# XMerge: Cross-Axis Selection and Reconstructive Layer Merging for LLM Depth Compression

arXiv: https://arxiv.org/abs/2609.02083 · HTML: https://mlsanity.com/paper/2609.02083/

## Key facts

- No citing papers in the map yet (data through 14 Oct 2026).
- Builds on 25 of its 27 references that are in the map.
- First posted on arXiv 2 Sep 2026.

## Builds on (most cited first, 25 of 25)

- [The Llama 3 Herd of Models](https://mlsanity.com/paper/2407.21783/) (2024)
- [Distilling the Knowledge in a Neural Network](https://mlsanity.com/paper/1503.02531/) (2015)
- [Qwen3 Technical Report](https://mlsanity.com/paper/2505.09388/) (2025)
- [Measuring Massive Multitask Language Understanding](https://mlsanity.com/paper/2009.03300/) (2020)
- [Qwen2.5 Technical Report](https://mlsanity.com/paper/2412.15115/) (2024)
- [Pointer Sentinel Mixture Models](https://mlsanity.com/paper/1609.07843/) (2016)
- [GPTQ: Accurate Post-Training Quantization for Generative Pre-trained Transformers](https://mlsanity.com/paper/2210.17323/) (2022)
- [Model soups: averaging weights of multiple fine-tuned models improves accuracy without increasing inference time](https://mlsanity.com/paper/2203.05482/) (2022)
- [SparseGPT: Massive Language Models Can Be Accurately Pruned in One-Shot](https://mlsanity.com/paper/2301.00774/) (2023)
- [Editing Models with Task Arithmetic](https://mlsanity.com/paper/2212.04089/) (2022)
- [A Simple and Effective Pruning Approach for Large Language Models](https://mlsanity.com/paper/2306.11695/) (2023)
- [LLM-Pruner: On the Structural Pruning of Large Language Models](https://mlsanity.com/paper/2305.11627/) (2023)
- [DataComp-LM: In search of the next generation of training sets for language models](https://mlsanity.com/paper/2406.11794/) (2024)
- [ShortGPT: Layers in Large Language Models are More Redundant Than You Expect](https://mlsanity.com/paper/2403.03853/) (2024)
- [LaCo: Large Language Model Pruning via Layer Collapse](https://mlsanity.com/paper/2402.11187/) (2024)
- [SLEB: Streamlining LLMs through Redundancy Verification and Elimination of Transformer Blocks](https://mlsanity.com/paper/2402.09025/) (2024)
- [The LLM Surgeon](https://mlsanity.com/paper/2312.17244/) (2023)
- [Streamlining Redundant Layers to Compress Large Language Models](https://mlsanity.com/paper/2403.19135/) (2024)
- [Pruning via Merging: Compressing LLMs via Manifold Alignment Based Layer Merging](https://mlsanity.com/paper/2406.16330/) (2024)
- [ReplaceMe: Network Simplification via Depth Pruning and Transformer Block Linearization](https://mlsanity.com/paper/2505.02819/) (2025)
- [Sliding-Window Merging for Compacting Patch-Redundant Layers in LLMs](https://mlsanity.com/paper/2502.19159/) (2025)
- [Layer as Puzzle Pieces: Compressing Large Language Models through Layer Concatenation](https://mlsanity.com/paper/2510.15304/) (2025)
- [Prune&Comp: Free Lunch for Layer-Pruned LLMs via Iterative Pruning with Magnitude Compensation](https://mlsanity.com/paper/2507.18212/) (2025)
- [A Free Lunch in LLM Compression: Revisiting Retraining after Pruning](https://mlsanity.com/paper/2510.14444/) (2025)
- [The Structure of Quantization Damage in LLMs: Why the Next Bit Should Be Spent Globally](https://mlsanity.com/paper/2609.01587/) (2026)

## Abstract (arXiv)

Removing complete transformer layers preserves a standard serving architecture, but existing depth-compression methods can lose substantial quality, and the loss varies unpredictably across models. We introduce XMerge, a post-training method with two components. Cross-axis selection identifies a block with low relative-magnitude and angular hidden-state change, and local boundary reconstruction re-fits the adjacent surviving block to match the original two-block output. XMerge uses no task labels or end-to-end fine-tuning, and it introduces neither architectural changes nor additional inference-time parameters. Across seven Llama and Qwen backbones (0.5B-8B), five published baselines, and three layer-reduction levels, its advantage over baselines is largest at the most aggressive removal: at k=4 it ranks first on six of seven backbones on CORE (a 22-task aggregate) and, separately, on six of seven on MMLU (five of seven on both at once), while avoiding the large perplexity increases of several competing operators. In a task-level bootstrap, the 95% confidence intervals for the three largest CORE margins exclude zero; the remaining margins are consistent with ties. Across the 14 (model, regime) cells it is also the only evaluated operator that never collapses, ranking top-2 in both zero-shot and in-context regimes; on a first calibration probe (one backbone) it is the best-calibrated operator. Ablations show that local reconstruction provides most of the gain, while cross-axis fusion helps when the two selection axes disagree. The additional construction cost is recovered through per-token decode savings after roughly tens of thousands of requests.