SAGE: Retain-Aware Post-Hoc Sanitization of Final Unlearning Vector

Large Language Model (LLM) unlearning aims to remove undesirable knowledge or behaviors while preserving retained capabilities. Current unlearning methods all involve a trade-off between unlearning and retention. We have found that the retention activation bias can also be used to quantify the damage an unlearning method inflicts on retention, without considering the specific implementation of the unlearning process. This allows us to restore retention performance for any unlearning method using a post-hoc approach. Therefore, we propose a complementary post-hoc setting to sanitize the final update vector without rerunning the original unlearning pipeline. In this setting, we design SAGE, Spectral Activation-GEometry Sanitization, a source-agnostic correction for final unlearning updates. SAGE collects real module inputs from a small retain proxy, extracts their dominant activation geometry, and solves a source-anchored optimization objective in closed form, which suppresses update components aligned with high-energy retained directions while preserving the source method's forgetting carrier. Across multiple unlearning methods, model scales, and benchmarks, SAGE consistently relieves the retain-forget trade-off, identifying post-hoc sanitization of final vectors as a practical and underexplored axis for machine unlearning.

Compress then Merge: From Multiple LoRAs into One Low-Rank Adapter

Low-rank adaptation (LoRA) enables parameter-efficient specialization of foundation models, but the proliferation of task-specific adapters fragments capabilities across many adapters, complicating reuse and deployment. We study the problem of merging $T$ LoRAs into a single rank-$r$ LoRA, thereby preserving the benefits of low-rank structure. Existing Merge-then-Compress pipelines treat the rank constraint as an afterthought: they merge adapters in the full parameter space, then compress the merged result to rank $r$ via truncated SVD. However, full-parameter merging may destroy the low-rank structure, making it difficult for subsequent compression to recover an effective rank-$r$ LoRA. We propose Compress-then-Merge (CtM), a reversed pipeline that enforces the rank-$r$ bottleneck before merging: CtM computes shared $r$-dimensional subspaces using only the LoRA weights to capture cross-adapter common structure, projects each adapter into the shared subspaces to obtain $r\times r$ coordinates, and then applies standard merging rules in this reduced space. CtM guarantees a rank-$r$ LoRA by construction, avoiding post-hoc truncation, and enables efficient computation in the core space spanned by concatenated LoRA factors. Experiments across multiple models and tasks show that CtM consistently outperforms existing single-LoRA-output baselines while narrowing the performance gap to full-parameter merging methods.

RAIN-Merging: A Gradient-Free Method to Enhance Instruction Following in Large Reasoning Models with Preserved Thinking Format

Large reasoning models (LRMs) excel at a long chain of reasoning but often fail to faithfully follow instructions regarding output format, constraints, or specific requirements. We investigate whether this gap can be closed by integrating an instruction-tuned model (ITM) into an LRM. Analyzing their differences in parameter space, namely task vectors, we find that their principal subspaces are nearly orthogonal across key modules, suggesting a lightweight merging with minimal interference. However, we also demonstrate that naive merges are fragile because they overlook the output format mismatch between LRMs (with explicit thinking and response segments) and ITMs (answers-only). We introduce RAIN-Merging (Reasoning-Aware Instruction-attention guided Null-space projection Merging), a gradient-free method that integrates instruction following while preserving thinking format and reasoning performance. First, with a small reasoning calibration set, we project the ITM task vector onto the null space of forward features at thinking special tokens, which preserves the LRM's structured reasoning mechanisms. Second, using a small instruction calibration set, we estimate instruction attention to derive module-specific scaling that amplifies instruction-relevant components and suppresses leakage. Across four instruction-following benchmarks and nine reasoning & general capability benchmarks, RAIN-Merging substantially improves instruction adherence while maintaining reasoning quality. The gains are consistent across model scales and architectures, translating to improved performance in agent settings.

Learn the Ropes, Then Trust the Wins: Self-imitation with Progressive Exploration for Agentic Reinforcement Learning

Reinforcement learning (RL) is the dominant paradigm for sharpening strategic tool use capabilities of LLMs on long-horizon, sparsely-rewarded agent tasks, yet it faces a fundamental challenge of exploration-exploitation trade-off. Existing studies stimulate exploration through the lens of policy entropy, but such mechanical entropy maximization is prone to RL training instability due to the multi-turn distribution shifting. In this paper, we target the progressive exploration-exploitation balance under the guidance of the agent own experiences without succumbing to either entropy collapsing or runaway divergence. We propose SPEAR, a curriculum-based self-imitation learning (SIL) recipe for training agentic LLMs. It extends the vanilla SIL framework, where a replay buffer stores self-generated promising trajectories for off-policy update, by gradually steering the policy evolution within a well-balanced range of entropy across stages. Specifically, our approach incorporates a curriculum to manage the exploration process, utilizing intrinsic rewards to foster skill-level exploration and facilitating action-level exploration through SIL. At first, the auxiliary tool call reward plays a critical role in the accumulation of tool-use skills, enabling broad exposure to the unfamiliar distributions of the environment feedback with an upward entropy trend. As training progresses, self-imitation gets strengthened to exploit existing successful patterns from replayed experiences for comparative action-level exploration, accelerating solution iteration without unbounded entropy growth. To further stabilize training, we recalibrate the advantages of experiences in the replay buffer to address the potential policy drift. Reugularizations such as the clipping of tokens with high covariance between probability and advantage are introduced to the trajectory-level entropy control to curb over-confidence.

T2I-ConBench: Text-to-Image Benchmark for Continual Post-training

Continual post-training adapts a single text-to-image diffusion model to learn new tasks without incurring the cost of separate models, but naive post-training causes forgetting of pretrained knowledge and undermines zero-shot compositionality. We observe that the absence of a standardized evaluation protocol hampers related research for continual post-training. To address this, we introduce T2I-ConBench, a unified benchmark for continual post-training of text-to-image models. T2I-ConBench focuses on two practical scenarios, item customization and domain enhancement, and analyzes four dimensions: (1) retention of generality, (2) target-task performance, (3) catastrophic forgetting, and (4) cross-task generalization. It combines automated metrics, human-preference modeling, and vision-language QA for comprehensive assessment. We benchmark ten representative methods across three realistic task sequences and find that no approach excels on all fronts. Even joint "oracle" training does not succeed for every task, and cross-task generalization remains unsolved. We release all datasets, code, and evaluation tools to accelerate research in continual post-training for text-to-image models.

A Unified Gradient-based Framework for Task-agnostic Continual Learning-Unlearning

Recent advancements in deep models have highlighted the need for intelligent systems that combine continual learning (CL) for knowledge acquisition with machine unlearning (MU) for data removal, forming the Continual Learning-Unlearning (CLU) paradigm. While existing work treats CL and MU as separate processes, we reveal their intrinsic connection through a unified optimization framework based on Kullback-Leibler divergence minimization. This framework decomposes gradient updates for approximate CLU into four components: learning new knowledge, unlearning targeted data, preserving existing knowledge, and modulation via weight saliency. A critical challenge lies in balancing knowledge update and retention during sequential learning-unlearning cycles. To resolve this stability-plasticity dilemma, we introduce a remain-preserved manifold constraint to induce a remaining Hessian compensation for CLU iterations. A fast-slow weight adaptation mechanism is designed to efficiently approximate the second-order optimization direction, combined with adaptive weighting coefficients and a balanced weight saliency mask, proposing a unified implementation framework for gradient-based CLU. Furthermore, we pioneer task-agnostic CLU scenarios that support fine-grained unlearning at the cross-task category and random sample levels beyond the traditional task-aware setups. Experiments demonstrate that the proposed UG-CLU framework effectively coordinates incremental learning, precise unlearning, and knowledge stability across multiple datasets and model architectures, providing a theoretical foundation and methodological support for dynamic, compliant intelligent systems.

MUSO: Achieving Exact Machine Unlearning in Over-Parameterized Regimes

Machine unlearning (MU) is to make a well-trained model behave as if it had never been trained on specific data. In today's over-parameterized models, dominated by neural networks, a common approach is to manually relabel data and fine-tune the well-trained model. It can approximate the MU model in the output space, but the question remains whether it can achieve exact MU, i.e., in the parameter space. We answer this question by employing random feature techniques to construct an analytical framework. Under the premise of model optimization via stochastic gradient descent, we theoretically demonstrated that over-parameterized linear models can achieve exact MU through relabeling specific data. We also extend this work to real-world nonlinear networks and propose an alternating optimization algorithm that unifies the tasks of unlearning and relabeling. The algorithm's effectiveness, confirmed through numerical experiments, highlights its superior performance in unlearning across various scenarios compared to current state-of-the-art methods, particularly excelling over similar relabeling-based MU approaches.

Unified Gradient-Based Machine Unlearning with Remain Geometry Enhancement

Machine unlearning (MU) has emerged to enhance the privacy and trustworthiness of deep neural networks. Approximate MU is a practical method for large-scale models. Our investigation into approximate MU starts with identifying the steepest descent direction, minimizing the output Kullback-Leibler divergence to exact MU inside a parameters' neighborhood. This probed direction decomposes into three components: weighted forgetting gradient ascent, fine-tuning retaining gradient descent, and a weight saliency matrix. Such decomposition derived from Euclidean metric encompasses most existing gradient-based MU methods. Nevertheless, adhering to Euclidean space may result in sub-optimal iterative trajectories due to the overlooked geometric structure of the output probability space. We suggest embedding the unlearning update into a manifold rendered by the remaining geometry, incorporating second-order Hessian from the remaining data. It helps prevent effective unlearning from interfering with the retained performance. However, computing the second-order Hessian for large-scale models is intractable. To efficiently leverage the benefits of Hessian modulation, we propose a fast-slow parameter update strategy to implicitly approximate the up-to-date salient unlearning direction. Free from specific modal constraints, our approach is adaptable across computer vision unlearning tasks, including classification and generation. Extensive experiments validate our efficacy and efficiency. Notably, our method successfully performs class-forgetting on ImageNet using DiT and forgets a class on CIFAR-10 using DDPM in just 50 steps, compared to thousands of steps required by previous methods.

Flat-LoRA: Low-Rank Adaption over a Flat Loss Landscape

Fine-tuning large-scale pre-trained models is prohibitively expensive in terms of computational and memory costs. Low-Rank Adaptation (LoRA), a popular Parameter-Efficient Fine-Tuning (PEFT) method, provides an efficient way to fine-tune models by optimizing only a low-rank matrix. Despite recent progress made in improving LoRA's performance, the connection between the LoRA optimization space and the original full parameter space is often overlooked. A solution that appears flat in the LoRA space may exist sharp directions in the full parameter space, potentially harming generalization performance. In this paper, we propose Flat-LoRA, an efficient approach that seeks a low-rank adaptation located in a flat region of the full parameter space.Instead of relying on the well-established sharpness-aware minimization approach, which can incur significant computational and memory burdens, we utilize random weight perturbation with a Bayesian expectation loss objective to maintain training efficiency and design a refined perturbation generation strategy for improved performance. Experiments on natural language processing and image classification tasks with various architectures demonstrate the effectiveness of our approach.

Pursuing Feature Separation based on Neural Collapse for Out-of-Distribution Detection

In the open world, detecting out-of-distribution (OOD) data, whose labels are disjoint with those of in-distribution (ID) samples, is important for reliable deep neural networks (DNNs). To achieve better detection performance, one type of approach proposes to fine-tune the model with auxiliary OOD datasets to amplify the difference between ID and OOD data through a separation loss defined on model outputs. However, none of these studies consider enlarging the feature disparity, which should be more effective compared to outputs. The main difficulty lies in the diversity of OOD samples, which makes it hard to describe their feature distribution, let alone design losses to separate them from ID features. In this paper, we neatly fence off the problem based on an aggregation property of ID features named Neural Collapse (NC). NC means that the penultimate features of ID samples within a class are nearly identical to the last layer weight of the corresponding class. Based on this property, we propose a simple but effective loss called OrthLoss, which binds the features of OOD data in a subspace orthogonal to the principal subspace of ID features formed by NC. In this way, the features of ID and OOD samples are separated by different dimensions. By optimizing the feature separation loss rather than purely enlarging output differences, our detection achieves SOTA performance on CIFAR benchmarks without any additional data augmentation or sampling, demonstrating the importance of feature separation in OOD detection. The code will be published.