The honesty of Large Language Models (LLMs) is increasingly important for safe deployment in high-stakes domains. However, this crucial trait is severely undermined by supervised fine-tuning (SFT), a common technique for model specialization. Existing recovery methods rely on data-intensive global parameter adjustments, implicitly assuming that SFT deeply corrupts the models' ability to recognize their knowledge boundaries. However, we observe that fine-tuned LLMs still preserve this ability; what is damaged is their capacity to faithfully express that awareness. Building on this, we propose Honesty-Critical Neurons Restoration (HCNR) to surgically repair this suppressed capacity. HCNR identifies and restores key expression-governing neurons to their pre-trained state while harmonizing them with task-oriented neurons via Hessian-guided compensation. Experiments on four QA tasks and five LLM families demonstrate that HCNR effectively recovers 33.25% of the compromised honesty while achieving at least 2.23x speedup with over 10x less data compared to baseline methods, offering a practical solution for trustworthy LLM deployment.

Retrieval-Augmented Generation (RAG) is a critical paradigm for building reliable, knowledge-intensive Large Language Model (LLM) applications. However, the multi-stage pipeline (retrieve, generate) and unique workload characteristics (e.g., knowledge dependency) of RAG systems pose significant challenges for serving performance optimization. Existing generic LLM inference traces fail to capture these RAG-specific dynamics, creating a significant performance gap between academic research and real-world deployment. To bridge this gap, this paper introduces RAGPulse, an open-source RAG workload trace dataset. This dataset was collected from an university-wide Q&A system serving that has served more than 40,000 students and faculties since April 2024. We detail RAGPulse's system architecture, its privacy-preserving hash-based data format, and provide an in-depth statistical analysis. Our analysis reveals that real-world RAG workloads exhibit significant temporal locality and a highly skewed hot document access pattern. RAGPulse provides a high-fidelity foundation for researchers to develop and validate novel optimization strategies for RAG systems, such as content-aware batching and retrieval caching, ultimately enhancing the efficiency and reliability of RAG services. The code is available at https://github.com/flashserve/RAGPulse.

The Instance Image Goal Navigation (IIN) problem requires mobile robots deployed i n unknown environments to search for specific objects or people of interest using only a single reference goal image of the target . This problem can be especially challenging when: 1) the reference image is captured from an arbitrary viewpoint, and 2) the robot must operate with sparse - view scene reconstructions. I n this paper, we address the IIN problem, by introduc ing SplatSearch, a novel architecture that leverages sparse - view 3D Gaussian Splatting (3DGS) reconstructions. SplatSearch renders multiple viewpoints around candidate objects using a sparse online 3DGS map, and uses a multi - view diffusion model to complete missing regions of the rendered image s, enabling robust feature matching against the goal image . A novel frontier exploration policy is introduced which uses visual context from the synthesized viewpoints with semantic context from the goal image to evaluate frontier locations, allowing the robot to prioritize frontiers that are semantically and visually relevant to the goal image . An ablation study confirms the design choices of SplatSearch.

Semi-supervised learning (SSL) has demonstrated high performance in image classification tasks by effectively utilizing both labeled and unlabeled data. However, existing SSL methods often suffer from poor calibration, with models yielding overconfident predictions that misrepresent actual prediction likelihoods. Recently, neural networks trained with {\tt mixup} that linearly interpolates random examples from the training set have shown better calibration in supervised settings. However, calibration of neural models remains under-explored in semi-supervised settings. Although effective in supervised model calibration, random mixup of pseudolabels in SSL presents challenges due to the overconfidence and unreliability of pseudolabels. In this work, we introduce CalibrateMix, a targeted mixup-based approach that aims to improve the calibration of SSL models while maintaining or even improving their classification accuracy. Our method leverages training dynamics of labeled and unlabeled samples to identify ``easy-to-learn'' and ``hard-to-learn'' samples, which in turn are utilized in a targeted mixup of easy and hard samples. Experimental results across several benchmark image datasets show that our method achieves lower expected calibration error (ECE) and superior accuracy compared to existing SSL approaches.

Precise and real-time detection of gastrointestinal polyps during endoscopic procedures is crucial for early diagnosis and prevention of colorectal cancer. This work presents En-doSight AI, a deep learning architecture developed and evaluated independently to enable accurate polyp localization and detailed boundary delineation. Leveraging the publicly available Hyper-Kvasir dataset, the system achieves a mean A verage Precision (mAP) of 88.3% for polyp detection and a Dice coefficient of up to 69% for segmentation, alongside real-time inference speeds exceeding 35 frames per second on GPU hardware. The training incorporates clinically relevant performance metrics and a novel thermal-aware procedure to ensure model robustness and efficiency. This integrated AI solution is designed for seamless deployment in endoscopy workflows, promising to advance diagnostic accuracy and clinical decision-making in gastrointestinal healthcare.

Large language models (LLMs) deployed in user-facing applications require long-horizon consistency: the ability to remember prior interactions, respect user preferences, and ground reasoning in past events. However, contemporary memory systems often adopt complex architectures such as knowledge graphs, multi-stage retrieval pipelines, and OS-style schedulers, which introduce engineering complexity and reproducibility challenges. We present ENGRAM, a lightweight memory system that organizes conversation into three canonical memory types (episodic, semantic, and procedural) through a single router and retriever. Each user turn is converted into typed memory records with normalized schemas and embeddings and stored in a database. At query time, the system retrieves top-k dense neighbors for each type, merges results with simple set operations, and provides the most relevant evidence as context to the model. ENGRAM attains state-of-the-art results on LoCoMo, a multi-session conversational QA benchmark for long-horizon memory, and exceeds the full-context baseline by 15 points on LongMemEval while using only about 1% of the tokens. These results show that careful memory typing and straightforward dense retrieval can enable effective long-term memory management in language models without requiring complex architectures.

Multivariate time series classification is increasingly investigated in space weather research as a means to predict intense solar flare events, which can cause widespread disruptions across modern technological systems. Magnetic field measurements of solar active regions are converted into structured multivariate time series, enabling predictive modeling across segmented observation windows. However, the inherently imbalanced nature of solar flare occurrences, where intense flares are rare compared to minor flare events, presents a significant barrier to effective learning. To address this challenge, we propose a novel Global Cross-Time Attention Fusion (GCTAF) architecture, a transformer-based model to enhance long-range temporal modeling. Unlike traditional self-attention mechanisms that rely solely on local interactions within time series, GCTAF injects a set of learnable cross-attentive global tokens that summarize salient temporal patterns across the entire sequence. These tokens are refined through cross-attention with the input sequence and fused back into the temporal representation, enabling the model to identify globally significant, non-contiguous time points that are critical for flare prediction. This mechanism functions as a dynamic attention-driven temporal summarizer that augments the model's capacity to capture discriminative flare-related dynamics. We evaluate our approach on the benchmark solar flare dataset and show that GCTAF effectively detects intense flares and improves predictive performance, demonstrating that refining transformer-based architectures presents a high-potential alternative for solar flare prediction tasks.

Time series forecasting under distribution shift remains challenging, as existing deep learning models often rely on local statistical normalization (e.g., mean and variance) that fails to capture global distribution shift. Methods like RevIN and its variants attempt to decouple distribution and pattern but still struggle with missing values, noisy observations, and invalid channel-wise affine transformation. To address these limitations, we propose Affine Prototype-Timestamp (APT), a lightweight and flexible plug-in module that injects global distribution features into the normalization-forecasting pipeline. By leveraging timestamp-conditioned prototype learning, APT dynamically generates affine parameters that modulate both input and output series, enabling the backbone to learn from self-supervised, distribution-aware clustered instances. APT is compatible with arbitrary forecasting backbones and normalization strategies while introducing minimal computational overhead. Extensive experiments across six benchmark datasets and multiple backbone-normalization combinations demonstrate that APT significantly improves forecasting performance under distribution shift.

Can multi-modal large language models (MLLMs) truly understand what they can see? Extending Searle's Chinese Room into the multi-modal domain, this paper proposes the Visual Room argument: MLLMs may describe every visual detail precisely yet fail to comprehend the underlying emotions and intentions, namely seeing is not understanding. Building on this, we introduce \textit{Visual Room} 2.0, a hierarchical benchmark for evaluating perception-cognition alignment of MLLMs. We model human perceptive and cognitive processes across three levels: low, middle, and high, covering 17 representative tasks. The perception component ranges from attribute recognition to scene understanding, while the cognition component extends from textual entailment to causal and social reasoning. The dataset contains 350 multi-modal samples, each with six progressive questions (2,100 in total) spanning perception to cognition. Evaluating 10 state-of-the-art (SoTA) MLLMs, we highlight three key findings: (1) MLLMs exhibit stronger perceptual competence than cognitive ability (8.0\%$\uparrow$); (2) cognition appears not causally dependent on perception-based reasoning; and (3) cognition scales with model size, but perception does not consistently improve with larger variants. This work operationalizes Seeing $\ne$ Understanding as a testable hypothesis, offering a new paradigm from perceptual processing to cognitive reasoning in MLLMs. Our dataset is available at https://huggingface.co/datasets/LHK2003/PCBench.

Recommending event schedules is a key issue in Event-based Social Networks (EBSNs) in order to maintain user activity. An effective recommendation is required to maximize the user's preference, subjecting to both time and geographical constraints. Existing methods face an inherent trade-off among efficiency, effectiveness, and generalization, due to the NP-hard nature of the problem. This paper proposes the Chain-of-Scheduling (CoS) framework, which activates the event scheduling capability of Large Language Models (LLMs) through a guided, efficient scheduling process. CoS enhances LLM by formulating the schedule task into three atomic stages, i.e., exploration, verification and integration. Then we enable the LLMs to generate CoS autonomously via Knowledge Distillation (KD). Experimental results show that CoS achieves near-theoretical optimal effectiveness with high efficiency on three real-world datasets in a interpretable manner. Moreover, it demonstrates strong zero-shot learning ability on out-of-domain data.