Authors present a new global dataset of 3691 pollen records that are transformed into biomes using the well-established method of Prentice et al (1996). The dataset covers the whole globe with some gaps in Africa, southern Asia and Australia, where the data are not available. They work at the level of 8 mega biomes which represent raw pattern of global vegetation. Indeed, it is difficult to work with finer biomes at a global scale (even if Prentice at al 2000 did for two time slices). The biomes reconstructed from 21 ka BP to present at 500 years’ time steps are compared with equivalent biomes simulated by three models (and two simulations each).

The main problem of the paper can be easily detected in the abstract (which reflect correctly the full paper). The paper is technical and too descriptive. Biomes are reconstructed, simulated with ESM and compared, some biases of the ESM simulations are pointed out and that is all. There are no general messages on the evolution of the vegetation through the 21 millennia. The ESM biases are not discussed by analysing which climatic variables are responsible of them. Why mediterranean biomes of the Holocene are simulated as steppes whereas data indicate TEDE? There are also biases in the reconstructed biomes. As an example, the glacial mediterranean biomes are sometimes reconstructed as TEDO while previous reconstructions (Elenga et al, 2000; Prentice et al, 2000) reconstructed STEP. The paper needs also to discuss why this megabiome approach is not always coherent with the standard biome approach of Prentice et al 2000. I have the feeling that for the regions I know better, this paper does a poorer reconstruction than previous attempts. If not, this should be argued.

In conclusion, the paper has potentialities, but it needs more work to be a good contribution to the discipline.

Specific comments

• Section 2.3: how are interpolated the biomes in sites from the grid of the ESM (the dots of Fig 1B)?
• 230-233: I disagree that there are no systematic mismatch: there are in the Med region, in the subarctic one … Note also that TUND and STEP have low agreement (50%)
• 233-235: I think that the good agreement comes to the fact that the comparison is restricted to 8 megabiomes. In the previous papers, finer biomes are considered and often the mismatch is between climatically neighbor biomes (this is less possible with megabiomes).
• 259-263: I do not understand this sentence.
• Section 3.1.2: It is necessary to try to explain the discrepancies between simulated biomes and potential vegetation by over or under-simulation of some climate variables.
• Figure 3: how is obtained the ice-sheet extension data from pollen?
• 320-331: Elenga et al 2000 and Prentice et al 2000 reconstructed steppes at 21 ka BP, which seems more realistic than tundra. Prentice et al 2000 is not cited despite the fact that they produced a full global reconstruction for the 21 ka BP period (and also mid-Holocene). It is a major paper that the authors cannot ignore.
• 325-332: the problem of assigning TUND to STEP and vice versa should also be discussed; previous papers reconstruct mainly STEP to the Med region and not TUND as here.
• 333-334: TUND seems more extended in Europe at 16-13ka than at 21ka, while the warming has already started, why?
• 342-353: Biomization starts to be realistic in the Holocene, much more than for the cold periods. But It is strange that there is no Mediterranean vegetation (WTFO) in the Med area during all the Holocene, as well for simulation as reconstruction. For simulation WTFO is replaced by STEP and for reconstruction, it is by TEFO
• 356-361: I read on the maps the opposite to what is claimed: I see a forest degradation in the simulations (maps at right) not in the reconstructions (maps at left). In the whole Holocene, reconstructions show a constant TEFO, while simulation shows steppes in the second part of the Holocene, tending to show that there is a bias towards aridity in the models. Pollen reconstructions may sometimes be influenced by human deforestation, but it has been shown in previous papers that biomization is robust as regards as this disturbance.
• Figure 4: It does not exist the possibility to computed significance levels for EDM?
• Caption of Fig.4: The sentence “The largest datamodel deviations occur during the LGM and early deglaciation periods” should not be put in the caption.
• 385-389: “the best data-model agreement occurs during the Bølling-Allerød interstadial (represented by the timeslice 14 cal. ka BP)”: This appears to be true with the global EMD but in the regional EMD, it does not appear that 14ka EDM was minimum in any regions.
• Section 3.3: it should be useful to summarize by a table the biases found in simulations and giving an interpretation of which climate variable is responsible of the biases
• Section 4 (conclusions): this conclusion is short and superficial. What is the origin of the ESM biases according to reconstruction at least for the main ones? Has this paper filled the initial objectives?

General comments

-The paper of Li et al. uses a global set of pollen data to reconstruct megabiomes since the last ice age.  This reconstruction is then compared with biomized outputs from Earth System Models in order to evaluate the fidelity of the model simulations.  The biome reconstructions are quite interesting, and I appreciate the large fossil datasets compiled by the authors, this is huge effort to put together.  However, I have some comments about the methods as well as about the scientific contribution this makes through the interpretation of the data-model comparison.  In particular, I find the conclusions to be very general and technical, and I hope that the authors are able to make the impacts of their study clearer.

Specific Comments

-In my comment above, I say that I find the interpretation of the data-model comparison to be overly general.  This is already observed in the abstract.  The abstract ends with the statement: To some extent, these mismatches could be attributed to systematic model biases in the simulated climate, as well as to the different plant representations and low taxonomic resolution of pollen in the reconstructions.

I find this to be so general that it makes it really hard for the reader to glean any nuance the help us understand specific insights about model bias or data issues.  This continues in the discussion with some very general statements about the sources of uncertainty in the comparison such as: We assume that the simulations used in this study share this rather common problem of a cold bias in boreal latitudes, resulting in the overestimation of tundra in simulations.  There is no reference here to support the assumption and no sign that this potential bias was evaluated for the simulations used in the study.  I suggest that the authors place more emphasis on the actual biome reconstruction from the pollen (which takes up more of the discussion, but is not much emphasized in the abstract) and take look deeper into the sources of mismatch to leave the reader with some key takeaways that relate directly to their stated goal (goal in abstract: to evaluate the paleosimulations from ESMs).

-Another point I would add here is that the authors acknowledge the limitation of their modern validation exercise in incorporating human land use impacts to ecosystems when comparing with models that don’t include such impacts.  Please expand on how this might also impact data-model comparisons of paleo-simulations.  Also what about the impact of fire, I assume this is not included in ESMs?  Could the lack of these processes in the models result in mismatch?

Very minor comments

-Abstract: line 31 I don’t understand term: global spatial megabiome.  Does this mean the megabiomes of any particular time slice?  I think there should be a way to simplify this.

-line 79: “8 of our own new records” there are no references for these.

-line 71: 3691 pollen records are in this compilation, but how many are included in the analysis after data filtration and quality control?  This question applies to the numbers of records in Table 1 as well.

-line 82-83: following the previous question, I know this paper which provides recommendations for data best practices, but it doesn’t specify specific practices for any study.  Could you please tell us specifically how you filtered data?  How many dates were required for age models, how were age models generated, how many pollen samples or counts were requires, etc?

-Line 90: great that dois for specific datasets were included!  This is great, helps attribute credit to individual record generators!

-Line 110-111 “Larix and Pinus were multiplied by factors of 15 and 0.5”  This is probably sensible for NA and Europe, but what about other overrepresented taxa in other regions?

-Line 121: “Furthermore, the assignment of pollen taxa to megabiomes and biomization routines were performed independently for each continent.”  Is there specific information on the differences for each continent in that supplementary material or somewhere else?  This is not clear to me. For example, different harmonization schemes have been published for different geographic areas, but I don’t see a reference to this or other geographically specific procedures.

-Line 159: regarding the tool of Dallmeyer et al., 2021, please provide a few details about how this works.