Negative Results in Cross-Fine-Tune LLM Merge: A Failure-Mode Catalog of SLERP, TIES, DARE, Observation-Driven Rescue, and Passthrough on Qwen2.5-Instruct x Qwen2.5-Coder

I run eight cross-fine-tune blend recipes on the Qwen2.5-7B-Instruct x Qwen2.5-Coder-7B-Instruct pair, evaluate on a five-prompt panel covering multilingual reasoning, structured argument, and code refactor, and report four reproducible failure modes the public model-merging literature has not catalogued:

(1) Conservative SLERP at t ≈ 0.20 produces complete collapse on all five inputs (multiple-choice loops, list collapse), worse than aggressive SLERP at t = 0.9.

(2) DARE-TIES at standard density 0.55 produces uninterpretable token noise on every input.

(3) Observation-driven TIES, setting per-block density from Mercury observation data, regresses on a code-bug input that uniform-density TIES handles correctly.

(4) Scalar-only post-merge anchor rescue patches layer-norm gain at observation-identified anchor coordinates but leaves downstream attention and MLP read paths unchanged; breaks generation on two of five inputs. Full-subspace rescue that also patches the read paths preserves coherence.

Among the eight recipes, simple passthrough Frankenmerge with conservative coder injection (24 Instruct blocks + 4 Coder blocks at tail, "v7b") is the empirical winner. Reading: a fine-tune adapts layer-norm and adjacent paths as an integrated subspace; averaging methods that touch only one component destroy the integration.

Open data, configurations, inference logs: https://github.com/norika1207-lab/mercury-paper-handoff