arXiv subjects

Junjie Wang

Publications and source records attributed to Junjie Wang.

5 recordsLinked to original sources

arXiv2026

PCoMoE: Shifting MoE Inference from Monolithic Expert Selection to Fine-Grained Path Composition

Mixture-of-Experts (MoE) architectures scale Large Language Model (LLM) capacity efficiently by activating a sparse subset of experts per token. However, modern MoE inference remains heavily constrained by the rigid, whole-expert abstraction. Existing frameworks manage, schedule, or prune experts as atomic execution units, which fixes the optimization boundary too early and leaves fine-grained intra-expert computational redundancy underexplored. In this work, we present PCoMoE, a path-compositional execution framework that shifts MoE inference from coarse-grained expert selection to fine-grained path composition. PCoMoE incorporates a path-level formulation of expert computation, a compatibility-aware layer-wise pruning strategy to suppress low-value path combinations, and a hardware-friendly execution engine to exploit reusable sub-expert structures under strictly bounded overheads. Experimental results demonstrate that PCoMoE achieves up to a 1.31x end-to-end inference speedup while enhancing model accuracy by 10%. The code is available at https://github.com/gzyyy0/PCoMoE

cs.CL↗

arXiv2026

On the Design of Qwen3.8-Next Architecture: Evaluation, Efficiency, and Training Stability

We describe the architecture and ablations of Qwen3.8-Flash-Next, a sparse mixture-of-experts model with 125B parameters, 6B activated per token, and additional 51B parameters of n-gram embedding tables held off the accelerator. On fourteen pre-training benchmarks the model leads the 397B-A17B predecessor on eight and trails it on the rest by at most 2.6 points, at 1/3 the activated parameters, 1/3 the training tokens, and roughly 1/9 the training FLOPs. Token mixing uses a layer-wise hybrid of Gated DeltaNet (GDN) and global attention, with one full-attention layer in every four; at continued-pretraining time those full-attention layers are replaced by Qwen Sparse Attention (QSA), which scores context at micro-block granularity with a compressed lightweight indexer. The residual stream is widened to four branches and read through an elementwise gate, a design we call the Gated Residual (GR). Capacity is added outside the backbone by a single n-gram embedding layer whose tables are prefetched from host memory. We evaluate every candidate change along three axes: loss together with downstream benchmarks; the cost of the change in training, prefill and decode; and its effect on the optimal hyperparameters and training stability. Loss and downstream accuracy do not always move together: enlarging the n-gram vocabulary lowers loss monotonically while downstream accuracy saturates. The architecture and the Muon optimizer together shift the optimal learning rate and batch size upwards, render batch-size warmup unnecessary, and substantially improve stability under stress tests. Loss, benchmarks, efficiency and stability form one design problem. Solved jointly, they yield a recipe that is simultaneously more efficient, more capable and more stable.

cs.CL↗

arXiv2026

Learning Where Outcomes Change:Credit-Addressable Reasoning for Multimodal Geometry

Multimodal geometry reasoning requires VLMs to extract precise visual relations and preserve them through multi-step deduction. Existing free-form traces obscure the decisions that determine the answer, and trajectory-level reinforcement learning distributes a single terminal signal across the entire response. We introduce credit-addressable reasoning, in which the semantic units exposed during inference also define where learning compares alternatives and assigns credit. We instantiate this principle with Code-CoT, which retains the diagram, represents visual relations as line-addressable executable code, and organizes reasoning into typed events, and CE-GRPO, which selects event boundaries using structural priors and type-normalized entropy, samples complete continuations from shared prefixes, and converts outcome differences into localized advantages. Across nine geometry benchmarks, CE-GRPO achieves an average accuracy of 76.04, outperforming Qwen3-VL-8B and trajectory-level GRPO by $8.09$ and 3.43 points, respectively. Its relative advantage increases with the number of intermediate events, demonstrating the value of representation--optimization co-design for long, dependency-heavy multimodal reasoning.

cs.LG↗

arXiv2026

Propagating construction-time knowledge quality into medical question answering: A framework grounded in clinical guidelines

Large language models have facilitated knowledge graph (KG) construction from clinical guidelines, but extracted triples vary in structural validity and evidential support. Meanwhile, graph-augmented question answering (QA) systems typically optimize query relevance during retrieval, with limited reuse of quality information produced during KG construction. This creates a disconnect between construction-time quality control and inference-time evidence use. We investigate whether construction-time triple quality can serve as a persistent signal for downstream evidence selection and presentation. We propose a quality-aware framework that models structural conformance (SchemaConf) and evidential support (EvidScore) as complementary dimensions and fuses them into a per-triple quality signal, Q(t). Rather than using quality solely for filtering, the framework retains Q(t) and derived quality tiers as graph attributes and propagates them into quality-weighted subgraph retrieval and tier-conditioned evidence prompting, while preserving passage-level provenance. Experiments on Chinese diabetes clinical guidelines show that the utility of the quality signal is distribution dependent. Under cross-version and cross-model shift, the fused Q(t) provides stronger triple-quality discrimination than either component alone (AUC 0.748 vs. 0.703 for EvidScore and 0.645 for SchemaConf). In guideline-grounded QA, propagating construction-time quality reduces required-knowledge omission from 16.3% to 5.3% and conflicting outputs from 16.3% to 2.7%, with an evidence-grounded precision of 81.6% and near-zero invalid citations. Blinded clinician ratings favor the full framework over no retrieval (4.68 vs. 4.21 on a five-point scale) and approach the oracle condition (4.80), while cross-generator experiments show consistent trends.

cs.AI↗

arXiv2026

Skim and Skip: Hierarchical Adaptive Inference for Efficient Multimodal Retrieval

Universal multimodal retrieval (UMR) increasingly adopts multimodal large language models (MLLMs) as unified embedding backbones, but their strong retrieval performance comes at substantial inference cost. Existing methods typically rely on uniformly dense inference, where all input tokens are processed through the entire model and matched using the final-layer [EOS] representation. However, this paradigm overlooks two key forms of heterogeneity in multimodal retrieval: token contributions to the final retrieval embedding are highly uneven, and different queries require markedly different amounts of inference depth. To address this, we propose Skim and Skip (SAS), a hierarchical adaptive inference framework for efficient multimodal retrieval. SAS first performs token-level evidence selection to preserve only the input information most relevant to the final retrieval embedding, and then performs depth-adaptive inference to determine whether the current representation is already sufficient for reliable matching. Experiments on 12 MMEB retrieval tasks show that SAS retains about 99% of the dense baseline's average retrieval performance while achieving up to 1.64 times end-to-end speedup and up to 66.3% FLOPs reduction.

cs.IR↗