Naive SLERP Fails Both Aggressively and Conservatively on Qwen2.5 Instruct x Coder Merges, and TIES Works: An Observation-Driven Case Study

© Copyright 2026 Chen, Ho Yiing (ORCID: 0009-0006-6816-9891), Independent researcher, charenix.com. All rights reserved by the author. This work is licensed under the Creative Commons Attribution 4.0.

Three merges of qwen2.5-7B-Instruct with qwen2.5-Coder-7B-Instruct are compared on a five-prompt diagnostic panel:

• v1 — aggressive SLERP (t ramped 0.1 to 0.9 across blocks). Fails with 5 distinct failure modes: persona drift, repetition loop, script drift (Traditional to Simplified Chinese), factual regression on a Python claim, register collapse.
• v2 — conservative SLERP (t ≈ 0.20 everywhere). Fails much harder, counterintuitively: every prompt collapses into meta-commentary ("You are a helpful assistant"), A-through-K listing enumeration, prompt repetition, or hallucinated personae ("dragonfly"). This is the unusual finding of the paper.
• v3 — TIES merge (density 0.55, weights 0.65/0.35). Wins on every prompt at quality comparable to the Instruct baseline, with no language drift, no repetition loop, no catastrophic identity collapse, and two minor artifacts (English code-switch in one Chinese output; code-bug answer diagnoses adjacent issue rather than exact issue).

Key claim: a low-t SLERP death valley exists between 0.15 and 0.25 for this donor pair, contradicting the mergekit-community working assumption that low t is the safer fallback. The right next step is observation-driven TIES parameter design (density from Mercury cell coverage, per-layer weight from lane membership, anchor protection from cross-model overlap), not further SLERP t tuning.

Companion records: Mercury method (10.5281/zenodo.20313154), Mercury discovery (10.5281/zenodo.20313748), Mercury-Viewer dataset (10.5281/zenodo.20313150), Mercury scaling (10.5281/zenodo.20325676).