mlsanityData to 14 Oct 2026

On Emergent Capabilities and Model Merging

arXiv 2609.24504 ↗First posted 21 Sep 2026cs.LG cs.AIcited by 0

Not on the map. Why?

Abstract (arXiv)

Fine-tuned checkpoints and adapters now fill public repositories, and the most common operation applied to these artifacts is model merging: arithmetic on their weights that assembles capabilities cheaply. We ask what this operation does to emergent capabilities: behaviors an artifact carries that were never an explicit training target. Studying two independent testbeds (activation oracles and emergent-misaligned models) across three model families, we find that the answer is threefold. First, merging preserves an emergent capability that both parents carry: merging two misaligned checkpoints retains most of their broad misalignment across the whole mixing range. Second, merging cannot create an emergent capability that is superadditive in its parents: no weighted merge of two single-task oracles reaches the jointly-trained oracle's auditing ability. Third, when only one parent carries the capability, merging dilutes it faster than the trained capability that accompanies it: the gap is significant in most settings. In short, emergent behaviors of an artifact do not compose the way its trained capability does.

Key facts. No citing papers in the map yet (data through 14 Oct 2026). Builds on 29 of its 33 references that are in the map. First posted on arXiv 21 Sep 2026.

Builds on

29 of 33 references.

References · 29

  1. 2020Language Models are Few-Shot Learnerscited by 32,419
  2. 2024The Llama 3 Herd of Modelscited by 14,715
  3. 2021LoRA: Low-Rank Adaptation of Large Language Modelscited by 14,380
  4. 2025Qwen3 Technical Reportcited by 8,820
  5. 2024Qwen2.5 Technical Reportcited by 4,942
  6. 2022Emergent Abilities of Large Language Modelscited by 2,512
  7. 2024Gemma 2: Improving Open Language Models at a Practical Sizecited by 1,998
  8. 2022Model soups: averaging weights of multiple fine-tuned models improves accuracy without increasing inference timecited by 1,396
  9. 2022Editing Models with Task Arithmeticcited by 1,120
  10. 2023TIES-Merging: Resolving Interference When Merging Modelscited by 818
  11. 2023Language Models are Super Mario: Absorbing Abilities from Homologous Models as a Free Lunchcited by 694
  12. 2023Are Emergent Abilities of Large Language Models a Mirage?cited by 523
17 more
  1. 2023LoraHub: Efficient Cross-Task Generalization via Dynamic LoRA Compositioncited by 364
  2. 2024Patchscopes: A Unifying Framework for Inspecting Hidden Representations of Language Modelscited by 235
  3. 2023Task Arithmetic in the Tangent Space: Improved Editing of Pre-Trained Modelscited by 221
  4. 2022Predictability and Surprise in Large Generative Modelscited by 210
  5. 2024Task Singular Vectors: Reducing Task Interference in Model Mergingcited by 147
  6. 2020Predicting Neural Network Accuracy from Weightscited by 110
  7. 2025Model Organisms for Emergent Misalignmentcited by 91
  8. 2025Convergent Linear Representations of Emergent Misalignmentcited by 71
  9. 2024SelfIE: Self-Interpretation of Large Language Model Embeddingscited by 67
  10. 2024LoRA Soups: Merging LoRAs for Practical Skill Composition Taskscited by 52
  11. 2025Activation Oracles: Training and Evaluating LLMs as General-Purpose Activation Explainerscited by 52
  12. 2024LatentQA: Teaching LLMs to Decode Activations Into Natural Languagecited by 51
  13. 2022Model Zoos: A Dataset of Diverse Populations of Neural Network Modelscited by 39
  14. 2025Towards eliciting latent knowledge from LLMs with mechanistic interpretabilitycited by 19
  15. 2025On Task Vectors and Gradientscited by 7
  16. 2024Meta-Models: An Architecture for Decoding LLM Behaviors Through Interpreted Embeddings and Natural Languagecited by 5
  17. 2026Activation Steering Induces Emergent Misalignment: A More Comprehensive Evaluationcited by 2

Built on by

Not cited yet.