Revisiting Model Interpolation for Efficient Reasoning

TL;DR

This work investigates efficient reasoning in large language models by revisiting a simple weight interpolation (MI) strategy that blends Thinking (long CoT) and Instruct (short responses) variants. It reveals a consistent three-stage dynamic as the interpolation coefficient $\lambda$ sweeps from 0 to 1, guiding when reasoning emerges and when overthinking occurs. Across 4B and 30B Qwen models, strategically chosen MI configurations surpass advanced merging baselines on math, instruction-following, and science-reasoning benchmarks, while also offering token-efficiency advantages. Extensive ablations map the contributions of layers, Transformer modules, and decoding choices, providing a practical framework for constructing reasoning behaviors with targeted token budgets. The findings offer a scalable, interpretable approach to model merging with broad implications for efficient, controllable reasoning in LLMs.

Abstract

Model merging, typically on Instruct and Thinking models, has shown remarkable performance for efficient reasoning. In this paper, we systematically revisit the simplest merging method that interpolates two weights directly. Particularly, we observe that model interpolation follows a three-stage evolutionary paradigm with distinct behaviors on the reasoning trajectory. These dynamics provide a principled guide for navigating the performance-cost trade-off. Empirical results demonstrate that a strategically interpolated model surprisingly surpasses sophisticated model merging baselines on both efficiency and effectiveness. We further validate our findings with extensive ablation studies on model layers, modules, and decoding strategies. Ultimately, this work demystifies model interpolation and offers a practical framework for crafting models with precisely targeted reasoning capabilities. Code is available at \href{https://github.com/wutaiqiang/MI}{Github}.

Figures (6)

• Figure 1: The performance dynamics for the model interpolation between Instruct and Thinking models. Think #R denotes the ratio of samples with </think> token in responses. Token #N denotes the number of tokens in responses.
• Figure 2: Performance of vanilla Instruct, Thinking, and model merging methods on AIME'25, IFEval, and GPQA-Diamond. MI denotes the model interpolation and the suffix for the interpolation coefficient $\lambda$. The results indicate that MI surpasses these baselines on both efficiency and effectiveness.
• Figure 3: The performance dynamics of model interpolation (MI) on Qwen3-4B-Instruct-2507 and Qwen3-4B-Thinking-2507. The dynamics follow a three-stage evolutionary paradigm colored in HTML]f6f6f6grey, HTML]ecfbeagreen, and HTML]f0f7fablue. $\lambda$ denotes the interpolation coefficient ranging from 0 to 1.
• Figure 4: Performance of MI-0.4 on IFEval and GPQA-Diamond under different decoding strategies on Qwen3-4B. We search for the temperature $T$ and Top-p.
• Figure 5: Ablation on modules to apply model interpolation. Attn denotes the MHA sub-layers and FFN for FFN sublayers. We report the results on AIME'25. Length Ratio denotes the ratio to the Thinking model.