Selective Fine-Tuning

Large language models require continuous adaptation to new tasks while preserving safety alignment. However, fine-tuning on even benign data often compromises safety behaviors, including refusal of harmful requests, truthfulness, and commonsense reasoning. We investigate which training samples cause alignment drift through a data-centric lens. Our empirical analysis shows samples contribute unequally: high-gradient samples cause greater safety degradation and drive models toward pretrained distributions, while moderate-gradient samples enable task learning with minimal alignment loss. We propose gradient-based sample selection that filters high-gradient samples during fine-tuning. Across multiple model families on continual domain tasks, our method substantially improves alignment preservation while maintaining competitive task performance, without requiring curated safe data or architectural modifications. Our method is robust across selection ratios, task orderings, and diverse attack benchmarks.

Large Language Models (LLMs) have demonstrated excellent performance in general language understanding, generation and other tasks. However, when fine-tuning for specific domain tasks, the general knowledge accumulated in the pre-training phase is often partially overwritten or forgotten due to parameter updates, which severely limits the generalization ability and transferability of LLMs. Traditional fine-tuning strategies mostly train on the entire parameter space, ignoring the heterogeneity of model parameters, that is, some parameters are extremely important for general tasks, while other parameters are more sensitive to specific tasks. To alleviate the above problems, this paper innovatively proposes a parameter element importance evaluation method, which divides parameters into "core parameters" and "non-core parameters" by distinguishing the importance of parameters for general language ability tasks and specific domain tasks, and fixes the core parameters during fine-tuning, and only fine-tunes the non-core parameters. Extensive experiments on scientific, medical and physical tasks using GPT-J and LLaMA-3 show that our method can mitigate catastrophic forgetting while enhancing the adaptability of the model.

Supervised Fine-Tuning (SFT) of large language models often suffers from task interference and catastrophic forgetting. Recent approaches alleviate this issue by isolating task-critical parameters during training. However, these methods represent a static solution to a dynamic problem, assuming that parameter importance remains fixed once identified. In this work, we empirically demonstrate that parameter importance exhibits temporal drift over the course of training. To address this, we propose Evolving Parameter Isolation (EPI), a fine-tuning framework that adapts isolation decisions based on online estimates of parameter importance. Instead of freezing a fixed subset of parameters, EPI periodically updates isolation masks using gradient-based signals, enabling the model to protect emerging task-critical parameters while releasing outdated ones to recover plasticity. Experiments on diverse multi-task benchmarks demonstrate that EPI consistently reduces interference and forgetting compared to static isolation and standard fine-tuning, while improving overall generalization. Our analysis highlights the necessity of synchronizing isolation mechanisms with the evolving dynamics of learning diverse abilities.

While critical for alignment, Supervised Fine-Tuning (SFT) incurs the risk of catastrophic forgetting, yet the layer-wise emergence of instruction-following capabilities remains elusive. We investigate this mechanism via a comprehensive analysis utilizing information-theoretic, geometric, and optimization metrics across model scales (1B-32B). Our experiments reveal a distinct depth-dependent pattern: middle layers (20%-80%) are stable, whereas final layers exhibit high sensitivity. Leveraging this insight, we propose Mid-Block Efficient Tuning, which selectively updates these critical intermediate layers. Empirically, our method outperforms standard LoRA up to 10.2% on GSM8K (OLMo2-7B) with reduced parameter overhead, demonstrating that effective alignment is architecturally localized rather than distributed. The code is publicly available at https://anonymous.4open.science/r/base_sft.

Standard LoRA fine-tuning of Mixture-of-Experts (MoE) models applies adapters to every expert, yet our profiling shows that per-layer expert routing is highly skewed: a small subset of experts handles most tokens in each layer, while many others are rarely activated ("cold"). We propose MoE-Sieve, a simple routing-guided framework for LoRA fine-tuning, and pair it with a systematic profiling study of expert routing across architectures and tasks. The method is simple: profile routing counts on a small calibration set, select the top-k most-routed experts per layer, and apply LoRA only to those experts. Across two architecturally distinct MoE models and three diverse tasks, tuning only the top 25% routed experts per layer remains competitive with full LoRA, with mean differences within +/-1 percentage point across all conditions. This reduces LoRA trainable parameters by 70-73%, adapter checkpoint size by 71-73%, and wall-clock training time by up to 50%. We also observe a non-monotonic relationship between expert count and seed-to-seed variance, consistent with the hypothesis that adapting cold experts can introduce gradient noise without improving accuracy. Further ablations show that random expert selection at matched budget is about 2.5 percentage points worse, indicating that the routing signal matters, while greedy per-layer budget optimization does not improve over uniform top-k.

Federated video action recognition enables collaborative model training without sharing raw video data, yet remains vulnerable to two key challenges: model exposure and communication overhead. Gradients exchanged between clients and the server can leak private motion patterns, while full-model synchronization of high-dimensional video networks causes significant bandwidth and communication costs. To address these issues, we propose Federated Differential Privacy with Selective Tuning and Efficient Communication for Action Recognition, namely FedDP-STECAR. Our FedDP-STECAR framework selectively fine-tunes and perturbs only a small subset of task-relevant layers under Differential Privacy (DP), reducing the surface of information leakage while preserving temporal coherence in video features. By transmitting only the tuned layers during aggregation, communication traffic is reduced by over 99% compared to full-model updates. Experiments on the UCF-101 dataset using the MViT-B-16x4 transformer show that FedDP-STECAR achieves up to 70.2% higher accuracy under strict privacy ($ε=0.65$) in centralized settings and 48% faster training with 73.1% accuracy in federated setups, enabling scalable and privacy-preserving video action recognition. Code available at https://github.com/izakariyya/mvit-federated-videodp

Deep learning models for meteorological forecasting often fail in rare but high-impact events such as typhoons, where relevant data is scarce. Existing fine-tuning methods typically face a trade-off between overlooking these extreme events and overfitting them at the expense of overall performance. We propose TaCT, an interpretable concept-gated fine-tuning framework that solves the aforementioned issue by selective model improvement: models are adapted specifically for failure cases while preserving performance in common scenarios. To this end, TaCT automatically discovers failure-related internal concepts using Sparse Autoencoders and counterfactual analysis, and updates parameters only when the corresponding concepts are activated, rather than applying uniform adaptation. Experiments show consistent improvements in typhoon forecasting across different regions without degrading other meteorological variables. The identified concepts correspond to physically meaningful circulation patterns, revealing model biases and supporting trustworthy adaptation in scientific forecasting tasks. The code is available at https://anonymous.4open.science/r/Concept-Gated-Fine-tune-62AC.

The rapid expansion of electronic health record (EHR) systems has generated large volumes of unstructured clinical narratives that contain valuable information for disease identification, patient cohort discovery, and clinical decision support. Extracting structured knowledge from these free-text documents remains challenging because clinical language is highly specialized, labeled datasets are limited, and full fine-tuning of large pretrained language models can require substantial computational resources. Efficient adaptation strategies are therefore essential for practical clinical natural language processing applications. This study proposes a parameter-efficient selective fine-tuning framework for adapting GPT-2 to clinical text classification tasks. Instead of updating the entire pretrained model, the majority of network parameters are frozen, and only the final Transformer block, the final layer normalization module, and a lightweight classification head are updated during training. This design substantially reduces the number of trainable parameters while preserving the contextual representation capabilities learned during pretraining. The proposed approach is evaluated using radiology reports from the MIMIC-IV-Note dataset with automatically derived CheXpert-style labels. Experiments on 50,000 radiology reports demonstrate that selective fine-tuning achieves approximately 91% classification accuracy while updating fewer than 6% of the model parameters. Comparative experiments with head-only training and full-model fine-tuning show that the proposed method provides a favorable balance between predictive performance and computational efficiency. These results indicate that selective fine-tuning offers an efficient and scalable framework for clinical text classification.

Parameter-Efficient Fine-Tuning (PEFT) has become a key strategy for adapting large language models, with recent advances in sparse tuning reducing overhead by selectively updating key parameters or subsets of data. Existing approaches generally focus on two distinct paradigms: layer-selective methods aiming to fine-tune critical layers to minimize computational load, and data-selective methods aiming to select effective training subsets to boost training. However, current methods typically overlook the fact that different data points contribute varying degrees to distinct model layers, and they often discard potentially valuable information from data perceived as of low quality. To address these limitations, we propose Gradient-aligned Sparse Tuning (GAST), an innovative method that simultaneously performs selective fine-tuning at both data and layer dimensions as integral components of a unified optimization strategy. GAST specifically targets redundancy in information by employing a layer-sparse strategy that adaptively selects the most impactful data points for each layer, providing a more comprehensive and sophisticated solution than approaches restricted to a single dimension. Experiments demonstrate that GAST consistently outperforms baseline methods, establishing a promising direction for future research in PEFT strategies.

Safety alignment is essential for the responsible deployment of Large Language Models (LLMs). Yet, existing approaches often rely on heavyweight fine-tuning that is costly to update, audit, and maintain across model families. Full fine-tuning incurs substantial computational and storage overhead, while parameter-efficient methods, e.g., Low-Rank Adaptation (LoRA), trade efficiency for inconsistent safety gains and sensitivity to design choices. Safety intervention mechanisms reduce unsafe outputs without modifying model weights, but do not directly shape or preserve the internal representations that govern safety behavior. We present NeST, a Neuron-Selective Tuning framework for efficient post-hoc safety alignment. NeST identifies safety-relevant feed-forward neurons via activation probing on vanilla harmful and benign prompts, clusters neurons with similar activation profiles, and trains shared cluster-level updates while freezing the rest of the model. Importantly, NeST is trained only on vanilla malicious prompts, without using jailbreak-specific attack data, yet generalizes robustly to diverse jailbreaks. The learned updates are then folded into the original weights, incurring no inference-time overhead. Evaluated on 14 open-weight language and multimodal models, NeST outperforms lightweight baselines and approaches full fine-tuning robustness with significantly fewer trainable parameters. On text-only models, NeST reduces average jailbreak attack success rate from 44.5% to 1.1% while training only 0.4M parameters on average. Across multimodal settings, it reduces ASR from 55.3% to 1.1%, and for downstream fine-tuned variants, it restores safety by reducing ASR from 53.8% to 0.8%. These results show that robust, maintainable safety alignment can be achieved by concentrating adaptation on localized, functionally coherent safety structures.

Text guided diffusion models are used by millions of users, but can be easily exploited to produce harmful content. Concept unlearning methods aim at reducing the models' likelihood of generating harmful content. Traditionally, this has been tackled at an individual concept level, with only a handful of recent works considering more realistic concept combinations. However, state of the art methods depend on full finetuning, which is computationally expensive. Concept localisation methods can facilitate selective finetuning, but existing techniques are static, resulting in suboptimal utility. In order to tackle these challenges, we propose TRUST (Targeted Robust Selective fine Tuning), a novel approach for dynamically estimating target concept neurons and unlearning them through selective finetuning, empowered by a Hessian based regularization. We show experimentally, against a number of SOTA baselines, that TRUST is robust against adversarial prompts, preserves generation quality to a significant degree, and is also significantly faster than the SOTA. Our method achieves unlearning of not only individual concepts but also combinations of concepts and conditional concepts, without any specific regularization.

Existing circuit discovery methods rely on templated tasks with clean counterfactuals, limiting their use on diverse natural text. We adapt Contextual Decomposition for Transformers (CD-T) for unstructured settings via label-balanced activation means and task-directional relevance scoring, enabling counterfactual-free circuit discovery. We leverage the discovered circuits for Circuit-Targeted Supervised Fine-Tuning (CT-SFT), restricting parameter updates to task-relevant heads and LayerNorm. Experiments on NusaX cross-lingual sentiment transfer show that CT-SFT is highly competitive for low-resource adaptation. While non-circuit sparse updates and full fine-tuning sometimes match target accuracy through capacity recruitment, CT-SFT most consistently avoids catastrophic forgetting, preserving source-language and related-task performance. Extensions to XNLI support the source-retention and intervention findings on a harder task and two model families, showing that circuit-targeted adaptation provides a more controlled, intervention-supported alternative to global fine-tuning.

We present Adjacent Possible Exploration (APE), a selective fine-tuning method for adapting large language models that systematically explores parameter modifications while maintaining model stability. Inspired by evolutionary optimization principles, APE evaluates multiple candidate parameter updates through fine-tuning on small data subsets and accepts only those exceeding a performance threshold. Unlike standard fine-tuning that follows single gradient directions, APE implements a filtered selection process that prevents destabilizing parameter changes while enabling systematic improvement. Our method achieves 33.9% BLEU improvement and 36.2% perplexity reduction on news summarization tasks while using minimal computational resources. The approach provides a practical framework for controlled model adaptation that balances performance gains with representational stability.