Skip to main content
Merge strategies are the algorithms that combine two or more models into one candidate. This page defines each strategy, its inputs, parameters, preconditions, failure modes, and known limitations. A decision tree at the end guides strategy selection.

Strategy Selection Decision Tree

Linear Merge

Purpose: Weighted average of two or more models trained on the same task with different seeds or data splits. Inputs: Two or more COMPATIBLE models. Parameters: Preconditions: All models COMPATIBLE. Same architecture, same task. Math: See Merge Math. Failure modes: Weights do not sum to 1.0; dtype mismatch. Known limitations: Assumes models are in the same basin. Models from different tasks may cancel skills.

SLERP

Purpose: Spherical linear interpolation between two models. Preserves weight vector direction better than linear interpolation. Inputs: Two COMPATIBLE models. Parameters: Preconditions: Models COMPATIBLE. Same parameter count. Math: See SLERP Math. Reference: Shoemake 1985. Failure modes: Theta near 0 (fallback to lerp); theta near pi (ambiguous shortest path). Known limitations: Only defined for two models. Extension to N models requires pairwise chaining or geodesic averaging.

TIES

Purpose: Trim, elect sign, and merge. Reduces interference by keeping only the most significant parameter changes and resolving sign conflicts. Inputs: Two or more COMPATIBLE models, or task vectors. Parameters: Preconditions: Models COMPATIBLE or task vectors precomputed. Math: See TIES Math. Reference: Yadav et al. 2023. Failure modes: Density too low (all parameters trimmed); sign consensus unreachable. Known limitations: Requires task vectors for best effect. Direct model merge may retain pretraining noise.

DARE

Purpose: Drop and rescale. Randomly sparsifies task vectors, then rescales survivors to preserve expected magnitude. Inputs: Two or more COMPATIBLE models, or task vectors. Parameters: Preconditions: Models COMPATIBLE or task vectors precomputed. Math: See DARE Math. Reference: Yu et al. 2023. Failure modes: p too high (all parameters dropped); p = 0 (no effect, degenerate). Known limitations: Random sparsification is non-deterministic without fixed seed. Best with multiple task vectors.

DARE+TIES

Purpose: Combine DARE sparsification with TIES sign election for robust multi-model merging. Inputs: Three or more COMPATIBLE models, or task vectors. Parameters: Preconditions: Models COMPATIBLE. Task vectors recommended. Math: See DARE Math and TIES Math. Reference: Yu et al. 2023. Failure modes: Combines DARE and TIES failure modes. Known limitations: More hyperparameters to tune. Best for 3+ models.

Task Arithmetic

Purpose: Add or subtract skills by combining task vectors (fine-tuned minus pre-trained). Inputs: Pre-trained base model, one or more fine-tuned models. Parameters: Preconditions: Base model and fine-tuned models COMPATIBLE. Task vectors computed as theta_ft - theta_base. Math: See Task Arithmetic Math. Reference: Ilharco et al. 2022. Failure modes: Alpha too large (catastrophic forgetting); base model missing. Known limitations: Requires a shared pre-trained base. Cannot merge two independent fine-tunes without a common ancestor.

Franken-Merge / Structural Merge

Purpose: Combine layers or blocks from different models to create a new architecture. Inputs: Two or more models, possibly cross-family. Parameters: Preconditions: Models CONDITIONALLY_COMPATIBLE with explicit adapters, or INCOMPATIBLE with franken-merge override. Failure modes: Layer map references non-existent layer; adapter shape mismatch; activation distribution mismatch at boundary. Known limitations: Requires manual layer map. No automatic search for optimal layer combinations in the initial core.

Passthrough

Purpose: Return the single input model unchanged. Used for baseline evaluation or when only one model is available. Inputs: One model. Parameters: None. Preconditions: Any model. Failure modes: None. Known limitations: Not a merge. Included for pipeline completeness.