ECCV2026 LLM Agent AI paper notes paper summaries Agents Multimodal/VLM Reasoning Model Compression Layout & Composition

🦾 LLM Agent¶

🎞️ ECCV2026 · 36 paper notes

📌 Same area in other venues: 🧠 NeurIPS2026 (8) · 📷 CVPR2026 (44) · 🔬 ICLR2026 (209) · 💬 ACL2026 (121) · 🧪 ICML2026 (83) · 🤖 AAAI2026 (59)

🔥 Top topics: Agents ×13 · Multimodal/VLM ×7 · Reasoning ×3 · Model Compression ×2 · Layout & Composition ×2

MolmoWeb: Open Visual Web Agent and Open Data for the Open Web: MolmoWeb combines trajectories from structured-web teachers, human demonstrations, and GUI perception supervision into screenshot-conditioned action learning, reaching 78.2% single-run success on WebVoyager with an 8B model and 94.7% pass@4 through parallel attempts; its main contribution is a reproducible data and training recipe rather than a new planning architecture.
QuantV2X: A Fully Quantized Multi-Agent System for Cooperative Perception: QuantV2X proposes the first fully quantized multi-agent V2X cooperative perception system. By utilizing end-to-end quantized neural network models and quantized communication message representations, it maintains 99.8% of the perception accuracy of full-precision models under low-bit INT4/INT8 quantization. Simultaneously, it reduces end-to-end system latency by 3.2 times and surpasses the full-precision baseline in mAP30 by +9.5 on the V2X-Real dataset.
GUI-AIMA: Aligning Intrinsic Multimodal Attention with a Context Anchor for GUI Grounding: GUI-AIMA turns a multimodal model's intrinsic attention into a GUI click localizer through a trailing anchor token, instruction-adaptive attention-head weighting, and patch-wise supervision; with 509k training instructions, its 3B model achieves 53.8% single-step and 61.5% crop-and-zoom two-step accuracy on ScreenSpot-Pro.
PPTArena: A Benchmark for PowerPoint Editing: PPTArena turns in-place editing of real presentations into a benchmark that checks both structural correctness and visual quality, while its code/OOXML hybrid agent PPTPilot achieves full-benchmark IF/VQ of 2.57/2.69 with reflection and verification, leaving a substantial gap to reliable complex editing.
A Benchmark and Multi-Agent System for Instruction-driven Cinematic Video Compilation: CineBench evaluates instruction-driven cinematic compilation, while the training-free CineAgents system analyzes source footage before planning and executing an edit, achieving a shot-level F1 of 64.13, temporal correctness of 52.09%, and adversarial rejection of 87.23% on this benchmark.
Automatic Method Illustration Generation for AI Scientific Papers via Drawing Middleware Creation, Evolution, and Orchestration: FigAgent turns recurring visual components into callable, evolvable drawing functions and orchestrates them through multi-agent lookahead search to produce editable SVGs, improving node-alignment F1 on FigAgentBench from PaperBanana's 46.3 to 51.4.
Narrative-Driven Paper-to-Slide Generation via ArcDeck: ArcDeck preserves argumentative dependencies through discourse trees, constrains multi-agent outline revision with a global commitment, and then constructs editable slides; with GPT-4o generation and GPT-5 judging on ArcBench, its Narrative Flow score is 43.25 versus SlideGen's 30.75, although overall quality remains below author-prepared slides.
GUIDE: Resolving Domain Bias in GUI Agents through Real-Time Web Video Retrieval and Plug-and-Play Annotation: GUIDE is a training-free, plug-and-play framework that automatically retrieves and extracts domain-specific planning and grounding knowledge from YouTube tutorial videos. By injecting this knowledge into corresponding modules of GUI agents to eliminate domain bias, it brings performance improvements of +4.47 to +7.48 percentage points for three agents with different architectures on OSWorld.
Exploratory, Communicative, and Deployable: Vision-Driven Embodied Agents for Open-World Mobile Manipulation: REAL trains Qwen3-VL-8B to explore, ask when necessary, and then manipulate through a tool interface that withholds privileged object state; SFT followed by online GSPO reaches 56.9% success on the interactive REAL-Bench split, and the high-level policy transfers zero-shot to a physical robot with 78.3% success over 60 episodes.
HippoCamp: Benchmarking Contextual Agents on Personal Computers: HippoCamp organizes heterogeneous personal-computer files into three isolated archetypal environments and evaluates factual retention and user profiling with 581 file-grounded questions; the strongest evaluated system achieves only 48.3% profiling accuracy, exposing a substantial gap between finding relevant files and reliably understanding a user.
Knowledge-Centric Agents for Workflow Generation in ComfyUI: The paper recovers strategies, reasoning explanations, and skeleton pseudocode from real ComfyUI workflows, then uses hierarchical SFT and self-refinement to reconstruct executable graphs from user instructions, improving average Pass / Resolve over ComfyAgent from 36.4% / 29.4% to 86.9% / 61.1% on its curated test set, although Resolve remains only 25.0% on the external ComfyBench subset.
Learn2Fold: Structured Origami Generation with World Model Planning: The paper casts origami folding-sequence generation as conditional program induction over a crease-pattern (CP) graph: an LLM proposes structured folding actions from a text goal, a learned graph-structured world model predicts residual states and per-edge violation probabilities by short-horizon rollout before execution, and a deterministic Level-0 constraint kernel performs hard verification inside an MPC search loop. On a 25-category origami benchmark it lifts step-level F1 from 0.266 (best baseline) to 0.739 and macro category success rate from 0.675 to 0.891.
NaLA: A 3D Native LLM Layout Agent for High-quality 3D Scene Generation: NaLA proposes a large language model layout agent that directly encodes 3D point clouds. Through a coarse-to-fine pose prediction mechanism, it achieves high-quality and highly efficient 3D scene generation, comprehensively outperforming existing methods in physical rationality, semantic consistency, and aesthetics.
ProMSA:Progressive Multimodal Search Agents for Knowledge-Based Visual Question Answering: ProMSA turns knowledge-based visual question answering into a budgeted loop of retrieval, evidence checking, and either further search or answering, using rejection-sampling SFT followed by TN-GSPO to bring Qwen3-VL-8B to overall scores of 52.6 on E-VQA and 53.4 on InfoSeek.
SceneOrchestra: Efficient Agentic 3D Scene Synthesis via Full Tool-Call Trajectory Generation: SceneOrchestra replaces SceneWeaver's step-by-step "execute—render—review—decide" agentic loop with a single forward pass: a fine-tuned Qwen3-4B orchestrator emits a complete tool-call trajectory at once, and a discriminator ranks multiple candidates during training and distills that ranking ability back into the orchestrator, beating SceneWeaver on scene quality across 10 room types while cutting average generation time from 91.0 to 39.7 minutes.
SymbOmni: Evolving Agentic Omni Models via Symbolic Concept Learning: SymbOmni does not generate end-to-end; it builds a memory-bearing agent on top of ComfyUI that abstracts every successful or failed attempt into a "symbolic concept" (a parameterized, executable workflow template) stored in an optimizable Symbolic Concept Box, retrieves and composes those concepts through an induction–transduction cycle, and refines the box with verbalized backpropagation — never updating a single model parameter — reaching 86.0% total resolve rate on ComfyBench (up from 73.0%), 0.98 on GenEval, and cutting total token consumption by 33.9%.
TIR-Agent: Training an Explorative and Efficient Agent for Image Restoration: TIR-Agent trains task ordering and tool selection as a closed-loop Qwen3-VL-8B policy, using perturbed demonstrations to broaden exploration and adaptive multi-metric rewards to coordinate fidelity and perceptual quality, reaching 22.80 dB PSNR on MiO100 Group B and at least 2.5x and 5x inference speedups over AgenticIR and 4KAgent, respectively.
WorldAgents: Can Foundation Image Models be Agents for 3D World Models?: WorldAgents closes the loop among a VLM director, a 2D image generator, and a two-stage 2D/3D verifier, using 3D-reprojection-constrained sequential inpainting to reduce reprojection error from the previous best of 0.39 to 0.36 without updating the generator's weights.
Are GUI Agents Focused Enough? Automated Distraction via Semantic-level UI Element Injection: The paper distracts GUI agents with real, content-harmless UI icons subject to spatial and semantic constraints, reaching a five-depth miss rate of 32.99% on UI-TARS-1.5-7B through cumulative multi-candidate search and using injected-icon click statistics to expose grounding weaknesses that content safety alignment alone does not address.
Think While You Map: Asynchronous Vision-Language Agents for Incremental 3D Scene Graphs: ThinkGraphs proposes an asynchronous architecture that decouples lightweight incremental 3D mapping from heavy VLM inference. By running a background Critic Agent (semantic loop closure detection to merge fragmented trajectories) and a Description Agent (multi-target frame scheduling to inject fine-grained attributes) without blocking online mapping, it continuously enriches scene graph semantics. It outperforms previous state-of-the-art by 15.3-18.8 [email protected] on three visual grounding benchmarks: Sr3D+, Nr3D, and ScanRefer.
VERITAS: A Multi-agent Co-scientist for Verifiable Image-Derived Hypothesis Testing: Veritas splits "natural-language clinical hypothesis to auditable conclusion" into four phases — analysis planning, segmentation, statistics, and interpretation — carried out by three role-specialized agents (PI, Imaging Specialist, Statistician) plus a Critic, and maps the observed statistics mechanically into Supported / Refuted / Underpowered / Invalid via an evidence classification operator that is independent of agent judgment; on 64 cardiac and brain-glioma MRI hypotheses spanning six complexity tiers it reaches 81.4% (frontier models) / 71.2% (locally deployed 8–30B open-weight models) majority-vote verdict accuracy and the highest rate of independently verifiable statistical outputs (86.6%).
ARGOS: Who, Where, and When in Agentic Multi-Camera Person Search: ARGOS turns person search into budgeted evidence gathering: a four-module agent asks questions and invokes camera-topology tools to isolate a target, with best spatial and temporal Turn-Weighted Success scores of 0.383 and 0.590 across a benchmark of 2,691 tasks.
CiQi-Agent: Aligning Vision, Tools and Aesthetics in Multimodal Agent for Cultural Reasoning on Chinese Porcelains: CiQi-Agent integrates porcelain expertise, local image zoom, and image/text retrieval through two phases of SFT and GRPO, bringing Qwen2.5-VL-7B to 81.5% average multiple-choice accuracy and a 66.7% free-form attribute score on CiQi-Bench; these results measure connoisseurship attributes, not authenticity detection.
Dotting the Eye: An Intent-Driven Image Retouching Agent for Visual Focus Enhancement: EyeControl interprets clicks or coarse strokes as coordinated global and local retouching plans, then trains a diffusion executor with pseudo-intent attention alignment and operation consistency to emphasize the intended region, achieving 21.8845 dB PSNR and an overall score of 8.7054 on ControlArt-Bench without simply maximizing editing strength.
GEAR-Seg: A Grounded Explainable Agent for Reasoning Segmentation and Data Engine: GEAR-Seg converts an image into detailed descriptions of indexed regions, then lets a text-only LLM select masks according to the user's intent, achieving 57.5 ReasonSeg gIoU without task-specific fine-tuning and extending the same pipeline into an explanation-producing data engine for end-to-end student training.
Learning from Reliable Negatives: Confidence-Anchored Test-Time Adaptation for GUI Grounding: The paper selects a pseudo-label from multiple predictions using coordinate-token confidence, then uses CANL to penalize only predictions far from that anchor; after independent label-free test-time training on each benchmark, Qwen-2.5-VL-7B improves from 88.1% to 92.1% on ScreenSpot-V2 and from 24.9% to 33.8% on ScreenSpot-Pro.
SpatiO: Adaptive Test-Time Orchestration of Vision-Language Agents for Spatial Reasoning: Rather than updating vision-language model weights, SpatiO uses a small labeled set to calibrate how reliable each model is under a particular role and spatial category, then selects specialists and integrates geometric evidence, reporting overall results of 72.4% on 3DSRBench, 88.2% on STVQA-7k, and 86.9% on CV-Bench in the paper's tables.
VisCritic: Visual State Comparison as Process Reward for GUI Agents: VisCritic proposes a process reward framework for GUI agents based on visual state comparison. Using a Siamese ViT to compare pre- and post-action screenshots in semantic feature space, it fuses action context to jointly predict action success probability, task progress, and error types. Serving as a plug-and-play inference-time verification module, it consistently improves the task success rate of various GUI agents across five benchmarks.
WebRetriever: A Large-Scale Comprehensive Benchmark for Efficient Web Agent Evaluation: WebRetriever combines 1,550 tasks across 800 websites, the interaction-aware NavEval judge, and three evaluation protocols, achieving 91.2% human agreement on its benchmark while revealing that six agents average only 11.8% human-assessed end-to-end task success.
AGE: Agentic Gaussian Editing in 3D Scenarios: AGE coordinates existing 3D editing tools through planning, execution, reflection, and backtracking agents with shared spatial memory, achieving a CLIP text–image direction similarity of 0.192 versus DGE's 0.133 on a custom benchmark, although component ablations are needed to isolate the sources of improvement.
From Gaze to Meaning: An AI Agent for Unified Zero-Shot Grounding and Explanation: Gaze Target Agent (GTA) renders predicted gaze as an arrow for a VLM, uses first-token confidence to trigger retrieval of labeled examples, and selects an object instance using the gaze heatmap, raising GazeFollow Acc@1 from GGVL's 0.450 to 0.493 without parameter updates, although its memory still uses training-set labels.
Guiding the Blind: Generalizing GUI Agents to Unseen Websites via Multimodal Tutorials: On WebOne's website-disjoint training and test split, WebLearner uses hierarchical memory and step-level GRPO rewards to learn to understand the current page, identify a relevant tutorial subtask, and execute an action, raising unseen-website task success from the Qwen3-VL-8B RAG baseline's 36.8% to 56.9%, a gain of 20.1 percentage points.
ICLAgent: Integrated Circuit Footprint Geometry Labeling via LMM-empowered Multi-Agent Framework: ICLAgent mirrors the cognitive workflow of PCB engineers by constructing a four-stage multi-agent framework encompassing diagram detection, topology planning, and parameter derivation, elevating IC footprint labeling accuracy to 79.0% IoU on a lightweight Qwen2-VL-7B backbone.
KATANA: Knowledge-Aligned Topology-Aware Neural Agents for RL-Driven Vision-Language Model Compression: KATANA treats the pruning algorithm itself as a learnable program, trains a code agent with rewards measured in a sandbox, and discovers KIRI's combination of cubic sparsity scheduling, Dual-Norm Activation scoring, and in-mask reconstruction; at 70% sparsity, LLaVA-1.5-7B reaches 84.6 MSCOCO CIDEr versus GSOP's 83.3, a 1.3-point gain, with approximately 2.8x decode acceleration on the specified sparse runtime.
OCTOPUS: Multi-Agentic Universal Compositional Visual Retrieval: OCTOPUS uses three task-training-free agents for perception, creation, and retrieval to turn a reference image or video and an editing instruction into visual and textual proxies, then improve ranking through bidirectional cross-modal retrieval and sequential candidate verification, reaching 67.15% WebVid R@1 and 37.00% average FashionIQ R@10 with ViT-L/14.
Self-Evolving MCP-GUI Agents via Automated Environment Generation and Experience Learning: The paper improves MCP-GUI agents through a loop of diagnosing weaknesses, generating executable tasks, collecting trajectories, and updating either the model or its experience, raising Chrome pass rate from 60.0% to 77.8% while finding that GUI-intensive VS Code benefits more from training-free experience augmentation than from directly imitating expert trajectories.