PaperlayerView exact revision on arXiv

Source-linked AI summary

Locate-then-Merge: Neuron-Level Parameter Fusion for Mitigating Catastrophic Forgetting in Multimodal LLMs

Zeping Yu, Sophia Ananiadou

arXiv:2505.16703v1cs.CL

TL;DR

Visual instruction tuning can cause catastrophic forgetting of an MLLM’s base-language abilities, creating a need to preserve language while retaining visual adaptation. Locate-then-Merge addresses this with training-free selective parameter fusion, and Neuron-Fusion consistently outperforms existing merging methods across 13 language and vision benchmarks while reducing context hallucination. The method is currently scoped to particular modalities, architectures, and base/fine-tuned model pairs.

  • Problem

    Visual instruction tuning can severely degrade the general language abilities of MLLMs, even on complex reasoning benchmarks and in strong models such as Llama3.

    Source: §1 · p. 1

  • Method

    Locate-then-Merge is a training-free framework that locates important parameter changes and selectively merges them, with Neuron-Fusion preserving large-change neurons while suppressing smaller changes.

    Source: §1 · p. 2, §4.1 · p. 4, §4.2 · p. 4

  • Results

    Neuron-Fusion consistently outperforms existing model-merging techniques across 13 language and vision benchmarks and reduces context hallucination in generation.

    Source: §1 · p. 2, §6 · p. 8

  • Takeaways & Limitations

    Neuron-level selective fusion can retain visual adaptation while mitigating language degradation and reducing context hallucination in MLLMs.

    Source: §1 · p. 2, §6 · p. 8

    The method is scoped to vision-focused, decoder-only settings and merging a base model with its same-architecture fine-tuned counterpart.

    Source: Limitations · p. 9

Abstract

from arXiv · show

Although multimodal large language models (MLLMs) have achieved impressive performance, the multimodal instruction tuning stage often causes catastrophic forgetting of the base LLM's language ability, even in strong models like Llama3. To address this, we propose Locate-then-Merge, a training-free parameter fusion framework that first locates important parameters and then selectively merges them. We further introduce Neuron-Fusion, a neuron-level strategy that preserves the influence of neurons with large parameter shifts--neurons likely responsible for newly acquired visual capabilities--while attenuating the influence of neurons with smaller changes that likely encode general-purpose language skills. This design enables better retention of visual adaptation while mitigating language degradation. Experiments on 13 benchmarks across both language and visual tasks show that Neuron-Fusion consistently outperforms existing model merging methods. Further analysis reveals that our method effectively reduces context hallucination in generation.

1 Introduction

MLLM visual instruction tuning can severely degrade base-language abilities, motivating a training-free framework that locates and selectively merges important parameter changes. Neuron-Fusion preserves large-change neurons and outperforms existing merging methods across 13 language and vision benchmarks.

Motivation

  • Visual instruction tuning can severely degrade general language abilities, including on complex reasoning benchmarks and in strong models such as Llama3.This catastrophic forgetting limits MLLM generalization and practical deployment across diverse language tasks.Source: §1 · p. 1
  • Existing model-merging methods were mainly developed for singlemodal LLMs, leaving their effectiveness for MLLMs and systematic design guidance unclear.The paper frames targeted recovery of language abilities while preserving visual adaptation as an open need.Source: §1 · pp. 1–2

Method

  • Locate-then-Merge is a training-free framework that separates locating important parameters from selectively merging them.It provides a general framework for parameter fusion rather than directly replacing the tuned model with the base model.Source: §1 · p. 2
  • Neuron-Fusion preserves neurons with large parameter changes, hypothesized to encode visual capabilities, while suppressing widespread small changes associated with language forgetting.The design aims to retain acquired visual skills while minimally affecting general language abilities.Source: Figure 1 · p. 1, §1 · p. 2

Results

  • Neuron-Fusion consistently outperforms five state-of-the-art merging methods across 13 language and vision benchmarks on two open-source MLLMs.The experiments evaluate both language retention and visual adaptation.Source: §1 · p. 2
  • Generation analysis shows that Neuron-Fusion reduces context hallucination, improving output quality and controllability.This analysis identifies a generation-level benefit beyond benchmark performance.Source: §1 · p. 2

2 Related Work

The related work describes MLLM architectures, catastrophic forgetting during multimodal adaptation, model merging as a data-free remedy, and evidence that capabilities can be localized to neurons.

2.1 Multimodal Large Language Models

  • MLLMs combine a modality encoder, connector, and pretrained LLM, which are jointly fine-tuned for vision-language tasks.The vision encoder produces embeddings, the connector maps them into the LLM feature space, and the LLM generates conditioned outputs.Source: §2.1 · p. 2

2.2 Catastrophic Forgetting and Model Merging

  • Catastrophic forgetting is the loss of previously acquired knowledge when a model learns new capabilities, and visual instruction tuning can impair the underlying LLM’s language abilities.Prior work reports that larger models may forget more information and relates forgetting to parameter deviation from previous-task optima.Source: §2.2 · p. 2
  • Model merging combines parameters without training-data access and has been used to improve performance across tasks while potentially mitigating catastrophic forgetting.Prior strategies include several forms of parameter fusion developed across different models and tasks.Source: §2.2 · pp. 2–3

2.3 Neuron-Level Ability Storage in LLMs

  • Prior neuron-level studies associate transformer neurons and FFN or attention representations with factual, arithmetic, multimodal, and controllable language abilities.These findings motivate locating parameter changes at neuron granularity.Source: §2.3 · p. 3

3 Background and Problem Formulation

The paper formulates MLLM adaptation as a trade-off between retaining visual gains and recovering language ability after visual instruction tuning. It seeks a merged model that preserves both capabilities without retraining.

3.1 Architectures of LLM and MLLM

  • An MLLM adds a vision encoder and connector to an LLM, mapping image embeddings into the tuned LLM before text generation.The tuned LLM parameters result from visual instruction tuning on image-text pairs.Source: §3.1 · p. 3
  • The base LLM processes text alone, whereas the MLLM combines textual inputs with mapped visual embeddings to produce textual outputs.The architecture difference is the basis for studying visual adaptation and language retention.Source: §3.1 · p. 3

3.2 Problem Formulation

  • The parameter delta is defined as Δ = θ_tuned − θ_base, representing changes introduced during visual instruction tuning.The formulation compares the tuned and base LLM weights directly.Source: §3.2 · p. 3
  • Visual tuning produces visual ability V_tuned but can reduce language ability from L_base to L_tuned because language-relevant parameters are inadvertently modified.The paper states that typically L_tuned < L_base.Source: §3.2 · p. 3
  • Hard-merge replaces the tuned LLM with the base LLM while retaining the vision encoder and connector.This can recover language parameters but often substantially reduces visual ability, with V_base significantly lower than V_tuned.Source: §3.2 · p. 3
  • The target merged model should achieve language ability comparable to L_base while keeping visual ability comparable to V_tuned.The desired outcome is simultaneous preservation of language and visual capabilities.Source: §3.2 · p. 3

4 Methodology

The methodology introduces Locate-then-Merge, which separates locating important parameter changes from selectively merging them, and instantiates it as Neuron-Fusion for preserving visual adaptation while mitigating catastrophic forgetting.

4.1 Locate-then-Merge Framework

  • Locate-then-Merge separates parameter fusion into locating a subset of task-vector changes and transforming and merging those changes into the base model.Existing merging methods differ mainly in their locating and rescaling choices within this framework.Source: §4.1 · p. 4
  • The framework views existing methods as special cases, including magnitude trimming and sign resolution in TIES, tail masking in Breadcrumbs, random dropping in DARE, and magnitude-dependent dropout in DELLA.Source: §4.1 · p. 4

4.2 Neuron-Level Motivation and Analysis

  • Neuron-level analysis finds that a small number of neurons undergo substantially larger changes than most neurons across representative FFN and attention matrices.Neuron change is measured by summing absolute parameter differences associated with each neuron; Figures 3 and 4 plot neuron indices against change scores.Source: §4.2 · pp. 4–5, Figure 3 · p. 5, Figure 4 · p. 5

4.3 Neuron-Fusion Method for Mitigating Catastrophic Forgetting

  • Neuron-Fusion preserves large-change neurons, which are hypothesized to encode visual capabilities, while suppressing widespread small-change parameters that may disrupt language abilities.The method is motivated by the structural role of neurons as units that encode distinct capabilities.Source: §4.2 · p. 4, §4.3 · p. 5
  • Neuron-Fusion first selects the top M% neurons by aggregated absolute change, then retains K% of module parameters and restores selected neurons through replacement or rescaling.Neuron-Suppress reduces small-change influences, while Neuron-Restore reinstates parameters or approximately restores each preserved neuron's original change score.Source: §4.3 · p. 5
  • The restoration step preserves important-neuron influence because suppressing parameters can weaken FFN coefficients that determine weighted contributions of key-value memory subvalues.Under roughly uniform inputs, rescaling can approximately restore the original coefficient while retaining the selected neuron's influence.Source: §4.3 · p. 6

5 Experiments

Experiments evaluate Neuron-Fusion across language and visual benchmarks, showing strong overall retention and targeted improvements in multimodal generation.

5.1 Experimental Settings

  • The evaluation covers seven language benchmarks and six visual benchmarks spanning reasoning, comprehension, perception, and visual question answering.Language evaluation uses exact-match accuracy, with 8-shot chain-of-thought for GSM8K and zero-shot evaluation elsewhere.Source: §5.1 · p. 6

5.2 Results of Neuron-Fusion Method

  • Neuron-Fusion consistently achieves the best overall ability across Llama3 and Mistral while preserving both language and visual abilities.Neu-P-TaskA and Neu-P-Bread obtain the highest scores on both models; DARE and DELLA perform worse, with language scores below the base LLMs.Source: §5.2 · pp. 6–7, Table 2 · p. 7, Table 3 · p. 7
  • At K=0.1, visual ability reaches its minimum before improving and declining again, while Task Arithmetic, TIES, and Breadcrumbs show decreasing language performance as density increases.DARE and DELLA remain below the MLLM on language ability across all tested densities, indicating they do not recover degraded language skills.Source: §5.2 · p. 7, Figure 6 · p. 7, Figure 7 · p. 7
  • Increasing the restored-neuron ratio from M=0.0 to 0.5 gradually reduces language ability, while visual ability peaks near M=0.3 before declining.The results support a moderate restoration ratio: restoring large-change neurons improves vision, whereas restoring additional neurons yields diminishing or harmful returns.Source: §5.2 · pp. 7–8, Figure 8 · p. 8
  • On ScienceQA, accuracy rises from 0% after hard-merge to 7% after suppression and 52.6% after restoration with M=0.3.The analysis examines transitions involving refusal-to-answer and context-hallucination errors.Source: §5.2 · p. 8, Table 4 · p. 8
  • Neuron-Suppress corrects 56.7% of corrected Not-Known cases, whereas Neuron-Restore corrects 97.2% of Context-Hallucination cases.These stages address different multimodal generation failure modes: answer uncertainty and content not grounded in the input.Source: §5.2 · p. 8

6 Conclusion

The paper introduces Locate-then-Merge and Neuron-Fusion to mitigate catastrophic forgetting in MLLMs by selectively preserving important neuron changes. Across language and vision benchmarks, the method improves capability retention and reduces generation failures.

6 Conclusion

  • Locate-then-Merge is a training-free framework that locates important parameters before selectively merging them.Neuron-Fusion preserves neurons with large parameter shifts while suppressing harmful widespread changes.Source: §6 · p. 8
  • Neuron-Fusion consistently outperforms existing merging techniques while retaining both language and visual capabilities.Generation analysis further reports fewer Not-Known and Context-Hallucination failures, producing more reliable and controllable outputs.Source: §6 · p. 8

Limitations

The proposed approach is evaluated within a limited scope: vision-based instruction tuning, decoder-only architectures, and merging a base model with its fine-tuned counterpart. Its applicability to other modalities, architectures, and mismatched model-merging settings remains untested.

  • The study focuses on vision and visual instruction tuning, leaving extension to audio or video uninvestigated.Source: Limitations · p. 9
  • Neuron-Fusion is developed specifically for decoder-only LLMs, so its generalizability across other multimodal architectures remains open.Source: Limitations · p. 9
  • Locate-then-Merge assumes a base model and fine-tuned counterpart share an architecture and differ only in learned weights.This limits direct applicability to architecturally different models or independently trained models with different objectives or tasks.Source: Limitations · p. 9

A.1 Catastrophic Forgetting in MLLMs

Visual instruction tuning substantially degrades language performance in Llama3 and Mistral, while directly restoring base parameters also harms visual performance. These results establish the need for a method that balances language recovery with visual retention.

  • Figures 9 and 10 report Llama3 and Mistral accuracy on language datasets after visual instruction tuning and on vision datasets after hard-merging into tuned MLLMs.Source: Figure 9 · p. 12, Figure 10 · p. 12
  • Language accuracy generally decreases after visual instruction tuning, with especially large drops on GSM8K and ARC-Challenge.Mistral’s CommonsenseQA result is the stated exception.Source: §A.1 · p. 12, Figure 9 · p. 12
  • Hard-merging original base-model parameters into tuned MLLMs causes significant accuracy drops across visual datasets for both Llama3 and Mistral.Replacing tuned parameters therefore does not preserve visual capabilities while recovering language ability.Source: §A.1 · p. 12, Figure 10 · p. 12

A.2 Results on All Datasets

The reported evaluation covers seven language datasets and six visual datasets for both Llama3-7B and Mistral-7B, with detailed results provided in separate tables.

  • The evaluation spans 6 visual datasets and 7 language datasets.Source: §A.2 · p. 12
  • Detailed Llama3 results are presented across the visual and language datasets in Table 5.Source: §A.2 · p. 12
  • Detailed Mistral results are presented across the visual and language datasets in Table 6.Source: §A.2 · p. 12

B Different Density in Mistral

For Mistral, density-dependent language and visual results show trends broadly similar to Llama3. At low density, several merging methods can exceed standalone-LLM accuracy on CommonsenseQA because the MLLM performs better there.

B Different Density in Mistral

Related papers

Loading 2505.16703v1…