{"id":"24110512-f51d-4d25-b1bb-18f7d7aac1a2","arxiv_id":"2505.12321","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"BeliefNest represents nested beliefs as hierarchical Minecraft simulators and shows LLM agents can use them to pass false-belief tasks.","lead":"BeliefNest is an open-source Minecraft-based simulator that represents agents' beliefs as nested virtual worlds. It lets LLM agents reason about what others think, and the paper demonstrates it on classic false-belief tasks.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Experiment confounds BeliefNest's belief construction with GPT-4o's prior knowledge of classic false-belief tasks; no direct verification of the constructed nested belief states is reported.","rationale":"The reader's weakest assumption—that Oi faithfully captures visibility and that direct observation fully determines belief—is plausible and relevant, and the direct state-dump test I propose would also probe it. However, I see a more load-bearing gap in the experimental logic: the reported demonstrations cannot distinguish the simulator's contribution from GPT-4o's well-documented priors on exactly these classic ToM tasks. This matters for the central claim because 'accurately constructing others' beliefs' and 'enabling appropriate inference' are two separate achievements; the end-to-end result does not localize either. The correct response is not to reject the system—BeliefNest is an open-source platform with a clear architecture, and the belief-update/visibility machinery is genuinely relevant—but to make acceptance conditional on a control that removes the LLM confound and on a direct check of the constructed belief states. Since the reader already returned CONDITIONAL, my stress-test does not change the verdict, but it sharpens what the revision must supply: an ablation or a probe that isolates belief construction from action generation.","tokens_in":6844,"tokens_out":6511,"duration_ms":72303,"concrete_test":"Run Condition 3 (and ideally all three Sally-Anne conditions) with GPT-4o under three prompt sources: (a) the current BeliefNest second-order belief prompt; (b) a prompt built from the observer's first-order belief state (which contains the diamond's true location but not Sally's movement); and (c) a prompt built from the real-world state (which includes Sally's true position and her view of Anne). If GPT-4o's predicted action is the same in (a), (b), and (c), the experiment is confounded by LLM priors and cannot support the claim that BeliefNest's nested belief state causes the correct inference. If instead the action flips in the predicted direction—e.g., left box for (a) but right box for (c)—then BeliefNest's belief construction is doing causal work.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim in Section 7 is that BeliefNest 'accurately constructs' others' beliefs and 'enables appropriate inference of belief-based actions.' But the reported evidence is only that GPT-4o, prompted with BeliefNest-generated text, produces the textbook-correct action in two canonical false-belief stories (Sally-Anne and Ice Cream Van). GPT-4o has almost certainly seen these tasks during training and can answer 'Sally goes to the original location' without any help from the simulator. Conditions 1 and 3 both terminate in the same expected action (left box), so a model relying on story priors would pass both even if the second-order belief state were empty, wrong, or identical to the observer's first-order belief. No ablation removes the nested-belief information, no baseline uses the real-world state or the observer's first-order belief, and no direct inspection of the belief simulators' contents (e.g., Sally's stored location, the diamond's position in the observer's belief about Sally's belief) is shown. Additionally, the belief-update rule 'unobserved elements remain unchanged' (Common-2) is a normative assumption: in Condition 3 the observer 'does not know that Sally observed the action' only if the observer truly has no way to infer Sally's movement from context or prior knowledge. Whether this matches actual agent beliefs is not empirically validated. Thus the strongest claim is not yet supported: the observed correct actions could be entirely driven by LLM priors rather than by BeliefNest's belief construction.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces BeliefNest, an open-source Minecraft-based simulator that supports embodied agents with Theory of Mind by maintaining a hierarchy of simulated worlds: a real-world simulator and nested belief simulators that represent agents' beliefs about other agents' beliefs. Each simulator updates beliefs from an observation function, supports control and follow modes, and can generate prompts for LLM-based agent control. The paper reports two demonstrations: a Sally-Anne false-belief task under three visibility conditions and an Ice Cream Van second-order false-belief task, both evaluated by using GPT-4o to generate actions from belief-informed prompts. The central claim is that BeliefNest accurately constructs others' beliefs and enables appropriate inference of belief-based actions.","tokens_in":7096,"tokens_out":3384,"duration_ms":36987,"significance":"If the claims are validated, BeliefNest would be a useful reusable platform: it has no fitted parameters, explicitly represents nested belief states in a deterministic way, is integrated with the widely used Minecraft environment, and provides prompt-generation support for LLM-based control. The paper's strengths are the clarity of the core mechanism and the open-source repository, which would allow other researchers to build on the system. However, the current demonstrations are qualitative and single-run, and they do not rule out the possibility that the LLM's prior knowledge of classic false-belief tasks produces the reported actions. The significance of the contribution therefore depends on stronger experimental evidence than the paper currently provides.","major_comments":[{"comment":"The claim that BeliefNest 'accurately constructs others' beliefs' and 'enables appropriate inference of belief-based actions' is not supported by the current experiment because GPT-4o has likely seen the classic Sally-Anne and Ice Cream Van tasks during training. Conditions 1 and 3 both terminate with the same expected action (Sally opens the left box), so a model relying on story priors would pass both even if the nested belief state were empty, incorrect, or identical to the observer's first-order belief. To support the claim, the authors should add at least: a baseline condition that uses the real world state or the observer's first-order belief without the nested belief information, an ablation condition that removes or corrupts the second-order belief state, and multiple runs with novel or randomized object layouts that the LLM cannot have memorized.","section":"Section 7 (results paragraph)"},{"comment":"The paper reports no direct verification of the constructed nested belief states. For each condition, the authors should show the actual contents or state summaries of the belief simulators (e.g., Sally's stored position, the recorded chest contents, and the visibility flags) and compare them against the ground-truth belief states. Without such inspection, the reader cannot confirm that the simulator produced the belief state that is claimed to explain the action, rather than the LLM independently inferring the textbook answer.","section":"Sections 7 and 8, and Algorithm 1"},{"comment":"The evaluation is qualitative and single-run: no metrics, no error bars, no number of trials, and no failure cases are reported. Because GPT-4o generation can be stochastic, multiple runs with different seeds are needed to verify that the observed actions are consistent. The Ice Cream Van section has only one scenario and no variation comparable to the three conditions in Section 7, so it does not by itself demonstrate that deeper nesting is handled correctly under conditions where the belief state would change.","section":"Sections 7 and 8"},{"comment":"The belief-update rule 'unobserved elements remain unchanged' is a normative assumption that is load-bearing for Condition 3. The observer's lack of knowledge about Sally's movement is modeled by leaving Sally's position unchanged, but the paper does not validate whether this matches actual human belief inference or whether the observer could infer unobserved events from context or prior knowledge. The authors should either provide evidence for this assumption or discuss its limitations and test conditions under which it fails.","section":"Section 4, Common-2 and Section 7, Condition 3"}],"minor_comments":[{"comment":"The author affiliation line has a spacing typo: 'Koichiro Terao2' should be 'Koichiro Terao 2' to match the other authors.","section":"Author line"},{"comment":"The terms 'box' and 'chest' are used inconsistently for the same objects in the Sally-Anne task; the paper should use one consistent term.","section":"Sections 7 and Appendix A.1"},{"comment":"The prompt generation mechanism is described only by example; it would be clearer to state explicitly which state variables are included and how they are serialized, since this matters for reproducibility.","section":"Section 5"},{"comment":"The statement 'A first-order false-belief task measures whether an agent can infer another agent's first-order belief. In this setting, the observer constructs a belief about Sally's belief, which corresponds to a second-order belief from the observer's perspective' is correct but could be clarified by explicitly labeling the level of each simulator in Figure 3.","section":"Section 7"},{"comment":"The conclusion claims the simulator is applicable to 'a broader range of research on joint action,' but the paper provides no example or argument beyond the false-belief tasks; a sentence explaining what joint-action scenarios are envisioned would strengthen the conclusion.","section":"Section 9"}],"recommendation":"major_revision","confidential_remarks":"The paper is best viewed as a system description with preliminary demonstrations. The main weakness is that the reported experiments do not isolate the contribution of the nested belief simulator from the LLM's prior knowledge of canonical false-belief tasks. The authors should either add the suggested control and ablation experiments or explicitly soften the claims to describe the system's functionality rather than its validated effectiveness. I would not reject the paper, because the platform itself is likely useful to the community, but the central claim requires substantially stronger evidence."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: BeliefNest is a genuinely new implementation—nested beliefs as stacked Minecraft simulators—and the code is open source, so I'd want it on my radar. But the paper's evidence for its central claim is much weaker than the claim itself.\n\nWhat's new and good: The architecture of control/follow modes and timeline branching is a sensible way to make nested belief states explicit and inspectable. The prompt-generation layer is flexible, letting different LLMs be plugged in. The authors build on Voyager and Minecraft, which is a reasonable testbed. The demonstrations on Sally-Anne and Ice Cream Van show the system can produce the expected answers. I believe the system is real and the design is coherent.\n\nSoft spots: The evaluation is the weak part. Two classic tasks, no quantitative metrics, no error bars, and—more seriously—the experiments as described cannot distinguish BeliefNest's belief construction from GPT-4o's prior knowledge. The stress-test note is right: GPT-4o has surely seen Sally-Anne, and Conditions 1 and 3 both predict 'left box,' so even a broken second-order belief state would pass. There is no ablation that removes the nested-belief information, no baseline that feeds the real-world state directly to the LLM, and no direct inspection of what the belief simulators actually contain (e.g., Sally's stored location). Also, the rule that 'unobserved elements remain unchanged' is a design assumption, not an empirical finding; it may break in open-ended tasks.\n\nIn fairness, the circularity burden is low—no fitted parameters, no self-referential validation—and the belief states are deterministic consequences of the observation function. The issue is that the paper doesn't verify those consequences. The authors claim Section 7 demonstrates 'accurately constructing others' beliefs'; the actual evidence is a couple of correct final actions from an LLM that could have produced them without the simulator.\n\nCitation pattern: relevant and honest. They cite the standard ToM literature and LLM ToM evaluation work; no self-citation problem.\n\nWhom is this for: researchers building embodied LLM agents who need a structured way to represent nested beliefs. As a tool paper, it's a useful contribution; as a scientific claim, it's not yet supported.\n\nRecommendation: send it to peer review, but the referee should require a proper experimental section: baselines, ablations, and ideally state logging. If the authors can show that BeliefNest's belief states change the LLM's answer, the tool becomes much more compelling.","headline":"BeliefNest is a genuinely new piece of ToM infrastructure whose demonstrations are too thin to support its headline claim—worth reviewing, but it needs a real experimental section.","tokens_in":7615,"tokens_out":3044,"would_cite":true,"duration_ms":30292,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"BeliefNest simulates nested beliefs by running a separate Minecraft world for each 'thinks that' level, and that lets embodied agents pass false-belief tasks.","keywords":["theory of mind","nested beliefs","belief simulation","embodied agents","false-belief tasks","large language models","Minecraft","joint action"],"falsifier":"Swap the wall in Condition 1 for a pane that is transparent in the rendered view but registered by the simulator's observation function as an occluder; if the system still predicts Sally opens the left box, belief states are driven by O_i's model of visibility rather than by what Sally could actually see, which would settle the load-bearing premise one way or the other.","tokens_in":6633,"feed_emoji":"🧠","tokens_out":6621,"duration_ms":66362,"temperature":0.7,"pith_summary":"This paper claims that explicit nested beliefs—'I think that Sally thinks the diamond is still in the left box'—can be made operational for embodied agents by physically instantiating each level of belief as its own simulator inside Minecraft. BeliefNest links these simulators so that observations in the real world propagate into shallower-level belief worlds, and deeper worlds can switch into a mode where an agent runs hypothetical actions. The authors demonstrate that an in-world observer agent, using GPT-4o for action generation, correctly infers Sally's first-order belief and resulting box-opening in the Sally-Anne task across three visibility conditions, and handles the second-order John–Mary–ice-cream-van scenario. If this works as claimed, the platform gives LLM-driven agents a concrete, inspectable substrate for Theory of Mind reasoning and a testbed for belief-based joint action.","feed_headline":"Nested Minecraft worlds let agents predict what others believe","feed_subtitle":"BeliefNest builds a separate simulator for every belief level, and GPT-4o agents act on the belief states they contain.","key_machinery":"The load-bearing mechanism is the nested hierarchy of Minecraft simulators, each identified by a belief path Z (a sequence of agent identifiers). For a simulator with state s_Z, agent i opens a deeper simulator whose state s_{Z||i} is meant to represent b_i, i's belief. Two operational modes do the work: control mode, where the agent acts and generates a hypothetical timeline, and follow mode, where the simulator passively adopts belief states received from the shallower level. Each simulator runs a common loop—observation from the agent's perspective O_i, belief update (observed elements change, unobserved elements stay), and propagation to deeper simulators—so belief content is exactly the residue of what has been observed. The branching timeline mechanism lets an agent try out several belief-based plans before choosing one in the real world.","core_discovery":"On the paper's own terms, the central discovery is that a hierarchy of physical simulators can serve as an explicit, inspectable representation of nested belief states. Each agent in the 'real-world' simulator has a belief simulator that is kept in follow mode: it receives the agent's current belief along with what the agent can observe at that moment. When an agent needs to reason about another agent's belief, it switches that simulator into control mode and runs a counterfactual timeline inside it; the resulting state is exactly 'what A believes about B's situation,' and it can be fed into an LLM-generated program for action. The Sally-Anne and Ice Cream Van experiments show the constructed second- and third-order belief states agree with the standard false-belief results: Sally believes the diamond is where she last saw it, and John believes Mary does not know the seller moved. The paper therefore claims that belief-level separation plus an observation-driven update rule is sufficient for an embodied agent to predict others' belief-based actions.","pith_inferences":["The paper does not develop it, but BeliefNest's per-level states could serve as automatic ground-truth labels for training belief-state trackers, since the simulator provides exact annotations for any scenario it can render.","The observation-update rule treats belief as a pure function of current visibility; adding probabilistic or episodic-memory updates would let the same machinery handle noisy or forgotten observations, a direction the paper leaves implicit.","Because the hierarchy is just nested simulators, the design could be lifted out of Minecraft onto any physics or game engine, turning BeliefNest into a general recipe for Theory of Mind in embodied robotics evaluation."],"forward_implications":["Any agent in the hierarchy can, in principle, reason about arbitrarily deep belief nesting by adding simulator layers, so first- and second-order false-belief tasks become a resource question rather than a representational one.","Because belief states are real Minecraft worlds, researchers can inspect, record, and replay exactly what an agent believes at each level, giving observable ground truth for belief tracking.","The same prompt-generation machinery can be pointed at any belief level, so an LLM can be asked to act 'as if I were Sally' with the actual belief content, not a paraphrase.","The simulator supports open-domain Minecraft tasks—movement, breaking, crafting, chests—so false-belief tests can scale to richer collaborative plans than puzzle-box scenarios."],"supporting_citations":[{"why":"Supplies the set of action primitives (movement, crafting, chests) that BeliefNest builds on for agent control.","marker":"[1]"},{"why":"Evidence that language models struggle to distinguish belief states grounded in different perspectives, the gap the simulator is designed to close.","marker":"[23]"},{"why":"Defines the Sally-Anne false-belief scenario used as the first experimental task.","marker":"[24]"},{"why":"The large language model used to generate action programs from belief-state prompts in both experiments.","marker":"[25]"},{"why":"Defines the second-order false-belief scenario ('John thinks Mary thinks') used as the second experimental task.","marker":"[26]"}],"fun_headline_variants":["BeliefNest simulators let agents predict false beliefs in Minecraft","Hierarchical belief sims enable AI agents to model others' minds","Layered Minecraft worlds track agents' beliefs about beliefs","Nested simulators give agents a testable theory of mind","BeliefNest's separate worlds help agents foresee others' actions"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The construction assumes that each agent's belief is fully determined by the observations its perception function provides at each step, so if agents infer unobserved facts from context, forget details, or perceive the world differently than the simulator's observation function does, the nested belief states and the predicted actions will diverge from genuine beliefs.","fun_headline_variants_meta":{"raw":{"variants":["BeliefNest simulators let agents predict false beliefs in Minecraft","Hierarchical belief sims enable AI agents to model others' minds","Layered Minecraft worlds track agents' beliefs about beliefs","Nested simulators give agents a testable theory of mind","BeliefNest's separate worlds help agents foresee others' actions"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00034,"raw_usage":{"total_tokens":1827,"prompt_tokens":848,"completion_tokens":979,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":464,"completion_tokens_details":{"reasoning_tokens":892}},"tokens_in":464,"tokens_out":979,"duration_ms":10561,"temperature":1.0,"reasoning_tokens":892,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T20:35:37.255259+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Swap the wall in Condition 1 for a pane that is transparent in the rendered view but registered by the simulator's observation function as an occluder; if the system still predicts Sally opens the left box, belief states are driven by O_i's model of visibility rather than by what Sally could actually see, which would settle the load-bearing premise one way or the other.","supporting_citations":[{"cited_title":"theory of mind","cited_arxiv_id":null,"evidence_quote":"Defines the Sally-Anne false-belief scenario used as the first experimental task."},{"cited_title":"John thinks that Mary thinks that","cited_arxiv_id":null,"evidence_quote":"Defines the second-order false-belief scenario ('John thinks Mary thinks') used as the second experimental task."}],"review_version":1}