Online Data Selection Is Implicit Alignment

Supervised fine-tuning (SFT) is often treated as a capability-adaptation step, while alignment is attributed to later preference optimization or reinforcement learning. This separation is incomplete: when examples are scored and kept online during fine-tuning, the choice of which data to train on already changes the model's behavioral preferences. We study online data selection as an implicit alignment mechanism. Given the same base model, optimizer, and selected-token budget, we compare random, loss-based, quality-based, and diversity-based online selectors and measure the behavioral drift they induce without any preference optimization. The proposed evaluation tracks helpfulness, refusal rate, verbosity, truthfulness, sycophancy, calibration, and jailbreak robustness, together with diagnostics for which behavioral modes are over-represented in the selected data. We formalize online selection as a reweighted SFT objective whose weights define an implicit preference over response styles and safety postures, so that an online scorer plays the role usually assigned to a reward model. This view predicts that high-scoring data can systematically favor longer, more assertive, more compliant, or more refusal-prone behaviors depending on how the online score is defined. Empirically, selectors that are statistically indistinguishable in task accuracy diverge sharply in refusal rate, verbosity, and sycophancy, and we show that the direction of the shift is predictable from the attribute mixture of the selected data. We introduce Alignment Drift Auditing (ADA), a controlled protocol for quantifying selection-induced behavioral movement, and Alignment-Aware Selection (AAS), a diagnostic online selector that retains data efficiency while constraining drift along safety and style axes.

The Long-Term Effects of Data Selection in LLM Fine-Tuning

Data selection is increasingly used to reduce the cost of large language model (LLM) fine-tuning, with recent methods prioritizing samples by current utility, diversity, quality, or influence. This paper studies a different question: when fine-tuning occurs over multiple stages, can selection strategies that look optimal now make the model less adaptable later? We introduce a long-horizon view of LLM data selection in which a selector is evaluated not only by immediate task performance, but also by future adaptation speed, forgetting, capability imbalance, and out-of-distribution robustness. We compare representative random, loss-based, gradient-based, diversity-based, quality-based, and utility-diversity selection families under a unified multi-stage protocol. Through controlled experiments designed to instantiate this protocol, we show how short-term selectors can exhibit rank reversal: they improve the current stage while slowing subsequent learning and increasing forgetting. We formalize this behavior as \emph{myopic selection}, provide a simple local analysis of why it can occur, and propose a diagnostic Long-Horizon Aware Selection (LHAS) objective that augments immediate utility with coverage, future-proxy transfer, and anti-concentration terms. The study argues that data selection should be evaluated as a training intervention that shapes the model's learning trajectory, rather than only as a local data-efficiency mechanism.

Disentangling Task Conflicts in Multi-Task LoRA via Orthogonal Gradient Projection

Multi-Task Learning (MTL) combined with Low-Rank Adaptation (LoRA) has emerged as a promising direction for parameter-efficient deployment of Large Language Models (LLMs). By sharing a single adapter across multiple tasks, one can significantly reduce storage overhead. However, this approach suffers from negative transfer, where conflicting gradient updates from distinct tasks degrade the performance of individual tasks compared to single-task fine-tuning. This problem is exacerbated in LoRA due to the low-rank constraint, which limits the optimization landscape's capacity to accommodate diverse task requirements. In this paper, we propose Ortho-LoRA, a gradient projection method specifically tailored for the bipartite structure of LoRA. Ortho-LoRA dynamically projects conflicting task gradients onto the orthogonal complement of each other within the intrinsic LoRA subspace. Extensive experiments on the GLUE benchmark demonstrate that Ortho-LoRA effectively mitigates task interference, outperforming standard joint training and recovering 95\% of the performance gap between multi-task and single-task baselines with negligible computational overhead.

Towards Specialized Generalists: A Multi-Task MoE-LoRA Framework for Domain-Specific LLM Adaptation

The rapid evolution of Large Language Models (LLMs) has shifted focus from general-purpose capabilities to domain-specific expertise. However, adapting LLMs to specialized fields such as medicine presents two challenge: (1) the "Stability-Plasticity Dilemma", where the model must acquire complex clinical knowledge without suffering from catastrophic forgetting of general world knowledge; and (2) "Task Interference", where disparate sub-tasks, such as medical diagnosis, report summarization, and drug-drug interaction prediction, compete for limited low-rank parameter space. In this paper, we propose Med-MoE-LoRA, a novel framework that integrates Mixture-of-Experts (MoE) with Low-Rank Adaptation (LoRA) to enable efficient multi-task domain adaptation, especially for medical scenarios. Drawing inspiration from recent advances, our framework employs an asymmetric expert distribution where deeper layers are equipped with a higher density of LoRA experts to capture complex semantic abstractions. We further introduce a "Knowledge-Preservation Plugin", inspired by LoRA MoE, to isolate and protect general-purpose reasoning. By utilizing soft merging with adaptive routing and rank-wise decoupling, Med-MoE-LoRA achieves superior performance in medical benchmarks while reducing interference. Experimental results demonstrate that our approach consistently outperforms standard LoRA and conventional MoE architectures across multiple clinical NLP tasks while retaining the model's general cognitive capabilities.