LP-SFT, a Local-Preserving Supervised Fine-Tuning objective designed to explicitly protect this inherent entropy structure, improves overall performance over vanilla SFT and recent SFT-enhancement baselines, suggesting that local preservation helps mitigate capability degradation without collapsing sampling-accessible diversity.
Abstract
Supervised fine-tuning (SFT) is the standard approach for adapting pretrained language models to downstream domains, yet it often improves target-domain behavior at the cost of degrading pre-existing capabilities. Standard cross-entropy fine-tuning promotes only the observed label token and leaves unconstrained how probability mass is redistributed over other plausible alternatives, potentially distorting the rich local preference structure learned during pretraining. We first analyze next-token predictions using Shannon and Renyi entropies, revealing that pretrained models exhibit a regular multimodal entropy structure. These entropy peaks correspond to varying numbers of plausible alternatives, indicating that the base model intrinsically encodes rich distributional knowledge beyond the single supervised token. Motivated by this observation, we propose LP-SFT, a Local-Preserving Supervised Fine-Tuning objective designed to explicitly protect this inherent entropy structure. At each step, LP-SFT constructs a local top-K support of alternative tokens from the frozen base distribution. Crucially, it removes the supervised target token from this set to avoid conflicting with the cross-entropy objective, and applies a locally normalized KL divergence to maintain the base model's relative preference structure among the remaining non-label alternatives. Across mixed-domain and single-domain fine-tuning experiments, LP-SFT improves overall performance over vanilla SFT and recent SFT-enhancement baselines, achieving the best balance between pass@1 accuracy and pass@k performance. These results suggest that local preservation helps mitigate capability degradation without collapsing sampling-accessible diversity.
Full fine-tuning remains a strong way to adapt pretrained LLMs, but it updates all weights and can be expensive. LoRA reduces the number of trainable parameters, but it does not directly answer which pretrained components should be trained and which can be frozen during adaptation. We introduce TopoTuner, a topology-guided fine-tuning framework for selective freezing of attention projection matrices. \method treats each projection matrix as a row cloud and uses Wasserstein distances between persistence diagrams to measure how its topology changes during fine-tuning. TopoTuner learns a reusable freezing profile from a source dataset and transfers it to efficiently fine-tune models on out-of-domain datasets, evaluating whether task-specific topological drift generalizes across question answering and sentiment analysis tasks. Across LLaMA-3.1-8B, Mistral-7B-v0.3, and Qwen3-8B-Base, TopoTuner is competitive with full fine-tuning while training only 1-2\% of the model parameters, and outperforms LoRA in 7 out of 9 model-dataset settings, which can change up to 39.57\% of the projection parameters. Along with minimized updates, TopoTuner reduces training time by 20.4\% relative to full fine-tuning and 5.5\% relative to LoRA on average. TopoTuner opens a new direction for reusable freezing profiles, where fine-tuning behavior learned on one dataset can be shared across multiple tasks.
Abdulkadir Erol, Yash Mahajan, Vepaul Hariprashad et al.· 0 citations
It is suggested that many reported downstream gains reflect instance and seed dependent elicitation match, rather than universally improving the attainable performance ceiling, in self-supervised fine-tuning.
Wangjin Zhou, Yizhou Zhang, Yichi Wang et al.· 0 citations
Fine-tuning pre-trained language models (PLMs) is crucial for achieving performance gains in downstream natural language processing tasks, but it is prone to overfitting in low-resource scenarios. Existing noise-based regularization methods typically rely on static parameter statistics or predefined heuristics to mitigate this issue. However, the noise magnitude and injection timing in these approaches depend on fixed hyperparameters or heuristic schedules, making them unable to adapt to the dynamic optimization state of the model during training. To address this limitation, this paper proposes a novel fine-tuning framework named gradient-guided noise injection (GNI). The core idea is to utilize gradient information generated during optimization to dynamically adjust the noise intensity in a layer-wise and real-time manner. Extensive experiments on multiple tasks from the GLUE and SuperGLUE benchmarks demonstrate that GNI generally improves average performance across different pre-trained models and task settings, while exhibiting stable generalizability and robustness. This work provides a simple yet effective solution for optimization-state aware dynamic regularization.
Qinglin Jiang, Cheng Zeng, Nan Chi· Scientific Reports· 0 citations
Shani et al. (2026) show that LLM representations broadly recover human category boundaries, while failing to reflect fine-grained typicality structure. Their analysis uses cosine similarity over dense model representations. We revisit their approach using overlap over active sparse autoencoder (SAE) latent sets as a more interpretable similarity measure. We first verify that this set-level measure is meaningful: SAE latent sets can recover union-like compositional structure in controlled toy models and induce semantically coherent neighborhoods in natural text. Extending the human-concepts analysis to SAE set similarities, we find that SAE activation sets do not recover human category boundaries or within-category typicality more faithfully than dense embeddings or residual-stream states, but instead track model-internal similarity structure. To probe this gap further, we study active latent sets under well-controlled semantic modifications, revealing a substantial mismatch between human judgements of conceptual change and change in the SAE active set. We interpret this as evidence that, outside idealised settings, SAE features do not compose via simple bag-of-features semantics.
Nikolai Bolik, Lennart Stöpler, Artur Andrzejak· 0 citations
: Dimensionality reduction is essential for analyzing and visualizing high-dimensional data, with t-distributed Stochastic Neighbor Embedding (t-SNE) being widely used due to its ability to preserve local neighborhood structures. However, its reliance on a single Kullback–Leibler (KL) divergence objective often leads to poor global structure preservation and sensitivity to local inconsistencies. In this paper, we propose a progressive multi-objective optimization framework that enhances t-SNE by integrating complementary loss functions, including a ranking-aware divergence (KLmax) and a Wasserstein-based term for global alignment. Rather than optimizing all objectives simultaneously, we introduce a progressive training strategy that gradually incorporates these components, enabling more stable convergence and improved embedding quality. Additionally, the framework is applied to latent representations learned via a neural encoder, providing a more structured feature space for dimensionality reduction. Experiments on the MNIST, Fashion-MNIST, CIFAR-10, and STL-10 datasets demonstrate that the proposed method improves clustering performance and yields more interpretable embeddings than standard and extended t-SNE approaches.
S. Belhaouari, Skander Bensegueni, Lyes Fennour et al.· Proceedings of the 15th Inte...· 0 citations
Efficiently aligning visual features with Large Language Models (LLMs) remains a critical bottleneck in Multimodal LLMs. Existing query-based alignment modules (e.g., Q-Former) rely on randomly initialized queries, resulting in an inefficient cold start exploration process. Furthermore, they enforce uniform cross-attention across all layers, leading to computational redundancy. Our empirical analysis reveals that query tokens initialized with language priors can rapidly capture global semantics, leading to early representation convergence after only a few layers. In this paper, we propose Cat-MoD , a Ca ption t oken Guided Asymmetric M ixture-o f-D epths framework. It incorporates a Hybrid Query Construction module where Guide Tokens initialized from coarse-grained linguistic priors rapidly anchor global semantic context, and randomly initialized Explorer Tokens remain active to capture fine-grained visual details. Leveraging this early convergence, we introduce an Asymmetric Mixture-of-Depths mechanism, where a similarity-aware router dynamically prunes redundant tokens from expensive cross-attention layers while preserving their context in self-attention. Experiments on multiple benchmarks demonstrate that Cat-MoD matches
Yijie Huang, Xiaocui Yang, Shi Feng et al.· Annual Meeting of the Associ...· 0 citations