Research map / Large language models / Instruction and fine-tuning

PortLLM: Personalizing Evolving Large Language Models with Training-Free and Portable Model Patches

ICLR 2025 · Large language models › Instruction and fine-tuning · approach: Fine-tuning generalization and delta compression · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes PortLLM, a training-free method that ports a LoRA-derived personalization patch from an old fine-tuned LLM onto an evolved (continually pretrained) LLM by simply adding the patch weights, with a theoretical argument that the residual is negligible.

Problem

Downstream users who fine-tuned an older LLM version must re-fine-tune whenever the provider updates the base model, which is costly and may be impossible when fine-tuning data is restricted.

Approach

Reuse the LoRA personalization patch Δθ_i trained on the old model by merging it directly into the newer continually pretrained model θ' (θ' + Δθ_i), with no retraining, justified by a claim that the patch difference between versions is negligible in norm relative to the merged model.

Key results

• On Mistral-7B with OpenOrca continued pretraining, the patched model improves zero-shot accuracy over the updated model by up to 26.16 points on GSM8K and 25.38 points on WNLI, and is within about 1 point of the fine-tuned model on most tasks.
• Compared with LoRA fine-tuning on SST-2, the merge uses about 12.2x less GPU memory (28.71 GB vs 350.61 GB) and no trainable parameters.
• Patches also work on full-weight continued pretraining (Mistral-7B), giving gains such as +20.69 on GSM8K and +16.90 on WNLI over the updated model.

Measured effects reported in the paper

Surprising findings

• The patched updated model sometimes outperforms the directly fine-tuned updated model, e.g. on GSM8K (+6.37) and WNLI (+1.02).

Limitations

• Theoretical argument relies on loose assumptions (patch rank much smaller than pretrained rank, singular-value heuristics), and the residual bound is not rigorous for real patches.
• Portability is shown only for the same architecture across versions; the method does not work across different model architectures as acknowledged in the conclusion.
• Evaluation is zero-shot on a limited set of tasks, and some results (e.g., Llama 2 WNLI) show the patch underperforming the fine-tuned model.

Experimental setup

Datasets: OpenOrca SlimOrca OpenPlatypus AlpacaGPT4 BoolQ SST-2 MRPC RTE WNLI WinoGrande GSM8K

Benchmarks: BoolQ SST-2 MRPC RTE WNLI WinoGrande GSM8K

Metrics: accuracy F1 GPU memory usage GPU hours

Baselines: Pretrained LLM Updated LLM (continued pretraining) LoRA fine-tuned model Fine-tuned updated model

Builds on: LoRA Transformers

Closest papers by meaning

• LensLLM: Unveiling Fine-Tuning Dynamics for LLM Selection (ICML 2025 · similarity 0.89)
• Tuning LayerNorm in Attention: Towards Efficient Multi-Modal LLM Finetuning (ICLR 2024 · similarity 0.89)
• Data Efficient Adaptation in Large Language Models via Continuous Low-Rank Fine-Tuning (NeurIPS 2025 · similarity 0.88)
• Knowledge Fusion of Large Language Models (ICLR 2024 · similarity 0.88)
• Generative Adapter: Contextualizing Language Models in Parameters with A Single Forward Pass (ICLR 2025 · similarity 0.88)
• Self-Play Fine-Tuning Converts Weak Language Models to Strong Language Models (ICML 2024 · similarity 0.87)