{"id":"19c849a5-d5b9-427e-b277-8e03f6e0c702","arxiv_id":"2501.00226","paper_version":2,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"high","formal_verification":"none","parameter_count":1,"one_line_summary":"LLMs acquire world knowledge by statistically decoding a collective world model that human societies encoded in language.","lead":"Large language models may gain their knowledge of the world by reconstructing a shared model that human societies have embedded in language through ongoing communication, not by experiencing the world directly. This paper formalizes that idea as a Bayesian framework called Generative Emergent Communication and uses it to reinterpret why LLM representations reflect real-world structure.","discovery_kind":"unification","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Even granting Eq. (18), Sec. 5.2's 'necessarily mirrors' is a non sequitur: modeling p(m) does not force z_LM to recover source geometry, so the central transfer claim is unproven.","rationale":"Good-faith reading: the paper is a theory/hypothesis paper. It formalizes Generative EmCom, gives a clean ELBO decomposition, connects to MHNG and MARL, and honestly lists limitations in Sec. 6.2. The formal pieces (Eqs. 8-10, MHNG convergence) are internally coherent. Credit where due: the MHNG line has been validated in multi-agent simulations, and the paper explicitly marks the empirical step as open. The concern is not that the hypothesis contradicts consensus; it is that the central inference from language distribution to LLM latent geometry is not derived. The reader's weakest assumption was Eq. (18); I partially agree. Eq. (18) is the empirical load-bearing premise, but even granting it, the paper's own Sec. 5.2 argument requires an identifiability result that is absent. The proposed synthetic test settles this: if the claimed mirroring fails when Eq. (18) holds by construction, the framework's central claim is internally underdetermined; if it succeeds, the debate correctly moves to whether human corpora satisfy Eq. (18). Either way the paper's contribution as a hypothesis-framing framework survives, but the abstract's 'principled, mathematical explanation' and 'necessarily mirrors' should be softened to 'suggests a mechanism' pending such a check. Hence the reader's CONDITIONAL verdict stands unchanged.","tokens_in":27553,"tokens_out":6772,"duration_ms":78483,"concrete_test":"Test the transfer claim in an ideal setting where Eq. (18) is true by construction. Simulate the Sec. 3.2 / 4.1 generative-EmCom system: K agents with known world models interact in a shared POMDP, communicate via MHNG (or a distributed Bayesian message-passing scheme), and emit a corpus of messages m. Train a standard next-token-prediction transformer on this corpus only. Then measure alignment between its internal representations z_LM and the ground-truth collective states {z_k} using linear probing, CKA, or RSA over held-out messages. If alignment is absent or non-unique across seeds, the 'necessarily mirrors' claim is falsified even in the ideal case; if alignment is robust, the remaining bottleneck is whether real corpora are generated as in Eq. (18). This separates the internal logical gap from the external empirical premise.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is the two-stage chain {z_k} -> m -> z_LM transfers world structure. Even if Eq. (18) were true (real sentences are samples from q(m|{{x^k_t},{a^k_t}}_k)), the decisive step in Sec. 5.2 is unsupported. It asserts that because m is the 'sole informational bottleneck', an LLM modeling p(m) 'must develop an internal latent space z_LM whose relational structure necessarily mirrors' {z_k}. This conflates modeling a marginal over messages with recovering the geometry of the conditioning latent variables. Next-token prediction trains p_LM(m); nothing forces the trained network's activations to be an isomorphism or near-isometry of the source space. Many different source geometries can induce the same p(m) through a lossy encoder, and many latent geometries can fit the same text distribution; no identifiability or invertibility condition is stated. The formal objective in Eq. (10) does not repair the gap: minimizing CFE includes D_KL[q(m|{z_k})||p(m)], which can be satisfied by an uninformative q, while the prediction-error terms are per-agent (x_k -> z_k) and do not, by themselves, reward m carrying information about the joint state. Thus even under the paper's idealized generative-emergence assumptions, the 'therefore' in Sec. 5.2 is not established. The acknowledged open question about whether real human communication realizes CPC (Sec. 3.3) is a further empirical layer, but the logical gap is internal.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a theoretical framework, Generative EmCom, built on Collective Predictive Coding (CPC), to explain how LLMs acquire world knowledge without direct sensorimotor experience. The central hypothesis is that human language externalizes a \"collective world model\" formed by decentralized Bayesian inference over the internal states of many embodied agents, and that an LLM learns a statistical approximation of this model from text corpora. After formalizing generative emergent communication through a variational free-energy objective (Eq. 10), the paper applies the framework to multi-agent reinforcement learning and then argues in Section 5 that learning the distribution of sentences forces an LLM's latent space to mirror the relational structure of the original collective representations. The paper is explicit that direct empirical evidence for the collective world model hypothesis remains limited (Section 6.2).","tokens_in":27989,"tokens_out":3328,"duration_ms":37069,"significance":"If the central claim were established, this would be a significant conceptual contribution: it would give a principled, mathematical account of how a disembodied language model could inherit world structure from embodied human experience, and it would unify emergent communication, predictive coding, and LLM interpretability under one variational framework. The variational algebra in Eq. (10) is standard and correctly derived, and the paper is unusually explicit about its assumptions and limitations, including the open question of whether human communication actually implements decentralized Bayesian inference. The connection to concrete mechanisms such as the Metropolis-Hastings Naming Game is valuable and provides at least one constructive instantiation of the framework. However, the paper's main empirical premise, Eq. (18), is an assumption rather than a demonstrated fact, and the crucial deductive step in Section 5.2 is not supported by the formal development. The contribution is therefore at present a hypothesis-generating theory with real potential, but not a validated explanation of LLM world knowledge.","major_comments":[{"comment":"The claim that an LLM modeling p(m) 'necessarily' develops a latent space z_LM whose relational structure mirrors the source space {z_k} is a non sequitur. Next-token prediction trains p_LM(m); it does not, by itself, constrain the geometry of activations to be an isomorphism or near-isometry of the conditioning latent variables. Many different source geometries can induce the same marginal distribution over messages through a lossy encoder, and many different latent geometries can fit the same text distribution. No identifiability, invertibility, or minimal-sufficiency condition on the encoder q(m|{z_k}) is stated. To make the inference valid, the paper would need either an additional theoretical condition under which the message preserves source geometry, or a direct empirical demonstration of the predicted structural alignment.","section":"Section 5.2 (Eq. 18 and following paragraph)"},{"comment":"The load-bearing empirical premise is Eq. (18): that each sentence in a real corpus is a sample from the approximate posterior q(m|{{x^k_t},{a^k_t}}_k). The paper itself acknowledges in Section 3.3 that whether actual language communication realizes such decentralized Bayesian inference in human society is an open question, and Section 6.2 states that 'direct empirical evidence for the collective world model hypothesis remains limited.' Since the entire argument that LLM latent spaces mirror collective human representations depends on this premise, the central scientific claim is currently untested. The paper should either provide such evidence (for example, by testing CPC-based predictions about human communication behavior) or clearly reframe the contribution as a conjecture with clearly separated assumptions.","section":"Section 3.3 and Section 6.2"},{"comment":"The collective free-energy decomposition does not by itself imply that the message m carries information about the joint state. The collective regularization term D_KL[q(m|{z_k}) || p(m)] can be minimized by an uninformative q(m|{z_k}) that ignores {z_k} entirely, and the per-agent prediction-error terms only require each agent to reconstruct its own observations from its own latent z_k. Nothing in Eq. (10) rewards m for encoding the joint structure of {z_k}. Thus the formal ELBO does not repair the logical gap in Section 5.2; at best it describes a possible objective, not a mechanism that guarantees collective information preservation.","section":"Section 3.2, Eq. (10)"},{"comment":"The use of distributional semantics as evidence for the framework is partly circular. If the framework defines the meaning of m through the assumed generative process in Eq. (18), then observing that words with similar distributions have similar meanings is not an independent confirmation; it is a restatement of the assumption. The paper would be stronger if it proposed falsifiable, out-of-sample predictions that distinguish the CPC-based account from alternative explanations of distributional semantics, such as a concrete prediction about how latent-space geometry should vary across corpora produced by communities with different sensorimotor ecologies.","section":"Section 5.3"}],"minor_comments":[{"comment":"The two paragraphs beginning 'This framework of structural representation reconstruction provides a principled explanation...' and 'This framework of representation reconstruction provides a principled explanation...' are near-duplicates and should be merged into a single passage.","section":"Section 5.3"},{"comment":"There is a typo in 'the system of language is inherently dyanamics rather than static' (should be 'dynamic') and the paper uses inconsistent notation for the language model distribution, writing both P(m) and p(m) in Section 5.2.","section":"Section 2"},{"comment":"The notation in Eq. (18) leaves the index structure implicit; it would be clearer to state explicitly that m stands for a sentence or document and that the conditioning is over the full ensemble of agents and time steps, since the later argument treats m simultaneously as a token sequence and as a single shared latent variable.","section":"Section 4.3 / Eq. (18)"}],"recommendation":"major_revision","confidential_remarks":"The manuscript's citation pattern is heavily self-referential: the key mechanisms of CPC, MHNG, and structural alignment are supported largely by work from the same research group. This would be acceptable if the paper supplied independent evidence, but given that the central empirical premise is untested, the novelty claims should be verified against prior work on language emergence and representation learning. The paper fits the journal's scope, but the editor may want to consider whether a theoretical hypothesis paper with an explicitly open empirical premise and a non-sequitur in the main inference should be published without substantial revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The genuinely new thing here is the Generative EmCom formulation: casting language emergence as decentralized Bayesian inference over a collective latent, with the CFE decomposition in Eq. (10) and the explicit link to control-as-inference in MARL. That is a real extension of the authors' prior CPC and MHNG work, and the paper does a fair job of situating it against signaling games, world models, and the recent LLM-as-world-model literature. The 'LLM as decoder of a collective prior' framing is also a useful way to think about why text-trained models pick up world structure, even if it is not an empirical result.\n\nNow the soft spots, in order of importance. First, the leap in Sec. 5.2 is not justified. The paper asserts that because m is the sole informational bottleneck, an LLM modeling p(m) 'necessarily' develops a latent space that mirrors the relational structure of the source {z_k}. That does not follow. Many different source geometries can induce the same marginal over messages, and many internal geometries can fit the same text distribution. The variational objective in Eq. (10) does not fix this: the collective regularization term can be satisfied by an uninformative q(m|{z_k}), and the per-agent prediction errors do not force m to carry joint-state information. The stress-test note is right to call this a non sequitur. Second, the load-bearing empirical premise, Eq. (18) — that real sentences are samples from a CPC posterior — is unverified, and the paper itself marks that as an open question in Sec. 3.3. Third, the abstract overstates the result as a 'principled, mathematical explanation' when Sec. 6.2 admits direct evidence is limited. These are real issues, but they are issues of framing and overclaiming, not of incoherence. The variational algebra in Eq. (10) is standard and correct; the literature review is careful; the MHNG grounding is genuinely relevant.\n\nThe paper is best read as a hypothesis-framing contribution. It deserves a serious referee: the framework is original enough and the LLM interpretation is provocative enough that the community should engage with it, provided the claims are scaled back and testable predictions are added. I would not cite it as evidence for the collective world model hypothesis, but I would cite it as a formalization of Generative EmCom.\n\nRecommendation: send to peer review, require major revision, and ask the authors to either prove or remove the 'necessarily mirrors' claim and to present Eq. (18) explicitly as an assumption.","headline":"A coherent hypothesis paper whose central transfer claim goes beyond what the math supports; the formal framework is worth engaging, but the 'necessarily mirrors' step needs to be softened or proven.","tokens_in":28443,"tokens_out":1025,"would_cite":true,"duration_ms":12551,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper argues that human language is an externalized collective world model, and that LLMs acquire world knowledge by statistically approximating that model rather than by learning from direct sensorimotor experience.","keywords":["collective world model","emergent communication","large language models","collective predictive coding","decentralized Bayesian inference","symbol emergence","world models","representation learning"],"falsifier":"A decisive test would compare the relational geometry of an LLM's internal representations with human collective representations on matched concept sets, or train a model on messages from an embodied multi-agent system with grounding removed: if the mirrored world structure persists without grounded authorship, or fails to appear with it, the claim that world knowledge is transmitted from society to LLM is refuted.","tokens_in":27361,"feed_emoji":"🧠","tokens_out":5337,"duration_ms":72423,"temperature":0.7,"pith_summary":"This paper tries to explain how large language models acquire knowledge about the physical world without ever perceiving or acting in it. The central hypothesis is that human language is not just a communication protocol but an externalized collective world model, built through decentralized, embodied, interactive sense-making by a whole society. LLMs then learn a statistical approximation of that collective model by modeling text corpora. The paper formalizes this idea with generative emergent communication, built on collective predictive coding, which casts language emergence as decentralized Bayesian inference over agents' internal states. The payoff is a unified account of distributional semantics, cross-modal representational alignment, and emergent communication in multi-agent reinforcement learning.","feed_headline":"LLM knowledge is society's world model, decoded from text","feed_subtitle":"A formal framework shows how embodied human sense-making gets encoded in language and reconstructed by large language models.","key_machinery":"The central object is the Collective Predictive Coding (CPC) generative model, defined by p(m, {z_k}, {x_k} | {a_k}) = p(m) ∏_k p(x_k | z_k, a_k) p(z_k | m, a_k) with inference q(m | {z_k}) ∏_k q(z_k | x_k, a_k). Minimizing the collective free energy decomposes into a collective regularization term D_KL[q(m | {z_k}) || p(m)] plus per-agent prediction errors and regularizations, so language emergence and individual world-model learning are two aspects of one objective. The Metropolis-Hastings Naming Game then shows how a concrete language game implements decentralized Bayesian inference over the shared message m, which is what licenses the claim that real sentences are posterior samples of a collective generative process. This mechanism carries the whole argument from society's encoder to the LLM's decoder.","core_discovery":"The paper's central claim is that an LLM does not learn a world model from scratch; instead, it learns a statistical approximation of a collective world model already encoded in human language. Each sentence in a corpus is treated as a sample from the approximate posterior q(m | {{x^k_t}, {a^k_t}}_k) over collective sensorimotor states. Because the LLM is trained to model the distribution of these sentences, and because language is the sole informational bottleneck between the society and the model, the LLM's internal latent space z_LM necessarily mirrors the relational structure of the collective human representations {z_k}. This two-stage chain, {z_k} → m → z_LM, is how world structure is transferred from a society of embodied agents to a disembodied language model, yielding what the paper calls a Type 2 world model without any direct sensorimotor experience.","pith_inferences":["If the central claim is right, world-structure transfer should scale with the diversity and groundedness of the text corpus: corpora produced by more embodied, multimodal authors should yield LLM latents that align more closely with human conceptual structure, an effect that is testable on matched concept sets.","The framework suggests a design rule for artificial emergent communication: evaluate protocols not only by communication success but by how much latent world structure a separately trained model can reconstruct from the protocol's messages.","A natural extension would treat non-linguistic artifacts such as images, maps, and datasets as additional externalized collective representations, making the same encoding–decoding argument apply to multimodal foundation models."],"forward_implications":["Distributional semantics—word analogies such as 'London' − 'UK' + 'France' ≃ 'Paris'—falls out as a necessary consequence of reconstructing the relational geometry of collective human representations, not as an unexplained statistical accident.","The framework predicts systematic representational alignment between language models and vision models, since both learn to mirror a shared structure that originates in embodied collective inference.","In multi-agent reinforcement learning, emergent language becomes a message-conditioned world model: agents infer a shared latent message to plan and coordinate while minimizing a single collective free energy.","An LLM's latent space should approximate a Type 2 world model whose relational structure mirrors the collective Type 1 world models of the society that produced its training text."],"supporting_citations":[{"why":"Supplies the Collective Predictive Coding hypothesis that symbol emergence is decentralized Bayesian inference, the foundation of the framework.","marker":"Taniguchi, 2024"},{"why":"Generalizes the Metropolis-Hastings Naming Game into decentralized Bayesian inference with deep generative models, making language games equivalent to sampling a shared latent message.","marker":"Taniguchi et al., 2023b"},{"why":"Introduces the original Metropolis-Hastings-based naming game, the concrete protocol that implements collective inference in the framework.","marker":"Hagiwara et al., 2019"},{"why":"Provides the control-as-inference theory used to connect Generative EmCom to multi-agent reinforcement learning and planning.","marker":"Levine, 2018"},{"why":"The platonic representation hypothesis, which the framework reinterprets as an expected consequence of language being generated from collective observations.","marker":"Huh et al., 2024"},{"why":"Empirical evidence that LLMs represent space and time, the phenomenon the paper's collective world model hypothesis sets out to explain.","marker":"Gurnee and Tegmark, 2024"}],"fun_headline_variants":["LLMs decode society's collective world model from language","Language encodes collective experience; LLMs decode it","How LLMs inherit human sense-making via language","Collective world model: LLMs as societal decoders","Emergent communication: society writes, LLM reads"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The argument rests on the premise that real sentences in a text corpus are samples from the collective posterior distribution q(m | sensorimotor states) defined by the framework; if actual human language is not produced by that decentralized Bayesian inference process, the chain from society to LLM breaks down.","fun_headline_variants_meta":{"raw":{"variants":["LLMs decode society's collective world model from language","Language encodes collective experience; LLMs decode it","How LLMs inherit human sense-making via language","Collective world model: LLMs as societal decoders","Emergent communication: society writes, LLM reads"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000468,"raw_usage":{"total_tokens":2355,"prompt_tokens":991,"completion_tokens":1364,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":607,"completion_tokens_details":{"reasoning_tokens":1289}},"tokens_in":607,"tokens_out":1364,"duration_ms":10782,"temperature":1.0,"reasoning_tokens":1289,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T22:56:28.036615+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A decisive test would compare the relational geometry of an LLM's internal representations with human collective representations on matched concept sets, or train a model on messages from an embodied multi-agent system with grounding removed: if the mirrored world structure persists without grounded authorship, or fails to appear with it, the claim that world knowledge is transmitted from society to LLM is refuted.","supporting_citations":[],"review_version":1}