Michael Brenndoerfer | Data & AI, Private Equity, Technology

Michael Brenndoerfer

Model Merging Applications › Visualizations

Michael BrenndoerferFebruary 28, 20268 min read

Language AI HandbookAll chapters

Part 41 of 67Model Compression

Chapter 312Model Merging Applications › Visualizations

Reading section 7 of 10 · Visualizations · 8 minSection 7 of 10 · 8 min

Section 7 of 10 · 8 min read

Model Merging Applications › VisualizationsOne section · Sections in this chapter7 / 10 ↓

Sections in this chapter10

Reading tools Definitions & inline links

Choose your expertise level to adjust how many terms are explained. Beginners see more tooltips, experts see fewer to maintain reading flow. Hover over underlined terms for instant definitions.

Article links

Make inline references clickable

Out[11]:

Visualization

Scatter plot of sentiment vs NLI performance for task arithmetic and TIES merging as lambda varies. — Pareto tradeoff between sentiment and NLI performance as the lambda scaling coefficient varies from 0.1 to 1.0. Task arithmetic (blue circles) and TIES merging (orange squares) trace distinct curves, with TIES achieving better combined scores at moderate lambda values. The filled star markers indicate each method's best combined configuration.

Out[12]:

Visualization

Line plot of normalized sentiment scores vs lambda for task arithmetic and TIES. — Normalized sentiment score across lambda values for task arithmetic (blue) and TIES merging (orange). Both curves rise from roughly 0.1, peak near lambda 0.5 at about one-third of the fine-tune gain, then fall as lambda approaches 1.0. The dotted 1.0 line shows the separate fine-tune ceiling and makes the remaining retention gap explicit.

Line plot of normalized NLI scores vs lambda for task arithmetic and TIES. — Normalized NLI score across lambda values. Both methods peak near lambda 0.45 to 0.5 at roughly 0.28, then decline toward zero for the largest lambda settings as cross-task interference overwhelms the NLI gain. The near-overlap shows little separation between these methods in this simulation.

Out[13]:

Visualization

Heatmap of cosine similarity values between four task vector pairs. — Pairwise cosine similarity between task vectors for four simulated tasks. Pairs with high positive similarity (dark blue) are candidates for clean task arithmetic merging. Values near zero (light) indicate orthogonal parameter usage and also merge well. Negative values (toward red) signal opposing weight changes where TIES or DARE conflict resolution is most beneficial.

Out[14]:

Visualization

2D plot showing LERP and SLERP interpolation paths between two orthogonal vectors. — Interpolation paths of LERP and SLERP between two orthogonal unit vectors in 2D weight space. SLERP (orange) follows the arc of the unit circle while LERP (blue) cuts straight through, producing shorter vectors at intermediate points. The dashed arc marks the unit sphere surface.

Line chart of vector norms vs interpolation coefficient for LERP and SLERP. — Vector norms throughout the interpolation for LERP and SLERP. LERP norms dip to approximately 0.71 at t=0.5, while SLERP norms remain constant at 1.0, showing norm-preserving interpolation that avoids the scale reduction inherent in linear blending.

Out[15]:

Visualization

Grouped bar chart of normalized performance for sentiment and NLI tasks across two merging methods. — Normalized performance scores for task arithmetic and TIES merging at their respective best-performing lambda values. Each bar shows the fraction of the fine-tuned model's gain over the base retained after merging, with 1.0 marking perfect preservation. In this simulation, task arithmetic is marginally higher on both sentiment (0.33 vs 0.32) and NLI (0.28 vs 0.27), while both methods remain well below the separate fine-tune ceiling.