Beyond Global Routing Aggregation: Phase-Aware Expert Merging for MoE Vision-Language Models

Mixture-of-experts vision-language models (MoE-VLMs) increase model capacity with sparse expert activation, yet deployment requires storing the full expert pool. Training-free expert merging reduces this burden, and many routing-based methods aggregate routing statistics across all tokens to determine merge compatibility. However, MoE-VLM inference is phase-structured: image-context tokens carry visual content, question tokens specify the query, and answer tokens produce the output, with different counts and routing distributions. Because image-context tokens are far more numerous, global aggregation can overemphasize image-context processing and obscure phase-conditioned expert roles, making experts serving different phases appear interchangeable and degrading model performance. We therefore argue that MoE-VLM expert merging should preserve phase-conditioned expert roles, judging compatibility by how experts serve different phases rather than globally aggregated routing statistics. Based on this view, we propose RoleMerge, a training-free method that constructs each expert's Routing Role Profile (RRP) from phase-normalized routing statistics, capturing its relative phase preference. Guided by expert-phase information loss, RoleMerge merges experts with compatible profiles and their corresponding router entries while preserving answer-decoding expert distinctions. Experiments on three models and multiple benchmarks show that RoleMerge preserves more of the full model's performance than alternative expert-merging methods at matched expert-retention ratios, with relative improvements of up to 9.6 percent in six-task macro-average performance. These results validate phase-conditioned expert roles as a more effective basis than global routing aggregation for MoE-VLM expert merging.

REFLEX: Rethinking MoE Inference as Refinement-Aware Compute Allocation in Diffusion Language Models

Mixture-of-experts (MoE) models increase parameter capacity by activating only a small subset of experts for each token. This conditional-computation paradigm has enabled autoregressive language models to scale model capacity without a proportional increase in per-token computation. In diffusion language models (DLMs), however, each denoising forward jointly revisits all token positions despite their sharply different refinement demands, while the default fixed token-choice routing assigns them a uniform expert budget, creating a mismatch between expert computation and refinement demand. We argue that MoE inference in DLMs should therefore be viewed as refinement-aware compute allocation across heterogeneous token refinement states. We propose REFLEX (\textbf{RE}finement-aware \textbf{FLEX}ible expert allocation), a training-free method that keeps the default router unchanged while reorganizing expert computation around the evolving refinement process. Specifically, REFLEX introduces a coarse-to-fine hierarchy for expert-budget allocation that aligns computation with block-relative refinement roles while using the Frontier-Progress Score to resolve active-block priorities. Across multiple widely used benchmarks on two representative MoE-based DLMs, LLaDA-MoE and LLaDA2.0-mini, REFLEX reduces allocated expert computation by 15\% on average while preserving or even improving generation quality on most benchmarks relative to default routing. Compared with autoregressive-style variable-expert routing methods, REFLEX also yields a more consistent quality--computation trade-off, further supporting the importance of allocating expert computation according to the heterogeneous refinement demands exposed within each denoising forward.

DAVET: Denoising-Aware Visual Evidence Trajectory Allocation for Diffusion Vision-Language Models

Diffusion vision-language models (dVLMs) iteratively denoise masked responses while conditioning each denoising step on visual evidence, making visual conditioning a substantial recurring inference cost. Unlike autoregressive decoding, diffusion generation repeatedly revisits the entire response as uncertainty evolves. Our analysis reveals that visual evidence demand is strongly step-dependent, motivating adaptive allocation across denoising steps. Existing inference acceleration methods operate through decoding-side strategies or visual token compression via pruning and merging, but do not explicitly treat visual evidence as a resource whose demand evolves across the diffusion process. Therefore, we present Denoising-Aware Visual Evidence Trajectory Allocation (DAVET), a training-free framework that allocates visual evidence according to the evolving generation state. Starting from a phase-conditioned evidence trajectory, the proposed allocation policy uses operation demand to set an evidence reserve whose allocation at each denoising step is modulated by trajectory risk. DAVET realizes the resulting budgets through a hierarchy of evidence views constructed from a single visual encoding, separating when and how much evidence is needed from how the evidence views are constructed. Evaluated on two representative dVLMs, LLaDA-V and LaViDa, across multiple visual-understanding benchmarks, DAVET achieves an average speedup of 1.55$\times$ with an average relative performance drop of 1.86\%, showing that denoising-aware visual evidence allocation can reduce visual conditioning cost while largely preserving generation quality.

Is Your Model Thinking or Just Stagnating? PUMA: Diagnosing Reasoning Pathology via Phase-Momentum Alignment

Test-time scaling empowers Large Reasoning Models (LRMs) to tackle complex tasks via extensive Chain-of-Thought (CoT). However, this often induces the "overthinking" paradox, where redundant reasoning increases computational overhead without guaranteeing accuracy. Existing test-time efficiency optimization methods primarily fall into two categories: information-theoretic approaches, which are prone to "deceptive convergence" where low uncertainty masks hallucinations, and latent representation analyses, which are often post-hoc, lacking the real-time sensitivity for dynamic reasoning. To bridge this gap, we first posit the Phase-Momentum Alignment Hypothesis, asserting that reasoning correctness hinges on the temporal synchronization between geometric momentum and uncertainty resolution. We then theoretically formulate the Cognitive-Energy Model to characterize these dynamics through two orthogonal dimensions: Geometric Cognitive Effort, quantified by latent velocity and tortuosity, and Entropic Cognitive Uncertainty. To operationalize this, we introduce PUMA (Phase-Uncertainty Momentum Alignment), a training-free framework employing a tiered diagnostic architecture. By coupling lightweight phase monitoring with event-triggered geometric analysis, PUMA effectively distinguishes active exploration from passive stagnation, enabling precise interventions through adaptive truncation or corrective measures. Extensive experiments on LRMs spanning 1.5B to 32B demonstrate that PUMA consistently outperforms state-of-the-art baselines across diverse benchmarks, achieving a superior accuracy-efficiency trade-off and robust cross-domain generalization.

Co-Scraper: query-aware DOM Pruning and Reusable Scraper Synthesis for Lightweight Web Data Extraction

The abundant and heterogeneous nature of web content necessitates automated information extraction, and generating scrapers that can be reused across similar web pages offers an effective solution for scalable data extraction. In this work, we propose Co-Scraper, a two-stage framework capable of handling the hierarchical complexity of long HTML documents. By integrating a query-aware DOM pruning mechanism with stable extraction strategy induction, Co-Scraper can effectively transforms web content into executable programmatic wrappers using a fine-tuned Qwen3-8B model. On the test set of SWDE, Co-Scraper achieves state-of-the-art performance with an F1 score of 94.78% and a reuse success rate of 90.39%. This framework significantly enhances the accuracy and resilience of data extraction, providing a highly efficient approach for web data acquisition tasks.

Your LLM Agent Can Leak Your Data: Data Exfiltration via Backdoored Tool Use

Tool-use large language model (LLM) agents are increasingly deployed to support sensitive workflows, relying on tool calls for retrieval, external API access, and session memory management. While prior research has examined various threats, the risk of systematic data exfiltration by backdoored agents remains underexplored. In this work, we present Back-Reveal, a data exfiltration attack that embeds semantic triggers into fine-tuned LLM agents. When triggered, the backdoored agent invokes memory-access tool calls to retrieve stored user context and exfiltrates it via disguised retrieval tool calls. We further demonstrate that multi-turn interaction amplifies the impact of data exfiltration, as attacker-controlled retrieval responses can subtly steer subsequent agent behavior and user interactions, enabling sustained and cumulative information leakage over time. Our experimental results expose a critical vulnerability in LLM agents with tool access and highlight the need for defenses against exfiltration-oriented backdoors.

FAVLA: A Force-Adaptive Fast-Slow VLA model for Contact-Rich Robotic Manipulation

Force/torque feedback can substantially improve Vision-Language-Action (VLA) models on contact-rich manipulation, but most existing approaches fuse all modalities at a single operating frequency. This design ignores the mismatched sampling rates of real robot sensors, forcing downsampling of the high-frequency contact cues needed for reactive correction. Combined with common VLM-action-expert (AE) pipelines that execute action chunks largely open loop between expensive VLM updates, unified-frequency fusion often yields delayed responses to impacts, stick-slip, and force spikes. We propose FAVLA, a force-adaptive fast-slow VLA that decouples slow perception planning from fast contact-aware control. FAVLA runs a slow VLM at a fixed low frequency to encode modalities to produce latent representations and to predict near-future force variation. A fast AE then executes at a variable high frequency, conditioning on the latest force sequence data to generate reactive actions. We further introduce a force adapter that injects high-frequency force features into multiple AE layers, and adaptively schedules the AE's execution frequency based on the VLM's predicted force variation. Extensive experiments on contact-rich tasks demonstrate that FAVLA significantly outperforms baselines, achieving superior reactivity and success rates, especially with a smaller contact force during manipulation.

Bridging the Compression-Precision Paradox: A Hybrid Architecture for Clinical EEG Report Generation with Guaranteed Measurement Accuracy

Automated EEG monitoring requires clinician-level precision for seizure detection and reporting. Clinical EEG recordings exceed LLM context windows, requiring extreme compression (400:1+ ratios) that destroys fine-grained temporal precision. A 0.5 Hz error distinguishes absence epilepsy from Lennox-Gastaut syndrome. LLMs lack inherent time-series comprehension and rely on statistical associations from compressed representations. This dual limitation causes systems to hallucinate clinically incorrect measurement values. We separate measurement extraction from text generation. Our hybrid architecture computes exact clinical values via signal processing before compression, employs a cross-modal bridge for EEG-to-language translation, and uses parameter-efficient fine-tuning with constrained decoding around frozen slots. Multirate sampling maintains long-range context while preserving event-level precision. Evaluation on TUH and CHB-MIT datasets achieves 60% fewer false alarms, 50% faster detection, and sub-clinical measurement precision. This is the first system guaranteeing clinical measurement accuracy in automated EEG reports.

SemanticForge: Repository-Level Code Generation through Semantic Knowledge Graphs and Constraint Satisfaction

Large language models (LLMs) have transformed software development by enabling automated code generation, yet they frequently suffer from systematic errors that limit practical deployment. We identify two critical failure modes: \textit{logical hallucination} (incorrect control/data-flow reasoning) and \textit{schematic hallucination} (type mismatches, signature violations, and architectural inconsistencies). These errors stem from the absence of explicit, queryable representations of repository-wide semantics. This paper presents \textbf{SemanticForge}, which introduces four fundamental algorithmic advances for semantically-aware code generation: (1) a novel automatic reconciliation algorithm for dual static-dynamic knowledge graphs, unifying compile-time and runtime program semantics; (2) a neural approach that learns to generate structured graph queries from natural language, achieving 73\% precision versus 51\% for traditional retrieval; (3) a novel beam search algorithm with integrated SMT solving, enabling real-time constraint verification during generation rather than post-hoc validation; and (4) an incremental maintenance algorithm that updates knowledge graphs in $O(|ΔR| \cdot \log n)$ time while maintaining semantic equivalence.

Knowledge Graph-Infused Fine-Tuning for Structured Reasoning in Large Language Models

This paper addresses the problems of missing reasoning chains and insufficient entity-level semantic understanding in large language models when dealing with tasks that require structured knowledge. It proposes a fine-tuning algorithm framework based on knowledge graph injection. The method builds on pretrained language models and introduces structured graph information for auxiliary learning. A graph neural network is used to encode entities and their relations, constructing a graph-based semantic representation. A fusion mechanism is then designed to jointly model the knowledge graph embeddings with the contextual representations from the language model. To enhance the robustness of knowledge integration, a gating mechanism is introduced to dynamically balance the contributions of linguistic semantics and structural knowledge. This effectively mitigates conflicts between different representational spaces. During training, a joint loss function is constructed to account for both task performance and structural alignment objectives. This helps improve the accuracy of entity prediction and semantic reasoning. The study also includes a series of systematic sensitivity experiments. It evaluates the effects of learning rate, graph coverage, and structural perturbations on model performance. The results further validate the effectiveness and stability of the proposed method across tasks such as entity recognition, question answering, and language generation. Experimental findings show that the proposed structure-aware fine-tuning framework significantly enhances the model's ability to represent complex semantic units. It demonstrates better semantic consistency and contextual logic modeling in scenarios involving structural reasoning and entity extraction.