SARA framework aligns MoE routing distributions to improve low-resource multilingual performance
Researchers introduce SARA (Semantically Anchored Routing Alignment), a framework that addresses cross-lingual routing divergence in sparse Mixture-of-Experts LLMs by aligning the internal routing distributions of low-resource language tokens to match those of high-resource semantic anchors via symmetric Jensen-Shannon divergence constraints. Unlike logit-level distillation, SARA operates directly on MoE routing layers to encourage mechanistic consistency in expert selection across languages. Experiments on Qwen3-30B-A3B and Phi-3.5-MoE-instruct across 5 low-resource languages show modest but consistent gains (up to +1.2%) on Global-MMLU over standard instruction tuning.
Related guides (2)
Related events (8)
SAERL: Using Sparse Autoencoders to Guide LLM Reinforcement Learning Data Engineering
SAERL is a post-training data engineering framework that uses Sparse Autoencoders (SAEs) — a mechanistic interpretability tool — to extract intrinsic model signals for controlling data diversity, difficulty, and quality during RL fine-tuning. The framework applies SAE-space clustering for batch diversity, a difficulty proxy for curriculum ordering, and a quality probe for data filtering. On Qwen2.5-Math-1.5B with GRPO, SAERL achieves 3% average accuracy improvement and reaches target accuracy with 20% fewer training steps. SAE representations transfer across model families and scales, suggesting broad applicability as a lightweight data engineering tool.
Calibrated Mixture-of-Experts under distribution shift: adversarial reweighting approach
A new arXiv preprint analyzes how mixture-of-experts (MoE) models maintain calibration under distribution shift, examining the interaction between routing mechanisms and expert-level calibration. The authors prove that expert calibration is sufficient for overall model calibration in hard-routed MoE but insufficient for soft-routed variants. To address the soft-routing gap, they propose an adversarial reweighting method that penalizes calibration errors of the routed aggregate under distribution shift, demonstrating improved accuracy-calibration tradeoffs across model classes and tasks.
Adversarial robustness and safety alignment in multilingual multimodal LLMs: cross-lingual vulnerability and 'safety-by-failure'
A systematic study evaluates adversarial robustness and safety alignment of multimodal LLMs across 12 languages, finding that adversarial images optimized in one language transfer to others (cross-lingual transferability). The paper introduces the concept of 'safety-by-failure': low-resource languages appear safer not due to genuine alignment but because models fail to comprehend harmful instructions in those languages. Models like Qwen3-VL that integrate multilingual capability throughout training (rather than only at instruction tuning) show genuine cross-lingual safety with active refusal. The findings challenge the assumption that low-resource language safety metrics reflect real alignment.
Expert Tying reduces MoE LLM memory footprint by ~2x with minimal quality loss
Researchers introduce Expert Tying, an architectural modification for Mixture-of-Experts LLMs that shares expert parameters across consecutive transformer layers while keeping routing and attention layer-independent. Evaluated on OLMoE, Qwen3, and DeepSeek-style MoE architectures, the method achieves nearly 2x memory reduction with negligible perplexity or downstream quality degradation. The approach exploits parameter redundancy in MoE pathways to improve the compute-to-memory trade-off for training and inference.
Adversarial Subspace Alignment for Robust Multimodal Knowledge Editing in MLLMs
This paper addresses the generalization gap in multimodal large language model (MLLM) knowledge editing, where edits fail to propagate across semantically equivalent visual and linguistic variations. The authors introduce Latent Adversarial Robustification (LAR), which generates adversarial but semantically coherent variants in joint latent space, and Rank-Constrained Subspace Learning (RCSL), which enforces low-rank alignment of adversarial representations at the edit layer. Together these form the ASAM framework, which formalizes robustness via knowledge units grouping semantically equivalent multimodal inputs. Empirical analysis demonstrates improved generality without sacrificing reliability or locality.
SMoA: Spectrum Modulation Adapter for Parameter-Efficient Fine-Tuning
SMoA is a new parameter-efficient fine-tuning method that addresses LoRA's trade-off between rank size and parameter budget. It partitions model layers into spectral blocks and applies Hadamard-modulated low-rank branches to each diagonal block, enabling broader coverage of pretrained spectral directions without proportionally increasing trainable parameters. Theoretical analysis and empirical results on multiple tasks show SMoA outperforms LoRA and competitive LoRA-style baselines in lower-budget settings.
Global-batch Load Balancing for MoE LLM Training from Qwen
Qwen Research introduces a global-batch load balancing technique for Mixture-of-Experts (MoE) LLM training, claiming it is nearly a 'free lunch' improvement. The method addresses expert load imbalance across training batches, a known efficiency and quality bottleneck in MoE architectures. The approach targets the router and expert activation dynamics in transformer-based MoE layers.
CRAM: Centroid-Routing and Adaptive MoE for Multimodal Continual Instruction Tuning
CRAM is a new method for Multimodal Continual Instruction Tuning (MCIT) that addresses the tension between catastrophic forgetting and parameter efficiency in MLLMs. It combines adaptive-rank instantiation to dynamically allocate parameters based on capability gaps, centroid-guided routing to reuse existing expert knowledge, and an orthogonality penalty to confine new updates to task-specific directions. The approach uses a Mixture-of-Experts architecture where task-specific patterns are isolated into independent modules, avoiding both the interference of shared updates and the parameter bloat of fully isolated expansion. Experiments across diverse benchmarks show consistent improvements over existing MCIT methods.