An arXiv preprint reports that GRIP, a model-blending approach, recorded similar benchmark accuracy with shorter generated answers in tests of matched reasoning and instruction models. Its average accuracy was 76.5, compared with 76.0 for Qwen3-Thinking, while average generation length was 27.0% lower. Against SLERP, it had the same average accuracy with 14.5% fewer tokens. No uncertainty intervals or significance tests were reported.
The blend gives each module its own setting
GRIP combines reasoning and instruction models with identical architectures, then trains only module-wise interpolation ratios while keeping both source models frozen. Its reward favors responses that are correct and concise. The study asks whether this setup can improve the accuracy-efficiency trade-off over fixed or search-based merging while revealing module-level patterns associated with efficient reasoning.
On Qwen3-4B, the setup used 74 fusion logits: 36 for attention modules, 36 for feed-forward network modules, one for the final RMSNorm and one shared by the tied input embedding and language-model head.
Training prompts came from DeepScaleR-preview. The evaluation covered AIME25, MATH500, GSM8K, GPQA-Diamond and LiveCodeBench.
The official evaluation sets contained 30 AIME25 problems, 500 MATH500 problems, 1,319 GSM8K test problems, 198 GPQA-Diamond questions and 175 LiveCodeBench v6 programming problems.
The comparisons included the original model modes, fixed-ratio merging methods and CMA-ES. The fixed-ratio interpolation baselines used a reasoning-model coefficient of 0.8. CMA-ES used the same prompts and module-wise parameterization as GRIP, with a fitness measure combining accuracy and output length.
The biggest differences appeared on selected tasks
On AIME25, GRIP’s accuracy was 6.7 percentage points higher than Qwen3-Thinking’s, while its output length was 39.7% lower. The paper also reports accuracy gains with shorter reasoning on GPQA-D and the lowest token use among strong reasoning variants on LiveCodeBench.
Targeted module sweeps showed a sharp contrast between attention and feed-forward network, or FFN, settings. Across the attention sweep, macro pass@1 stayed between 0.690 and 0.722 while response length rose 26%. Across the FFN sweep, macro pass@1 rose from 0.612 to 0.782, while length rose 178%. In these tests, FFN coefficients were associated with larger shifts in both accuracy and response length than attention coefficients.
The learned settings also became more varied from layer to layer during training. Inter-layer standard deviation for both attention and FFN coefficients started near zero, rose rapidly during the first approximately 300 steps and saturated around 0.30. The reported pattern indicates that the trained ratios differentiated by layer rather than staying globally uniform.
In a separate ablation, layer-wise interpolation recorded 73.5 average accuracy and 7,571 tokens, while module-wise interpolation recorded 76.5 average accuracy and 7,930 tokens. The reported difference was 3.0 accuracy points. The layer-wise result was evaluated at step 230, whereas the module-wise result was the reported GRIP configuration.
The reported optimization traces were also different. GRIP’s median adjacent-update RMS shift was 0.018, compared with 0.083 for CMA-ES; its path-to-net-displacement ratio was 4.9 times, compared with 21.8 times for CMA-ES, which used 600 fitness evaluations. These measures describe the shape of the optimization path, not uncertainty around the benchmark scores.
What the tests do not settle
The evidence is limited to 4B dense Transformer models and matched pairs with identical architectures within one model family. The study did not evaluate transfer to models at 30B or above, Mixture-of-Experts systems or fusing models from different families.
That leaves open whether the reported accuracy-token pattern would hold at larger scales, with other architectures or across model families.
The document is an arXiv v1 preprint dated 26 August 2026. The authors state that they will release the GRIP source code to support reproducibility.
Paper data and sources
Original title: GRIP: Granular Reward-Guided Parameter Interpolation for Efficient Reasoning
Authors: Lam So, Canhui Wu, Han Lin
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-26
DOI: Not available
Original paper · Full text