This study develops a biophysical attribution framework to quantify the sources of divergence in aerosol-induced gross primary production (GPP) anomalies across five CMIP6 Earth System Models (ESMs). The results showed that ecosystem climate sensitivities drive the inter-model spread, rather than aerosol radiative and climatic effects alone. The finding is highly valuable. The manuscript's focus aligns perfectly with GMD's scope of model evaluation and diagnostics. However, there are still some problems with the Methodology and Discussion. Therefore, I recommend that the manuscript be accepted after a major revision.

Major comments:

1. The DFE varies significantly with vegetation type. However, the paper does not adequately consider the impacts of vegetation type on DFE. It would be better to show the model simulations influenced by plant functional types (PFTs) in the paper.
2. The paper presents validation results against FLUXNET and FLUXCOM-X, but lacks detailed discussion of these results in the main text. Please add a brief discussion on how model performance affects the reliability of the attribution analysis, particularly noting the systematic underestimation of high GPP values.
3. Table 1 indicates that IPSL-CM6A-LR does not consider the DFE, while the Discussion (Section 4.1) mentions that the ORCHIDEE_DF version can simulate DFE. Please clarify which version of IPSL-CM6A-LR was used in this study and explain why the ORCHIDEE_DF version was not included for comparison.

Specific comments:

1. The abstract does not specify the study period (1850-2014). Please add this information. Additionally, please change "These findings are very important for fully understanding..." to a more objective expression like "These findings provide critical insights into understanding...".
2. The introduction mentions four pathways through which aerosols affect GPP, but the subsequent analysis mainly focuses on radiation and temperature effects, with less discussion on precipitation effects. Either expand the analysis to include precipitation or adjust the text.
3. Figure 2a: The time series shows positive GPP anomalies for some models between 1850 and 1890. Given the extremely low aerosol loadings during this pre-industrial to early-industrial period, discuss whether these fluctuations are driven by internal climate variability rather than actual aerosol forcing.
4. The terms "diffuse fertilization effect" and "diffuse fertilization effects" are used inconsistently throughout the paper.
5. Please check and ensure consistent formatting of all references, including journal names, volume numbers, page numbers, and DOI formats.

The authors analyse CMIP6 simulations regarding the GPP response to aerosol emissions and present a neat decomposition framework to identify the major differences between the modelled responses. The study highlights an important aspect regarding the importance of vegetation dynamics and responses in driving global carbon dynamics, climate interactions, and differences in model projections. As such, the study is very well suited for publication in GMD. However, there are several aspects that require more thorough discussion and methodological explanation before the paper can be published.

General comments:

The derivation of the attribution framework should be explained in more detail. Could you provide a rational for your steps of using a first-order Taylor expansion and how eq (2) is derived (might be worth writing it similar to eq. 4 first)? Is it equivalent to the total differential? Is there a methodological reason that prevents you from explicitly testing variation in fPAR? And does fitting the regression per PFT remove the fPAR in eq. 6?

The authors state that there are large differences in the ecosystem response to the climatic changes (e.g., photosynthesis to temperature), but could expand in the discussion on the underlying drivers of these differences between vegetation models and how they relate to the observed differences in sensitivities. They do state differences between the vegetation models (e.g., L462-489), but they do not necessarily relate how these differences match with the observed sensitivity differences. Beyond discussion, one way of doing this might be to show how the model fits vary between different PFTs across models. Another suggestion, is it possible to do the analysis spatially and for each year, rather than through time?

The authors state that their framework captures >40% of the differences in the intermodal spread. A discussion of what the remaining variation represents is necessary.

Detailed comments:

Introduction: Could you state what are the main aerosols and aerosol drivers during the considered period?

L171. Please provide a definition of what 𝛿GPP represents and how it is derived from the available simulations. Also, it might help the reader to state at what time resolution you do this decomposition? E.g., the difference in GPP across the entire period 1850-2014 as derived from model and multi-model mean.

L211 and Fig. 1: Could you specify what time period this represents. Also, I would suggest to change the colour scheme, with red indicating loss and blue gain (even more intuitive would probably be a green-brown colour scale).

Fig. 2: Please provide a second y axis label instead of mixing labels from figures. I think it increases clarity. The same applies to Fig. 1 (and other plots of similar style), I would suggest to give an axis label in all panels and at the bottom of the latitudinal mean side panels to make it straightforward to the reader. It would also be good to put labels with the model names at the top of each panel, and not just in the caption.

L400: Please expand on what you mean with ‘light quality’.

L414-416: You provide a mechanistic link to the vegetation model, which is great. If possible, this would be highly informative for the other sensitivities as well.

L491-510: The authors suggest pathways to improve future modelling, but it is not clear based on what they evaluate “accuracy” of any given model? So, either it needs to be clearly presented what they base their conclusion on that “the study suggests approaches to reduce uncertainty”, or this needs to be reformulated in the sense that they identify important mechanisms that need further evaluation.