#01Oct 29, 2026
cs.LG
Characterizing Recoverable Expert-Utilization Signatures in Fine-Tuned Mixture-of-Experts LLMs
Fei Wang, Baochun Li, Bo Li
As Mixture-of-Experts (MoE) large language models are increasingly deployed in distributed environments, from federated learning to edge-cloud systems, their architecture harbors an underexplored landscape of security vulnerabilities. Moving beyond textual analysis, we uncover structural memorization: a novel vulnerability where a fine-tuned model memorizes not only its training data but also the internal pathway used to process it, creating a characteristic routing fingerprint unique to the training process. We provide comprehensive empirical evidence that structural memorization emerges consistently even from a few samples, and that its intensity is directly influenced by the choice of trainable modules. We show that these memorized expert pathways can be stealthily recovered through the model's unconditional generation, using only routing and output logits, with no knowledge of the fine-tuning data or procedure. In distributed deployments, this vulnerability is amplified as routing information is exposed at multiple trust boundaries, including federated aggregation points, edge-cloud interfaces, and multi-tenant monitoring systems. Recovery exposes a critical blind spot: traditional membership inference metrics, viewed through the lens of expert routing, fail to identify memorized data corresponding to these structural pathways. To close this gap, we devise composite metrics combining reference model tail likelihood with complexity-aware normalization, significantly improving detection of memorized routing structures. This enables adversaries to pinpoint critical experts shaped by fine-tuning, whether dominant or dormant, and execute targeted attacks such as router manipulation or backdoor injection in realistic distributed settings where attackers lack training data but can observe routing patterns across system components.
#02Oct 30, 2026
cs.LG
TalentRL: Strategy-Guided Reinforcement Learning for Hierarchical Skill Routing
Huancheng Chen, Xiaodi Sun, Zhaoqiong Huang and 2 more
Natural-language skills enable large language model (LLM) agents to reuse procedures across tasks, yet most systems retrieve individual skills from a flat SkillBank without explicitly modeling the high-level strategy that governs their selection and use. We introduce TalentRL, a strategy-guided reinforcement learning framework for hierarchical routing over procedural memory. Coarse-grained talents represent recurring strategies for skill use, while fine-grained skills provide executable procedures. For each task, the policy model selects a talent that guides execution and softly prioritizes its associated skills during similarity-based retrieval. However, since task outcomes depend jointly on talent selection, retrieved skill context, and downstream execution, terminal rewards alone cannot isolate the contribution of the routing decision. TalentRL therefore trains talent routing with reward contrasts between rollouts of the same task under matched interventions on the talent-conditioned context, and optimizes downstream execution with group-relative advantages. Across seven open-domain and multi-hop question-answering benchmarks, talent-aware supervised fine-tuning improves average success rate by 7.53%, 6.71%, and 9.21% over the corresponding zero-shot baselines for Qwen2.5-7B, Qwen3-8B, and Llama-3.1-8B, respectively. Reinforcement learning further improves these SFT checkpoints by 7.58% to 9.32%, and TalentRL achieves the highest average success rate among all compared methods on all three backbones. On the agentic benchmarks ALFWorld and WebShop, TalentRL achieves the highest ALFWorld success rate (99.3%) and the highest WebShop score (90.7) among all compared methods.
#03Oct 29, 2026
cs.LG
Where, When, and How Many Bits? Scaling Laws for SGD with Stochastic Rounding
Zhichao Wang, Michael W. Mahoney
Low-precision training reduces computation but introduces additional rounding error. We theoretically study how stochastic rounding (SR) affects prediction risk as model size, training time, and precision change. We consider a two-layer linear network with a fixed random first layer, trained by mini-batch SGD on fresh Gaussian data with power-law spectra. At each SGD step, the features used in forward and backward computations are rounded independently. We prove matched upper and lower bounds on prediction error for the final model and for uniform parameter averaging, with general learning-rate and precision schedules chosen before training. For fixed precisions and constant learning rates with parameter averaging, we derive how the smallest prediction error decreases with compute, and we identify model sizes and training times that achieve this optimal-compute scaling. Rounding only the forward features preserves the compute scaling of exact arithmetic, while rounding backward features changes it. Averaging parameters can change which computation needs higher precision and when higher precision is useful. Our results also give bit requirements for a given accuracy tolerance and show how backward rounding can change learning-rate schedule design. At a fixed compute budget, lower precision can improve prediction when the benefit of a larger model or longer training outweighs the additional SR-induced error. Model simulations illustrate these tradeoffs, and experiments with an MLP and a small transformer examine rounding choices beyond the theoretical model.
#04Oct 30, 2026
cs.LG
Corroborative Learning: Mutual Enhancement through Task-Response Reconstruction
Zexing Zhang, Jichao Li, Tianyang Lei and 2 more
In pattern analysis, models trained independently for the same task often develop complementary capabilities through differences in architecture, data, and optimization. Exploiting complementarity while retaining independent deployment is central to collaborative learning. Most methods establish transfer by reducing differences in predictions, relations, or representations between target and collaborator, although these differences also underlie complementarity. This creates a structural trade-off between reducing differences for knowledge exchange and preserving them for complementarity. We therefore propose Corroborative Learning (CL). Each model forms task-driven responses, with cross-group support from its collaborator. Source-excluded reconstruction of response increments from collaborator representations separates fitting and evaluation across disjoint groups and selectively constrains directions by cross-group reconstructability. The process requires no direct output or feature matching or additional inference dependency. Furthermore, with linear reconstruction and fixed collaborator representations, the loss equals a positive semidefinite directional regularizer, constraining less reconstructable directions more strongly. Signal and noise conditions characterize when directional shrinkage reduces response-estimation error. Across six language tasks, 24 models, and 36 pairs, CL improves over capacity-matched independent adaptation on every task, with positive gains across most directed model combinations. Pairwise comparisons show broad advantages over transfer and mutual-learning methods and different gains from complementarity across collaboration rules. Low-label language and vision results support cross-domain applicability. Task-relevant differences can support mutual enhancement without eliminating disagreement.
#05Oct 30, 2026
cs.LG
Role-Aware Server Optimization for Multimodal Federated Learning with Missing Modalities
Liuyin Shi, Wenbo He
Multimodal federated learning (MFL) trains a shared model by aggregating client updates without exchanging raw data. Under missing modalities, clients adapt active modality encoders and a shared prediction block (fusion and classifier) to different input subsets. This raises two server-side questions: which clients should contribute to each component, and how should their aggregates be incorporated across rounds? We observe that an absent-modality encoder can still adapt through missing-input handling. Availability- based aggregation can therefore omit its update while retaining jointly learned prediction-block updates. Contributor selection also shapes server momentum, a weighted history of aggregated updates. We observe that prediction-block momentum can improve or reduce performance depending on the encoder update policy. We therefore propose Role-Aware Server Optimization (RASO), which coordinates contributor selection with role-specific server updates. RASO selects contributors through prediction dependence and inclusion in local optimization, retains encoder momentum, and synchronizes the shared prediction block to its current aggregate. Experiments across datasets and fusion architectures demonstrate the effectiveness of RASO under missing modalities.