NeurIPS 2026 论文列表

Mixture-of-Experts (MoE) architectures have become essential for building capable large language models, with recent work demonstrating the benefits of fine-grained expert designs. However, training MoE models from scratch is computationally expensive, and existing upcycling methods that convert dense models to MoE face limitations: they either initialize experts as identical copies (limiting routing diversity) or use random perturbations (risking knowledge loss). We propose \textsc{DivMoE}, a framework that addresses these challenges through two innovations. First, we introduce \emph{domain-specialized expert initialization}, deriving fine-grained experts from models fine-tuned on distinct domains (mathematics, code, science, commonsense), providing meaningful diversity while preserving pre-trained knowledge. Second, we propose \emph{diversity-constrained routing}, which enforces that each token selects at most one expert per domain group, structurally preventing routing collapse and enabling cross-domain knowledge composition. Experiments on two base models demonstrate that \textsc{DivMoE} outperforms all upcycling baselines (47.0% vs.\ 44.4% average accuracy) and achieves competitive performance with MoE models trained from scratch, matching Moonlight-MoE at 64.5% average while being constructed via efficient upcycling.

Mixture-of-Experts (MoE) transformers scale capacity by activating only a few experts per token, but this sparsity creates a hidden reliability problem: when routing is imperfect, load-balanced models may send tokens to experts that are insufficiently trained for the assigned inputs. We propose Distributionally Robust MoE Training (DRMoET), a drop-in objective that treats layer-wise experts as endogenous robustness groups and optimizes high-loss routing outcomes rather than merely equalizing traffic. DRMoET updates a per-layer expert distribution by an entropy-regularized softmax rule on EMA-smoothed, activation-weighted expert losses, strengthening plausible non-top routing paths while preserving standard MoE computation. Under the FLAME-MoE recipe at 746M-total and 10.3B-total scales, DRMoET improves downstream averages over both standard FLAME-MoE and auxiliary-loss-free balancing. Mechanistic analyses show lower expert-loss variance with nearly unchanged mean loss, 4.3\% lower degradation under forced mid-$k$ misrouting, and improved domain--expert specialization. These results position routing robustness--not only utilization balance--as a practical objective for reliable sparse MoE scaling.

Multimodal Large Language Models (MLLMs) have shown strong multimodal instruction-following ability, but adapting them to diverse visual-language domains typically assumes centralized data access and costly joint training. This is restrictive when data is distributed across private, domain-specific, or permission-limited clients. To this end, we propose DISTMOE, a mixture-of-experts (MoE) approach for distributed visual instruction tuning. In each layer of the language decoder it augments the public feedforward network (FFN) with a client-specific private FFN expert, with the goal to acquire domain-specific knowledge. However, independent expert training causes the private FFNs to learn representation of different scale and magnitudes, making merging the experts difficult. To reduce client-specific drift, we introduce a public-anchored expert composition stage that updates only routers and lightweight private projection adapters on a mix of local client data and public data, via an isotropic regularization loss, therefore making it rehearsal-free. During inference, DISTMOE performs modular routing over public and private experts, enabling token-wise domain composition without explicit domain labels. Experiments across diverse visual-language benchmarks show that DISTMOE enables flexible expert reuse, effective domain adaptation, and competitive performance while preserving modular control over client-specific knowledge.

Large-scale pre-trained vision models provide powerful representations for class-incremental learning, yet continuously adapting them to new classes without historical samples, task identifiers, or unrestricted parameter growth remains a fundamental challenge. Existing LoRA-based approaches typically follow either a one-adapter-per-task expansion paradigm or a fixed parameter-sharing strategy. The former leads to linear parameter growth as the task sequence expands, while the latter often induces severe cross-task interference. To address these limitations, we propose ${\bf E}^3$ (Elastic Ensemble of Experts), a parameter-efficient framework for class-incremental learning. ${\bf E}^3$ treats LoRA modules as elastic experts whose task capacity is dynamically scheduled according to representational saturation. Within each expert, we further introduce Group-Aware Parameter Partitioning, which allocates LoRA parameters into disjoint task-specific subspaces using magnitude-based importance estimation, thereby mitigating parameter overwriting without additional gradient-sensitivity computation. Moreover, ${\bf E}^3$ incorporates hierarchical moment regularization and orthogonal gradient projection to constrain classifier drift and prevent new-task updates from disrupting historical feature subspaces. Extensive experiments on standard class-incremental learning benchmarks demonstrate that E³ achieves state-of-the-art or competitive performance, especially in long-horizon task sequences, while maintaining a favorable balance between stability, plasticity, and parameter efficiency.

Despite growing interest in on-device LLM deployment, large-scale Mixture-of-Experts (MoE) models remain impractical on resource-constrained devices, as sparse expert activation still requires all expert weights to be memory-resident. To address this, we identify three key observations for on-device MoE inference: (a) cache hit rates are fundamentally bounded even under Belady’s optimal policy; (b) these bounds are insufficient for low-latency on-device inference; and (c) prior cache-aware routing based on binary cache presence can induce routing instability. Based on these insights, we propose CECAR (Cache & Expert Co-Aware Routing), a MoE inference system that jointly considers expert routing and cache management. Unlike prior cache-aware routing methods, CECAR uses an ML-based cache policy to predict the expected reuse distance of each expert and prioritizes routing and execution toward experts with smaller predicted reuse distances. On the Qwen3-30B-A3B model, CECAR achieves 12.5 tokens/s on a consumer-grade GPU with only 16 GB of VRAM, where non-resident experts are fetched from SSD, delivering a 4.5× speedup over conventional MoE inference and exceeding the decode speed of baselines that keep the entire model fully resident in DRAM.

Large language model (LLM) agents often keep the needed value in their prompt and still fail to pass it to a later tool call. We study this failure at the level of \textit{field routing}: an earlier observation may expose an \texttt{order\_id}, policy flag, or nested item, and a later call must bind that field to the right argument. We formalize tool use as a boundary-indexed field-routing graph and introduce Information Utilization Rate (IUR), which measures whether each required dependency edge is realized in the later tool call. A 400-dependency expert audit, task-success correlations, and discriminant checks validate IUR as a field-use signal. On a 357-task public evaluation from ToolSandbox \texttt{STATE\_DEPENDENCY} (192 named scenario variants) and the original-repository $\tau$-bench retail/airline files (115+50 tasks), aggregate dependency-level IUR across GPT-4o, Claude-3.5-Sonnet, Llama-3.1-70B, and Qwen-2.5-72B falls from 451/548 (82.3\%) at boundary distance $\delta=1$ to 144/283 (50.9\%) at $\delta=5$, a 31.4-point drop, with mean context length only 3,094 tokens. Token-matched analyses and boundary transformations show that raw token length and lost-in-the-middle do not explain the curve by themselves. The same field-level view also leads to a compact intervention. Boundary-Aware Field Memory (BAFM) predicts live fields without gold dependencies at inference time and injects field reminders rather than whole observations. In paired GPT-4o runs, BAFM improves task success from 144/357 (40.3\%) to 192/357 (53.8\%), a 13.4-point gain over the base agent and 2.8 points over self-predicted retrieval-augmented generation. The results recast long-chain tool use as state propagation over fields: the central question is which prior values remain live for the next action.

Vision-language models (VLMs), such as CLIP, have shown remarkable transferability in the biomedical domain, and prompt tuning enables efficient adaptation with limited supervision. In practice, biomedical images exhibit pronounced acquisition-induced style shifts, especially across imaging sites and protocols, while cross-modality transfer further introduces compound shifts from different physical acquisition processes. However, existing prompt-tuning methods ignore acquisition-induced style shifts, making the prompts sensitive to source-specific style and limiting the transferability to unseen medical domains. To address this, we formulate acquisition-induced style shifts as mixture shifts, where each domain is viewed as a mixture of latent style components. Under this formulation, the target risk depends on both mixture weights and latent component risks, motivating prompt adaptation that accounts for style-specific variation rather than relying on a single source-specific prompt. In this paper, we propose SaMoE, a Style-aware Mixture-of-Experts Multimodal Prompt Learning framework for adapting VLMs to biomedical domains under acquisition-induced style shifts. To account for latent component risks in prompt adaptation, SaMoE represents multimodal prompts with latent style-conditioned prompt transformations and composes them through image-conditioned routing, enabling each image to induce style-adapted visual prompts. By extracting style cues from intermediate visual representations, the sparse expert composition adapts style-related prompt weights without domain labels and is injected across multiple VLM layers to align with hierarchical visual-language representations. Extensive experiments on 15 medical datasets across 12 modalities and 10 organs demonstrate significant improvements in both accuracy and generalizability over state-of-the-art methods.

Recent literature on fine-tuning Large Language Models highlights a fundamental debate. While Full Fine-Tuning (FFT) provides the representational plasticity required for high-entropy knowledge injection, Low-Rank Adaptation (LoRA) can match or surpass FFT performance because many tasks only require updates in a low-rank space and benefit from LoRA's additional regularization. Through empirical evaluation across diverse tasks (SQL, Medical QA, and Counterfactual Knowledge) and varying language models (Gemma-3-1B, Qwen2.5-1.5B, and Qwen2.5-3B), we verify both trends and demonstrate that relying solely on either static architecture is structurally limited. To address this challenge, we propose a Mixture of LoRA and Full (MoLF) Fine-Tuning, a unified framework that enables continuous navigation between both training regimes. MoLF dynamically routes updates between FFT and LoRA at the optimizer level to ensure that exact gradient signals are available to both experts throughout training, yielding stable training dynamics. For memory-constrained environments, we also introduce MoLF-Efficient, which freezes base weights and only routes updates among a pair of LoRA experts of potentially varying rank. Our evaluations show that MoLF either improves on or stays within $1.5$% of the better of FFT and LoRA across all settings, while MoLF-Efficient outperforms prior adaptive LoRA approaches by up to $20$% on Fact and $9$% on Med and SQL.

Medical Vision-Language Pre-training (VLP) leverages the semantic richness of radiology reports to learn 3D representations, yet current architectures remain fundamentally anatomy-agnostic. While clinical diagnostics rely on organ-specific context, standard encoders apply a uniform, dense parameterization to all volumetric patches, and existing Mixture-of-Experts (MoE) routers gate on isolated local features without explicit access to the organ composition of the scan or crop. We propose Anatomy-Activated Mixture-of-Experts (A$^{2}$MoE), a framework that incorporates anatomical inductive biases within the model's computational path. A$^{2}$MoE introduces two key components: (1) a Histogram Representation Router (HRR) that conditions expert selection on global visual context and a predicted anatomy-presence histogram, ensuring that parameters specialize in specific anatomical domains; and (2) a mask-free Organ-Query Multi-Scale Projector (OQMSP) that utilizes learnable queries to extract organ-level embeddings for fine-grained alignment without requiring manual annotations at inference. By routing at the patch level rather than the token level, A$^{2}$MoE reduces routing complexity by orders of magnitude while satisfying load balancing by construction. Pre-trained on over 60,000 multi-phase CT scans, A$^{2}$MoE achieves state-of-the-art performance across segmentation, classification, and detection benchmarks. Furthermore, we demonstrate that our anatomically-grounded inductive biases support robust generalization to MRI and ultrasound, establishing a scalable foundation for medical representation learning.

Mixture-of-Experts (MoE) architectures scale model capacity through sparse expert activation, but their deployment remains memory-bound because all expert weights must reside in memory. Mixed-precision quantization can substantially reduce this footprint by assigning different bit-widths to different experts. Existing approaches, however, typically rely on calibration data to estimate expert importance and determine bit allocation. For frontier MoE LLMs, the original training data, and hence the true training distribution, is proprietary and inaccessible. As a result, calibration sets are inevitably imperfect surrogates, which can misestimate expert utilization and lead to suboptimal bit allocation. Motivated by the substantial cross-expert quality variability observed in modern MoE models, we propose AlphaQ, a calibration-free bit-allocation method for MoE quantization. AlphaQ draws on Heavy-Tailed Self-Regularization (HT-SR) theory and follows a simple principle: experts with more heavy-tailed weight spectra are typically better trained and should receive higher bit-widths, while experts with weaker heavy-tailed structure can be quantized more aggressively. AlphaQ operationalizes this principle by measuring expert-wise spectral heavy-tailedness and solving a budget-constrained optimization problem that minimizes total quantization error under a global bit-budget constraint. Across several MoE models, AlphaQ consistently outperforms calibration-based baselines under matched bit budgets. Notably, on Qwen1.5-MoE, AlphaQ achieves near full-precision accuracy with an average expert precision of only 3.5 bits, while delivering more than 4$\times$ memory compression.

In training Mixture-of-Experts models, sharded data parallelism splits each expert’s parameters across GPUs. Before each layer runs, GPUs must use Allgather to rebuild the full weight matrix. This communication can take a large part of each training iteration. Prior work reduces this cost with lossy compression, but lossy methods can affect accuracy. We propose Adaptive Exponent Reuse Allgather, or AdERA, a lossless Allgather compression method for sharded MoE training. Our key observation is that, after a short warmup phase, most weight exponents stay the same across iterations. AdERA stores these exponents locally and sends only the sign and mantissa when the exponent has not changed. The receiver then rebuilds the exact original weight by combining the cached exponent with the received sign and mantissa. Because parameter matrices have different shapes across layers, AdERA compresses only when compression saves time. It also overlaps compression with Allgather communication and computation. Our real experiments and large-scale trace-driven simulator show that, compared with the lossless baseline, AdERA achieves a 3.70× speedup on 16 GPUs and is projected to reach 4.28× on 128 GPUs, while preserving bitwise-exact parameter reconstruction. Compared with the lossy baseline, AdERA achieves a 3.93× speedup on 8 GPUs and is projected to reach 4.42× on 128 GPUs. The source code is publicly available.

Time series forecasting models are increasingly scaled through large Transformer backbones, yet most existing approaches process all series through a shared dense computation path despite substantial heterogeneity in temporal structure. Mixture-of-Experts (MoE) offers a natural alternative by enabling conditional computation, but standard MoE routing leaves expert specialization weakly identified and often unstable during downstream adaptation. We propose AME-TS, a structure-guided sparse time series foundation model that aligns expert routing with interpretable temporal structure. AME-TS first uses a lightweight regime predictor to estimate series-level descriptors, including forecastability, seasonality, trend, and sparsity, and maps them to a soft structural prior over experts. This series-level prior guides token-level routing during training, encouraging structure-aligned specialization. On the GIFT-Eval benchmark, AME-TS delivers a strong accuracy-efficiency tradeoff across model scales: it substantially outperforms existing time series foundation models at small model scales and remains competitive with the strongest models at larger scales, while activating substantially fewer parameters through sparse routing. We further show that AME-TS learns more interpretable routing geometry and substantially more stable expert specialization than standard MoE during fine-tuning on the M5 dataset. These results suggest that structure-aware routing is an effective and reliable way to realize the benefits of sparse expert models for time series forecasting.

Self-attention has greatly contributed to the success of the widely used Transformer architecture by enabling learning from data with long-range dependencies. In an effort to improve performance, a gated attention model that leverages a gating mechanism within the multi-head self-attention has recently been proposed as a promising alternative. Gated attention has been empirically demonstrated to increase the expressiveness of low-rank mapping in standard attention and even to eliminate the attention sink phenomenon. Despite its efficacy, a clear theoretical understanding of gated attention's benefits remains lacking in the literature. To close this gap, we rigorously show that each entry in a gated attention matrix or a multi-head self-attention matrix can be written as a hierarchical mixture of experts. By recasting learning as an expert estimation problem, we demonstrate that gated attention is more sample-efficient than multi-head self-attention. In particular, while the former needs only a polynomial number of data points to estimate an expert, the latter requires exponentially many data points to achieve the same estimation error. Furthermore, our analysis also provides a theoretical justification for why gated attention yields higher performance when a gate is placed at the output of the scaled dot product attention or the value map rather than at other positions in the multi-head self-attention architecture.

Optimal transport (OT) provides a principled framework for mapping between probability distributions. Despite extensive progress, applying OT to large-scale data remains computationally demanding, and the resulting pointwise transport plans are often difficult to interpret. We introduce Optimal Mixture Transport (OMT), a scalable framework that shifts the transport paradigm from individual samples to mixtures of subpopulations, reformulating the transport problem as a strictly biconvex optimization with a unique global minimizer. We further establish theoretical guarantees on the stability of the OMT map, showing that bounded perturbations of the underlying distributions lead to bounded changes in the transport plan. By formulating subpopulations as exponential-family distributions, OMT decouples computational complexity from the sample size, scaling solely with the number of mixture components. We demonstrate the effectiveness and practicality of OMT on a wide range of synthetic benchmarks and real-world datasets, including image data and large-scale single-cell RNA sequencing measurements.