention.To address these limitations, we propose a training-free, Adaptive Visual Token Pruning (AVTP) framework, applicable to diverse LVLM architectures. We strategically determine pruning layers based on empirical analysis of visual attention distributions across various LVLMs, and implement adaptive pruning ratios in multi-image contexts where images of higher importance retain proportionally more tokens. We conduct extensive experiments across multiple LVLMs with different architectures to demonstrate the effectiveness and robustness of AVTP. Specifically, Qwen3VL-8B achieves 2× inference speedup while maintaining 96.1% of its original accuracy on multiple multi-image benchmarks, InternVL3.5-8B retains 94.1% accuracy, and LLaVA-OneVision-7B even exceeds its original baseline performance. Our code is available at https://github.com/zry13/AVTP.

Chain-of-Thought (CoT) reasoning is widely used to improve LLM performance, and recent foundation recommender models adopt it by generating textual reasoning before predicting target items represented by Semantic IDs (SIDs). However, we observe that enabling thinking mode in models such as OpenOneRec can degrade recommendation quality by up to 25%. We investigate this failure and identify Linguistic Inertia: when a textual CoT segment is inserted before SID generation, the model relies more on natural-language context and less on historical SID evidence. Further analyses show that this effect is amplified by reduced access to historical information and longer CoT lengths. To mitigate it, we propose Linguistic-Inertia-Calibrated Decoding (LICD), a training-free framework that combines Reasoning-Chain Compression and Bias-Subtracted Contrastive Inference. Experiments on three large-scale benchmarks show that LICD consistently outperforms both no-thinking and original-thinking baselines. Our code is available at https://github.com/USTC-StarTeam/LICD.

The evolution of Large Language Models (LLMs) is shifting the focus from single, verifiable tasks toward complex, open-ended real-world scenarios, imposing significant challenges on the post-training phase. In these settings, the scale and complexity of reward systems have grown significantly, transitioning toward multi-objective formulations that encompass a comprehensive spectrum of model capabilities and application contexts. However, traditional methods typically rely on fixed reward weights, ignoring non-stationary learning dynamics and struggling with data heterogeneity across dimensions. To address these issues, we propose SPARD, a framework that establishes an automated, self-paced curriculum by perceiving learning progress to dynamically adjust multi-objective reward weights and data importance, thereby synchronizing learning intent with data utility for optimal performance. Extensive experiments across multiple benchmarks demonstrate that SPARD significantly enhances model capabilities across all domains. Our code is publicly available at https://github.com/USTC-StarTeam/SPARD.

Recent advancements in large language models (LLMs) have revolutionized natural language processing through their remarkable capabilities in understanding and executing diverse tasks. While supervised fine-tuning, particularly in Retrieval-Augmented Generation (RAG) scenarios, effectively enhances task-specific performance, it often leads to catastrophic forgetting, where models lose their previously acquired knowledge and general capabilities. Existing solutions either require access to general instruction data or face limitations in preserving the model’s original distribution. To overcome these limitations, we propose SelfAug, a self-distribution alignment method that aligns input sequence logits to preserve the model’s semantic distribution, thereby mitigating catastrophic forgetting and improving downstream performance. Extensive experiments demonstrate that SelfAug achieves a superior balance between downstream learning and general capability retention. Our comprehensive empirical analysis reveals a direct correlation between distribution shifts and the severity of catastrophic forgetting in RAG scenarios, highlighting how the absence of RAG capabilities in general instruction tuning leads to significant distribution shifts during fine-tuning. Our findings not only advance the understanding of catastrophic forgetting in RAG contexts but also provide a practical solution applicable across diverse fine-tuning scenarios.