Conference · Published ·
SPEAR-MM: Selective Parameter Evaluation and Restoration via Model Merging for Efficient Financial LLM Adaptation
Berkcan Kapusuzoglu, Supriyo Chakraborty, Renkun Ni, Stephen Rawls, Sambit Sahu
2025 IEEE International Conference on Big Data · Conference
Summary
SPEAR-MM uses post-hoc layer evaluation and spherical interpolation merging to preserve general capabilities during financial-domain adaptation of language models.
Research contribution
The work presents a selective evaluation and restoration framework for balancing domain adaptation with retention of general capabilities.
Scroll the figure sideways to read its labels, or open the full-size image below.
Lead figure
Evaluate which layers changed, then choose what to restore from the base model and validate the merged configuration.
SPEAR-MM evaluates layer importance and selectively restores parameters through spherical merging. Scores combine signal-to-noise and parameter-change metrics from the base and adapted models.
Source: Berkcan Kapusuzoglu et al., SPEAR-MM (2025), Figure 1. License: CC BY 4.0; rasterized without scientific edits.
Open full-size figure for labels and details.
What the results show
Scroll the figure sideways to read its labels, or open the full-size image below.
Result figure
Choose a configuration for the balance you need: the reported SPEAR-MM settings offer a stronger adaptation-retention trade-off than Baseline A and Baseline B in this evaluation.
Domain adaptation and retention of general capabilities form a controllable trade-off. Each SPEAR-MM point represents a different freezing configuration; the y-axis reports domain performance relative to the non-adapted baseline.
Source: Berkcan Kapusuzoglu et al., SPEAR-MM (2025), Figure 2. License: CC BY 4.0; rasterized without scientific edits.
Open full-size figure for labels and details.
