Mixture-of-Experts (MoE) architectures decouple model capacity from per-token computation, enabling scaling beyond the computational limits imposed by dense scaling laws. Yet how MoE architectures shape knowledge acquisition during pre-training—and how this process differs from dense architectures—remains unknown. To address this issue, we introduce Gated-LPI (Log-Probability Increase), a neuron-level attribution metric that decomposes log-probability increase across neurons. We present a time-resolved comparison of knowledge acquisition dynamics in MoE and dense architectures, tracking checkpoints over 1.2M (~ 5.0T tokens) and 600K (~ 2.5T tokens) training steps, respectively. Our experiments uncover three patterns: (1) Low-entropy backbone. The top approximately 1% of MoE neurons capture over 45% of positive updates, forming a high-utility core, which is absent in the dense baseline. (2) Early consolidation. The MoE model locks into a stable importance profile within < 100K steps, whereas the dense model remains volatile throughout training. (3) Functional robustness. Masking the ten most important MoE attention heads reduces relational HIT@10 by < 10%, compared with > 50% for the dense model, showing that sparsity fosters distributed—rather than brittle—knowledge storage. These patterns collectively demonstrate that sparsity fosters an intrinsically stable and distributed computational backbone from early in training, helping bridge the gap between sparse architectures and training-time interpretability.

The interpretability of Mixture-of-Experts (MoE) models, especially those with heterogeneous designs, remains underexplored.Existing attribution methods for dense models fail to capture dynamic routing-expert interactions in sparse MoE architectures.To address this issue, we propose a cross-level attribution algorithm to analyze sparse MoE architectures (Qwen 1.5-MoE, OLMoE, Mixtral-8x7B) against dense models (Qwen 1.5-7B, Llama-7B, Mistral-7B).Results show MoE models achieve 31% higher per-layer efficiency via a "mid-activation, lateamplification" pattern: early layers screen experts, while late layers refine knowledge collaboratively.Ablation studies reveal a "basicrefinement" framework-shared experts handle general tasks (entity recognition), while routed experts specialize in domain-specific processing (geographic attributes).Semanticdriven routing is evidenced by strong correlations between attention heads and experts (r = 0.68), enabling task-aware coordination.Notably, architectural depth dictates robustness: deep Qwen 1.5-MoE mitigates expert failures (e.g., 43% MRR drop in geographic tasks when blocking top-10 experts) through shared expert redundancy, whereas shallow OLMoE suffers severe degradation (76% drop).Task sensitivity further guides design: core-sensitive tasks (geography) require concentrated expertise, while distributed-tolerant tasks (object attributes) leverage broader participation.These insights advance MoE interpretability, offering principles to balance efficiency, specialization, and robustness.

Mixture-of-Experts (MoE) models offer immense capacity via sparsely gated expert subnetworks, yet adapting them to multiple domains without catastrophic forgetting remains an open challenge.Existing approaches either incur prohibitive computation, suffer crossdomain interference, or require separate runs per domain.We propose DES-MoE, a dynamic expert specialization framework for multi-domain adaptation of Mixture-of-Experts models.DES-MoE addresses catastrophic forgetting through three innovations: (1) an adaptive router balancing pre-trained knowledge retention and task-specific updates via distillation, (2) real-time expert-domain correlation mapping to isolate domain-specific gradients, and (3) a three-phase adaptive fine-tuning schedule that progressively freezes non-specialized parameters.Evaluated on six domains (math, code, law, etc.), DES-MoE matches singledomain ESFT performance while training one unified model, reduces forgetting by 89% compared to full fine-tuning as domains scale from 2 to 6, and achieves 68% faster convergence than conventional methods.Our work establishes dynamic expert isolation as a scalable paradigm for multi-task MoE adaptation.

We show that large language models (LLMs) exhibit an internal chain-of-thought: they sequentially decompose and execute composite tasks layer-by-layer.Two claims ground our study: (i) distinct subtasks are learned at different network depths, and (ii) these subtasks are executed sequentially across layers.On a benchmark of 15 two-step composite tasks, we employ layer-from context-masking and propose a novel cross-task patching method, confirming (i).To examine claim (ii), we apply LogitLens to decode hidden states, revealing a consistent layerwise execution pattern.We further replicate our analysis on the real-world TRACE benchmark, observing the same stepwise dynamics.Together, our results enhance LLMs transparency by showing their capacity to internally plan and execute subtasks (or instructions), opening avenues for fine-grained, instruction-level activation steering.

Quantization and fine-tuning are crucial for deploying large language models (LLMs) on resource-constrained edge devices. However, fine-tuning quantized models presents significant challenges, primarily stemming from: First, the mismatch in data types between the low-precision quantized weights (e.g., 4-bit) and the high-precision adaptation weights (e.g., 16-bit). This mismatch limits the computational efficiency advantage offered by quantized weights during inference. Second, potential accuracy degradation when merging these high-precision adaptation weights into the low-precision quantized weights, as the adaptation weights often necessitate approximation or truncation. Third, as far as we know, no existing methods support the lossless merging of adaptation while adjusting all quantized weights. To address these challenges, we introduce lossless ternary adaptation for quantization-aware fine-tuning (LoTA-QAF). This is a novel fine-tuning method specifically designed for quantized LLMs, enabling the lossless merging of ternary adaptation weights into quantized weights and the adjustment of all quantized weights. LoTA-QAF operates through a combination of: i) A custom-designed ternary adaptation (TA) that aligns ternary weights with the quantization grid and uses these ternary weights to adjust quantized weights. ii) A TA-based mechanism that enables the lossless merging of adaptation weights. iii) Ternary signed gradient descent (t-SignSGD) for updating the TA weights. We apply LoTA-QAF to Llama-3.1/3.3 and Qwen-2.5 model families and validate its effectiveness on several downstream tasks. On the MMLU benchmark, our method effectively recovers performance for quantized models, surpassing 16-bit LoRA by up to 5.14\%. For task-specific fine-tuning, 16-bit LoRA achieves superior results, but LoTA-QAF still outperforms other methods. Code is available in github.com/KingdalfGoodman/LoTA-QAF.

Large language models pretrained on a huge amount of data capture rich knowledge and information in the training data.The ability of data memorization and regurgitation in pretrained language models, revealed in previous studies, brings the risk of data leakage.In order to effectively reduce these risks, we propose a framework DEPN to Detect and Edit Privacy Neurons in pretrained language models, partially inspired by knowledge neurons and model editing.In DEPN, we introduce a novel method, termed as privacy neuron detector, to locate neurons associated with private information, and then edit these detected privacy neurons by setting their activations to zero.Furthermore, we propose a privacy neuron aggregator dememorize private information in a batch processing manner.Experimental results show that our method can significantly and efficiently reduce the exposure of private data leakage without deteriorating the performance of the model.Additionally, we empirically demonstrate the relationship between model memorization and privacy neurons, from multiple perspectives, including model size, training time, prompts, privacy neuron distribution, illustrating the robustness of our approach.

Multilingual pretrained language models serve as repositories of multilingual factual knowledge.Nevertheless, a substantial performance gap of factual knowledge probing exists between high-resource languages and lowresource languages, suggesting limited implicit factual knowledge transfer across languages in multilingual pretrained language models.This paper investigates the feasibility of explicitly transferring relatively rich factual knowledge from English to non-English languages.To accomplish this, we propose two parameter-free Language Representation Projection modules (LRP2).The first module converts non-English representations into English-like equivalents, while the second module reverts English-like representations back into representations of the corresponding non-English language.Experimental results on the mLAMA dataset demonstrate that LRP2 significantly improves factual knowledge retrieval accuracy and facilitates knowledge transferability across diverse non-English languages.We further investigate the working mechanism of LRP2 from the perspectives of representation space and cross-lingual knowledge neuron.

In this paper, we study two issues of semantic parsing approaches to conversational question answering over a large-scale knowledge base:(1) The actions defined in grammar are not sufficient to handle uncertain reasoning common in real-world scenarios.(2) Knowledge base information is not well exploited and incorporated into semantic parsing.To mitigate the two issues, we propose a knowledge-aware fuzzy semantic parsing framework (KaFSP).It defines fuzzy comparison operations in the grammar system for uncertain reasoning based on the fuzzy set theory.In order to enhance the interaction between semantic parsing and knowledge base, we incorporate entity triples from the knowledge base into a knowledgeaware entity disambiguation module.Additionally, we propose a multi-label classification framework to not only capture correlations between entity types and relations but also detect knowledge base information relevant to the current utterance.Both enhancements are based on pre-trained language models.Experiments on a large-scale conversational question answering benchmark demonstrate that the proposed KaFSP achieves significant improvements over previous state-of-the-art models, setting new SOTA results on 8 out of 10 question types, gaining improvements of over 10% F1 or accuracy on 3 question types, and improving overall F1 from 83.01% to 85.33%.The source code of KaFSP is available at https: //github.com/tjunlp-lab/KaFSP.