Poster Session 1
(11/6, Fri, 14:15~15:30)
(11/6, Fri, 14:15~15:30)
[Part 1] 11/6(Fri.) 14:15~15:30
[1-1] WavNAF: Learning Wave Propagation Priors for Neural Acoustic Fields (NeurIPS 2026)
김태호(석박통합과정)
Room acoustics modeling requires capturing intricate wave phenomena such as reflections, refractions, and diffractions beyond direct sound propagation. Recent neural acoustic synthesis methods have improved acoustic realism but typically focus only on straight sound paths and coarse reverberation, missing detailed interactions like diffraction or multi-order reflections. We propose WavNAF, a neural framework that leverages physically-informed wave propagation priors to explicitly capture complex acoustic interactions. We generate these priors by numerically solving the wave equation with the Finite-Difference Time-Domain (FDTD) method, which directly simulates wave-based acoustic behavior that geometric methods cannot capture. Specifically, we extract essential acoustic parameters for FDTD, such as wave speed and density, from visual scene geometry encoded by Neural Radiance Fields (NeRF). We then generate physically-informed pressure maps and encode them via a feature extractor to learn wave propagation priors that capture intricate acoustic phenomena. To address the inherent computational cost issue of FDTD, we introduce a novel Neural Acoustic Scaling Module, inspired by traditional acoustic scale model. This module adaptively recalibrates encoded pressure map features from temporally compressed simulations to efficiently estimate accurate full-scale Room Impulse Responses. Experimental results demonstrate that WavNAF achieves substantial improvements in acoustic quality across various evaluation metrics compared to existing state-of-the-art methods.
[1-2] Masking Matters: Unlocking the Spatial Reasoning Capabilities of LLMs for 3D Scene-Language Understanding (CVPR 2026)
전예림(박사과정)
Recent advances in 3D scene-language understanding have leveraged Large Language Models (LLMs) for 3D reasoning by transferring their general reasoning ability to 3D multi-modal contexts. However, existing methods typically adopt standard decoders from language modeling, which rely on a causal attention mask. This design introduces two fundamental conflicts in 3D scene understanding: sequential bias among order-agnostic 3D objects and restricted object-instruction attention, hindering task-specific reasoning. To overcome these limitations, we propose 3D Spatial Language Instruction Mask (3D-SLIM), an effective masking strategy that replaces the causal mask with an adaptive attention mask tailored to the spatial structure of 3D scenes. Our 3D-SLIM introduces two key components: a Geometry-adaptive Mask that constrains attention based on spatial density rather than token order, and an Instruction-aware Mask that enables object tokens to directly access instruction context. This design allows the model to process objects based on their spatial relationships while being guided by the user's task. 3D-SLIM is simple, requires no architectural modifications, and adds no extra parameters, yet it yields substantial performance improvements across diverse 3D scene-language tasks. Extensive experiments across multiple benchmarks and LLM baselines validate its effectiveness and underscore the critical role of decoder design in 3D multi-modal reasoning.
[1-3] Unified Loss-Aware Density Control for 3D Gaussian Splatting (NeurlPS 2026)
김태영(박사과정)
The efficiency and fidelity of 3D Gaussian Splatting critically depend on adaptive density control, which grows and suppresses Gaussian primitives throughout optimization. However, the standard pipeline still relies on heuristic view-space gradients, scale thresholds, and global opacity resets, rather than explicitly assessing the loss-reducing effect of each split, clone, or opacity update. This mismatch can allocate primitives to already-sufficient regions, miss under-reconstructed structures, and destabilize training through blanket opacity changes. In this work, we recast adaptive density control as a collection of local loss-reduction decisions. We derive efficient Taylor-based per-Gaussian surrogates that predict the objective change caused by three primitive-level operations: splitting, cloning, and opacity decay. This yields unified loss-aware framework with three components. First, we extend second-order splitting to a normalized joint position-scale space, enabling children to adapt not only their locations but also their anisotropic extents. Second, we formulate duplicate cloning as a surrogate minimization problem, selecting only Gaussians whose duplicated contribution is predicted to improve reconstruction and assigning each child a principled opacity. Third, we replace global opacity reset with selective soft opacity decay that weakens harmful primitives while preserving useful ones. Across standard 3DGS benchmarks, our method produces more targeted primitive allocation and achieves improved rendering quality with comparable or smaller Gaussian budgets compared to baselines.
[1-4] Improving Sparse-View 3DGS Generalization via Flat Minima Optimization (ECCV 2026)
서강민(석사과정)
Recent advances in neural rendering have established 3D Gaussian Splatting (3DGS) as a highly efficient representation for novel view synthesis, enabling fast training and real-time rendering with strong fidelity. However, when supervision is limited to sparse input views, 3DGS tends to overfit to the observed images and generalize poorly to unseen viewpoints. We address this challenge from the perspective of flat minima (FM) optimization, which seeks solutions that remain stable under small parameter perturbations. Viewing Gaussian parameters as trainable weights, we adapt FM principles to the geometric and dynamic nature of 3DGS with a lightweight training framework. Our method regularizes optimization with controlled Gaussian perturbations that account for each Gaussian's anisotropy and the training progress, preserving fine details while improving robustness to sparse-view overfitting. To further stabilize this flat minima optimization process, we introduce periodic reinitialization, which temporarily returns non-positional parameters to their initial states for a short window. Together, these techniques integrate seamlessly into existing 3DGS pipelines without architectural changes. Experiments on LLFF and Mip-NeRF360 datasets demonstrate improved quantitative metrics and perceptual quality under sparse-view supervision, producing reconstructions that are sharper, more stable, and better generalized to novel viewpoints.
[1-5] SeaCache: Spectral-Evolution-Aware Cache for Accelerating Diffusion Models (CVPR 2026)
정지우(박사과정)
Diffusion models are a strong backbone for visual generation, but their inherently sequential denoising process leads to slow inference. Previous methods accelerate sampling by caching and reusing intermediate outputs based on feature distances between adjacent timesteps. However, existing caching strategies typically rely on raw feature differences that entangle content and noise. This design overlooks spectral evolution, where low-frequency structure appears early and high-frequency detail is refined later. We introduce Spectral-Evolution-Aware Cache (SeaCache), a training-free cache schedule that bases reuse decisions on a spectrally aligned representation. Through theoretical and empirical analysis, we derive a Spectral-Evolution-Aware (SEA) filter that preserves content-relevant components while suppressing noise. Employing SEA-filtered input features to estimate redundancy leads to dynamic schedules that adapt to content while respecting the spectral priors underlying the diffusion model. Extensive experiments on diverse visual generative models and the baselines show that SeaCache achieves state-of-the-art latency-quality trade-offs.
[1-6] Mitigating Modality and Language-Style Gaps for Zero-Shot Video Moment Retrieval (ECCV 2026)
조철호(석박통합과정)
Zero-shot video moment retrieval (ZMR) aims to overcome the limitations of traditional approaches that require large-scale datasets annotated with text and its relevant temporal spans. Despite advances in pre-trained vision–language models (VLMs) and multimodal large language models (MLLMs), existing ZMR methods still heavily depend on query-to-video content similarity, making them vulnerable to modality and language-style gaps. These gaps lead to unreliable span proposals and unstable moment retrieval results. To address this issue, we propose Self-Similarity-based Moment proposal and Scoring (Self-SiMS) that instead exploits intrinsic relationships within videos, enabling robust span generation and scoring. By deriving self-similarity only from the video content, we circumvent the noisy and mismatched patterns of query–frame or query–caption similarities, thereby mitigating both modality and language-style gaps. Furthermore, we introduce a queryaware MLLM-based reasoning stage to further sharpen alignment between text and video. Extensive experiments demonstrate that Self-SiMS achieves the state-of-the-art performance across ZMR benchmarks.
[1-7] G-MIXER: Geodesic Mixup-based Implicit Semantic Expansion and Explicit Semantic Re-ranking for Zero-Shot Composed Image Retrieval (CVPR 2026)
임지영(석박통합과정)
Composed Image Retrieval (CIR) aims to retrieve target images by integrating a reference image with a corresponding modification text. CIR requires jointly considering the explicit semantics specified in the query and the implicit semantics embedded within its bi-modal composition. Recent training-free Zero-Shot CIR (ZS-CIR) methods leverage Multimodal Large Language Models (MLLMs) to generate detailed target descriptions, converting the implicit information into explicit textual expressions. However, these methods rely heavily on the textual modality and fail to capture the fuzzy retrieval nature that requires considering diverse combinations of candidates. This leads to reduced diversity and accuracy in retrieval results. To address this limitation, we propose a novel trainingfree method, Geodesic Mixup-based Implicit semantic eX-pansion and Explicit semantic Re-ranking for ZS-CIR (G-MIXER). G-MIXER constructs composed query features that reflect the implicit semantics of reference image-text pairs through geodesic mixup over a range of mixup ratios, and builds a diverse candidate set. The generated candidates are then re-ranked using explicit semantics derived from MLLMs, improving both retrieval diversity and accuracy. Our proposed G-MIXER achieves state-of-the-art performance across multiple ZS-CIR benchmarks, effectively handling both implicit and explicit semantics without additional training. Our code will be available at https://github.com/maya0395/gmixer.
[1-8] Prevention over Correction: Learning Aligned Representations in One-Shot Federated Learning (ECCV 2026)
강용훈(박사과정)
One-shot Federated Learning (OFL) reduces communication costs by requiring only a single round of model exchange. However, when clients train independently on heterogeneous data, their learned representations develop different means and variances, creating distributional mismatch that prevents effective server-side aggregation. Existing methods attempt post-calibration at the server, but struggle to reconcile representations that have already diverged into incompatible statistical spaces. We propose FACE (Federated Aligned-Consistent Ensemble) to prevent distributional divergence during client training rather than correcting it afterward. FACE uses a shared fixed reference matrix to guide representation means toward consistent directions, applies a variance stabilizer to maintain consistent feature variance, and performs feature-fusion ensemble inference. Extensive experiments on CIFAR-10/100 and Tiny-ImageNet demonstrate that FACE achieves state-of-the-art performance, with particularly strong improvements under extreme data heterogeneity. Comprehensive ablation studies validate the necessity of both components for achieving distributional alignment.
[1-9] Remember Your Trace: Memory-Guided Long-Horizon Agentic Framework for Consistent and Hierarchical Repository-Level Code Documentation (NeurIPS 2026)
배수영(박사과정)
Automated code documentation is essential for modern software development, providing the contextual grounding that both human developers and coding agents rely on to navigate large codebases. Existing repository-level approaches process components independently, causing redundant retrieval and conflicting descriptions across documents while producing outputs that lack hierarchical structure. Therefore, we propose MemDocAgent, a long-horizon agentic framework that generates documentation within a single, integrated context spanning the entire repository. It combines two components: (i) Dependency-Aware Traversal Guiding that predetermines a traversal order respecting dependency and granularity hierarchies; (ii) Memory-Guided Agentic Interaction, in which the agent interacts with RepoMemory, a shared memory accumulating prior work traces through read, write, and verify operations. Through an in-depth multi-criteria evaluation, MemDocAgent achieves the best performance over both open and closed-source baselines and demonstrates practical applicability in real software development workflows.
[1-10] Causally Redundant Reasoning: A Dual-Disruption Test Framework for Measuring Shortcut Tendencies in Chain-of-Thought (EMNLP 2026 Findings)
김지우(박사과정)
A chain-of-thought (CoT) rationale can be coherent and answer-consistent while remaining causally unused. We call this phenomenon \emph{shortcut reasoning}: the model appears to reason through CoT, but its answer does not depend on the expressed trace. We formalize shortcut reasoning as \emph{causal redundancy} and propose a dual-disruption framework for measuring shortcut tendencies in CoT by testing two complementary aspects of causal explanation: necessity and sufficiency. We operationalize them as two behavioral signals: whether the answer changes under CoT disruption, and whether the answer can be recovered from the CoT after masking the original input. We instantiate this framework with \emph{CORD} (\emph{Causal Redundancy of Reasoning via Disruption}), an instance-level score that combines both signals. The two axes reduce complementary ambiguities in single-axis tests: answer preservation under CoT disruption can reflect shortcut use or redundant truthful reasoning, while early answer recovery can reflect shortcut use or a compact truthful rationale. Across 6 models and 2 datasets, CORD reveals a training-paradigm effect: R1-Distill models exhibit roughly twice the shortcut tendency of non-distilled models, while the two axes capture related but distinct signals. Together, these results motivate joint tests of necessity and sufficiency for detecting shortcut tendencies that can remain hidden behind correct answers.
[1-11] Real-Time LiDAR Gaussian Splatting SLAM via Geometry-Aware Covariance Coupling (ECCV 2026)
탁승준(석박통합과정)
We present a real-time LiDAR-based framework for Gaussian Splatting SLAM that tightly couples fast G-ICP registration with spherical rasterization-based dense mapping for large-scale sequences. Leveraging LiDAR geometry rather than appearance, we reuse tracking-estimated local covariances to initialize Gaussians with range-aware scales and to derive surface normals for geometry-aware map optimization. We further introduce a covariance-derived geometry score that measures local complexity and drives pruning in planar regions and selective densification in structurally rich areas, while optimized Gaussians and LiDAR-specific confidence cues are fed back to improve tracking robustness. On the Newer College dataset, our method achieves an F-score of 86.78% using purely online trajectories at real-time speed (>20 FPS), and additional experiments on other datasets confirm its stability and scalability. Code is available at github.com/Lab-of-AI-and-Robotics/LiDAR-GS-SLAM. Project Page: lab-of-ai-and-robotics.github.io/GS-SLAM-Family
[1-12] MA-Mesh: Resilient Multi-Agent Collaboration via Role-based Self-Healing (EMNLP 2026)
이준용(석사과정)
Effective collaboration in multi-agent systems requires mechanisms for coordination and adaptability in dynamic and uncertain environments. However, existing multi-agent systems rarely model resilience explicitly, instead relying on ad-hoc output correction that limits systematic, runtime management of robustness. In this work, we propose Multi-Agent Mesh (MA-Mesh), a framework that constructs a role-based self-healing layer on top of a collaboration pattern. Specifically, MA-Mesh generates resilience-aware system specifications and synthesizes them into executable multi-agent policy code. To enhance resilience, we introduce two self-healing mechanisms: Cross-agent reflection, a mechanism that allows agents to evaluate not only their own outcomes but also the performance of their peers, enabling informed behavioral adjustments and collective policy refinement; Role-based failover, a strategy which dynamically reassigns roles and enables agents to substitute for one another when sub-tasks fail or agents underperform, ensuring continuity of task execution. Experimental results across diverse tasks (e.g., software development, tool use, scientific reasoning) demonstrate that MA-Mesh significantly outperforms existing multi-agent systems in both task performance and robustness under system uncertainty.
[1-13] Cerberus: Cross-Layer ECC Co-Design for Robust and Efficient Memory Protection (ISCA 2026)
김승현(석사과정)
As DRAM scales to higher density and I/O speeds, ensuring data correctness becomes increasingly difficult. Industry has responded with a three-layer stack—on-die ECC (O-ECC), link ECC (L-ECC), and system ECC (S-ECC)—but these layers have evolved independently, often duplicating redundancy, leaving coverage gaps, and occasionally interfering. We propose Cerberus, a cross-layer ECC co-design that unifies protection across device, link, and system while preserving each layer’s native role. At its core is an Encode-Once, Decode-Many (EODM) architecture: the controller performs a single encoding whose redundancy is reused by L-ECC for immediate write-path detection/retry, by O-ECC for in-device repair on reads, and by S-ECC for strong end-to-end recovery. Cerberus jointly designs complementary parity/syndrome structures, orders decoders and allocates the correction budget to prevent miscorrection amplification, and enables selective correction under tight redundancy constraints. Our evaluations show improved resilience to clustered and peripheral faults while reducing redundancy overhead, underscoring the importance of coordinated crosslayer protection for next-generation memory systems, such as custom HBMs.
[1-14] Pro-Step: Step-level Process Reward Optimization for Retrieval-Agumented Generation (EMNLP 2026)
김민건(석사과정)
Retrieval-Augmented Generation enhances Large Language Models by grounding responses in external knowledge, but multi-hop reasoning remains vulnerable to error propagation, where early retrieval failures confound subsequent steps. Standard outcome-based optimization only rewards the final answer, leaving intermediate retrieval and reasoning errors undetected. While existing process-based methods introduce step-level signals, they still score each step against the final answer, rewarding spurious successes where flawed retrieval coincidentally produces the correct answer. Steplevel supervision in RAG requires evaluating both logical validity and evidential grounding at each step. We introduce PRO-STEP: we train a generative PRM that evaluates both dimensions, employ PRM-guided value tree search to construct preference pairs contrasting valid steps against flawed ones, and optimize the policy via step-level Direct Preference Optimization. Experiments on singleand multi-hop QA datasets demonstrate that PRO-STEP achieves the best average EM and F1 across five benchmarks.
[1-15] Learning Generalizable Skill Policy with Data-Efficient Unsupervised RL(ICML 2026)
박종찬(박사과정)
Unsupervised Reinforcement Learning (URL) aims to pre-train scalable, skill-conditioned policies without extrinsic rewards, serving as a foundation for downstream control tasks. Despite recent progress, we argue that current off-policy URL methods are limited by two critical, overlooked bottlenecks: (1) non-stationary skill semantics and (2) brittle generalization. To address these challenges, we propose GenDa (Generalizable Data-efficient Agent), a unified framework for robust unsupervised reinforcement learning. First, we introduce a skill relabeling mechanism to mitigate non-stationarity and significantly improve data efficiency for pre-training. Second, we propose a Complementary Information Bottleneck (CIB), encouraging the learned skill policy to focus on ego-centric features and become robust to distribution shifts for downstream tasks. Through various experiments, we demonstrate that GenDa significantly enhances the scalability of URL with superior generalizability and data efficiency. Our code and videos are available at this https URL.
[1-16] Decoding Dense Embeddings: Sparse Autoencoders for Interpreting and Discretizing Dense Retrieval(EMNLP 2025)
김택림(석박통합과정)
Despite their strong performance, Dense Passage Retrieval (DPR) models suffer from a lack of interpretability. In this work, we propose a novel interpretability framework that leverages Sparse Autoencoders (SAEs) to decompose previously uninterpretable dense embeddings from DPR models into distinct, interpretable latent concepts. We generate natural language descriptions for each latent concept, enabling human interpretations of both the dense embeddings and the query-document similarity scores of DPR models. We further introduce Concept- Level Sparse Retrieval (CL-SR), a retrieval framework that directly utilizes the extracted latent concepts as indexing units. CL-SR effectively combines the semantic expressiveness of dense embeddings with the transparency and efficiency of sparse representations. We show that CL-SR achieves high index-space and computational efficiency while maintaining robust performance across vocabulary and semantic mismatches.
[1-17] ConvX: A Lightweight Converter to Bridge Indexed Dense Representations and Large Language Models for Retrieval-Augmented Generation (ACL 2026 Findings)
한준호(석박통합과정)
Retrieval-Augmented Generation (RAG) has significantly advanced open-domain question answering systems by incorporating external knowledge into large language models. Despite its effectiveness, existing RAG pipelines suffer from critical efficiency limitations. In particular, modern transformer-based generators exhibit quadratic or higher computational complexity with respect to input sequence length and hidden dimensionality, leading to substantial inference latency as model scales and contextual inputs increase. This issue is exacerbated in RAG settings, where retrieved contexts substantially expand the input prompt. To alleviate this challenge, we propose an effective compression-based RAG framework, ConvX, that directly leverages indexed dense representations produced by a retriever, entirely substituting to long text contexts. Our approach expands a single dense representation into a fixed number of memory slots using a lightweight converter to provide rich lexical information. This design enables efficient knowledge integration while significantly reducing input length and computational overhead. Empirical evaluations demonstrate that the proposed model achieves competitive performances compared to the existing state-of-the-art model that uses a large ad-hoc context compressor, while offering substantially improved inference efficiency.
[1-18] ACA-GS: Adaptive-Capacity Anchored Gaussian Splatting for Compact Dynamic Radiance Fields (ACM MM 2026)
송승현(석박통합과정)
Recent advances in 4D Gaussian Splatting (4DGS) enable high-fidelity, real-time spatiotemporal rendering, but expose a fundamental trade-off between motion expressiveness and storage efficiency. While anchor-based designs achieve compactness through anchor-level parameter sharing, their rigid uniform parametrization enforces fixed Neural Gaussian counts and feature budgets per anchor. Consequently, insufficient fidelity is addressed by excessive anchor density, rather than lightweight, targeted increases in Neural Gaussian count or feature capacity, resulting in memory waste. To overcome this rigidity, we introduce an adaptive-capacity anchor-based framework that dynamically allocates the representational capacity based on local spatiotemporal demands. Adaptive Anchor Cardinality varies the number of Neural Gaussians per anchor, concentrating primitives in regions of high geometric or motion complexity while suppressing redundancy. In parallel, Adaptive Anchor Feature Masking modulates anchor-level feature channels, assigning rich features to complex regions and lightweight representations to simpler ones. Experiments on MPEG, Panoptic Sports, and N3DV datasets demonstrate substantial storage reduction without degrading visual quality. Notably, on challenging MPEG sequences with complex motion, our method achieves up to 1.5× higher compression than state-of-the-art anchor-based methods while preserving comparable quality.
[1-19] Less Is More: Knowledge Graph Construction and Schema-Guided Canonicalization with LLMs (CIKM 2026)
송명석(석사과정)
Knowledge graphs (KGs) are widely used in domains such as healthcare and finance, but existing LLM-based construction methods often produce redundant triples and semantically inconsistent representations. We propose a unified framework that integrates LLMbased triple extraction with schema-guided canonicalization for knowledge graph construction. Our approach first extracts triples from text and induces a schema, which is then used to constrain canonicalization candidates and enable semantically consistent entity and relation merging. Experiments on five benchmark datasets show that our method consistently produces more compact KGs while achieving higher F1 scores than existing approaches. Ablation studies further demonstrate that schema-guided canonicalization effectively reduces redundancy with minimal loss of semantic fidelity. In addition, our canonicalized KGs improve performance on retrieval-augmented generation (RAG) question answering. Our code is publicly available at https://anonymous.4open.science/r/ less-is-more-kg-3875
[1-20] Edge-Induced Subgraph Representation Learning (KDD 2026)
백승렬(석사과정)
A variety of approaches have been proposed for subgraph-level representation learning. However, these approaches have primarily been developed and evaluated under node-induced subgraph settings, where each subgraph is defined by a selected set of nodes. In contrast, subgraph prediction tasks in which subgraphs are induced by selected sets of edges remain largely unexplored, despite arising naturally in domains such as knowledge graph reasoning, scene graph understanding, and functional connectivity analysis in network neuroscience. Edge-induced subgraph prediction introduces two technical requirements beyond those of the node-induced setting: (1) sensitivity to subgraph-internal edge structure and (2) isolation of subgraph-specific information within a mini-batch. To address these requirements, we introduce the segregated graph, a construction that represents the internal structure of each subgraph via subgraph-specific copies of base-graph nodes connected only by the edges selected for that subgraph. We perform message passing in parallel on the segregated graph and the base graph, and fuse the resulting representations at each layer through identitybased mixing, thereby combining internal-structure awareness with boundary and global contextual information. Experiments on three benchmarks derived from DocRED, Visual Genome, and the Human Connectome demonstrate that our method consistently outperforms existing subgraph prediction approaches, confirming the effectiveness of jointly modeling the segregated graph and the base graph.
[1-21] Out-of-Distribution Robust Explainer for Graph Neural Networks (KDD 2026)
한건희(박사과정)
Graph Neural Networks (GNNs) have become widely used for analyzing graph-structured data, motivating post hoc explanation methods for interpreting pre-trained GNNs. However, most existing explainers have primarily been designed and evaluated in settings where the inference graph is structurally or distributionally similar to the graph used to train the GNN. This setting can be fragile in practical deployment, where inference time graphs may contain newly introduced nodes and edges that deviate substantially from the training distribution. Such Out-of-Distribution (OOD) entities can introduce spurious signals during explanation generation, causing explainers to highlight unreliable graph components. Nevertheless, the robustness of post hoc GNN explainers under OOD-contaminated inference graphs remains largely unexplored. To address this gap, we propose ORExplainer (Outof-Distribution Robust Explainer), a post-hoc node-level explanation framework for inference graphs containing OOD nodes. ORExplainer quantifies node-level distributional deviation using an energy-based score computed from a pre-trained GNN, and propagates this score to learn explanations that emphasize In-Distribution (ID) nodes while suppressing OOD influence. Experiments on synthetic and real-world datasets show that ORExplainer consistently provides more robust explanations across diverse node-level OOD settings, outperforming existing baselines. Our code is available at https://github.com/gunhee8178/ORExplainer.
[1-22] S2R-Merge: Soft-Gated Routing with Structural Node Merge Optimization (CIKM 2026)
박세현(석사과정)
Large Language Model (LLM)-based Multi-Agent Systems (MAS) have shown strong potential on complex reasoning tasks, but their scalability is limited by substantial communication overhead and token consumption. Recent methods reduce these costs by pruning redundant agents or edges, but such deletion-based optimization can discard role-specific information and often operates only after activating the full agent pool. To address these limitations, we propose S2R-Merge, a unified routing-and-merging framework that reframes MAS structural optimization from node deletion to information-preserving consolidation. S2R-Merge first uses an uncertainty-aware 2-Stage Soft-Gated Router to select a task-relevant agent subgraph, avoiding unnecessary full-pool activation. It then applies graph-based node merging to identify structurally inefficient agents from policy-gradient-learned communication pathways, absorbing them into strongly related partners while preserving their role-level information through Prompt-level Fusion. Experiments on six reasoning and code-generation benchmarks with Llama-3-8B-Instruct and GPT-4 show that S2R-Merge achieves the best average accuracy among all compared methods, reaching 77.22% and 92.37%, respectively. It also reduces average prompt token usage by up to 12.3× compared with full multi-agent collaboration and achieves the lowest token consumption among multi-agent graph-optimization baselines. These results demonstrate that S2R-Merge provides an effective cost-performance trade-off for scalable LLM-based MAS.