The immense memory requirements of state-of-the-art Mixture-of-Experts (MoE) models present a significant challenge for inference, often exceeding the capacity of a single accelerator. While offloading experts to host memory is a common solution, it introduces a severe I/O bottleneck over the PCIe bus, as the data-dependent nature of expert selection places these synchronous transfers directly on the critical path of execution, crippling performance. This paper argues that the I/O bottleneck can be overcome by trading a small amount of cheap, on-device computation to hide the immense cost of data movement. We present MoE-SpeQ, a new inference system built on a novel co-design of speculative execution and expert offloading. MoE-SpeQ employs a small, on-device draft model to predict the sequence of required experts for future tokens. This foresight enables a runtime orchestrator to prefetch these experts from host memory, effectively overlapping the expensive I/O with useful computation and hiding the latency from the critical path. To maximize performance, an adaptive governor, guided by an Amortization Roofline Model, dynamically tunes the speculation strategy to the underlying hardware. Our evaluation on memory-constrained devices shows that for the Phi-MoE model, MoE-SpeQ achieves at most 2.34x speedup over the state-of-the-art offloading framework. Our work establishes a new, principled approach for managing data-dependent memory access in resource-limited environments, making MoE inference more accessible on commodity hardware.

In this paper, we model and analyze how a network of interacting LLMs performs collaborative question-answering (CQA) in order to estimate a ground truth given a distributed set of documents. This problem is interesting because LLMs often hallucinate when direct evidence to answer a question is lacking, and these effects become more pronounced in a network of interacting LLMs. The hallucination spreads, causing previously accurate LLMs to hallucinate. We study interacting LLMs and their hallucination by combining novel ideas of mean-field dynamics (MFD) from network science and the randomized utility model from economics to construct a useful generative model. We model the LLM with a latent state that indicates if it is truthful or not with respect to the ground truth, and extend a tractable analytical model considering an MFD to model the diffusion of information in a directed network of LLMs. To specify the probabilities that govern the dynamics of the MFD, we propose a randomized utility model. For a network of LLMs, where each LLM has two possible latent states, we posit sufficient conditions for the existence and uniqueness of a fixed point and analyze the behavior of the fixed point in terms of the incentive (e.g., test-time compute) given to individual LLMs. We experimentally study and analyze the behavior of a network of $100$ open-source LLMs with respect to data heterogeneity, node capability, network structure, and sensitivity to framing on multiple semi-synthetic datasets.

Retrieval-augmented generation (RAG) greatly enhances large language models (LLMs) performance in knowledge-intensive tasks. However, naive RAG methods struggle with multi-hop question answering due to their limited capacity to capture complex dependencies across documents. Recent studies employ graph-based RAG to capture document connections. However, these approaches often result in a loss of semantic coherence and introduce irrelevant noise during node matching and subgraph construction. To address these limitations, we propose NeuroPath, an LLM-driven semantic path tracking RAG framework inspired by the path navigational planning of place cells in neurobiology. It consists of two steps: Dynamic Path Tracking and Post-retrieval Completion. Dynamic Path Tracking performs goal-directed semantic path tracking and pruning over the constructed knowledge graph (KG), improving noise reduction and semantic coherence. Post-retrieval Completion further reinforces these benefits by conducting second-stage retrieval using intermediate reasoning and the original query to refine the query goal and complete missing information in the reasoning path. NeuroPath surpasses current state-of-the-art baselines on three multi-hop QA datasets, achieving average improvements of 16.3% on recall@2 and 13.5% on recall@5 over advanced graph-based RAG methods. Moreover, compared to existing iter-based RAG methods, NeuroPath achieves higher accuracy and reduces token consumption by 22.8%. Finally, we demonstrate the robustness of NeuroPath across four smaller LLMs (Llama3.1, GLM4, Mistral0.3, and Gemma3), and further validate its scalability across tasks of varying complexity. Code is available at https://github.com/KennyCaty/NeuroPath.

Despite recent progress in 3D-LLMs, they remain limited in accurately grounding language to visual and spatial elements in 3D environments. This limitation stems in part from training data that focuses on language reasoning rather than spatial understanding due to scarce 3D resources, leaving inherent grounding biases unresolved. T o address this, we propose 3D scene editing as a key mechanism to generate precise visual counterfactuals that mitigate these biases through fine-grained spatial manipulation, without requiring costly scene reconstruction or large-scale 3D data collection. Furthermore, to make these edits targeted and directly address the specific weaknesses of the model, we introduce DEER-3D, an error-driven framework following a structured "Decompose, Diagnostic Evaluation, Edit, and Re-train" workflow, rather than broadly or randomly augmenting data as in conventional approaches. Specifically, upon identifying a grounding failure of the 3D-LLM, our framework first diagnoses the exact predicate-level error (e.g., attribute or spatial relation). It then executes minimal, predicate-aligned 3D scene edits, such as recoloring or repositioning, to produce targeted counterfactual supervision for iterative model fine-tuning, significantly enhancing grounding accuracy. W e evaluate our editing pipeline across multiple benchmarks for 3D grounding and scene understanding tasks, consistently demonstrating improvements across all evaluated datasets through iterative refinement. DEER-3D underscores the effectiveness of targeted, error-driven scene editing in bridging linguistic reasoning capabilities with spatial grounding in 3D LLMs.

Membership inference attacks (MIAs) on large language models (LLMs) pose significant privacy risks across various stages of model training. Recent advances in Reinforcement Learning with Verifiable Rewards (RLVR) have brought a profound paradigm shift in LLM training, particularly for complex reasoning tasks. However, the on-policy nature of RLVR introduces a unique privacy leakage pattern: since training relies on self-generated responses without fixed ground-truth outputs, membership inference must now determine whether a given prompt (independent of any specific response) is used during fine-tuning. This creates a threat where leakage arises not from answer memorization. To audit this novel privacy risk, we propose Divergence-in-Behavior Attack (DIBA), the first membership inference framework specifically designed for RLVR. DIBA shifts the focus from memorization to behavioral change, leveraging measurable shifts in model behavior across two axes: advantage-side improvement (e.g., correctness gain) and logit-side divergence (e.g., policy drift). Through comprehensive evaluations, we demonstrate that DIBA significantly outperforms existing baselines, achieving around 0.8 AUC and an order-of-magnitude higher TPR@0.1%FPR. We validate DIBA's superiority across multiple settings--including in-distribution, cross-dataset, cross-algorithm, black-box scenarios, and extensions to vision-language models. Furthermore, our attack remains robust under moderate defensive measures. To the best of our knowledge, this is the first work to systematically analyze privacy vulnerabilities in RLVR, revealing that even in the absence of explicit supervision, training data exposure can be reliably inferred through behavioral traces.

Recent advancements in multimodal out-of-context (OOC) misinformation detection have made remarkable progress in checking the consistencies between different modalities for supporting or refuting image-text pairs. However, existing OOC misinformation detection methods tend to emphasize the role of internal consistency, ignoring the significant of external consistency between image-text pairs and external evidence. In this paper, we propose HiEAG, a novel Hierarchical Evidence-Augmented Generation framework to refine external consistency checking through leveraging the extensive knowledge of multimodal large language models (MLLMs). Our approach decomposes external consistency checking into a comprehensive engine pipeline, which integrates reranking and rewriting, apart from retrieval. Evidence reranking module utilizes Automatic Evidence Selection Prompting (AESP) that acquires the relevant evidence item from the products of evidence retrieval. Subsequently, evidence rewriting module leverages Automatic Evidence Generation Prompting (AEGP) to improve task adaptation on MLLM-based OOC misinformation detectors. Furthermore, our approach enables explanation for judgment, and achieves impressive performance with instruction tuning. Experimental results on different benchmark datasets demonstrate that our proposed HiEAG surpasses previous state-of-the-art (SOTA) methods in the accuracy over all samples.

Triage is a critically important decision-making process in mass casualty incidents (MCIs) to maximize victim survival rates. While the role of AI in such situations is gaining attention for making optimal decisions within limited resources and time, its development and performance evaluation require benchmark datasets of sufficient quantity and quality. However, MCIs occur infrequently, and sufficient records are difficult to accumulate at the scene, making it challenging to collect large-scale real-world data for research use. Therefore, we developed Syn-STARTS, a framework that uses LLMs to generate triage cases, and verified its effectiveness. The results showed that the triage cases generated by Syn-STARTS were qualitatively indistinguishable from the TRIAGE open dataset generated by manual curation from training materials. Furthermore, when evaluating the LLM accuracy using hundreds of cases each from the green, yellow, red, and black categories defined by the standard triage method START, the results were found to be highly stable. This strongly indicates the possibility of synthetic data in developing high-performance AI models for severe and critical medical situations.

Modern codebases evolve continuously: files are renamed or deleted; public APIs drift; behavior shifts within otherwise familiar modules. A model trained yesterday to map a developer's natural-language question to the exact set of repository file paths that matter will degrade tomorrow, even if the questions themselves look unchanged. In this paper we study, at system scale and across several widely used repositories, how to keep such a model fresh without surrendering retention on earlier code. We frame freshness as a form of domain drift between a base snapshot and the current HEAD, and we compare three families of update strategies: (A) Full Refresh, retraining the entire model at the new snapshot; (B) In-Context Learning (ICL) that injects recent deltas (raw git diffs or concise English summaries) at inference; and (C) Incremental Fine-Tuning (Inc-FT) on delta-derived training sets, with carefully controlled NEW:OLD mixing to mitigate catastrophic forgetting. We contribute an alias-aware evaluation protocol that credits rename while never rewarding deleted paths, and a practical Forgetting Probe that quantifies residual emissions of obsolete paths. Across Flask, SQLAlchemy, Pandas, and Poetry, Inc-FT with old-aware mixes delivers the best overall balance on mixed sets, ICL with English delta summaries delivers the fastest new-code lift when training is not feasible, and Full Refresh remains the ceiling when maximum NEW accuracy matters. We also compare Git-diff Inc-FT to full-file Inc-FT, showing that diffs excel in rename/delete-heavy windows while full-file context wins in behavior-change-heavy windows.

The static nature of knowledge within Large Language Models (LLMs) makes it difficult for them to adapt to evolving information, rendering knowledge editing a critical task. However, existing methods struggle with challenges of scalability and retrieval efficiency, particularly when handling complex, multi-hop questions that require multi-step reasoning. To address these challenges, this paper introduces ALEX (A Light Editing-knowledge Extractor), a lightweight knowledge editing framework. The core innovation of ALEX is its hierarchical memory architecture, which organizes knowledge updates (edits) into semantic clusters. This design fundamentally reduces retrieval complexity from a linear O(N) to a highly scalable O(K + N/C). Furthermore, the framework integrates an Inferential Query Synthesis (IQS) module to bridge the semantic gap between queries and facts, and a Dynamic Evidence Adjudication (DEA) engine that executes an efficient two-stage retrieval process. Experiments on the MQUAKE benchmark demonstrate that ALEX significantly improves both the accuracy of multi-hop answers (MultiHop-ACC) and the reliability of reasoning paths (HopWise-ACC). It also reduces the required search space by over 80%, presenting a promising path toward building scalable, efficient, and accurate knowledge editing systems.

Recent work has shown that fine-tuning on insecure code data can trigger an emergent misalignment (EMA) phenomenon, where models generate malicious responses even to prompts unrelated to the original insecure code-writing task. Such cross-domain generalization of harmful behavior underscores the need for a deeper understanding of the algorithms, tasks, and datasets that induce emergent misalignment. In this work, we extend this study by demonstrating that emergent misalignment can also arise from narrow refusal unlearning in specific domains. We perform refusal unlearning on Cybersecurity and Safety concept, and evaluate EMA by monitoring refusal scores across seven responsible AI (RAI) domains, Cybersecurity, Safety, Toxicity, Bias, Sensitive Content, Medical/Legal, and Privacy. Our work shows that narrow domain unlearning can yield compliance responses for the targeted concept, however, it may also propagate EMA to unrelated domains. Among the two intervened concepts, Cybersecurity and Safety, we find that the safety concept can have larger EMA impact, i.e, causing lower refusal scores, across other unrelated domains such as bias. We observe this effect consistently across two model families, Mistral-7b-0.3v, and Qwen-7b-2.5. Further, we show that refusal unlearning augmented with cross-entropy loss function on a small set of retain data from the affected domains can largely, if not fully, restore alignment across the impacted domains while having lower refusal rate on the concept we perform unlearning on. To investigate the underlying causes of EMA, we analyze concept entanglements at the representation level via concept vectors. Our analysis reveals that concepts with higher representation similarity in earlier layers are more susceptible to EMA after intervention when the refusal stream is altered through targeted refusal unlearning.