{"id":"5a1bc7f1-4f6c-499f-96c4-f8309803e11d","arxiv_id":"2501.01638","paper_version":2,"verdict":"REJECT","confidence":"HIGH","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":4,"one_line_summary":"Claims that LLMs are non-ergodic and that capability emergence obeys a resource-constrained 'adjacent possible' equation, but the derivation is an analogy and the experiments are too small to validate it.","lead":"This paper proposes that large language models are non-ergodic and that emergent abilities follow a resource-limited version of Kauffman's adjacent possible equation. The framework is a conceptual analogy, but the mathematical proofs and experiments are too weak to support the strong claims.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central equation rests on Eq. 18, which equates a d-dimensional attention output to a scalar combinatorial sum; without a valid isomorphism, Eqs. 33–36 are not derived, and the experiments do not instantiate Eq. 36.","rationale":"The reader's weakest-assumption analysis identified the same load-bearing joint: Lemma 2's Eq. 18 is an analogy, not a proof. My stress-test sharpens this concern: the equation is not merely unproven but dimensionally incoherent as written, since the attention output is a vector while the TAP sum is a scalar. Because the subsequent derivation from Eq. 33 through Eq. 36 depends on this purported isomorphism, the central claim that capability emergence follows Eq. 36 is unsupported. The experimental sections do not close this gap: they report entropy, PCA dimensionality, accuracy, and path metrics, none of which are instantiations of A_t or α(i,t) from Eq. 36. The paper even acknowledges the need for independent evidence in Section 10.2. I grant the paper's positive contributions: the conceptual framing of LLMs as path-dependent, non-ergodic systems is reasonable, the literature review is broad, and the empirical observations of architecture-dependent behavior and path dependence are mildly suggestive. However, the central claim is a mathematical derivation, and the missing derivation is not a minor gap. Therefore I agree with the reader's REJECT verdict, with the same underlying concern identified as the most load-bearing issue.","tokens_in":29230,"tokens_out":3399,"duration_ms":35779,"concrete_test":"Check Eq. 18 directly in a real or minimal transformer. For a fixed model and context, compute the attention output A = softmax(QK^T/√d)V ∈ R^d. State the intended vector-valued meaning of Σ_i α_i (|V_t| choose i); then search over α_i ∈ [0,1] for a match to A. If the claim is per-output-dimension, require the same α_i to satisfy all d equations simultaneously. For d>1 and more than one context, this is overdetermined; if no α_i exist, Eq. 18 is false as written and Eqs. 33–36 do not follow. Alternatively, ask for an explicit derivation of Eq. 36 from the transformer definition without using Eq. 18; absent such a derivation, the central equation lacks a proof.","verdict_should_be":"REJECT","load_bearing_attack":"Equation 36 is the paper's central predictive object, but it is never derived from transformer mechanics. The only bridge is Lemma 2 (Eq. 18, Section 8.2.1), which states Attention(Q,K,V)=softmax(QK^T/√d)V ~= Σ_i α_i (|M_t| choose i). This is not an isomorphism: the left side is a d-dimensional vector (or matrix for multiple heads), while the right side is a scalar combinatorial count. No bijection, dimension-matching, or preservation of operations is specified; the justifications given—softmax outputs lie in [0,1], and attention patterns reflect combinatorial selection—are analogies, not mathematical conditions. Steps 33–36 then replace M_t with |V_t|, bundle constraints into α(i,t), and insert hierarchical maps g_l, but none of these steps repairs the undefined equality. Moreover, the experiments in Section 9 measure accuracy, attention entropy, PCA dimensionality, and path metrics; they never measure A_t or α(i,t), so they cannot validate Eq. 36. The paper itself states in Section 10.2 that 'we need additional independent evidence to validate whether this emphasis accurately represents the real dynamics of language models.' Thus the central claim—that LLM capability emergence follows Eq. 36—is unsupported at its load-bearing joint. Non-ergodicity per se is plausible, but it is not a proof of Eq. 36.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a theoretical framework, based on Stuart Kauffman's theory of the adjacent possible (TAP), for explaining emergent capabilities in large language models. It claims to prove that LLMs are non-ergodic systems and to derive a 'resource-bounded TAP equation' (Eq. 36) that governs the growth of an accessible state space through constrained combinatorial expansion. The empirical section reports analyses of three open-source models (gpt2-xl, opt-1.3B, pythia-1.4B) on the MMLU high-school mathematics subset, measuring accuracy, attention entropy, PCA-based effective dimensionality, constraint-like quantities, and path-dependence metrics, and claims support for three hypotheses about phase transitions, multiplicative constraint interactions, and path dependence.","tokens_in":29641,"tokens_out":6147,"duration_ms":62324,"significance":"The ambition is substantial: a unifying quantitative theory of emergent LLM capabilities would be a major contribution, and the paper is to be credited for making the non-ergodicity intuition precise enough to state specific hypotheses and for evaluating them on multiple open models with a public benchmark. However, the central theoretical object (Eq. 36) is never rigorously derived, the asserted attention-to-TAP isomorphism is not established, and the experiments do not measure the quantities appearing in Eq. 36. The paper also contains explicit self-acknowledged limitations, notably in §10.2 ('we need additional independent evidence to validate whether this emphasis accurately represents the real dynamics of language models') and §9.2 (90 questions per difficulty level, a simple difficulty split). Consequently, although the direction is thought-provoking, the current manuscript does not support its central claims as stated.","major_comments":[{"comment":"The asserted identification Attention(Q,K,V) ~= sum_i alpha_i (|M_t| choose i) is not an isomorphism or a derivation. The left-hand side is a d-dimensional vector (or a matrix for multiple heads), while the right-hand side is a scalar combinatorial sum; no dimension-matching, bijection, or preservation of operations is supplied. The justifications given (softmax outputs lie in [0,1]; attention patterns reflect combinatorial selection; the 1/sqrt(d) scale is a natural limit) are analogies, not mathematical conditions. Because Eqs. (33)-(36) rely on this step to replace the TAP combinatorial sum with attention, the central equation of the paper is not derived from transformer mechanics. Lemma 4 in §8.6 restates the same claim in category-theoretic language but again provides no construction or proof.","section":"§8.2.1, Lemma 2, Eq. (18)"},{"comment":"The derivation of the 'resource-bounded TAP equation' consists of asserted substitutions: M_t is replaced by |V_t|, the constraint pair (beta_i, gamma_i delta_i) is replaced by alpha(i,t) = min(beta_i, gamma_i delta_i, R(C_t)), and hierarchical functions g_l are inserted with a norm bound but without specifying their domain or codomain consistently. In particular, g_l is said to map P(V) to R^n, but its argument in Eq. (36) is a scalar sum; the equality A_{t+1} = A_t + sum_l g_l(...) therefore mixes a scalar input with a vector-valued output. The initial condition A_0 = |V_0| also conflates the accessible state space with the vocabulary size. No theorem states that this equation follows from transformer architecture, attention, or training; it is a proposed model, not a derived one.","section":"§8.5.3, Eqs. (33)-(36)"},{"comment":"The experiments measure accuracy, attention entropy, PCA effective dimensionality, and path-difference metrics, but never measure A_t or alpha(i,t) as defined in Eq. (36). The 'phase transitions' reported in Figures 1-3 and 8 are transitions in attention entropy, PCA dimensionality, and a heuristic combined constraint product; no evidence connects these to the combinatorial phase transitions of the TAP equation. The paper itself acknowledges this gap in §10.2, stating that additional independent evidence is needed to validate whether the framework's emphasis accurately represents the real dynamics of language models. As a result, the empirical sections do not constitute a test of the paper's central equation.","section":"§9-§10, experimental validation of Eq. (36)"},{"comment":"The statistical evidence is too weak for the strength of the claims. Each difficulty level contains only 90 questions, no error bars or repeated-seed variability are reported, and §9.5 uses only 30 questions. Figure 5 reports R^2 values of -0.866 and 0.500 with p-values computed on what appear to be three data points; negative R^2 and values such as R^2=1.0 with p=0.000 indicate that the reported quantity is not the standard coefficient of determination but a signed correlation coefficient, and the text does not clarify the estimator. The thresholds in Figure 8 are drawn by inspection without uncertainty quantification. These issues undermine the specific claims about 'precise positive correlation' and 'perfect negative correlation' that are used to support the multiplicative-constraint hypothesis.","section":"§9.2, §9.5, Fig. 5, Tables 4-5"},{"comment":"The abstract and Section 7.3 claim that the paper 'prove[s] that language models are actually non-ergodic systems,' but no formal proof is provided. The argument consists of asserting path-dependent transition probabilities (Eq. 3, Eq. 27) and training-induced state-space restrictions, then equating path dependence with non-ergodicity. A rigorous proof would need a precise dynamical system or stochastic process model of LLM inference, a definition of the relevant invariant measure, and a demonstration that the Birkhoff ergodic theorem or its equivalent fails in that setting. The current text offers a plausible hypothesis and empirical illustrations, not a proof. This matters because the paper's title and abstract make the proof claim central.","section":"Abstract; §7.3; §8.4.5"}],"minor_comments":[{"comment":"The heading 'Constraint interactions shape capability boundaries (H3)' appears to be a copy-paste error; the section content tests Hypothesis 3, which is about path dependence, not constraint interactions.","section":"§10.3 heading"},{"comment":"The formula for effective dimensionality, d_eff = min{k : sum_{i=1}^k lambda_i / sum_{i=1}^k lambda_i >= 0.9}, has the same index in numerator and denominator, so the ratio is identically 1; the denominator should presumably be the total variance sum_{i=1}^n lambda_i.","section":"§9.3, Eq. (51)"},{"comment":"The cross-references to equations are inconsistent: the text says architectural constraints are measured 'as in Equation 47' and later refers to 'Equation 46' and 'Equation 50', but the displayed formulas are Eqs. (52) and (53); the numbering appears to have shifted during revision.","section":"§9.4"},{"comment":"Equations (5) and (6) are displayed with identical notation, although the text says Eq. (5) uses a constant alpha and Eq. (6) uses a sequence alpha_i; the distinction is not visible, making the discussion of Kauffman's reformulation confusing.","section":"§8.4.4, Eqs. (5)-(6)"},{"comment":"There are frequent typographical and grammatical issues, including 'op-1.3B' vs 'opt-1.3B' in Table 1, 'sof tmax' in Eq. (18), 'Kaufman' vs 'Kauffman', and inconsistent citation formats (e.g., '[28,65,76,86]' vs '[87]'); a thorough editing pass is needed.","section":"Throughout"},{"comment":"The experimental setup does not state whether any code, prompts, or preprocessing scripts will be released, and it reports no hardware details, random seeds, or number of runs; a reproducibility statement would strengthen the empirical contribution.","section":"§9.2"}],"recommendation":"reject","confidential_remarks":"This is a case where the central contribution—a derived governing equation for emergent LLM capabilities—is not supported by the arguments provided, and the experiments do not test that equation. The issues are not local presentation defects: repairing them would require either a genuine derivation from transformer mechanics or a substantial reframing of the paper's claims from 'proof and derivation' to 'analogy and exploratory evidence'. I do not see a feasible revision within the current scope that would make the central claim sound. The paper does contain interesting ideas and a clear set of hypotheses, which might form the basis of a future, more carefully scoped study, but in its present form it is not publishable in a serious journal."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this paper has a genuinely suggestive analogy and almost none of the load-bearing structure needed to back it. The new thing is the explicit mapping of Kauffman's TAP equation onto LLM capability growth, culminating in Eq. 36. The paper also does a real service by collecting non-ergodicity arguments and connecting them to emergence, and it is honest enough to include a passage in Section 10.2 admitting that independent evidence is still needed.\n\nWhat is not there: the jump from attention to binomially constrained expansion. Lemma 2 (Eq. 18) states an isomorphism between softmax attention and a scalar combinatorial sum, but the left side is a vector-valued expectation and the right side is a scalar count; no bijection or dimension matching is given. The paper's own justifications—softmax outputs in [0,1], attention as selection—are analogies. Equations 33–36 then ride on that undefined equality. The conservation law in Eq. 37 and the hierarchy condition are more like consistency checks than derivations; they do not establish that transformer dynamics follow the TAP equation. The experiments use 90 questions per level, three models, no error bars, and the metrics (accuracy, attention entropy, PCA dimensions, path differences) never actually measure A_t or alpha(i,t), so they cannot validate Eq. 36 no matter how nicely the curves look. The admitted limitation in Section 10.2 is an honest caveat, but it applies to the entire empirical section, not just one emphasis.\n\nOne correction to the reader's take: the negative R² values in Fig. 5 are not impossible. R² can be negative when the fitted model is worse than the horizontal mean. That particular criticism should be dropped; the general small-sample/no-error-bar point stands.\n\nWould I cite it? Probably not in the next year; the non-ergodicity point is already in the literature it cites (Kauffman 2022, Ziemann et al. 2024). Would I bring it to reading group? Maybe, as a case study in how a suggestive framework can overreach its formal support.\n\nRecommendation: do not send to a serious referee in its present form. The central equation is unsupported at the joint, and the experiments do not instantiate it. A desk rejection with a note pointing to Eq. 18 and the unmeasured A_t would be the honest outcome.","headline":"A suggestive analogy with a promising non-ergodicity angle, but the central isomorphism is asserted and the experiments never measure the quantities in the central equation.","tokens_in":30087,"tokens_out":3251,"would_cite":false,"duration_ms":31769,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper argues that large language models are non-ergodic systems whose capabilities emerge through discrete phase transitions, formalized as a resource-bounded growth equation over token combinations.","keywords":["large language models","non-ergodic systems","theory of the adjacent possible","emergent capabilities","phase transitions","constraint interactions","path dependence","resource-bounded TAP equation"],"falsifier":"One concrete check: take a fixed context, compute the actual attention distribution, and compare the model's next-token predictions against the predictions of the TAP combination sum with any choice of constraint coefficients; if there exist contexts where attention is sharply concentrated but the combinatorially predicted spread is wide, the isomorphism fails. A second check would be to measure capability curves on a controlled benchmark and see whether jumps occur at thresholds predicted by Eq. 36 or whether the curves are continuous and scale-smooth.","tokens_in":29030,"feed_emoji":"🧩","tokens_out":6877,"duration_ms":64135,"temperature":0.7,"pith_summary":"This paper sets out to prove that large language models are non-ergodic systems: their behavior at any moment depends on the path of context and training that led there, so the usual assumption that time averages equal ensemble averages does not hold. It then argues that capabilities do not grow smoothly, but through discrete phase transitions in semantic space, and offers a mathematical framework for this growth based on the theory of the adjacent possible (TAP), adapted to computational constraints. The central object is a resource-bounded TAP equation in which the growth of the accessible state space is a constrained combinatorial expansion over the active vocabulary, with a single constraint function combining architectural, training, and contextual limits. If this is right, emergent capabilities are not mysterious accidents of scale; they are phase transitions in a constrained possibility space, with thresholds shaped by how constraints interact.","feed_headline":"LLM abilities emerge through constrained combinatorial jumps","feed_subtitle":"A resource-bounded TAP equation ties capability growth to architecture, training data, and context limits.","key_machinery":"The central object is the resource-bounded TAP equation, adapted from the theory of the adjacent possible, a framework describing how a system expands its space of possibilities by combining currently available elements under constraints. The load-bearing identity is the claimed isomorphism between softmax attention and TAP's combinatorial sum, $$\\text{Attention}(Q,K,V) \\approx \\sum_i \\alpha_i \\binom{|M_t|}{i}$$, which lets the paper carry the combinatorial expansion from the original TAP equation into transformer architectures. The equation says that the next state of the model's accessible semantic space is built by taking the current state and adding hierarchical transformations of all constrained token combinations, with the whole increment capped by a resource bound $R(C_t)$. It is this isomorphism, rather than any empirical regularity, that carries the derivation from the original TAP equation to Eq. 36.","core_discovery":"The paper's central claim is that language models are non-ergodic and that their capability emergence is governed by the resource-bounded TAP equation $$A_{t+1} = A_t + \\sum_{l=1}^{L} g_l\\left( \\sum_{i=1}^{|V_t|} \\$\\alpha$(i,t) \\binom{|V_t|}{i} \\right)$$ where $A_t$ is the accessible semantic state space, $|V_t|$ the active vocabulary, $\\binom{|V_t|}{i}$ counts token combinations, $\\alpha(i,t)$ fuses architectural, training, and contextual constraints multiplicatively, and $g_l$ are hierarchical functions that map token combinations into semantic space while bounded by computational resources. The paper claims to prove that this equation follows by mapping the token-prediction probability space onto the TAP combinatorial structure, treating the softmax attention mechanism as an isomorphism to combinatorial selection. It reports experiments on three open models that show phase-transition-like jumps in performance, multiplicative constraint effects with distinct thresholds, and path-dependent problem-solving behavior, which it reads as empirical support for the framework.","pith_inferences":["Editorial inference: if Eq. 36 is right, transition thresholds could in principle be estimated before training from the architecture's resource bounds and the vocabulary size, making at least some emergent capabilities predictable at design time.","Editorial inference: the attention-to-TAP isomorphism suggests that changing attention variants (sparse, linear, sliding-window) changes the effective constraint profile $\\alpha(i,t)$, which predicts that different attention designs will shift phase-transition points; this is a testable extension the paper does not run.","Editorial inference: the framework implies a new scaling variable, namely the growth rate of the accessible state space under constraints, rather than parameter count alone; comparing models on effective-dimensionality growth could be a sharper predictor of benchmarks than raw scale."],"forward_implications":["Capability emergence should appear as discrete jumps at critical points rather than smooth improvements, so continuous scaling curves would indicate that a constraint threshold has not been reached.","Relaxing any single constraint (architecture, training data, or context) will produce only limited gains while the other constraints are binding; improvements need balanced relaxation of all three.","Because the system is path-dependent, the same prompt presented in different orders can lead to measurably different solution trajectories, so evaluation and alignment work must track trajectories rather than averages.","Architecture design can target phase-transition thresholds, since the hierarchical functions $g_l$ and the resource bound $R(C_t)$ set the limits of capability growth and can make emergence more predictable."],"supporting_citations":[{"why":"Supplies the theory of the adjacent possible and the original combinatorial equation for possibility-space expansion.","marker":"[67]"},{"why":"Presents TAP as a framework for non-ergodic evolution and constrained exploration, the conceptual bridge to language models.","marker":"[68]"},{"why":"Provides the reformulated TAP equation with a sequence of constraint constants, which the paper adapts into the integrated constraint function.","marker":"[29]"},{"why":"Analyzes combinatorial innovation dynamics that motivate the resource-bounded growth process in Eq. 36.","marker":"[116]"},{"why":"Defines the softmax attention mechanism that the paper claims is isomorphic to TAP's combinatorial sum.","marker":"[121]"},{"why":"Documents the sudden emergence of capabilities at scale that the framework is built to explain.","marker":"[126]"},{"why":"Argues that non-ergodic systems construct and expand their own phase space, grounding the non-ergodicity claim.","marker":"[65]"},{"why":"One of the three open models used in the experiments, contributing evidence for phase-transition behavior.","marker":"[99]"},{"why":"One of the three open models used in the experiments, providing the path-dependence and constraint measurements.","marker":"[17]"},{"why":"Supplies the MMLU high-school mathematics benchmark on which capability jumps, constraints, and path dependence are measured.","marker":"[56]"}],"fun_headline_variants":["LLM emergence: non-ergodic paths to new capabilities","Why LLMs gain abilities: non-ergodic, constrained jumps","Emergent LLM skills explained by non-ergodic math","Capability leaps in LLMs from non-ergodic dynamics","Non-ergodic framework shows how LLM abilities emerge"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that softmax attention really is a combinatorial selection process, so an attention computation can be rewritten as a constrained sum over token combinations; if that isomorphism fails, Equations 33 through 36 do not follow.","fun_headline_variants_meta":{"raw":{"variants":["LLM emergence: non-ergodic paths to new capabilities","Why LLMs gain abilities: non-ergodic, constrained jumps","Emergent LLM skills explained by non-ergodic math","Capability leaps in LLMs from non-ergodic dynamics","Non-ergodic framework shows how LLM abilities emerge"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000626,"raw_usage":{"total_tokens":2864,"prompt_tokens":878,"completion_tokens":1986,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":494,"completion_tokens_details":{"reasoning_tokens":1900}},"tokens_in":494,"tokens_out":1986,"duration_ms":12835,"temperature":1.0,"reasoning_tokens":1900,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T22:23:03.114221+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"One concrete check: take a fixed context, compute the actual attention distribution, and compare the model's next-token predictions against the predictions of the TAP combination sum with any choice of constraint coefficients; if there exist contexts where attention is sharply concentrated but the combinatorially predicted spread is wide, the isomorphism fails. A second check would be to measure capability curves on a controlled benchmark and see whether jumps occur at thresholds predicted by Eq. 36 or whether the curves are continuous and scale-smooth.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the theory of the adjacent possible and the original combinatorial equation for possibility-space expansion."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Presents TAP as a framework for non-ergodic evolution and constrained exploration, the conceptual bridge to language models."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Analyzes combinatorial innovation dynamics that motivate the resource-bounded growth process in Eq. 36."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the softmax attention mechanism that the paper claims is isomorphic to TAP's combinatorial sum."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"One of the three open models used in the experiments, contributing evidence for phase-transition behavior."}],"review_version":1}