STEVE-1: A Generative Model for Text-to-Behavior in Minecraft

Constructing AI models that respond to text instructions is challenging, especially for sequential decision-making tasks. This work introduces a methodology, inspired by unCLIP, for instruction-tuning generative models of behavior without relying on a large dataset of instruction-labeled trajectories. Using this methodology, we create an instruction-tuned Video Pretraining (VPT) model called STEVE-1, which can follow short-horizon open-ended text and visual instructions in Minecraft. STEVE-1 is trained in two steps: adapting the pretrained VPT model to follow commands in MineCLIP's latent space, then training a prior to predict latent codes from text. This allows us to finetune VPT through self-supervised behavioral cloning and hindsight relabeling, reducing the need for costly human text annotations, and all for only $60 of compute. By leveraging pretrained models like VPT and MineCLIP and employing best practices from text-conditioned image generation, STEVE-1 sets a new bar for open-ended instruction-following in Minecraft with low-level controls (mouse and keyboard) and raw pixel inputs, far outperforming previous baselines and robustly completing 12 of 13 tasks in our early-game evaluation suite. We provide experimental evidence highlighting key factors for downstream performance, including pretraining, classifier-free guidance, and data scaling. All resources, including our model weights, training scripts, and evaluation tools are made available for further research.

Paper

References (70)

Scroll for more · 38 remaining

Similar papers

Peer review

Reviewer 7QfK5/10 · confidence 5/52023-06-27

Summary

This paper proposes a generative pretraining method to learn an instruction-following agent in Minecraft by fine-tuning the VPT agent. It first trains an image-goal-conditioned policy and then leverages the foundation model MineCLIP as the bridge to map the language instruction into the image-goal space. Using this method, this paper claims that it can solve any short-horizon open-ended text and visual task in Minecraft.

Strengths

(1) This paper explores the feasibility of jointly using the foundation models (such as the vision-language model MineCLIP and pre-trained policy model VPT) for solving decision-making problems and shows a preliminary conclusion. (2) Using unCLIP approach to solve instruction-following decision-making problems is interesting. (3) This paper proposes to improve the instruction sensitivity with the classifier-free guidance technique, which is interesting and reasonable. (4) The paper is well-written and easy to follow.

Weaknesses

(1) The paper overstates its performance by claiming that "STEVE-1 can follow nearly **any short-horizon** open-ended text and visual task in Minecraft." To verify this, I downloaded the provided code and ran the attached checkpoints on some basic short-horizon tasks like "kill sheep", "kill cow", and "kill pig". Before testing these tasks, I had already summoned 10 "pigs", "cows", and "sheep" nearby. While I found that the agent actually killed some animals, it did not seem aware of which particular animal it was targeting. The behavior was more like "kill cow if there is a cow" rather than following the specific instruction. I tried several other prompt variants such as "hunt cow" and "hunt cow in Minecraft", but they were unsuccessful. Additionally, I tested STEVE-1 on some short-horizon crafting tasks with sufficient materials in the initial inventory, like "craft stick", "craft oak plank", and "craft torch". The agent merely opened its inventory and then acted randomly. Given these observations, I do not believe that STEVE-1 has solved all short-horizon tasks, as it cannot even differentiate between "cows" and "sheep". This could lead others to misconstrue the extent of the research in the Minecraft environment. (2) This paper lacks generalization experiments on unseen text instructions. As the primary aim of this method is to support open-ended text instructions in STEVE-1, it is crucial to demonstrate its performance on unseen text instructions. The authors collected approximately 10,000 instruction-trajectory pairs, with 2,000 of them hand-labeled and an additional 8,000 instructions generated by GPT. It remains unclear whether the tasks ("dig dirt", "build a tower", "make wooden planks") in the experiments were already included in these instructions. If so, the experiments may not be particularly convincing, as neural networks can easily map the text instructions to corresponding visual embeddings in the training set. That is, pre-training STEVE-1 on visual instructions already achieves this goal and the later generative model training may be not important, which is not impressive. So, I strongly recommend including generalization experiments if the paper wants to show support for open-ended text instructions. (3) The evaluation metric is insufficient. The MineCLIP evaluation metric is not convincingly effective in measuring whether a trajectory corresponds to a given task. This is because the MineCLIP latent space cannot distinguish some trajectories well. For instance, we have found that it was unable to differentiate between two trajectories where one involved approaching a tree and the other involved moving away from it (MineCLIP was not sensitive to distance). To address this issue, I recommend that the paper incorporates a success-rate-based confusion matrix, in which the values in Figure 3 (b) are replaced with the task success rate, conditioned on the given instruction.

Questions

As stated in the weakness part.

Rating

5: Borderline accept: Technically solid paper where reasons to accept outweigh reasons to reject, e.g., limited evaluation. Please use sparingly.

Confidence

5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.

Soundness

3 good

Presentation

3 good

Contribution

2 fair

Limitations

The main limitations include (1) prompt design being too sensitive and (2) chain-of-thought prompting is not automated. The authors have adequately stated the limitations.

Reviewer vziy6/10 · confidence 4/52023-07-07

Summary

This paper proposed an instruction-tuned model for Minecraft, which turns previous RL model like VPT into goal-conditioned model. The experiment shows that the proposed method can follow nearly any short-horizon open-ended text and visual task in Minecraft.

Strengths

1. This paper proposes a novel technique for instruction-tuning video game RL-agent which is inspired by previous unCLIP method in the text-to-image generation field. This idea is interesting and the method can turn an RL agent into a goal-conditioned one which shows great potential for designing a more general RL agent. 2. The experiment part shows that the tuned agent can follow short-horizon goal instructions very well. When given the appropriate textual instruction, the proposed method collects 66x more dirt, 4.5x more wood, 28x more seeds, and travels 3x further than the unconditional agent.

Weaknesses

This method is not applicable for long-horizon tasks in the open world. For example, in Minecraft, if the goal is obtaining a diamond, the instruction may be too difficult for the proposed method since it is not trained for long-horizon trajectories. Regarding this, it is not sure if this method can be applied to challenging tasks in the open world which may contain extremely long planning steps. In this paper, no complicated tasks like obtaining diamond are studied. Thus I am concerned the scalability of the proposed method to much complicated tasks.

Questions

I am curious about how is the prompt designed in this paper? The detail of the prompt should be explained in the main text of the paper.

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

The limitation is adequately addressed in the paper. No discussion about negative social impact.

Reviewer Q7hP7/10 · confidence 3/52023-07-07

Summary

The paper introduces STEVE-1, a sequential decision-making agent designed to follow textual instructions and accomplish goals in the Minecraft environment. The authors utilize two pre-trained models, VPT (Video Pretraining Transformer) and MineCLIP, to facilitate this process. VPT is a transformer model trained to predict action sequences from aligned video sequences, while MineCLIP aligns consecutive video timesteps with corresponding transcripts in Minecraft. To finetune VPT, the authors employ self-supervised behavioral cloning conditioned on latent visual goals. They generate goal-conditioned data by randomly selecting timesteps from episodes and using hindsight relabeling to set intermediate goals. MineCLIP is used to map textual goals to visual goal embeddings in the unCLIP-based approach. This mapping is achieved through a conditional Variational Autoencoder (VAE) with Gaussian prior and posterior, conditioned on MineCLIP text representations. The training dataset comprises Minecraft gameplay data, combined with the OpenAI contractor dataset, and additional data generated using VPT. The authors curate 2,000 instruction-labeled trajectory segments, each consisting of 16 frames, and augment this dataset by identifying similar gameplay segments. Additionally, 8,000 additional instructions are generated using GPT-3.5 Turbo. Classifier-free guidance is employed to balance the logits between unconditional and conditional behavior. The results demonstrate that STEVE-1 successfully solves various short-horizon open-ended text and visual tasks, with a training cost of only $60. The performance of the agent is evaluated using programmatic evaluation and MineCLIP evaluation, revealing significant improvements with goal conditioning, especially visual. The authors also observe that prompt chaining, as opposed to direct prompting, proves advantageous for complex tasks like building towers or creating planks. A thorough ablation study is conducted on the classifier-free guidance hyperparameter, pretrained VPT weights, and prompt engineering, providing valuable insights into the optimal settings for these components. Overall, the paper showcases the effectiveness of STEVE-1 in achieving goals based on textual instructions in the Minecraft environment. The combination of VPT and MineCLIP, along with the conditioning techniques and ablation studies, contribute to a comprehensive understanding of the agent's performance and its potential applications.

Strengths

- This unique adaptation - STEVE-1 - of the unCLIP method demonstrates the versatility and effectiveness of the approach in the context of sequential decision making in Minecraft. - The experiments and analysis conducted in the paper are highly novel and insightful. The authors provide valuable findings, such as the benefits of prompt chaining, the potential for scaling to improve certain metrics, and the limitations observed in complex tasks. These insights enhance our understanding of the proposed approach and its implications. - The paper demonstrates that prompt chaining is effective in accomplishing complex tasks, such as building towers or making wooden planks. The results show that success rates and programmatic evaluation metrics plateau after a certain number of frames, highlighting the potential and limitations of the approach. Additionally, the comparison to direct prompting reveals the superiority of prompt chaining in achieving satisfactory performance. - The thorough ablation study conducted on various components, including the classifier-free guidance hyperparameter, pretrained VPT weights, and prompt engineering, provides valuable insights into the optimal settings for these elements. - The inclusion of additional ablations, such as goal chunk sizes, VAE variants, and text augmentation, in the appendix further enhances reproducibility and facilitates a deeper understanding of the approach.

Weaknesses

- The paper would benefit from clearer explanations regarding certain aspects, such as the distinction between packed hindsight relabeling and hindsight relabeling. Additionally, providing a more detailed explanation of the interpretation of the heatmap, specifically the meaning of the 1/0 values, would enhance understanding. - Some questions (see section) during the review could be addressed in the main work or in an appendix to provide a more comprehensive understanding of the approach. - The proposed approach combines existing ideas from well-established approaches, such as CVAE, MineCLIP, and VPT, which may limit the novelty of the work. However, the extensive experiments and detailed study conducted in the paper contribute to its significance and overall value. - The scalability of the approach to other environments or datasets is not discussed or addressed in the paper. Considering the dependency on pretrained video-text aligned models and large-scale pretrained transformers like VPT, the applicability of the approach to environments without such models is unclear. Addressing this limitation and discussing potential scalability issues would enhance the practical relevance and broader applicability of the approach.

Questions

- Can you provide insights into why random timesteps from episodes are selected as goals? Are there alternative approaches that could be considered, and why is this method deemed the most effective? - Does randomly resetting the agent's memory and turning the agent to a random direction during data generation provide necessary benefits? What would be the implications if this step were omitted? - How does the utilization of pre-trained VPT contribute to the approach when the input is modified with conditional goal embeddings? Can you elaborate on the specific advantages and improvements gained from incorporating VPT in this manner? - Could you explain the process of generating additional text instructions based on GPT 3.5? How are these instructions generated and integrated into the training process? - It is not clear why the graph from Baker et al. is not included in the text-conditioning results. Can you provide clarification or discuss the reasoning behind its exclusion? - In Figure 4 (right), what is "relevant" and "irrelevant" in the graph? Clarifying the interpretation of these components would improve the understanding of the presented results.

Rating

7: Accept: Technically solid paper, with high impact on at least one sub-area, or moderate-to-high impact on more than one areas, with good-to-excellent evaluation, resources, reproducibility, and no unaddressed ethical considerations.

Confidence

3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

4 excellent

Presentation

4 excellent

Contribution

4 excellent

Limitations

- The authors extensively discuss the limitations of the approach, including challenges with multiple steps of reasoning, prompt engineering, and potential negative societal impacts. While further discussion on these aspects could be beneficial, there are no explicit omissions or significant limitations that require specific mention. The paper adequately addresses and acknowledges its limitations.

Reviewer 3aRk8/10 · confidence 4/52023-07-08

Summary

Paper presents a method to create instruction following agents in Minecraft. It starts with collecting trajectories using OpenAI’s VPT Minecraft agents, which cannot be controlled through instructions. Then some intermediate visual observations are randomly selected as visual goals. These visual goals are then encoded by the visual encoder of a pretrained text-video contrastive foundation model in the Minecraft domain called MineCLIP. VPT is then fine-tuned to achieve these visual goals (specified as MineCLIP visual embeddings) using the same trajectories. Finally, a CVAE is trained to produce textual embedding that matches the visual goal embeddings, therefore the agents can be piloted by natural language instructions. Experiments highlight the performances of presented method on some entry-level tasks including dirt, log and seed collection. Additional results on prompt chaining, scaling, prompt engineering and other hyperparameters confirm the effectiveness of the proposed approach.

Strengths

+The paper is overall clear and well-written. The presentation should be very friendly to readers without RL or LfD background by foundation model in general. The research problem here is relevant to the scope of NeurIPS and its proximity to many emerging topics including open-endedness, generalist agents and large models should be of interest to large audience of this conference. +I find the approach presented here is technically sound with good results. Although the original VPT agent did offer some preliminary results on instruction following, but no enough details were provided so the method here seems to make solid contributions on building the first open-ended language-piloted agents in Minecraft. Some tricks like baseline subtraction seems helpful. Thanks for bringing this to the attention of the community. +The authors did a good job highlighting go/no-go about their method. Just list a few that I find most interesting: - MineCLIP evaluation, which help explain why their method works better on short-horizon task - Prompt-chaining and metrics over time, which clearly demonstrate how the agent progresses under different prompt conditions. - Prompt-engineering, which showcases how different instruction can have significant impact at the agent.

Weaknesses

At this point, I don’t have any major concerns but here are a few suggestions: -Baselines; in the main paper, I couldn’t find any baselines or ablations other than the main ”Steve-1” model and VPT(without instruction-following). I agree that there might not be many proper counterparts available but some single-task RL and imitation learning baseline, ex. [10] could still help with better sense on the actual gain the proposed model has in Minecraft. Moreover, the baseline subtraction trick should also be ablated in the main paper with more evidences on other tasks, it seems to be a working technique, but it must be better validated. -Error bar; In Figure 6, I am not sure if the numbers in parentheses indicate error range. If they are not, error bar is needed for these results. -Scaling: please elaborate more on “we see evidence of tasks that do not require much data for STEVE-1 to learn, tasks that steadily get more reliable as the agent is trained longer, and tasks where capability suddenly spikes after the agent reaches some threshold.” If not all tasks can benefit from scaling, why “Put together, this suggests that further scaling would likely significantly improve the agent” -After reading the paper, it’s still unclear to me how the $60 budget is allocated. Please provide detailed explanation to justify this point.

Questions

See [Weaknesses]

Rating

8: Strong Accept: Technically strong paper, with novel ideas, excellent impact on at least one area, or high-to-excellent impact on multiple areas, with excellent evaluation, resources, and reproducibility, and no unaddressed ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

4 excellent

Presentation

4 excellent

Contribution

4 excellent

Limitations

One thing that could be missing in the limitation statement is how well does the cVAE model perform on producing goal embedding and whether it could be a bottleneck to the overall performances on more challenging tasks like long-horizon planning, etc. The authors seem to blame MineCLIP for this but since in the original unCLIP paper, diffusion is used instead of a cVAE, more comparisons might be necessary.

Reviewer 7QfK2023-08-11

Thanks for the author's rebuttal. I appreciate the author's honesty and additional experiments to solve my concern. I'm going to raise the final rating. I strongly recommend the final revision of this paper should include a more detailed limitation part (illustrate what steve-1 is not capable of and why).

Reviewer 3aRk2023-08-12

Thank you for the reply. Some of my concerns have been addressed. But after reading other reviews as well, some additional questions pop up. Here are some follow-ups: -It would be much better off if you can pull some results from [10] and compare them with yours in the current submission. It seems to be a very recent and close counterpart (goal-conditioned control in Minecraft), and these numbers will help readers understand how open-world control can be improved with your techniques, especially with modern neural network architectures. Some additional discussions might be needed as well. -It's interesting to know you think of long-term tasks with steve-1, we may have to use visual instruction instead. Why is this the case? I suppose specifying a long-term task with text should be simpler, no? Also, what do you think is the bottleneck of steve-1 solving long-term tasks? I understand you've made it clear on the capacity of steve-1, which is mostly about short-term but I'm sure some analysis & prospect about solving long-term goals (this could be relevant to the prompt-chaining experiments in the main paper as well) in the limitation section should be of interest to reviewers and potential future readers. -I've read other reviews and I agree with Reviewer 7QfK that saying "STEVE-1 can follow nearly any short-horizon open-ended text and visual task in Minecraft" is improper given the current state of evaluation. I don't even think "a wide range of tasks" seems to be appropriate either. My suggestion is to make it clear what kind of tasks steve-1 is able to robustly accomplish by adding a table to the main paper, even better, putting it here in the comment section for the reviewer to examine as your code&model is also provided, and change your statement to something like "steve-1 is able to robustly complete <number of tasks> of tasks in Minecraft". Please avoid using vague and non-academic terms.

Authorsrebuttal2023-08-15

Response to Reviewer 3aRk [1/2]

> It would be much better off if you can pull some results from [10] and compare them with yours in the current submission. It seems to be a very recent and close counterpart (goal-conditioned control in Minecraft), and these numbers will help readers understand how open-world control can be improved with your techniques, especially with modern neural network architectures. Some additional discussions might be needed as well. > After reading through [10] carefully, we agree that it is an important related work and we thank you for bringing it to our attention. Both of our works focus on goal-conditioned control in Minecraft, with the major difference being that [10] trains on a fixed set of goals and STEVE-1 uses unCLIP, hindsight relabeling, and MineCLIP to learn goal-reaching behavior from a large dataset in a self-supervised way. We believe that many of the techniques in our work improve open-world control with modern neural network architectures, including our efficient packed hindsight relabeling implementation, classifier-free guidance, effectively using the knowledge in pretrained models with unCLIP, and our overall scalable recipe that learns to reach goals in a self-supervised way. We will be sure to include a discussion of this and the details of [10] (including its different goal-conditioning architecture) in the final version of the paper. Regarding comparing numbers, the results in [10] include the success rate for single-task training for “chopping trees” (50%), “combat cow” (58%), and “combat sheep” (60%). Since the task (both the world and the reward function), action space, observation space, and training data are very different, comparing to our own success rate numbers in Figure A of the rebuttal pdf is likely to mislead. Regardless, for most of the above mentioned techniques, we include ablation studies that show the improvements they can bring to open-world control. > It's interesting to know you think of long-term tasks with steve-1, we may have to use visual instruction instead. Why is this the case? I suppose specifying a long-term task with text should be simpler, no? > We are afraid there may have been a misunderstanding since we do not believe that there is a difference between visual and text instructions for long-term planning. We agree that specifying a long-term task with text should be simpler. > What do you think is the bottleneck of steve-1 solving long-term tasks? I understand you've made it clear on the capacity of steve-1, which is mostly about short-term but I'm sure some analysis & prospect about solving long-term goals (this could be relevant to the prompt-chaining experiments in the main paper as well) in the limitation section should be of interest to reviewers and potential future readers. > We think one bottleneck of solving long-term tasks is that during our packed hindsight relabeling, we limit the hindsight goals to at most 200 timesteps in the future (10 seconds). Due to this, tasks that require more than 200 timesteps to complete are technically out-of-distribution for STEVE-1. We experimented with using longer goal lengths (denoted ‘chunk size’) in Figure 10 in the appendix and found that the performance on all of our evaluation tasks tends to decrease if we increase this hyperparameter too much. We suspect that while an increased chunk size may be able to improve long-horizon performance, it also increases noise and comes at the cost of reducing performance on short-horizon goals. It is an important avenue for future work to investigate whether it is possible to achieve a better tradeoff. Improved performance on long-horizon tasks is one of the most important follow-up directions to improve upon STEVE-1. We point you to our response to reviewer vziy, where we enumerate a few other approaches that may improve long-horizon performance in future work: **1) Scaling:** Our scaling results indicate that as we train the agent on more data, more tasks become achievable. So scaling up the amount of data could enable the agent to complete longer-horizon tasks. **2) Finetuning with RL:** We also think that finetuning the STEVE-1 pretrained agent with RL is another very interesting avenue. For example, VPT finetuned their pretrained agent with RL to learn to mine diamonds. So we think it’s likely that RL finetuning on top of the pretrained STEVE-1 could enhance long-horizon task completion abilities. **3) Using LLMs or VLMs:** Using LLMs or Visual-Language Models (VLMs) to automatically provide prompt chains to the STEVE-1 agent could be an effective way to improve long-horizon performance. We started to investigate this possibility with our prompt chaining experiments and we think that this is a very exciting direction for follow-up work.

Authorsrebuttal2023-08-15

Response to Reviewer 3aRk [2/2]

> I've read other reviews and I agree with Reviewer 7QfK that saying "STEVE-1 can follow nearly any short-horizon open-ended text and visual task in Minecraft" is improper given the current state of evaluation. I don't even think "a wide range of tasks" seems to be appropriate either. My suggestion is to make it clear what kind of tasks steve-1 is able to robustly accomplish by adding a table to the main paper, even better, putting it here in the comment section for the reviewer to examine as your code&model is also provided, and change your statement to something like "steve-1 is able to robustly complete <number of tasks> of tasks in Minecraft". Please avoid using vague and non-academic terms. > Thanks for raising this and we definitely agree that, especially in the context of providing a precise baseline to future works that compare with STEVE-1, avoiding vague statements is important. We will change the statement to “Our testing shows that STEVE-1 is able to robustly complete 12 of 13 goal-conditioned control tasks in our early-game evaluation suite.” (based on Figure A in the rebuttal pdf). It is important to recognize that these 13 tasks included in our evaluation are not strictly all the agent can do and that evaluating and describing the capabilities of open-ended models is an open research problem itself since capability depends strongly on preconditions, prompt engineering, and our own ability to come up with varied and challenging tasks (see the many recent works on the evaluation of LLMs for examples). We will add this discussion to our limitations section, change the language describing STEVE-1’s capability as described above, and include the following table where we summarize what, in our experience, STEVE-1 can do and how to reproduce it. Work-in-progress version of the table: | Task | Prompts | Notes | | --- | --- | --- | | Early-Game Evaluation Suite: dig a hole, get dirt, look at the sky, break leaves, get wood, get seeds, break a flower, go explore, go swimming, go underwater, open inventory | Use the best-performing prompts in Table 3 in the Appendix. | Precondition: For grass and flowers, it works best when grass and flowers are in the current biome. We found breaking flowers to be less reliable than the others. | | make a tower | “build a tower” | Precondition: building blocks in hotbar. To obtain building blocks, you can use the dig a hole or get wood prompts from Table 3. | | craft wooden planks | “make wooden planks, craft wooden planks” (*) | Precondition: wooden logs. To obtain wooden logs, use the get wood prompt from Table 3. | | place torches, place a crafting table, place wooden planks | “place [torches/a crafting table/wooden planks]” | Precondition: [torches/crafting table/wooden planks] in the hotbar | | make a crafting table | “make a crafting table” | Precondition: wooden planks. | | break stone | “mine stone, go mining, get stone” (*) | | | get cobblestone | “mine stone, go mining, get cobblestone” (*) | Precondition: pickaxe in hotbar. | | create a wooden pickaxe | “craft a wooden pickaxe, make a wooden pickaxe” (*) | Precondition: already looking at a placed crafting table, has necessary materials. | | hit a sheep, hit a pig, hit a cow | “kill a [sheep/pig/cow]” | Precondition: agent is close to and looking at the sheep/pig/cow. (See our response to Reviewer 7QfK.) | (*) note that this is a single prompt-engineered prompt

Reviewer 3aRk2023-08-16

Thank you for the very detailed reply. The comparison with [10], additional analysis on long-term tasks, and revised statement on the capacity of steve-1 all look good to me. Please ensure these contents will be included in the final version. I am happy to raise my score to 8.

Reviewer vziy2023-08-21

I will keep my rating.

Thanks to the authors' detailed feedback. After reading the rebuttal, I still have concerns about long-horizon ability but I do agree with some possible solutions proposed by the authors.

Program Chairsdecision2023-09-21

Decision

Accept (spotlight)

© 2026 NYSGPT2525 LLC