mlsanityData to 14 Oct 2026

LED-Merging: Mitigating Safety-Utility Conflicts in Model Merging with Location-Election-Disjoint

arXiv 2502.16770 ↗First posted 24 Feb 2025cs.CL cs.AIcited by 7

Not on the map. Why?

Abstract (arXiv)

Fine-tuning pre-trained Large Language Models (LLMs) for specialized tasks incurs substantial computational and data costs. While model merging offers a training-free solution to integrate multiple task-specific models, existing methods suffer from safety-utility conflicts where enhanced general capabilities degrade safety safeguards. We identify two root causes: neuron misidentification due to simplistic parameter magnitude-based selection, and cross-task neuron interference during merging. To address these challenges, we propose LED-Merging, a three-stage framework that Locates task-specific neurons via gradient-based attribution, dynamically Elects critical neurons through multi-model importance fusion, and Disjoints conflicting updates through parameter isolation. Extensive experiments on Llama-3-8B, Mistral-7B, and Llama2-13B demonstrate that LED-Merging effectively reduces harmful response rates, showing a 31.4% decrease on Llama-3-8B-Instruct on HarmBench, while simultaneously preserving 95% of utility performance, such as achieving 52.39% accuracy on GSM8K. LED-Merging resolves safety-utility conflicts and provides a lightweight, training-free paradigm for constructing reliable multi-task LLMs. Code is available at hrefhttps://github.com/MqLeet/LED-MergingGitHub.

Key facts. Cited by 7 papers in the map (data through 14 Oct 2026). Builds on 51 of its 79 references that are in the map. First posted on arXiv 24 Feb 2025.

Builds on

51 of 79 references.

References · 51

  1. 2020Language Models are Few-Shot Learnerscited by 32,419
  2. 2022Training language models to follow instructions with human feedbackcited by 15,971
  3. 2023Llama 2: Open Foundation and Fine-Tuned Chat Modelscited by 13,028
  4. 2021Training Verifiers to Solve Math Word Problemscited by 10,109
  5. 2023Direct Preference Optimization: Your Language Model is Secretly a Reward Modelcited by 8,065
  6. 2023Efficient Memory Management for Large Language Model Serving with PagedAttentioncited by 5,944
  7. 2021Measuring Mathematical Problem Solving With the MATH Datasetcited by 5,371
  8. 2021Program Synthesis with Large Language Modelscited by 3,637
  9. 2017Axiomatic Attribution for Deep Networkscited by 3,384
  10. 2017Learning Important Features Through Propagating Activation Differencescited by 1,458
  11. 2024HarmBench: A Standardized Evaluation Framework for Automated Red Teaming and Robust Refusalcited by 1,429
  12. 2022Model soups: averaging weights of multiple fine-tuned models improves accuracy without increasing inference timecited by 1,396
38 more
  1. 2023Inference-Time Intervention: Eliciting Truthful Answers from a Language Modelcited by 1,302
  2. 2023WizardLM: Empowering large pre-trained language models to follow complex instructionscited by 1,125
  3. 2023SparseGPT: Massive Language Models Can Be Accurately Pruned in One-Shotcited by 1,121
  4. 2014Striving for Simplicity: The All Convolutional Netcited by 990
  5. 2024SimPO: Simple Preference Optimization with a Reference-Free Rewardcited by 963
  6. 2019Are Sixteen Heads Really Better than One?cited by 816
  7. 2018SNIP: Single-shot Network Pruning based on Connection Sensitivitycited by 812
  8. 2022Discovering Latent Knowledge in Language Models Without Supervisioncited by 767
  9. 2023Language Models are Super Mario: Absorbing Abilities from Homologous Models as a Free Lunchcited by 694
  10. 2023WizardMath: Empowering Mathematical Reasoning for Large Language Models via Reinforced Evol-Instructcited by 625
  11. 2021Merging Models with Fisher-Weighted Averagingcited by 621
  12. 2024The WMDP Benchmark: Measuring and Reducing Malicious Use With Unlearningcited by 509
  13. 2024The Instruction Hierarchy: Training LLMs to Prioritize Privileged Instructionscited by 493
  14. 2023Trustworthy LLMs: a Survey and Guideline for Evaluating Large Language Models' Alignmentcited by 401
  15. 2022Dataless Knowledge Fusion by Merging Weights of Language Modelscited by 387
  16. 2023LoraHub: Efficient Cross-Task Generalization via Dynamic LoRA Compositioncited by 364
  17. 2023Label Words are Anchors: An Information Flow Perspective for Understanding In-Context Learningcited by 236
  18. 2024Assessing the Brittleness of Safety Alignment via Pruning and Low-Rank Modificationscited by 231
  19. 2024Evolutionary Optimization of Model Merging Recipescited by 208
  20. 2022Patching open-vocabulary models by interpolating weightscited by 190
  21. 2024A Mechanistic Understanding of Alignment Algorithms: A Case Study on DPO and Toxicitycited by 185
  22. 2024MAmmoTH2: Scaling Instructions from the Webcited by 145
  23. 2021Ensemble of Averages: Improving Model Selection and Boosting Performance in Domain Generalizationcited by 144
  24. 2023Model Breadcrumbs: Scaling Multi-Task Model Merging with Sparse Maskscited by 133
  25. 2022Fusing finetuned models for better pretrainingcited by 110
  26. 2024EMR-Merging: Tuning-Free High-Performance Model Mergingcited by 110
  27. 2024Aligner: Efficient Alignment by Learning to Correctcited by 98
  28. 2024Model Stock: All we need is just a few fine-tuned modelscited by 92
  29. 2024Detoxifying Large Language Models via Knowledge Editingcited by 87
  30. 2024VLSBench: Unveiling Visual Leakage in Multimodal Safetycited by 63
  31. 2023Breaking Language Barriers in Multilingual Mathematical Reasoning: Insights and Observationscited by 59
  32. 2020Stochastic Weight Averaging in Parallel: Large-Batch Training that Generalizes Wellcited by 57
  33. 2024Explainable and Interpretable Multimodal Large Language Models: A Comprehensive Surveycited by 52
  34. 2024Identifying Semantic Induction Heads to Understand In-Context Learningcited by 47
  35. 2023LM-Cocktail: Resilient Tuning of Language Models via Model Mergingcited by 44
  36. 2024Model Merging and Safety Alignment: One Bad Model Spoils the Bunchcited by 43
  37. 2024Model Surgery: Modulating LLM's Behavior Via Simple Parameter Editingcited by 23
  38. 2024SoFA: Shielded On-the-fly Alignment via Priority Rule Followingcited by 18

50 most cited of 51 All 51 (JSON)

Built on by

Cited by year (2026 to 14 Oct)

Citing papers per year: 2025: 1; 2026: 6.

7 papers.

  1. 2026Paper2Rebuttal: A Multi-Agent Framework for Transparent Author Response Assistancecited by 9
  2. 2025Human-Agent Collaborative Paper-to-Page Craftingcited by 7
  3. 2026Beyond Parameter Arithmetic: Sparse Complementary Fusion for Distribution-Aware Model Mergingcited by 0
  4. 2026A Data-Efficient Path to Multilingual LLMs: Language Expansion via Post-training PARAMDelta Integration into Upcycled MoEcited by 0
  5. 2026Asymmetric Collapse in Model Merging: When Refusal Over- writes Recognitioncited by 0
  6. 2026When Skills Meet Safety: Benchmarking and Characterizing the Adaptive Jailbreak Robustness of Skill-Merged LLMscited by 0
  7. 2026DiffSafeMerge: Mitigating Backdoor Inheritance in Diffusion Model Mergingcited by 0