While Retrieval-Augmented Generation (RAG) has become a standard paradigm for mitigating hallucinations in Large Language Models (LLMs), its effectiveness in complex medical reasoning remains limited. Existing RAG methods suffer from two main challenges: First, Semantic Drift: without explicit domain constraints, LLM-driven query decomposition often deviates from the original clinical intent, introducing substantial noise that degrades retrieval relevance. Second, Concatenation Fallacy: retrieved evidence from different semantic aspects is aggregated in a naive, unstructured manner, without modeling their inter-dependencies and potential conflicts, which ultimately undermines downstream reasoning. To address these challenges, we propose Med-SRAF, a multi-agent retrieval augmentation framework guided by medical domain knowledge. This framework reconstructs the traditional RAG process through two core mechanisms: (1) Intent-driven Semantic Routing, where a UMLS-based NavigationAgent dynamically maps queries to medical dimensions for strategic search space pruning; and (2) Evidence-based Agentic Fusion, where a FusionAgent resolves conflicts among dimension-specific evidence to build logically consistent reasoning chains. Extensive experiments on five widely used medical benchmarks show that Med-SRAF consistently outperforms existing general RAG baselines, achieving an average accuracy improvement of over 4.9%, highlighting its effectiveness in robust and interpretable medical reasoning. Our code is at https://anonymous.4open.science/r/MultiAgent_RAG-F6DC.

Automated homework assessment requires not only recognizing student answers but accurately locating complete answers and intermediate reasoning steps in noisy, multi-page handwritten work. We address the missing evaluation setting of *page-aware, two-level answer-region grounding*: given homework page images, a model must localize complete answer regions and their ordered step-level subregions. We introduce **HG-Bench**, comprising 500 human-annotated K–12 homework samples curated from a 1,489,278-image source pool, with hierarchically linked question- and step-level boxes. Its page-aware protocol separately measures complete-answer localization (ℱ_A) and step-level decomposition (ℱ_S^𝜇), distinguishing structured spatial grounding from visible-text parsing. Across frontier closed-source and open-weight VLMs, no zero-shot system exceeds 55.22% on ℱ_A or 48.22% on ℱ_S^𝜇, whereas a GLM-4.6V 9B reference fine-tuned on approximately 10k in-domain examples reaches 74.97/72.26. These results expose step-level handwritten grounding as a concrete capability gap and provide a reproducible benchmark, evaluation protocol, and trained reference for future work on automated homework assessment.

Long-context LLMs are constrained by the linear growth of key–value (KV) caches during autoregressive decoding, which incurs pronounced latency and memory overhead. KV eviction mitigates this issue, with existing efforts fall into offline policies with fixed eviction patterns and online policies that adaptively discard cache based on attention scores. While online eviction typically preserves accuracy under standard benchmarks, its performance can collapse in practical multi-turn dialogue scenarios where the query positions vary, and integration with pre-fill acceleration remains challenging. In contrast, offline eviction is infrastructure-friendly and generalizable but commonly sacrifices more accuracy. In this paper, we explore Supervised Fine-Tuning (SFT) for offline KV eviction and demonstrate its efficacy as a simple and powerful alternative to the design of complex online eviction metrics. We further propose Hybrid Sparse Supervised Fine-Tuning (HS-SFT) to explore the optimal offline design of KV eviction within SFT. In particular, HS-SFT employs a straight-through estimator to learn discrete local-window allocations of streaming heads across layers with budget-aware balancing loss, such that under high compression ratios—where dense-head capacity is constrained—the budget can be more effectively skewed to capture critical information. Across extensive evaluations on a wide array of LLMs and long-context tasks, HS-SFT delivers substantial performance gains over state-of-the-art eviction baselines. For example, with fewer than 4 hours of SFT of LLaMA-3-8B-1048K using a single 8-GPU node, HS-SFT achieves 5.86% and 38.3% higher average accuracy than Duo-attn on Longbench and Ruler-16K at 10% KV budget, respectively. These results position training-aware offline eviction—achieved with simple SFT—as an effective and practical path to scalable long-context inference.

Multimodal protein language models (pLMs) learn joint protein sequence-structure distributions, and their generation performance should also depend critically on inference-time sampling strategies. Yet prior work has focused more on model training than on how inference-time strategies behave. In this paper, we establish a three-stage investigation framework to empirically study the inference design space of multimodal pLMs across three representative pLMs and four fundamental tasks. We evaluate vanilla sampling, task-specific classifier-free guidance, and reward-guided beam search on multimodal pLMs, corresponding to controls over sampling distributions, per-step logits, and parallel trajectories. Throughout the complementary advancements centered on exploration-exploitation trade-off, we (1) reveal the suboptimality of default inference protocols and identify task-oriented sampling preferences; (2) observe substantial quantitative gains across tasks, consistently boosting the upper bound performance of multimodal pLMs without updating model parameters; (3) derive conclusions about base models that differ from prior consensus. Our code is available at: https://github.com/EchoChou990919/mplm_inference.

Although contextualized embeddings generated from large-scale pre-trained models perform well in many tasks, traditional static embeddings (e.g., Skip-gram, Word2Vec) still play an important role in low-resource and lightweight settings due to their low computational cost, ease of deployment, and stability. In this paper, we aim to improve word embeddings by 1) incorporating more contextual information from existing pre-trained models into the Skip-gram framework, which we call Context-to-Vec; 2) proposing a post-processing retrofitting method for static embeddings independent of training by employing priori synonym knowledge and weighted vector distribution. Through extrinsic and intrinsic tasks, our methods are well proven to outperform the baselines by a large margin.