Overall assessment

This manuscript presents a comprehensive and ambitious evaluation of Northern Hemisphere tree-cover dynamics over the last ~20 kyr, combining a transient MPI-ESM1.2 simulation with the recently developed hemispheric REVEALS-based tree-cover reconstruction by Schild et al. (2025). The study goes well beyond a descriptive model–data comparison by systematically diagnosing spatial patterns of agreement and disagreement, disentangling climatic drivers using emulation and GAM approaches, and explicitly discussing structural limitations of current dynamic vegetation models (DVMs).

The manuscript is clearly written, methodologically sophisticated, and addresses questions that are highly relevant to the paleoclimate community, particularly those working at the interface of palaeodata synthesis and Earth system modelling. The use of a transient simulation rather than time-slice experiments is a major strength, as is the explicit treatment of non-linearity in climate–vegetation relationships.

Overall, I find this to be a strong and publishable contribution, but in its current form it would benefit from clarifications, tighter framing of some conclusions, and a more critical separation between (1) climate biases, (2) vegetation model structure, and (3) reconstruction limitations. My comments below are intended to strengthen the robustness and interpretability of the results rather than to challenge the major findings.

1. Interpretation of REVEALS tree cover versus modelled absolute cover

A central issue in this manuscript is the comparison between REVEALS-derived tree cover (which sums to 100% vegetation) and MPI-ESM absolute tree cover including bare ground. The authors correctly acknowledge this mismatch and justify their choice to compare absolute PFT area (fi) rather than relative fractions (ci).

However, this choice has far-reaching implications for the interpretation of MAE, variance differences, and the systematic bias pattern (overestimation at low tree cover, underestimation at high tree cover). At present, these implications are discussed mainly qualitatively.

I suggest that the authors (1) add a concise conceptual clarification (possibly a schematic or boxed explanation) explicitly explaining how REVEALS tree cover should be interpreted in open landscapes, and how this affects MAE and variance metrics; (2) Clarify more explicitly that part of the diagnosed non-linear bias pattern (Fig. 6) is methodological rather than purely ecological or model-structural, especially in sparsely vegetated regions; (3) Consider whether at least a sensitivity comparison using relative cover (ci) for selected well-vegetated regions (e.g. mid-Holocene Europe) could help bound the uncertainty. This clarification is important because many readers may otherwise interpret MAE patterns too directly as "model error".

2. Mid-Holocene forest maximum: model limitation or forcing limitation?

A key result is the failure of MPI-ESM to reproduce the reconstructed mid-Holocene tree-cover maximum across large parts of the Northern Hemisphere, with the model instead peaking in the early Holocene. This is an important and robust finding. However, the manuscript currently blends several possible explanations such as overly strong warm-season temperature control linked to insolation, missing processes (permafrost, soils, disturbances), climate biases (e.g. absence of a mid-Holocene thermal maximum), and possible reconstruction artefacts.

I suggest sharpening this discussion by more explicitly separating (a) deficiencies in simulated climate trajectories (e.g. lack of MH warmth or hydroclimate persistence) from (b) deficiencies in vegetation sensitivity to that climate.

Also clarify whether the emulator experiments indicate that even with a corrected climate, JSBACH would still fail to produce a mid-Holocene maximum in boreal regions (which would strongly support a structural vegetation limitation). And explicitly position this result in relation to other transient Holocene simulations (e.g. PMIP-style experiments), even if only qualitatively.

3. Interpretation of CO₂ dominance and linearity

The manuscript convincingly shows that strong linear alignment of tree cover with CO₂ is associated with high temporal correlation but inflated variance and MAE, leading to the conclusion that CO₂ sensitivity is likely too strong in this model version. This is an important point, but it requires careful wording to avoid over-interpretation. The strong correlation between REVEALS tree cover and CO₂ is plausibly a proxy effect reflecting hemispheric deglacial trends rather than a physiological signal. This distinction should be emphasised more clearly in the Results and Discussion. It would be helpful to explicitly state that the emulator diagnoses model-internal sensitivities, not real-world sensitivities, and that the mismatch with REVEALS may arise from both sides. The conclusion that "CO₂ sensitivity is too strong" should be framed as relative to reconstructed variability patterns, not as an absolute statement about palaeo-CO₂ fertilisation.

4. Regional interpretation of disagreement: local dynamics versus model failure

The analysis of high- vs low-agreement grid cells is one of the strongest parts of the manuscript. The conclusion that poor agreement often coincides with non-linear, summer-temperature-dominated responses is compelling. However, I encourage the authors to make clearer that poor model–data agreement does not automatically imply "model failure", but may indicate regions where local ecological processes, migration lags, disturbance regimes, or microclimates dominate. This distinction is particularly important for Siberia, forest–tundra ecotones, and forest–steppe transition zones.

Minor comments and technical suggestions

P5, Line 149, "We consider only grid cells that indicate a significant (p<0.1) correlation between the simulated and reconstructed tree cover." The choice of p < 0.1 for correlation significance should be briefly justified, especially given the large number of grid cells.

P4, Line 116, it states that both model output and reconstructions are binned into 500-year intervals, but this temporal aggregation is not discussed further. Given that many key diagnostics (variance differences, non-linearity from GAMs, emulator performance, and correlation strength) are sensitive to temporal smoothing, the authors should briefly discuss how the 500-year binning may influence the detected strength of non-linear responses and the interpretation of model–data agreement, particularly in regions with rapid postglacial changes.

P12, 372-373, and P18, Line 553, the use of "energy-limited" vs "water-limited", would be helpful to briefly define these terms (e.g. dominant driver in emulator/GAM sense) to avoid confusion.

Fig. 5 and Fig. 6 are central but hard to understand; slightly stronger guidance in the captions on how to read them would help non-specialist readers.

Figure 11 is central for interpreting the effect of bias correction, but it is not immediately intuitive that the colours represent percentile-based changes relative to the original simulation rather than absolute agreement. I recommend clarifying this more explicitly in the caption (e.g. blue indicates relative improvement compared to MPI-ESM, not necessarily good agreement) and possibly adding a short explanatory sentence in the main text to guide readers.

Regional time-series examples (Fig. 12) are very effective and could be referenced more explicitly earlier in the text.

Summary: Dallmeyer et al. attempt to unravel the climate drivers of Northern Hemisphere tree cover dynamics over the last 20,000 years using the pollen-based REVEALS reconstruction, a transient MPI-ESM simulation, and a statistical emulator with bias-corrected climate forcing. I commend the authors on the addressing several complex scientific topics, from proxy-model agreement to climate drivers of vegetation dynamics to dynamic vegetation modeled processes, as well as using several independent linear and non-linear statistical measures. They have several interesting findings, including the ability of MPI-ESM to capture broad trends in tree cover and water- vs energy-limited conditions over the last deglaciation and Holocene, but substantial regional data-model mismatches that cannot be resolved with bias correction of climate information, and the dominance of summer temperature as a climate driver in simulated high-latitude forests. Overall, I feel that the scope of the paper may be too large and therefore is difficult to follow as a reader, but with some clarifications and improvements it will be appropriate for the Climate of the Past. I explain some minor and major comments in more detail below.

Minor Comments:

• Line 163: grid should be capitalized
• Line 233 and Figure 8: The panels are not labeled with letters but are referenced in the text (Fig. 8d)

Major Comments:

• This study has a high number of acromyns and shortened words (e.g., BiasCorr_Variance, F_CO2, r-group-1) that make it very difficult to know what exactly is being discussed. I recommend lengthening and clarifying the group names to make them more self-explanatory so that readers do not need to study the main text and/or figure captions before looking at each plot. Two examples of this: (1) Table 1: why not use the “Group” as “group_name” so that the meaning of the groups is more clear in the text? Naming a group “Positive Correlation Group” rather than “r_group 1” would be easier to understand. (2) Figure C1: why do you use 1, 2, and 3 as x-axis labels rather than writing out Positive, None, Negative Correlations? This adds an unnecessary layer of complexity to an already complex paper.
• A video animation of the forest cover through time would be incredibly powerful and helpful for following the description of results, a great example of this is Shafer et al. (2021)
• Lines 327-329: How did you determine this list of climate drivers to consider? I am wondering why other variables, such as incident solar insolation and mean annual temperature, were not included.
• Line 453-454: What climate variables/regions changed the most dramatically with bias correction? It would be very useful to see a couple time slice plots where the bias-corrected climate makes a large difference in simulated tree cover. For example, on lines 588-591, you mention that the model is too cold in the northern high-latitudes. This would be helpful to understand your interpretation of impacts of bias correction on simulated tree cover.
• Lines 582-583: I’m not entirely convinced that the tree cover response is too sensitive to CO2-level. This implies that tree cover is responding very strongly to the physiological impacts of CO2 concentrations, which should play some role, but I am not sure that you have included all of the potential climate predictors that capture the indirect, temperature-related effects of CO2. You have included two temperature-related predictors, Tw and Tc. Do these strongly correlate with CO2 level, or does mean annual temperature more closely vary with CO2 level? Can you clarify the implications of “the sensitivity of the tree cover response to CO2-changes seem to be too strong in this model version” on the way that photosynthesis/stomatal conductance is represented in MPI-ESM and JSBACH?
• Section 3.3: The emulator approach to disentangle the importance of different climate drivers is very interesting. In Figure 7, you visualize the most important climate driver in each grid cell, but I’m wondering how often there is a mixture of variables that have substantial importance (e.g., let’s say in N America, P and Tw are always the highest correlations so in many grid cells that are light blue or orange there is only a small difference their relative importance and which one is “most important”) and that information is missing from this figure. Would it be possible to add this information by modifying the shade of the colors? For example, if P is the most important predictor in one grid cell and far outcompetes other predictors, it is light blue, but if it is only slightly more important than other predictors, it is closer to white.

Works Cited

Shafer, S.L., Bartlein, P.J., Sommers, A.N., Otto-Bliesner, B.L., Lipscomb, W.H., Lofverstrom, M., Brady, E.C., Kluzek, E., Leguy, G., Thayer-Calder, K., and Tomas, R.A., 2021, Global biomes for the Last Interglacial period (127-119 ka) simulated by BIOME4