Parameter-Efficient Fine-Tuning (PEFT), particularly Low-Rank Adaptation (LoRA), has become a standard approach for adapting Large Language Models (LLMs) under limited compute. However, in continual settings where models are updated sequentially with small datasets, conventional LoRA updates struggle to balance rapid adaptation and knowledge retention. Existing methods typically treat the low-rank space as a homogeneous update region, lacking mechanisms to regulate how short-term updates are consolidated over time. We propose a continual LoRA framework with Program memory, inspired by Complementary Learning Systems in neuroscience. Our approach, dubbed ProCL, organizes LoRA adapters into structured program memory slots that are dynamically retrieved through input-conditioned attention. This enables rapid and localized adaptation, encouraging similar inputs to reuse shared adapter regions while reserving unused capacity for future data. The slots are then combined with the underlying adapter, which maintains a distributed representation that gradually accumulates knowledge across tasks to balance plasticity and stability. Our method operates entirely within the LoRA parameterization and incurs no additional inference cost. Experiments on diverse benchmarks demonstrate improved retention and reduced catastrophic forgetting over other continual LoRA strategies.

Large language models (LLMs) have shown strong reasoning capabilities when fine-tuned with reinforcement learning (RL). However, such methods require extensive data and compute, making them impractical under many realistic training budgets. Many existing pipelines sample training examples uniformly across steps or epochs, ignoring differences in difficulty, redundancy, and learning value, which slows learning and wastes computation. We propose SPaCe, a self-paced learning framework that enables efficient learning based on the capability of the model being trained through optimizing which data to use and when. First, we apply cluster-based data reduction to partition training data by semantics and difficulty, extracting a compact yet diverse subset that reduces redundancy. Then, a multi-armed bandit treats data clusters as arms, allocating training samples based on the model’s solve rates and learning progress. Experiments across multiple reasoning benchmarks show that SPaCe achieves comparable or better accuracy than state-of-the-art baselines while using up to (100 times) fewer samples. Ablation studies and analyses further highlight the importance of both data clustering and adaptive selection. Our results demonstrate that carefully curated, performance-driven training curricula can unlock strong reasoning abilities in LLMs with minimal resources.

In the era of Large Language Models (LLMs), the Mixture of Experts (MoE) architecture has emerged as an effective approach for training extremely large models with improved computational efficiency. This success builds upon extensive prior research aimed at enhancing expert specialization in MoE-based LLMs. However, the nature of such specializations and how they can be systematically interpreted remain open research challenges. In this work, we investigate this gap by posing a fundamental question: Do domain-specific experts exist in MoE-based LLMs? To answer the question, we evaluate ten advanced MoE-based LLMs ranging from 3.8B to 120B parameters and provide empirical evidence for the existence of domain-specific experts. Building on this finding, we propose Domain Steering Mixture of Experts (DSMoE), a training-free framework that introduces zero additional inference cost and outperforms both well-trained MoE-based LLMs and strong baselines, including Supervised Fine-Tuning (SFT). Experiments on four advanced open-source MoE-based LLMs across both target and non-target domains demonstrate that our method achieves strong performance and robust generalization without increasing inference cost or requiring additional retraining.

Mixture-of-Experts (MoE) routing collapses in parameter-efficient fine-tuning: on LLaMA-3 8B with MixLoRA, a single expert absorbs 82.5% of tokens. Prior MoE-PEFT remedies intervene at the output: routing reformulations, output-similarity penalties, or shared-expert capacity. All leave each routed expert’s input view identical. We propose SIMPA (Sparse-Input Magnitude Partitioning Adaptation), which restricts what experts *see* rather than what they produce: a magnitude-thresholded input drives a sparse routed-LoRA branch, a parallel dense LoRA branch (adapting DeepSeekMoE’s (Dai et al., 2024) shared-expert topology to PEFT) provides an information safeguard, and a column-wise gate fuses the two. At a matched 2.9% trainable-parameter budget, SIMPA leads compared baselines on reasoning (86.2% on seven LLaMA-3 8B tasks, +3.9 over MixDoRA), code and arithmetic (38.4 HumanEval Pass@1 and 59.4 GSM8K, matching full fine-tuning under identical data and compute), and NLU (88.7 GLUE / 77.7 SuperGLUE on T5-base); against direct MoE-PEFT counterparts HydraLoRA and MoLoRA at matched setup, SIMPA leads HumanEval by +14.2 and +13.6 and GSM8K by +2.1 and +2.8. Direct measurement returns an expert-input Jaccard of 0.184 versus MixLoRA’s 1.000, a 5.4× reduction. The same block-aligned mask drives a Triton kernel measuring 1.46× end-to-end inference speedup on A100. Closed-form partitioning and retention bounds characterize the mechanism, and a tail-index-driven closed form predicts the optimal sparsity p^* from the pretrained backbone. The mechanism, its predicted optimum, and its bound all transfer to a held-out Qwen2.5-7B backbone. Code: https://github.com/triquang26/simpa.

Creative generation tasks, such as narrative writing and scientific ideation, demand both high-quality outputs and distinct responses across independent runs to maximize exploration. Multi-Agent Debate (MAD) has shown strong quality gains on factual and reasoning tasks, making it a natural candidate for creative generation. However, we find its convergence-driven design actively suppresses output diversity across independent runs, creating an inherent trade-off with creative tasks. We theoretically show that preserving diversity among agents within each debate session is a necessary condition for achieving diverse outputs across independent runs. Building on this finding, we propose Creative-MAD, which introduces two synergistic interventions to sustain agent divergence. Specifically, Cognitive Lens Assignment counters identity drift by anchoring each agent to a distinct and persistent cognitive mode, while Embedding-based Peer Selection counters majority pull by limiting each agent’s context to its most semantically distant peers. Experiments across four creative benchmarks demonstrate that Creative-MAD significantly enhances both lexical and semantic diversity while maintaining MAD’s output quality.

This paper introduces layout-aware graph modeling for multimodal RAG. Different from traditional RAG methods that only deal with flat text chunks, the proposed method takes into account the relationship of multimodalities by using a graph structure. To do that, a graph modeling structure is defined based on document layout parsing. The structure of an input document is retained with the connection of text chunks, tables, and figures. This representation allows the method to handle complex questions that require information from multimodalities. To confirm the efficiency of the graph modeling, a flexible RAG pipeline is developed using robust components. Experimental results on four benchmark test sets confirm the contribution of the layout-aware modeling for performance improvement of the RAG pipeline.

Sparse mixture of experts (SMoE) have emerged as an effective approach for scaling large language models while keeping a constant computational cost. Regardless of several notable successes of SMoE, effective training such architecture remains elusive due to the representation collapse problem, which in turn harms model performance and causes parameter redundancy. In this work, we present Similarity-based Sparse Mixture of Experts (SimSMoE), a novel similarity of neural network algorithm, that guarantees a solution to address the representation collapse issue between experts given a fixed FLOPs budget. We conduct extensive empirical evaluations on three large language models for both Pre-training and Fine-tuning tasks to illustrate the efficacy, robustness, and scalability of our method. The results demonstrate that SimSMoE significantly enhances existing routing policy and outperforms other SMoE routing methods in performance for the tasks. Our implementation is publicly available at https://github.com/giangdip2410/SimSMoE.

Aligning large language models (LLMs) with specific task objectives is challenging, especially when access to feedback signals for guiding the model is limited. While existing parametric methods perform reasonably, they rely heavily on large datasets and frequent feedback, making them impractical in scenarios with limited human feedback. We introduce Alignment Learning with Episodic Control (ALEC), a non-parametric framework that aligns LLM outputs during inference without fine-tuning. ALEC employs a key-value memory to store the associations between generated text and its corresponding values. It leverages a novel confidence-based writing scheme to update these stored values, maximizing the use of available data. During inference, ALEC utilizes a nearest-neighbor mechanism to estimate the values of generated texts, enabling the selection of the optimal text for decoding. Our method outperforms state-of-the-art baselines on harmless, helpful, and summarization tasks, demonstrating improved alignment with minimal interactions with the true reward model.

Large Language Models (LLMs) exhibit emergent in-context learning (ICL) capabilities, allowing them to adapt to unseen tasks based on example demonstrations. Traditional ICL embeds examples within the prompt, while activation steering, uses a vector derived from examples to guide the latent states of LLMs toward desired behaviors. However, traditional ICL is difficult to control quantitatively and consumes valuable context space. Existing activation steering methods apply a single sentence-level steering vector uniformly across all tokens, ignoring LLMs’ token-wise, auto-regressive nature. This coarse control can lead to inconsistencies and suboptimal adjustments during generation. To address this problem, we introduce Dynamic Steering with Episodic Memory (DSEM), a novel training-free framework that aligns LLMs to given demonstrations by steering at the token level conditioned on the input query. DSEM employs a key-value memory to store associations between generated tokens and steering vectors. During inference, it uses a nearest-neighbor mechanism to dynamically compute steering vectors for each token chunk, enabling more precise and adaptive guidance. Our method surpasses strong baselines across diverse alignment tasks - including safety, style transfer, and role-playing - demonstrating improved alignment as demonstration size scales.

For summarization, human preferences is critical to tame outputs of the summarizer in favor of human interests, as ground-truth summaries are scarce and ambiguous. Practical settings require dynamic exchanges between humans and AI agents wherein feedback is provided in an online manner, a few at a time. In this paper, we introduce a new framework to train summarization models with preference feedback interactively. By properly leveraging offline data and a novel reward model, we improve the performance regarding ROUGE scores and sample-efficiency. Our experiments on three various datasets confirm the benefit of the proposed framework in active, few-shot and online settings of preference learning.