Retrieval-augmented generation systems face a context problem where undisciplined loops lead to the inclusion of unused information. Knowledge compilation, an ingestion pipeline enriching documents with structures like entity graphs and ontologies, is proposed as the missing layer. This compilation serves as a supply layer rather than a better retriever, as structural artifacts do not inherently boost retrieval accuracy. A navigation tree based on claim-evidence pairs can significantly reduce query tokens, offering efficiency gains. The proposed RAGFlow pipeline creates an agent-native database of compiled, governed context for improved system performance.
Vision transformer models like CLIP can achieve few-shot defect detection using limited images, but structured decision models like Jev and Clef-Flash offer improvements. The research introduces a framework for k-shot learning, an agentic approach to refine text descriptions, and a calibration method for defect probabilities. Testing across five manufacturing cases showed Jev and Clef-Flash outperformed CLIP's accuracy with fewer references. Agentic text refinement significantly improved accuracy, and calibration reduced the Brier score in many configurations.
Effective image editing evaluation requires identifying local failures to guide subsequent model improvements. WeGenEditBench introduces 2,740 evaluation instances across various editing tasks, incorporating item-level deductions and visual evidence. WeGenEvoStudio is a multi-agent harness designed for recursive self-improvement by integrating failure diagnosis and adaptation. Experiments show that this approach leads to substantial targeted gains in image editing model performance. The work provides a foundation for developing more reliable image models through interpretable evaluation and feedback-driven learning.
Coarse3D generates controllable and faithful sparse 3D structure from arbitrary images for image-to-3D tasks. It addresses limitations in existing coarse generators by controlling orientation and ensuring structural consistency with input images. The method employs a first-frame anchored canonical space, a dual-tower DiT, and spatial-prior guided attention for structure generation. Coarse3D can replace the coarse stage in existing pipelines to produce high-quality sparse structure. Experiments demonstrate that this approach substantially improves structure quality and enhances final 3D generation, achieving state-of-the-art results.
Web agents are vulnerable to adversarial content injection on websites, which existing evaluations fail to capture by only assessing text-level attacks. The research introduces REEF, a benchmark of executable environments with verification to measure web agent security vulnerabilities. ANGLER is an adaptive framework that uses execution feedback from verifiers to guide LLM-based mutation for red-teaming. ANGLER achieved an end-to-end attack success rate of 20.36% against eight VLM-based agents. Attacks generated by ANGLER significantly reduce the success rate of benign tasks, dropping it from 61.9% to 44.4%.
Agentic Soft-Token Compression (ASTC) is a training-free method that compresses context in long-horizon LLM agents using soft tokens instead of hard tokens. ASTC utilizes the model's recurrent dynamics and hybrid GDN-attention to capture absorbed information. This approach allows the system to carry context forward as agentic soft tokens while preserving exact values as hard tokens. Experiments show ASTC achieves accuracy matching or exceeding uncompressed agents with fewer Attention Pairs. This method demonstrates a model-native way to manage evolving task representations efficiently.
Particle image velocimetry (PIV) accuracy is hindered by subjective choices in the estimation pipeline, especially when used for real-time flow control. ART of PIV introduces an agent that automates pipeline configuration selection for each image pair based on a measurement objective. This learned agent reduces end-point error in various estimators by 12.2-22.4% compared to single estimators at matched speeds. In a simulated task, ART of PIV reduced the mean normalized closed-loop cost by 13.24-13.59 percentage points. Experimental results show ART of PIV effectively controls flow in physical systems, such as tracking prescribed flow fields.
Multimodal prompt injection embeds malicious instructions in external media, posing a challenge for existing defenses that often impair task utility. ERASE is an on-policy self-distillation method that defends against this by erasing the injected region and replacing it with modality-appropriate blank content. The method trains a student model by observing feedback from a frozen base model reading the erased view through token-level feedback. Experiments show ERASE achieves near-zero observed attack success while retaining task utility close to the undefended input level. This approach significantly reduces injection dependence, demonstrating improved robustness across different modalities.
GenDSR transfers representations from video generation models to vision-language models to enable dynamic spatial reasoning. The framework creates two branches from video model features: one for holistic context and one for temporal variation. Task-conditioned tokenwise gates fuse these signals to refine the visual representation. On DSR-Bench, GenDSR achieved 67.2%, setting a new state-of-the-art result.
Generating small molecules for protein-protein interactions is challenging because PPI interfaces are broad, unlike small-molecule binding pockets. The PIFT framework transfers peptide-target interaction patterns to small molecules using joint guidance of diffusion and flow matching. This method maps peptide and small-molecule interfaces into a shared space to capture transferable interface information across molecular types. The process uses peptide-derived references to steer molecule generation, resulting in improved interface recovery and favorable properties. Experiments confirm that this guidance enhances cross-modal compatibility and docking performance for small molecules.