Searcharxiv⌕ Search

arXiv subjects

Yibo Zhao

Publications and source records attributed to Yibo Zhao.

At least 19 recordsLinked to original sources
arXiv2026

NORmal: Architecting 3D Ferroelectric NOR Flash as Main Memory -- From Co-Fabricated BEOL Arrays to 200B-Class Multi-Agent LLMs in a Smartphone

Smartphone agents increasingly work as teams, in which a large planner directs parallel workers across long sessions. Running such a team on the device requires 200B-class models and contexts that exceed phone DRAM many times over, yet existing tiers treat flash as storage: UFS starves per-token weight streaming, and mobile High-Bandwidth Flash (HBF) needs a redesigned SoC and cannot absorb continuous KV writes. We show that a non-volatile memory can join an unmodified LPDDR6 bus as a plain rank only if its access latency fits the DRAM timing window, a condition no NAND tier meets. We present NORmal, to our knowledge the first main-memory architecture for 3D ferroelectric NOR flash. We have fabricated and experimentally characterized back-end-of-line (BEOL) 3D FeNOR arrays that co-integrate dense 3-bit MLC and fast, high-endurance in-place-writable SLC with sub-7.5 ns sensing. NORmal turns these co-fabricated arrays into a byte-addressable, deterministic-timing load/store rank on an unmodified LPDDR6 interface through an under-array DRAM-emulation bridge, staggered bank reads with zero-cycle weight swizzling, and posted SLC writes. Across this rank and DRAM, NORmal places agent state by mutability and lifetime rather than by hotness, serving 200B-class models and persistent planner contexts in place while dynamically partitioning subchannels across agent phases. In serving-level simulation on production agent traces, NORmal is 8.9$\times$--28.0$\times$ faster than UFS offloading, within 0.4% of an unlimited-DRAM ideal, and 2.8$\times$--6.3$\times$ faster than HBF on sessions beyond 256K tokens, with 8.1$\times$--17.8$\times$ less memory energy per token.

cs.AR↗
arXiv2026

FlatCake: A Flat KVCache Storage Architecture for LLM Serving

KV cache reuse accelerates LLM serving, but existing systems manage host DRAM and NVMe SSD as a tiered hierarchy where DRAM acts as an intermediate cache. Production LLM workloads break both premises of this hierarchy. First, bimodal reuse defeats DRAM temporal filtering, since workloads are dominated either by one-shot blocks with long reuse distances or by multi-gigabyte agentic sessions that exceed DRAM capacity, putting SSD on the critical read path. Second, DRAM staging and eviction-ordered layout bottleneck SSD restoration. Holding host staging buffers for in-flight prefixes causes contention under load that aborts prefetches, while incremental eviction scatters a session's blocks across the SSD and prevents large sequential reads. We present FlatCake, a flat KV storage architecture that exposes host DRAM and NVMe SSD as peer media in a unified 64-bit address space. Exploiting the fact that KV blocks are write-once, immutable, and reused at session scope, FlatCake places each block directly onto its target medium in session-contiguous, layer-major order at creation time and never migrates it. FlatCake streams coalesced SSD reads directly into a bounded GPU window via GPUDirect Storage and decouples DRAM and SSD request queues to prevent head-of-line blocking. On a production Claude Code trace with DeepSeek-V4-Flash, FlatCake sustains a 94% cache hit rate as load increases (versus 64% for Mooncake), eliminates Mooncake's $6.5\times$--$6.7\times$ SSD read amplification, and reduces peak-load mean TTFT by $1.6\times$ while improving total token throughput by $1.6\times$. Across 8K--128K contexts on Llama-3.1-70B, FlatCake further reduces mean TTFT by up to $2.4\times$ over Mooncake and $1.8\times$ over Mooncake-GDS.

cs.DC↗
arXiv2026

CrossPair: Unlocking Generalist Visual Transformation Learning in Image Editors

Language models switch tasks from demonstrations, but vision still often relies on separate systems for restoration, editing, and stylization. A before--after example offers a shared instruction only if the transferable change can be separated from the scene that demonstrates it. We introduce CrossPair, which adapts a pretrained multimodal image editor to transfer diverse transformations from one exemplar pair. Exemplar-before, exemplar-after, and query images take fixed roles in its native VLM and VAE pathways: the VLM extracts operation semantics, while the VAE retains fine appearance and structure. A training-only, change-aware DINOv3 objective aligns feature-change directions across content-distinct scenes, guiding the output to follow the demonstrated relation without an inference-time encoder. We construct Transform-360K by grouping filtered pairs from conditional representations, restoration, editing, virtual try-on, and stylization under shared transformations. On Relation252K-derived seen and held-out transformations, \method improves target agreement and edit-direction similarity over released systems. Ablations show that the semantic and visual exemplar routes are complementary and that DINO relation alignment further strengthens transfer; Transform-360K training also improves aggregate held-out-operation scores.

cs.CV↗
arXiv2026

OS-Wizard: Scaling Computer-Using Agents to Specialized Software via Knowledge-Guided Exploration

Computer-using agents are increasingly capable of performing complex tasks in the digital world. However, high-quality interaction trajectories remain scarce for specialized desktop applications, where limited application knowledge, cumbersome GUI workflows, and incomplete exploration hinder automated collection. We introduce OS-Wizard, an automated trajectory collection framework. It combines tutorial-guided exploration with a hybrid GUI--Code action space, providing application-specific knowledge and simplifying GUI workflows through direct execution of application-native commands. The framework also checkpoints and revisits pivotal states during tree exploration to explore alternative workflows, complementing linear rollouts for broader coverage. We collect 10.9K trajectories comprising 54.9K state--action samples. Fine-tuning open-source base models on this corpus yields a best-performing model that improves performance on a benchmark of specialized desktop workflows by 58.2 percentage points to 64.1\%, surpassing reported human performance of 59.8\% and establishing a new state of the art. The fine-tuned model also improves performance on general desktop workflows using GUI actions only, suggesting that the collected supervision benefits general desktop tasks as well. Comparative experiments and ablations demonstrate the efficiency of trajectory collection and the importance of code actions.

cs.AI↗
arXiv2026

CAVC: Convergence-Adaptive Audio-Visual Cache for Modality-Asymmetric Diffusion Generation

Diffusion-based audio-visual generation methods typically perform joint denoising of audio and video latents throughout the entire sampling trajectory, repeatedly executing modality-specific denoising branches and cross-modal attention at every timestep. However, we observe that such uniform joint computation is not always necessary. The key is a modality-asymmetric recoverability: audio-side drift introduced by mid-stage caching can be substantially reduced by only a few terminal joint denoising steps, whereas video-side degradation caused by caching leads to persistent loss of motion and temporal details. This asymmetry suggests that inference acceleration should selectively cache recoverable audio-side computation while preserving full video refinement. Motivated by this observation, we propose \textbf{Convergence-Adaptive Audio-Visual Cache (CAVC)}, a training-free modality-asymmetric caching framework. CAVC adopts a sandwich-structured computation allocation: after detecting that audio denoiser outputs enter a cache-tolerant regime, it caches and reuses the audio denoiser output to bypass audio denoising and cross-modal attention in the middle stage, while the video branch continues independent spatio-temporal refinement; a few joint steps are then restored at the end to reduce accumulated cross-modal drift and improve synchronization. Experiments across multiple public benchmarks demonstrate that CAVC achieves 1.10--1.35x wall-clock acceleration, improves several visual dynamics and temporal statistics, and maintains competitive audio fidelity and audio-visual alignment.

cs.CV↗
arXiv2026

Relation-Aware Graph Foundation Model

In recent years, large language models (LLMs) have demonstrated remarkable capability to generalize across diverse natural language processing tasks, inspiring the development of graph foundation models (GFMs) for large-scale pre-training. However, unlike language models with explicit token units, graphs lack a well-defined unit for generalization, making it challenging to design effective pre-training strategies. In this work, we propose REEF, a novel GFM framework that leverages relation tokens as the fundamental units. We construct a vocabulary of relation tokens to encode relational information within graphs. To accommodate diverse relations, we introduce two hypernetworks that adaptively generate the parameters of aggregators and classifiers in graph neural networks based on relation tokens. In addition, we design another hypernetwork to construct dataset-specific projectors and incorporate a dataset-level feature bias into the initial node representations, enhancing flexibility across different datasets with the same relation. Extensive experiments demonstrate that REEF consistently outperforms existing methods in both pre-training and transfer learning, highlighting its potential as a general-purpose graph foundation model.

cs.LG↗
arXiv2026

Not Every Token Is Worth Distilling: Selective Supervision for Direct-OPD

Direct On-Policy Distillation (Direct-OPD) transfers reinforcement-learning-induced policy improvements from a small model to a larger student by using the token-level log-ratio between post-RL and pre-RL checkpoints as dense supervision on the student's own rollouts. This transfer rewards the policy shift at every state, yet the log-ratio measures only relative change: it can stay fixed even as the probability mass that both checkpoints assign to the student's candidate tokens vanishes. Through an exact construction, we show that the Direct-OPD reward and its update can remain unchanged while the Jensen-Shannon divergence (JSD) and both KL directions between the checkpoints vanish with this mass, and we note that a small JSD bounds how much the teacher's behavior changed. Motivated by this analysis, we propose Selective Supervision for Direct-OPD (S$^2$D-OPD), which ranks student-sampled states by their teacher-reference JSD and masks Direct-OPD supervision at low-divergence states, retaining only the top 10% of states per response. Across two teacher pairs and four student models ranging from 1.7B to 8B parameters, S$^2$D-OPD improves held-out accuracy over dense Direct-OPD on AIME and HMMT benchmarks in seven of eight settings and matches it in the eighth, without extra forward passes.

cs.LG↗
arXiv2026

HBF Sucks? A Full-Stack Characterization of High-Bandwidth Flash for KV-Centric LLM Serving

A faster storage device should make serving faster. We find the opposite. High-Bandwidth Flash (HBF) stacks NAND behind a wide, package-local interface, promising flash-scale capacity with far lower read latency and higher bandwidth than an SSD. The obvious move is to keep an SSD-style Mooncake KV-offloading stack and swap in HBF underneath. We built that system and measured it: an extended TokenSim, four complete two-hour Qwen-Bailian production traces, five dense and mixture-of-experts models, and H100/B200 profiles. The upgrade backfires. Average end-to-end latency rises 2--5.5$\times$ and maximum SLO goodput falls 1.1--2.7$\times$ across H100 and B200, so the faster device yields a slower system. A cost-benefit model explains the paradox: a faster far tier pays off only when read I/O is the bottleneck, reads outweigh writes, and delivered bandwidth is sustainable. Transient KV violates all three at once. Buying flash through the package costs GPU near-tier capacity and bandwidth, while HBF's own read/write latency barely matters: scaling it 3.75$\times$ moves latency less than 1\%. Worse, the two-tier hierarchy keeps reuse in the near tier and hands HBF a relentless write-heavy stream. Writes outnumber reads on every trace, so a 3D-ICE model shows the stack hits its thermal limit well below peak bandwidth, and a TLC tier wears out sooner than the SSD pool it replaced. The device is fine; the drop-in deployment is not. HBF sucks as an SSD replacement for transient KV, but earns its place in LLM serving when used selectively with reuse-aware placement, write budgeting, and thermal coordination.

cs.AR↗
arXiv2026

MERIT: Matching Expertise via Rubric-Informed Training for Reviewer Assignment

Matching submissions with suitable reviewers at scale is a growing challenge for major venues, yet existing approaches either rely on coarse proxy signals that conflate general relatedness with true suitability, or require expensive human annotations that are difficult to scale for training. We propose MERIT, a two-stage framework that bridges this gap by converting criterion-level expertise matching into scalable suitability supervision. In the first stage, we train a reviewer assessor via reinforcement learning to identify the expertise dimensions a paper requires, match them against the reviewer's prior work, and produce a suitability decision, with rewards provided by an LLM judge guided by paper-specific expertise rubrics. In the second stage, we distill the assessor's predictions into an embedding-based retriever for efficient large-scale assignment. Experiments show that our 4B reviewer assessor outperforms larger general-purpose LLMs on suitability classification, and the resulting retriever achieves state-of-the-art performance across LR-Bench and the CMU Gold dataset. Our code is available at https://github.com/Luli3220/MERIT.

cs.CL↗
arXiv2026

Preserving Long-Tailed Expert Information in Mixture-of-Experts Tuning

Despite MoE models leading many benchmarks, supervised fine-tuning (SFT) for the MoE architectures remains difficult because its router layers are fragile. Methods such as DenseMixer and ESFT mitigate router collapse with dense mixing or auxiliary load-balancing losses, but these introduce noisy gradients that often degrade performance. In preliminary experiments, we systematically pruned experts and observed that while certain super experts are activated far more frequently, discarding less used experts still leads to notable performance degradation. This suggests that even rarely activated experts encode non-trivial knowledge useful for downstream tasks. Motivated by this, we propose an auxiliary-loss-free MoE SFT framework that combines bias-driven sparsification with always-active gated condenser experts. Rather than enforcing balanced activation across all experts, our method encourages task-relevant experts to remain active while pushing long-tailed experts toward inactivity. The condenser experts provide a persistent, learnable pathway that alleviates gradient starvation and facilitates consolidation of information that would otherwise remain fragmented across sparsely activated experts. Analysis further suggest that this design better preserves long-tailed expert information under sparse routing. Experiments on large-scale MoE models demonstrate that our approach outperforms state-of-the-art SFT baselines such as DenseMixer and ESFT, achieving average gain of 2.5%+ on both mathematical reasoning and commonsenseQA benchmarks.

cs.LG↗
arXiv2026

AnyGS2Mesh: Feed-Forward Mesh Reconstruction from 3D Gaussian Splatting with Arbitrary-Resolution Views

Existing 3D mesh reconstruction methods from Gaussian scene representations predominantly rely on iterative optimization, resulting in slow inference and limited scalability to high-resolution inputs. In this paper, we present AnyGS2Mesh, the first feed-forward framework for directly reconstructing 3D meshes from 3D Gaussian Splatting representations with support for arbitrary input image resolutions. Our approach incorporates a Gaussian-Guided Transformer architecture that exploits explicit 3D geometric priors for efficient mesh generation. We introduce three key components: (1) a Gaussian-Guided Spatial Reasoning Transformer represents Gaussian primitives as structured 3D tokens and jointly reasons over Gaussian and image features; (2) a Streaming and Patchwise Geometry Encoder processes native-resolution views sequentially and aggregates information across variable-length view sets; (3) a Scale-Aligned Hybrid Depth Refiner uses a PatchFusion-style encoder--decoder to fuse RGB-conditioned predicted depth with Gaussian-rendered metric depth, combining fine local structures with globally consistent metric scale. The refined depth maps are integrated through TSDF fusion, followed by Marching Cubes for deterministic mesh extraction. Extensive experiments show that AnyGS2Mesh achieves state-of-the-art reconstruction quality while significantly reducing inference time compared with optimization-based baselines, enabling near-real-time, high-quality mesh reconstruction. Our results demonstrate the potential of combining Gaussian representations and feed-forward Transformer architectures for scalable 3D geometry reconstruction. The code will be made publicly available upon acceptance.

cs.CG↗
arXiv2026

TokenStack: A Heterogeneous HBM-PIM Architecture and Runtime for Efficient LLM Inference

Large language model (LLM) serving is now limited by the key-value (KV) cache. During decode, each new token rereads prior KV state, so attention becomes a bandwidth- and capacity-heavy memory task. HBM-PIM helps by moving attention closer to memory, but current stack organizations still waste resources. In practice, only hot KV blocks benefit from near-memory compute. Weights, activations, and cold KV mainly need dense storage and GPU-visible bandwidth. A uniform HBM-PIM stack makes all layers pay for PIM logic, while a dedicated-PIM design such as AttAcc recovers capacity but shrinks the HBM bandwidth left for GPU-side work. We propose TokenStack, a vertically heterogeneous HBM-PIM architecture for KV-centric LLM serving that leverages HBM4's logic-die substrate. TokenStack separates each stack into dense capacity layers and PIM-enabled compute layers, then uses the logic base die as a stack-local control point that manages cross-layer movement without host-side overhead. The base-die controller handles cross-layer DMA, layered address translation, attention-side gather/broadcast coordination, and inline quantization during migration. On top of this hardware, TokenStack uses topology-aware KV placement, workload-aware eviction, and bounded replication to keep hot KV near PIM compute while moving colder state to dense layers. Using production-derived traces across four models, completed multi-QPS runs show that TokenStack increases geometric-mean token throughput by 1.62x and SLO-compliant serving capacity by 1.70x over AttAcc, and reduces per-token energy by 30-47%.

cs.AR↗
arXiv2026

LazyMem: Retrieve Broadly, Construct Selectively for Efficient Long-Term Agent Memory

Long-term memory enables LLM agents to leverage past interactions, but dialogue histories quickly exceed the context window, forcing agents to retrieve relevant subsets at query time. Because useful evidence is sparse and scattered across verbose conversations, retrieval faces a fundamental tension: broadening recall improves coverage but floods downstream reasoning with noise, while compressing memories at write time eases retrieval but irreversibly discards details that future queries may need. We introduce LazyMem, which resolves this tension by deferring all memory construction to query time. Given a retrieved candidate pool, a lightweight model processes it in overlapping parallel windows, selectively retaining and compressing only query-relevant content. The model is trained with supervised fine-tuning followed by reinforcement learning, using a reward that jointly encourages the identification of relevant messages and the generation of compressions that are faithful to the source and useful for answering the query. On LongMemEval, LazyMem-4B achieves an LLM-judge accuracy of 0.85, outperforming the strongest non-oracle baseline while using only 213 answer-context memory tokens, 21.0 times fewer than the baseline. It further generalizes to LoCoMo without target-domain training and reduces mean latency relative to the prior query-time baseline. Code is available at https://github.com/allacnobug/LazyMem.

cs.AI↗
arXiv2026

PRGCN: A Graph Memory Network for Cross-Sequence Pattern Reuse in 3D Human Pose Estimation

Monocular 3D human pose estimation remains a fundamentally ill-posed inverse problem due to the inherent depth ambiguity in 2D-to-3D lifting. While contemporary video-based methods leverage temporal context to enhance spatial reasoning, they operate under a critical paradigm limitation: processing each sequence in isolation, thereby failing to exploit the strong structural regularities and repetitive motion patterns that pervade human movement across sequences. This work introduces the Pattern Reuse Graph Convolutional Network (PRGCN), a novel framework that formalizes pose estimation as a problem of pattern retrieval and adaptation. At its core, PRGCN features a graph memory bank that learns and stores a compact set of pose prototypes, encoded as relational graphs, which are dynamically retrieved via an attention mechanism to provide structured priors. These priors are adaptively fused with hard-coded anatomical constraints through a memory-driven graph convolution, ensuring geometrical plausibility. To underpin this retrieval process with robust spatiotemporal features, we design a dual-stream hybrid architecture that synergistically combines the linear-complexity, local temporal modeling of Mamba-based state-space models with the global relational capacity of self-attention. Extensive evaluations on Human3.6M and MPI-INF-3DHP benchmarks demonstrate that PRGCN establishes a new state-of-the-art, achieving an MPJPE of 37.1mm and 13.4mm, respectively, while exhibiting enhanced cross-domain generalization capability. Our work posits that the long-overlooked mechanism of cross-sequence pattern reuse is pivotal to advancing the field, shifting the paradigm from per-sequence optimization towards cumulative knowledge learning.

cs.CV↗
arXiv2026

Think Big, Search Small: Where Capacity Matters in Hierarchical Search Agents?

Large language model based search agents increasingly adopt multi-agent architectures in which a main agent decomposes a complex question into sub-queries and dispatches them to parallel sub-agents. However, existing systems instantiate all roles from a single model of identical scale, leaving open how model capacity should be distributed across roles. We factorize hierarchical search into three roles: a delegation role responsible for task decomposition, an execution role responsible for retrieval and evidence extraction, and an answer generation role held fixed as a confound control. We then conduct controlled capacity sweeps along the delegation and execution axes on five multi-hop QA benchmarks. The experiments yield three findings. First, role factorization consistently outperforms a single-agent baseline, improving exact match from 4.5 to 8.6 points across six model scales. Second, capacity sensitivity is asymmetric: scaling the delegation backbone improves EM by ~11 points, whereas scaling the execution sub-agent moves EM by only ~2.6 points, identifying decomposition as the capability bottleneck. Third, a 1.7B-parameter executor trained via quality-filtered trajectory distillation matches a frontier sub-agent in accuracy while consuming 37% fewer sub-agent tokens, advancing the Pareto frontier. These results suggest a concrete recipe for building hierarchical search agents: concentrate capacity at delegation and downsize execution without sacrificing accuracy. Our code is available at https://github.com/QinnanCai0115/role-factorized-search.

cs.CL↗
arXiv2026

MacAgentBench: Benchmarking AI Agents on Real-World macOS Desktop

Computer use agents (CUAs) have advanced rapidly in desktop automation, and a growing number of users deploy CUAs such as OpenClaw on Mac Mini for always-on automation. However, existing benchmarks, including those for macOS, evaluate agents without framework augmentation and rely on binary evaluation. As a result, they fail to capture both the framework capabilities leveraged by modern CUAs and the partial progress on long-horizon, multi-application tasks. We present MacAgentBench, a comprehensive macOS agent benchmark comprising 676 tasks across 25 applications, with nearly 60% involving both GUI and CLI interaction. The benchmark adopts deterministic rule-based evaluation and introduces fine-grained multi-checkpoint scoring with capability annotations for multi-application tasks. Experiments across three frameworks and 16 models show that the best configuration, Claude Opus 4.6 on OpenClaw, attains 73.7% Pass@1, while this advantage is primarily driven by the skill library rather than by framework design. Fine-grained metrics further reveal that models with similar Pass@1 can differ substantially in sub-goal completion. Our code and data are publicly available at https://github.com/JetAstra/MacAgentBench.

cs.AI↗
arXiv2026

From Patchwork to Network: A Comprehensive Framework for Demand Analysis and Fleet Optimization of Urban Air Mobility

Urban Air Mobility (UAM) presents a transformative vision for metropolitan transportation, but its practical implementation is hindered by substantial infrastructure costs and operational complexities. We address these challenges by modeling a UAM network that leverages existing regional airports and operates with an optimized, heterogeneous fleet of aircraft. We introduce LPSim, a Large-Scale Parallel Simulation framework that utilizes multi-GPU computing to co-optimize UAM demand, fleet operations, and ground transportation interactions simultaneously. Our equilibrium search algorithm is extended to accurately forecast demand and determine the most efficient fleet composition. Applied to a case study of the San Francisco Bay Area, our results demonstrate that this UAM model can yield over 20 minutes' travel time savings for 230,000 selected trips. However, the analysis also reveals that system-wide success is critically dependent on seamless integration with ground access and dynamic scheduling.

cs.DC↗
arXiv2026

Retrieval, Reward, and Training Protocols: What Matters in Training Search Agents?

Search agents powered by large language models can autonomously decompose queries, retrieve information, and synthesize answers through multi-step reasoning. However, the rapid growth of training methods has outpaced controlled comparison: existing works differ in retrieval corpora, reward designs, and training protocols, making it unclear what actually drives improvements. We present a controlled empirical study that isolates three under-explored dimensions of search agent training. First, we identify a critical data-coverage issue in the widely used Wikipedia 2018 corpus and show that correcting it alone yields larger gains than the differences between training algorithms. Second, we systematically compare outcome-based and process-based reward methods across three base models, finding that the simplest outcome-based approach achieves competitive or superior performance in most settings, and that process-level credit assignment can over-correct agent behavior. Third, we analyze training data diversity, off-policy data utilization, and search budget scaling, distilling practical guidelines for training effective search agents. Our code is available at https://github.com/YiboZhao624/SearchAgentReview.

cs.CL↗