arXiv:2606. 00357v1 Announce Type: new Abstract: Training strong large language models (LLMs) requires high-quality supervision, which is often scarce.
arXiv:2607. 18026v1 Announce Type: new Abstract: Can large language models with substantially different parameter spaces be merged by direct weighted averaging, without training or semantic alignment?
CoMerge is a conflict‑driven preference optimization framework for merging multiple expert language models into a single multi‑task model without full retraining. It treats model merging as a preference optimization problem, using self‑supervised, conflict‑driven hard negative samples derived from naive merging defects to refine lightweight, tensor‑wise merging coefficients. Experiments show CoMerge achieves near‑perfect performance on MergeBench and improves instruction‑following and safety on Llama‑3.1‑8B‑Instruct while optimizing only 1,445 scalar coefficients.
arXiv:2606. 19549v1 Announce Type: new Abstract: Low-rank adaptation (LoRA) makes it cheap to train many domain- and task-specific language model adapters, but whether two adapters can be merged is usually discovered only after both have been fully trained and evaluated.
arXiv:2602. 12323v2 Announce Type: replace Abstract: The widespread availability of fine-tuned LoRA modules for open pre-trained models has led to an interest in methods that can adaptively merge LoRAs to improve performance.
CoMerge introduces a conflict‑driven preference optimization framework for merging multi‑task large language models, reframing merging as a preference problem that uses self‑supervised hard negative samples derived from naive merging defects. By optimizing lightweight, tensor‑wise merging coefficients, the method mitigates parameter‑space conflicts while preserving task‑specific capabilities. Experiments show CoMerge achieves an average normalized performance of 0.9968 on MergeBench and improves conflict‑sensitive tasks on Llama‑3.1‑8B‑Instruct, outperforming both data‑free and data‑driven baselines while optimizing only 1,445 scalar coefficients.
arXiv:2605.07111v3 Announce Type: replace-cross Abstract: Recent literature on fine-tuning Large Language Models highlights a fundamental debate. While Full Fine-Tuning (FFT) provides greater represe...
The paper introduces Align‑LoRA, a unified LoRA framework for multi‑task learning that replaces complex, isolated adapter designs with a single‑adapter model enhanced by a higher rank and an explicit alignment loss. It demonstrates that a router‑free, multi‑head model with high inter‑head redundancy can outperform more elaborate baselines, and that a unified LoRA can achieve competitive performance while enabling weight merging and zero inference latency. Extensive experiments and theoretical analysis confirm that Align‑LoRA surpasses prevailing approaches, offering a simpler, production‑friendly paradigm for parameter‑efficient fine‑tuning of large language models.
arXiv:2509. 23982v2 Announce Type: replace-cross Abstract: Preference alignment is a critical step in making Large Language Models (LLMs) useful and aligned with (human) preferences.
InfiFPO is a new preference optimization technique for implicit model fusion that replaces the reference model in Direct Preference Optimization with a fused source model, preserving probability information and avoiding complex vocabulary alignment. By incorporating probability clipping and max‑margin fusion, it aligns a pivot model with human preferences while distilling knowledge from multiple source models. Experiments on 11 benchmarks show that InfiFPO consistently outperforms existing fusion and preference optimization methods, raising Phi‑4’s average score from 79.95 to 83.33 and improving performance in mathematics, coding, and reasoning tasks.
arXiv:2606. 07690v1 Announce Type: cross Abstract: Finetuning data selection requires balancing two competing goals: selecting examples that improve the downstream objective, and doing so without repeatedly finetuning models.
The paper introduces READ, a method for composing low‑rank adapters (LoRA) in large language models. By rewriting each adapter into a balanced canonical form and enforcing a one‑directional coupling, READ allows new skills to read but never write into the output subspaces of existing skills, eliminating interference. Experiments on four benchmark suites and two model families show that READ consistently outperforms existing baselines, improving SuperGLUE scores by over twenty points and domain suite scores by more than seven points.
arXiv:2606. 01717v1 Announce Type: new Abstract: Instruction tuning aligns large language models, including multimodal ones, with diverse user intents, but scaling to heterogeneous mixtures is hindered by gradient interference and bandwidth-heavy synchronization.