MindForge: Empowering Embodied Agents with Theory of Mind for Lifelong Cultural Learning

Thank you for the review rsUz, we appreciate your comments and suggestions!

1. The current design requires a weaker agent to fail a task before initiating dialogue.

To clarify, there is no limiting factor in the design of the MindForge framework itself with respect to the communication protocol as long as communication is done in natural language. Note that we only adopt this approach (where failure is required before initiating dialogue and then performing communication and task actions in an interleaved manner) for our experiments, to be able to compare the effect of communication on the base behavior for task completion.

Flexbile communication protocol

To showcase that our MindForge framework is not tied to a specific communication protocol, we perform additional experiments on two MindForge tasks where we allow for a more flexible communication setting. Specifically, we consider the following setup: in an instructive setting, we allow at each turn for the weak agent to decide if it wants to start communication ( in the case it is unsure how to solve the task to begin with ) or try the action itself.

If the agents decides to try the action and does not feel the need to communicate, in case of failure to complete the task, we follow the standard protocol considered in the manuscript. Below you can see the results of this experiment together with some useful statistics.

Mixtral-8x7B

Mistral-7B

Observations in a flexible communication scenario:

• Mixtral-8x7B correctly assess that it needs help to mine dirt without needing to fail first, which results in an increase in task success
• Mistral-7B is extremely confident in its ability to solve the task without help and never asks for help before failure even if this option is given.
• When dealing with a task resembling something they solved in the past, both models correctly identify they do not require additional help as they should already possess the required knowledge in semantic and procedural memory.

2. Since the paper claims it is a general framework for diverse settings...mechanisms transfer effectively. Would the framework generalize beyond Minecraft?

To begin with a clarification, in terms of measuring generalization, different Minecraft biomes already serve as diverse environments for testing, making the tech-tree experiments a significant reflection of continual learning capabilities, similar to Voyager’s experiments (Wang et al., 2023a).

That said, our proposed innovations (perspective taking, natural communication) and the MindForge architecture itself are game agnostic. Generalization beyond Minecraft would only change the base prompt for the individual modules to adapt to the specific environment (or the real world). However, incorporating different games for experimentation are non-trivial on academic budgets; the experiments in Minecraft alone in the current paper cost around 600 GPU hours.

3. its core research contributions, particularly how it differentiates from other collaboration frameworks, are not clearly presented.

To clarify, the manuscript tackles the following research questions:

• (i) can we bridge the capabilities between closed-source and open-weight models in embodied settings through a natural language interface that is also compatible with humans

• (ii) similar to humans, can we use perspective-taking and theory of mind to enhance the communication between multiple agents

• (iii) what type of memory subsystems are beneficial for embodied agents in a lifelong learning setting

To the best of our knowledge, we are the first to enable agents to take perspective-taking using a structured representation. Moreover, we consider a lifelong cultural learning setting where agents communicate and evolve across a curriculum of tasks inside Minecraft.

Contributions with respect to the Research Questions

• (i) we show in Figure 1 b) how our MindForge agents an in instructive setting achieve new milestones compared to the Voyager baseline in a lifelong setting inside Minecraft, bringing the performance closer to closed source models like GPT-4. Besides the immediate performance improvements, our MindForge framework offers a potential solution to tackle false beliefs through communication in multi-agent settings.

• (ii) We show through an ablation in Table 5 Appendix B that using perspective taking and the BigTOM representation results in more effective communication between agents. The text snippet in Appendix B shows an example how perspective taking incentives agents to think from the perspective of the other agents, tailoring their advice to their needs.

• (iii) we add 2 more memory components compared to Voyager: episodic and semantic memory. We consider semantic memory an important component in MindForge as it represents a store for correct beliefs about Minecraft tasks, gathered through multi-agent interactions. This component functions as a way to override the inherent false beliefs that stem from LLM pretraining. Moreover, **we show how episodic memory enables embodied agents to remember their mistakes and avoid making them again. **

MindForge functions more as an engineering framework than a typical research paper.

The majority of the scientific work in this foundational model powered embodied agents can be considered engineering frameworks [1] [2].

Moreover, Reviewer ADK5 compares our belief and memory systems to a number of single-agent papers (ExpeL, CLIN, SSO, DEPS, ADAPT), highlighting that most of the papers in this space are engineered around foundational models. Please refer to our response for Reviewer ADK5 for a more in-depth breakdown.

However, the goal of MindForge is not to propose an engineering framework to solve Minecraft, but rather to provide a solution towards human-AI and multi-agent collaboration in cultural learning settings purely through natural language and in-context learning. Minecraft simply provides us with the environmental substrate required to assess the efficacy while the core of MindForge remains generally applicable

4. Can the compute-performance tradeoff be more precisely quantified?

Focusing on the compute-performance trade-off would erroneously consider MindForge as an engineering goal (solving the task efficiently) rather than the manuscript's stated scientific goal: to investigate the process of in-situ learning and skill acquisition in an agent that is otherwise incapable, using social collaboration as the mechanism.

Nonetheless, the effect of compute-performance tradeoff is best seen in the context of the Minecraft tech tree experiments as shown in Fig. 1 b) and Table 3.

MindForge agents manage to achieve more milestones than Voyager open-weight alternatives in the same number of actions. Moreover it reaches certain milestones faster ( as counted by the number of actions) than the Voyager counterpart.

The efficiency of MindForge agents can be quantified in two separate ways:

• (1) number of effective tokens generated by the model across communication, action, belief creation and perspective taking

• (2) number of actions in the Minecraft environment taken by the agent.

We consider the latter as a more appropriate way of measuring the compute-performance trade off since it is agnostic to the underlying model, as opposed to counting the effective number of tokens which is heavily dependent on the prompting, type of model ( base, instruct, reasoning model ) and model provider training recipe.

Please refer to the last point mentioned in the rebuttal for Reviewer qoax.

References:

[1] Zhang, Hongxin, et al. "COMBO: compositional world models for embodied multi-agent cooperation." arXiv preprint arXiv:2404.10775 (2024). [2] Sumers, Theodore, et al. "Cognitive architectures for language agents." Transactions on Machine Learning Research (2023).

Thank you for the review and overall positive assessment of our paper, rCMn! We’re grateful that you found the manuscript effectively organized and easy to follow, clear exposition was and is a primary goal, so your comment affirms that effort. We also appreciate your endorsement of the ablation suite: by isolating the Theory-of-Mind graph and the memory, we aimed to give readers a transparent view of where the gains truly arise. Your acknowledgment that these studies are both detailed and well-structured reinforces the value of that analytical depth.

1. Figure 6 is unclear.

We will fix this for the camera-ready version. We will add a legend of the starting population success for each colored line, currently this is denoted on the curve itself in the figure.

2. Does the strong agent actively interact with the environment?

Yes, the strong agent is actively interacting with the environment while helping the weaker agent. The complexity and open-endedness of environments like Minecraft require agents to get continuous environment feedback as a way to revalidate or correct their pre-training induced beliefs. Even strong models like GPT-4 struggle to complete tasks in a zero-shot manner, often requiring additional environment signals, as shown in Voyager [60]. Besides stronger grounding in the environment, the lack of environment interaction would hinder the ability of the other agents to take perspective, potentially making communication less effective, as shown in Appendix B Table 5.

Moreover, requiring both agents to interact with the environment offers interesting perspectives on how they co-evolve (e.g do their belief structure converge over time?). While we consider this out-of-scope for this submission, we believe it is interesting to experiment how two or more agents influence each other’s curriculum, which becomes even more meaningful in a purely collaborative setting.

Expert agent as an oracle

It is important to note that, while the weak agent is indeed asking for help from the strong agent, the strong agent is not an oracle. While rare in our experiments, the weak agent can influence the beliefs of the strong agent given the way our MindForge agents are constructed.

3. The agent beliefs are represented using a {question:answer} format. How are these questions generated? What mechanism determines which aspects of the environment are relevant or worth querying about?

To clarify, we employ the {question:answer} format for task-related beliefs, while for perception beliefs and partner beliefs we use a list like format {perception_beliefs: list} and {partner_beliefs:list}.

The questions for the task-related beliefs are generated similar to the base Voyager [60] setup by the autocurriculum module. Once the curriculum generator (also an LLM) decides which task to tackle next (e.g collect wood block) it generates a question for the beliefs about solving the task (e.g. "How do you collect a wood block?"). This becomes the question corresponding to the {question:answer} format. Note that the answer to the question might change over time as a result of communication with another agent, which results in a belief correction.

A concrete example of task beliefs in the {question:answer} format can be seen in Appendix C.3.

The auto-curriculum generator is also responsible for parsing the environmental state based on the agent sensory input and the items in the current inventory for deciding which task to pursue next based on resources available in proximity. This is also similar to the base Voyager setup.

4. Have the authors tested scenarios involving more than two agents with similar capabilities level?

While we have not specifically tested this setting for this manuscript, we consider this as a promising future research direction. The way our framework is structured allows for easy scalability in terms of number of agents. This would however add a layer of complexity on the communication side, and given the principle investigation needed for the two-agent case, we didn't have sufficient space to do the 3-and-more-agents case adequate justice.

5. In case where two agents of similar ability interact, do their beliefs converge over time, or does each agent maintain its own belief system?

Throughout interaction in a lifelong collaborative setting (which refers to the symmetric ability scenario), we do observe that in most cases their beliefs (especially task related beliefs) match after a certain number of communication rounds. Note that this is not always beneficial; as we discuss in Sec. 5.5, below a certain individual performance threshold, agents can also reinforce false beliefs (see the discussion on Jury Theorem in 5.5). However, we observe that when performance improves for specific initial competence starting points (Fig. 6), the false beliefs related to the tasks do get corrected and converge. We will add concrete qualitative examples to the Appendix in the camera ready version.

6. Have the authors explored settings where GPT-4 interacts with only one of the two agents in a collaborative task?

We run this experiment as part of this rebuttal, showing that indeed the uninformed agent benefits indirectly from its partner interaction with GPT-4. Thanks for this suggestion! We will add this to the paper.

In this experimental setup, we let a Mixtral-8x7B agent collaborate in an instructive setting with GPT-4 as shown in Figure 4 and then pair it with an uninitiated (base) Mixtral-8x7B. As the table below shows, the performance of the base Mixtral improves by 17% (37% $\rightarrow$ 54%) if its partner previously communicated with GPT-4.

This is an advantage of our structured belief representation as well as memory components, allowing agent to recall correct beliefs from memory and share them further with weaker agents.

Dear reviewer rCMn,

Thank you for the review and suggestions! We've addressed your questions and incorporated your suggestions. Since we have less than 2 days in the extended author discussion period, we would like to follow up to see if there is anything further we can do in the remaining time to move towards a more positive assessment.

Best,

The authors

Thank you for your review! We appreciate your recognition that (i) our agent-centric design enables smalelr LLMs to punch above their weight, (ii) the systematic ablations isolate how components such as the Theory-of-Mind graph and memory each contribute to the overall lift, and (iii) the architecture proves its value on challenging, high-complexity settings where parameter count alone typically stalls progress. Your acknowledgment underscores the practical importance of making lightweight models more capable through principled architectural choices.

To clarify, MindForge occupies a distinctive point in the design space for multi-agent interactions because it unifies three capabilities seldom found together.

1. Explicit Theory-of-Mind graph. Each agent carries a causal belief network with four disjoint buckets—task, perception, interaction, partner—and can inspect or update another agent’s beliefs during dialogue. This makes perspective-taking and mis-belief repair first-class operations rather than emergent side effects.

2. Tri-partitite long-term memory. MindForge separates experience into episodic traces, semantic summaries, and procedural code skills, each with tailored retrieval. A reward-gated consolidation step promotes useful skills and corrects false beliefs across episodes, so knowledge compounds over time instead of remaining in context.

3. Native multi-agent collaboration. Because both pillars are baked in, agents can teach, coordinate, and critique one another, not merely self-reflect. Collaboration is therefore a built-in driver of learning, not an after-thought.

To begin, all the methods you mention (EXPEL, CLIN, SSO, DEPS, ADAPT) focus on single-agent setups. Nevertheless, each provide valuable pieces: self-reflection, skill libraries, or causal abstractions but omits at least one of these three elements. The following are the concrete differences.

• EXPEL
  Targets non-parametric experiential learning, yet its representational substrate differs sharply from MindForge’s. The only enduring artefacts are (i) a vector database of full task trajectories and (ii) a free-text “insight list” distilled in offline batches. At run-time the system simply retrieves the top-k similar trajectories and concatenates them—together with any matching insights—into the current prompt. There is no explicit belief graph (social or otherwise): the chain-of-thought produced by ReAct/Reflexion is discarded once the episode ends, so the agent cannot query or revise another agent’s mental state. In MindForge’s terms, ExpeL instantiates just a two-store design (episodic-like vectors plus semantic-like insights), omitting (a) a procedural skill library, (b) reward-gated consolidation across memory types, and (c) any Theory-of-Mind reasoning. MindForge, by contrast, maintains all three stores and exposes belief-revision operations as first-class actions during dialogue, making collaboration a driver of learning rather than a post-hoc prompt augmentation.

• CLIN
  CLIN stores experience as sentence-level causal abstractions such as “Action → Effect (may/should).” After every trial a dedicated “memory-generator” LLM rewrites this list, pruning to the most salient rules; a single Controller LLM then retrieves the top-k rules (keyword + embedding match) to pick the next goal, yielding fast within-task adaptation. Yet the system remains strictly single-agent: (i) it builds no explicit belief graph, (ii) the causal rules are not promoted into any long-lived semantic or procedural store, and (iii) there is no recursion over other minds—the agent cannot inspect or repair a partner’s beliefs. Cross-task generalisation is limited to a lightweight “meta-memory” text summary, so CLIN lacks the multi-store memory and Theory-of-Mind mechanisms that drive MindForge’s collaborative learning.

• SSO
  Skill-Set Optimisation (SSO) is best understood as the procedural-memory slice of MindForge: it keeps a library of short, reward-validated code trajectories and retrieves the top-k matches by embedding similarity at inference time, precisely mirroring our procedural store. Yet SSO goes no further—there are no episodic or semantic layers, so any experience that is not rewritten into a surviving skill vanishes when the context window resets; its consolidation is purely numeric, retaining a skill only while its running return stays above ε, which trims dead weight efficiently but cannot migrate corrected beliefs or factual insights across tasks; and because the skills are treated as opaque black boxes, the framework never models why a collaborator failed or which beliefs need updating, leaving it unable to perform perspective-taking or mis-belief repair. MindForge therefore subsumes SSO’s economical skill caching while augmenting it with episodic and semantic memories and, crucially, an explicit belief graph that makes social teaching and belief revision first-class operations.

• DEPS
  DEPS excels at in-episode self-repair—each cycle it re-describes the world, explains any failure, replans, and selects the next step—yet this textual Descriptor + Explainer remains an ad-hoc string that vanishes with the context window. Because it never crystallises into a graph-structured belief model or persists beyond the episode, DEPS lacks both long-term memory (episodic, semantic, or procedural) and any representation of a partner’s mind. MindForge, by contrast, preserves experience in multi-store memory and manipulates an explicit Theory-of-Mind graph, enabling agents not only to debug themselves but also to query, revise, and teach one another’s beliefs across tasks.

• ADAPT
  ADAPT offers a neat “plan-on-failure” loop but remains strictly ephemeral: its executor relies on a fixed prompt of hand-written atomic skills, and each decomposition tree evaporates once the episode ends. Because beliefs live only as transient chain-of-thought text, the agent neither stores episodic traces nor forms a structured Theory-of-Mind, and successful sub-plans are never promoted to new skills. MindForge supplies precisely what ADAPT leaves on the table—an explicit BigToM belief graph for self- and other-modeling, tri-partitioned long-term memory that survives across tasks, and reward-gated consolidation that turns ad-hoc sub-tasks into reusable procedures—thereby shifting recovery from a one-off patch to a cumulative, collaborative learning process.

Summary Across the spectrum, MindForge is the only system that simultaneously (i) reasons over explicit causal beliefs about self and others, (ii) separates memory into episodic, semantic, and procedural components with tailored retrieval, and (iii) treats collaboration as a primary driver of belief revision. The other frameworks focus on one or two of these dimensions but do not combine all three, which is why we believe MindForge occupies a distinctive place in the current literature.

Please let us know if there are further clarifications or information you require that could lead to a increase in the score!

Thank you for your review qoax! We believe several of your comments contain factual inaccuracies (such as experiments and ablations we have already included in the paper) and misinterpretations of our contributions, which we clarify below.

1. Broaden the task set. include success-rate or reward curves on a wider range of MineDojo goals spanning different difficulties.

In fact, our main experiment (Fig.1b and Table 3) demonstrates MindForge agents’ effectiveness in an open-ended setting, for multiple tasks along the Minecraft tech-tree with increasing complexity, corresponding to MineDojo goals spanning different difficulties. Specifically, we do not restrict the set of tasks MindForge agents are performing during lifelong learning but rather let the curriculum continuously choose diverse tasks in the environment. The success-rate and reward curves presented indicate MineForge agents effectiveness on these diverse tasks over the Voyager counterparts.

We specifically report detailed results later in the paper on the simplest Minecraft tasks (dirt, wood) since the base Voyager agents failed at even these simple tasks when powered by open-weight LLMs.

Nevertheless, for further clarity for more complex tasks up the tech-tree (crafting pickaxe, mining iron), we ran additional experiments quantifying the effect of communication on task success rate ( like in Figure 4) with a MindForge agent (LLaMA 70B):

In the instructive setting, MindForge agents improve task success rate by +25% for crafting a pickaxe and 21% for mining an iron ore. Together with the existing results in Figure 4, Figure 1b) and Table 3, these demonstrate that MindForge agents can learn through natural language communication in lifelong learning settings, where tasks become increasingly hard.

In the collaborative setting two MindForge agents, we see that in a medium complex task like crafting a pickaxe, collaboration leads to an increase from $20$% to $33$%.

2. Deeper ablations: Provide separate ablations that disable each memory subsystem (episodic, semantic, procedural) and the ToM causal graph to quantify individual contribution.

In fact, we already provide ablations for our novel contributions (specific memory components and perspective-taking) (L269-72):

Ablations on memory subsystems

• episodic memory (Appendix C Table 6): MindForge agents perform worse when there is no episodic memory to remember the agent's failures.
• semantic memory (Appendix A Table 4): positive contribution; the MindForge agent uses semantic knowledge gained through prior collaboration to solve a task either in-distribution or out-of-distribution.
• procedural memory: we do not repeat the ablation since Voyager [60] already performs this experiment and shows the dramatic decrease in lifelong learning performance when this component is removed.

Perspective Taking

In Table 5, Appendix B we have already performed an ablation of perspective-taking as a whole in the MindForge framework, which together with the structured representations yields a boost in task success as the number of communication rounds increases. A visually intuitive example of how partner perspective evolves with communication is shown in Fig. 5.

Please note that performing an ablation on an individual component within the Belief-Desire-Intention framework by Bratman [R1] is not conceptually viable: a lifelong learning agent needs to have a goal (desire) and intention (action) to perform any task in the environment.

Nevertheless, we have additionally performed an ablation comparing our structured representation of Theory of Mind in MindForge with unstructured perspective-taking through a simple 2-step prompt as in Think Twice [64]. Results below indicate that structured representations provide a meaningful advantage.

3. Questionable effectiveness under purely collaborative use. The paper reports a performance drop after communication.

Potential misunderstanding about the goal: we are not looking to provide agents for 'use' towards the best improvement in the Minecraft tasks. Rather, we are investigating social interactions (perspective-taking specifically) as a means for for continual improvement of agents that fail in isolation.

In Fig. 6, we show that in the collaborative setting with extended communication regime, MindForge agents actually improve their task success rate depending on the initial expertise. This is still a positive scientific result characterizing the conditions when collaboration with symmetric expertise can work over the instructive setting. As we discuss in Sec. 5.5, the “blind leading the blind” scenario where agents below a certain individual performance threshold can reinforce mistakes is well known within the Jury theorem [3], and our results should be interpreted within this context.

Solving Minecraft tasks requires two core capabilities:

• i: factual knowledge about the game

• ii: the ability to translate that knowledge into appropriate actions.

In a collaborative setting, two agents of similar level cannot obtain new factual knowledge (unless there is an external environment signal), yielding the question: can two agents of similar level cooperate to correct their code capabilities with the right facts, reducing hallucination?

Intuitively, both communication and belief systems are as powerful as the capabilities of the underlying models allow it. While a purely collaborative scenario does not work out of the box with the set of open weights models, Figure 6 shows that collaborative MindForge agents can thrive given better initial expertise.

We show both quantitatively (see Figure 4) as well as quantitatively (see Figure 5) that the proposed system allows for a gradual increase of task performance as more communication is allowed between agents.

4. Limited novelty.. The main performance gains come from distilling reasoning produced by a stronger expert model; the distillation pipeline itself follows standard practice.

In fact, our incorporation of the Cultural Learning [56] for multi-agent embodied interactions has never been attempted in literature to the best of our knowledge. We also tackle two modes of Cultural Learning throughout the manuscript: instructive and collaborative learning.

Further, our mode of knowledge distillation doesn't follow standard practice since we employ structured theory of mind representations along with memory. In current literature, there are two main approaches to distill knowledge between different models: knowledge distillation by matching the distribution of the expert [22] and generating task specific data with the teacher model and performing supervised fine-tuning [R2]. We specifically tackle the latter case as a baseline (see Table 1) and find that distillation through supervised fine tuning just marginally improves the performance of failing Voyager [60] agents on the simplest tasks inside Minecraft: collecting dirt and wood blocks.

The approach we propose in MindForge uses the in-context learning abilities of the models to learn more efficiently in terms of computation, avoiding expensive and data intensive weight updates. Additionally, owing to natural language communication, MindForge works in human-agent settings too. We have also tested this scenario (Table 2), where interacting with a human results in improved task success rate (Mixtral-8x7B goes from 29.15% to 87% task success).

5. Compare communication cost to expert cost: Multi-round peer communication may involve more cumulative LLM tokens than a single expert rollout. Please report token counts, latency...

Reporting the absolute token count would not be a meaningful way of measuring cost since token count is depended on the prompt, model variant, type of model (base, instruct, reasoning) and API provider, to which our framework is agnostic. Instead, we consider the number of actions needed to complete an objective as the go-to measure when it comes to performance and trade-off, which we have reported in Fig. 1b.

Potential misinterpretation of goal: we are not seeking to reduce cost compared to expert rollouts in an engineering sense, but rather to get rid of the reliance on large, pretrained experts in favor of in-deployment skill acquisition.

This means that if we look at things like token counts, we must also account for the computational cost of pretraining the experts for a fair comparison overall. Instead, as we show in Fig. 1b, MindForge agents take fewer number of actions to reach a tech-tree milestone as well as reach new milestones that where previously unattainable for the agent.

References:

[R1] Bratman, M. E. (1987). Intention, Plans, and Practical Reason.

[R2] Radford, Alec, et al. "Improving language understanding by generative pre-training." (2018)

Thank you for the continued engagement, qoax, we appreciate the discussion. Since we are at the end of the phase, can we assume that our earlier clarifications have addressed your earlier concerns on:

(i) fairness of Voyager comparison,

(ii) novelty claims and foregrounding of BDI,

(iii) computation cost comparison, and

(iv) effectiveness in collaborative setting.

For these remaining points that mainly deal with task complexity and task-specific comparisons, we:

1. Clarify a potential missed point about task complexity and our claims there
2. Outline what the proposed comparisons would entail, to hopefully show that i. we have considered them in good faith, ii. explain why they are not trivial to operationalize, and iii. why they do not alter our valid basis for core contribution claims.

Task complexity

A more diverse task set seems to be necessary (ex. killing a sheep/chicken, craft a wooden axe ...)

Please note that we already ran and reported the detailed results for you on crafting a pickaxe and on another complex task mining iron above in the rebuttal.

Moreover, in Fig. 1b our agents actually reach all the way up to crafting a chainmail helmet that actually requires succeeding at all the dependent tasks earlier like get coal, craft a stone pickaxe, etc. while the corresponding Voyager does not manage to go beyond the wooden pickaxe.

Killing a chicken is not fundamentally complex, a wood axe and a single hit is enough. In contrast, our ladder of tasks are more compositional.

I personally believe "dirt" and "wood" are too very simple tasks in the Minecraft settings

As humans we also agree with you, but they remain non-trivial for open-weight agents even after finetuning on Minecraft Wiki and GPT-4 logs (L236–241). This is why we need to report these in greater detail: they are genuine stumbling blocks in our evaluation. To further emphasize the points, we describe what an experiment of the nature you are suggesting would require…

On adding non-lifelong baselines

A fair per-task comparison like you suggest would involve:

1. Manually ordering tasks by prerequisite dependencies.
2. Ensuring feasibility: controlling the spawn environment so each task is possible.
3. Evaluate each task for both agents starting with the simplest:
   • 2a (success) — record performance and move to the next task.
   • 2b (failure) — “bootstrap” missing skills without lifelong learning (e.g., grant prerequisite items, inject a single demo into memory, or fine-tune briefly), then rerun.
4. Stop when a task cannot be mastered even after bootstrapping.

Please note that Step 2b is quite non-trivial: Minecraft skills are compositional. Failing get wood but attempting craft wooden axe is like trying to playing a song on the piano without first learning the notes; it requires manual “cognitive surgery” on multiple subsystems which isn't always feasible.

For 2a, Our BDI ablation already approximates this branch and reflects what a comparison against CoELA is likely to look like for such a baseline.

Crucially, please consider that running full 2a/2b chains for all tasks would require substantial extra engineering and compute (our current experiments already exceed 600 GPU-hrs).

On “weak” models

our method significantly improves the performance of vgg-16 and ResNet

A more accurate analogy to our claim would be "our method significantly improves the performance of vgg-16 and ResNet without any pretraining directly at test". This would still be significant for the community because it

• changes the research question and contribution away from absolute performance gains to constraints about test time improvement.
• ofcourse, the core methodological components (e.g. explicit BDI through the prompts we have provided in this case) can still be integrated in other more powerful models. Our method actually provides increasing performance returns with increase in base in-context capabilities of these open-weight LLMs. Please do see this phenomena demonstrated with our experiments using Mistral-7b and LlaMa-70b.
• Please specify the broader claim are you referring to? We make no case that this is the best way to "solve" Minecraft which it seems you are alluding to with your analogy involving vgg and ResNet-this not the case we make.