The central tension in continual learning (CL) is the trade-off between plasticity (acquiring new knowledge) and stability (retaining prior knowledge). We study how a pre-trained backbone can be continually updated to absorb new knowledge while preserving existing capabilities, via capacity control: regulating the effective rank of each parameter update, a per-step quantity directly controllable inside a LoRA update. A controlled probe of LoRA rank and placement across modules and tasks reveals a consistent trade-off, with a moderate-rank sweet spot that varies by placement and task, leaving no universally optimal fixed rank; a formal bound shows forgetting grows with rank. Building on these findings, we propose Continual Dynamic Rank-Selective LoRA (CoDyRA), which jointly trains each LoRA update with rank minimization via sparsity-promoting regularization on per-component importance weights. The supervised objective drives plasticity; rank minimization regularizes forgetting. We show that rank minimization serves as an implicit forgetting regularizer in the CL regime, protecting general capability and prior-task knowledge simultaneously by controlling forgetting against the current model state. Across MTIL, X-TAIL, and TRACE (CLIP, LLaMA, Gemma), CoDyRA outperforms prior CL methods on new knowledge learning and forgetting, achieving a strong plasticity–stability balance. Project page: https://artificer-ai-lab.github.io/CoDyRA.

Low-rank adaptation (LoRA) has become the standard for parameter-efficient fine-tuning of large language models. Most LoRA variants follow a uniform-LR convention, applying a single global learning rate across every rank-one component of every adapter. We show that this convention overlooks substantial within-module heterogeneity, where the rank-one components of a LoRA adapter update at highly uneven rates and low-velocity modules converge to concentrated singular spectra that underutilize the nominal rank budget. To address this, we propose an adaptive anisotropic learning-rate model that assigns each rank-one component its own effective learning rate, computed online from training-time signals and mean-normalized per module to preserve the global LR budget. AnLR-LoRA instantiates this model with two signals available during AdamW optimization, namely function-space velocity and Adam SNR, as a lightweight scheme with no extra trainable parameters. Across commonsense reasoning, natural language generation and visual instruction-tuning benchmarks, AnLR-LoRA consistently improves over LoRA while encouraging broader use of rank capacity, with gains that remain robust across a wide range of global learning rates and transfer cleanly to other LoRA variants.

Chain-of-thought (CoT) reasoning in large language models often produces redundant steps that increase generation cost and obscure when reasoning has effectively completed. Existing stopping methods rely on output-level or token-level signals, overlooking internal representation dynamics that reflect reasoning progress. We propose STARS (STate-Aware Reasoning Stopping), a framework that leverages spectral analysis of hidden-state representations to detect transitions in reasoning phases. Our key insight is that structural changes in the representation space capture shifts in reasoning states, enabling early and reliable stopping decisions. To address representation instability, where semantically similar steps can induce different hidden states, we focus on spectral structure rather than raw activations, improving robustness to context-dependent drift. Experiments on four mathematical and logical reasoning benchmarks show that STARS reduces generation cost by over 45% under fixed token budgets while improving accuracy over standard CoT baselines.

Large language model-based agents operating in long-horizon interactions require memory systems that support temporal consistency, multi-hop reasoning, and evidence-grounded reuse across sessions. Existing approaches largely rely on unstructured retrieval or coarse abstractions, which often lead to temporal conflicts, brittle reasoning, and limited traceability. We propose MemWeaver, a unified memory framework that consolidates long-term agent experiences into three interconnected components: a temporally grounded graph memory for structured relational reasoning, an experience memory that abstracts recurring interaction patterns from repeated observations, and a passage memory that preserves original textual evidence. MemWeaver employs a dual-channel retrieval strategy that jointly retrieves structured knowledge and supporting evidence to construct compact yet information-dense contexts for reasoning. Experiments on the LoCoMo benchmark demonstrate that MemWeaver substantially improves multi-hop and temporal reasoning accuracy while reducing input context length by over 95% compared to long-context baselines.

Reinforcement Learning (RL) has proven effective in eliciting reasoning abilities of Multimodal Large Language Models (MLLMs) and is emerging in document understanding. However, existing approaches usually lack fine-grained region-level grounding supervision, while the key challenge in document reasoning lies in linking visual evidence within precise regions directly to question–answer semantics. We introduce RegionDoc-R1, a novel framework for document understanding that enhances MLLMs’ reasoning with step-wise feedback. Instead of previous OCR-based text positions, we encourage the model to predict question-relevant Regions of Interest (RoIs) and ground the answer in these regions. Specifically, we first propose the Region-Aware Group Relative Policy Optimization (RA-GRPO), which encourages models to utilize region-level spatial information in documents for reasoning. Correspondingly, we construct a hybrid training corpus, named SR-Doc, containing Semantic Reasoning (SR) examples enriched with cross-page and region-level reasoning layout annotations. Meanwhile, we also introduce an Adaptive Chain-of-Thought (Ada-CoT) strategy, which dynamically adjusts the reasoning process according to different tasks, enabling more efficient and flexible step-wise document understanding. Experiments on several document reasoning benchmarks demonstrate that RegionDoc-R1 achieves state-of-the-art performance across tasks such as form document understanding, table-based QA, and layout-sensitive information extraction.

Multimodal large language models often struggle with faithful reasoning in complex visual scenes, where intricate entities and relations require precise visual grounding at each step. This reasoning unfaithfulness frequently manifests as hallucinated entities, mis-grounded relations, skipped steps, and over-specified reasoning. Existing preference-based approaches, typically relying on textual perturbations or answer-conditioned rationales, fail to address this challenge as they allow models to exploit language priors to bypass visual grounding. To address this, we propose SceneAlign, a framework that leverages scene graphs as structured visual information to perform controllable structural interventions. By identifying reasoning-critical nodes and perturbing them through four targeted strategies that mimic typical grounding failures, SceneAlign constructs hard negative rationales that remain linguistically plausible but are grounded in inaccurate visual facts. These contrastive pairs are used in Direct Preference Optimization to steer models toward fine-grained, structure-faithful reasoning. Across seven visual reasoning benchmarks, SceneAlign consistently improves answer accuracy and reasoning faithfulness, highlighting the effectiveness of grounding-aware alignment for multimodal reasoning.

Large Language Models (LLMs) are susceptible to indirect prompt injection attack, where the model inadvertently responds to instructions injected into the prompt context. This vulnerability stems from LLMs’ inability to distinguish between data and instructions within a prompt. We propose CachePrune that defends against this attack by identifying and pruning neurons associated with instruction-following, during KV cache encoding of the prompt context. The pruning steers the LLM toward interpreting the context purely as data rather than as instructions to follow. To identify these neurons, we introduce a neural attribution mechanism guided by a preferential attribution loss, and theoretically connect this loss to an upper bound of the Direct Preference Optimization (DPO) objective. Further, we improve on the fidelity of neural attribution by leveraging an observed triggering effect in instruction-following. Our approach does not interfere with prompt formatting or incur test-time overhead in response generation. Experiments show that CachePrune significantly reduces the attack success rate while preserving the LLM’s ability to follow user instructions.

Recent MLLMs have demonstrated strong visual understanding and reasoning after large-scale multimodal pre-training. However, instruction-tuning is typically text-driven with limited visual supervision, leading to significant visual forgetting and degradation of pre-trained visual knowledge. Existing fine-tuning and continual learning methods compress visual representations and emphasize task alignment over visual retention, failing to address this challenge. We present a novel perspective using effective rank to quantify the loss of visual representation richness, framing visual forgetting as excessive compression under the information bottleneck principle. To address this, we propose modality-decoupled gradient descent (MDGD), which regulates gradient updates to preserve the effective rank of visual features and explicitly disentangles visual learning from task-specific alignment. We further introduce a memory-efficient fine-tuning variant using gradient masking for parameter-efficient adaptation. Extensive experiments show that MDGD effectively mitigates visual forgetting across downstream tasks and models, maintaining pre-trained visual knowledge while supporting strong task adaptation.

Graphical User Interface (GUI) agents, powered by Large Foundation Models, have emerged as a transformative approach to automating human-computer interaction. These agents autonomously interact with digital systems via GUIs, emulating human actions such as clicking, typing, and navigating visual elements across diverse platforms. Motivated by the growing interest and fundamental importance of GUI agents, we provide a comprehensive survey that categorizes their benchmarks, evaluation metrics, architectures, and training methods. We propose a unified framework that delineates their perception, reasoning, planning, and acting capabilities. Furthermore, we identify important open challenges and discuss key future directions. Finally, this work serves as a basis for practitioners and researchers to gain an intuitive understanding of current progress, techniques, benchmarks, and critical open problems that remain to be addressed.

Instruction tuning in multimodal large language models (MLLMs) generally involves cooperative learning between a backbone LLM and a feature encoder of non-text input modalities. The major challenge is how to efficiently find the synergy between the two modules so that LLMs can adapt their reasoning abilities to downstream tasks while feature encoders can adjust to provide more task-specific information about its modality. In this paper, we analyze the MLLM instruction tuning from both theoretical and empirical perspectives, where we find the unbalanced learning between the feature encoder and the LLM can cause problems of oscillation and biased learning that lead to sub-optimal convergence. Inspired by our findings, we propose a Multimodal Balance Coefficient that enables quantitative measurement of the balance of learning. Based on this, we further design a dynamic learning scheduler that better coordinates the learning between the LLM and feature encoder, alleviating the problems of oscillation and biased learning. In addition, we introduce an auxiliary regularization on the gradient to promote updating with larger step sizes, which potentially allows for a more accurate estimation of the proposed MultiModal Balance Coefficient and further improves the training sufficiency. Our proposed approach is agnostic to the architecture of LLM and feature encoder, so it can be generically integrated with various MLLMs. We conduct experiments on multiple downstream tasks with various MLLMs, demonstrating that the proposed method is more effective than the baselines in MLLM instruction tuning.

Large Language Model (LLM) agents are commonly tuned with supervised finetuning on ReAct-style expert trajectories or preference optimization over pairwise rollouts. Most of these methods focus on imitating specific expert behaviors or promoting chosen reasoning thoughts and actions over rejected ones. However, without reasoning and comparing over alternatives actions, LLM agents finetuned with these methods may over-commit towards seemingly plausible but suboptimal actions due to limited action space exploration. To address this, in this paper we propose Self-taught ActioN Deliberation (SAND) framework, enabling LLM agents to explicitly deliberate over candidate actions before committing to one. To tackle the challenges of when and what to deliberate given large action space and step-level action evaluation, we incorporate self-consistency action sampling and execution-guided action critique to help synthesize step-wise action deliberation thoughts using the base model of the LLM agent. In an iterative manner, the deliberation trajectories are then used to finetune the LLM agent itself. Evaluating on two representative interactive agent tasks, SAND achieves an average 20% improvement over initial supervised finetuning and also outperforms state-of-the-art agent tuning approaches.

Retrieval-augmented large language models (LLMs) have been remarkably competent in various NLP tasks. However, it was observed by previous works that retrieval is not always helpful, especially when the LLM is already knowledgable on the query to answer. Motivated by this, Adaptive Retrieval-Augmented Generation (ARAG) studies retrieving only when the knowledge asked by the query is absent in the LLM. Previous works of ARAG either require accessing the pre-training corpus or prompting with additional model inferences. Aiming to avoid such drawbacks, we propose to determine whether the model is knowledgeable on a query via inspecting the (contextualized) pre-trained token embeddings of LLMs. We hypothesize that such embeddings capture rich information on the model’s intrinsic knowledge base, which enables an efficient way of judging the necessity to retrieve from an external corpus. Extensive experiments demonstrate our ARAG approach’s superior performance across various benchmarks.

In this paper, we study personalized federated learning for text classification with Pretrained Language Models (PLMs). We identify two challenges in efficiently leveraging PLMs for personalized federated learning: 1) Communication. PLMs are usually large in size, e.g., with hundreds of millions of parameters, inducing huge communication cost in a federated setting. 2) Local Training. Training with PLMs generally requires back-propagation, during which memory consumption can be several times that of the forward-propagation. This may not be affordable when the PLMs are trained locally on the clients that are resource constrained, e.g., mobile devices with limited access to memory resources. Additionally, the proprietary PLMs can be provided as concealed APIs, for which the back-propagation operations may not be available. In solving these, we propose a training framework that includes an approach of discrete local search for gradient-free local training, along with a compression mechanism inspired from the linear word analogy that allows communicating with discretely indexed tokens, thus significantly reducing the communication cost. Experiments show that our gradient-free framework achieves superior performance compared with baselines.

Recently, Denoising Diffusion Probabilistic Models (DDPMs) have attained leading performances across a diverse range of generative tasks. However, in the field of speech synthesis, although DDPMs exhibit impressive performance, their prolonged training duration and substantial inference costs hinder practical deployment. Existing approaches primarily focus on enhancing inference speed, while approaches to accelerate training—a key factor in the costs associated with adding or customizing voices—often necessitate complex modifications to the model, compromising their universal applicability. To address the aforementioned challenges, we propose an inquiry: is it possible to enhance the training/inference speed and performance of DDPMs by modifying the speech signal itself? In this paper, we double the training and inference speed of Speech DDPMs by simply redirecting the generative target to the wavelet domain. This method not only achieves comparable or superior performance to the original model in speech synthesis tasks but also demonstrates its versatility. By investigating and utilizing different wavelet bases, our approach proves effective not just in speech synthesis, but also in speech enhancement.