{"id":"a2d456fb-7960-4b14-b8a0-9f1cf1c212f2","arxiv_id":"2511.06697","paper_version":2,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"In the random Generalized Lotka-Volterra model, a stable state's likelihood of being reached increases sharply with its biomass, and a coarse-grained theory predicts state likelihoods using only biomass and average diversity.","lead":"Complex ecosystems can settle into many different stable states; this paper shows in a standard mathematical model that the probability of ending up in a given state is set by its total biomass, which acts as a collective self-inhibition. If this pattern transfers to real ecosystems, ecologists could predict likely ecological outcomes from two easy measurements instead of the full web of species interactions.","discovery_kind":"first_principles","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The uncontrolled block/mean-field replacement of the disordered interaction matrix is the load-bearing step: Eq. (5) is derived for non-overlapping blocks with D > A_eff, a condition not checked for random-matrix states, and the approximation is admittedly worst for the rare states that the 'entire","rationale":"The reader's weakest assumption identified the same load-bearing premise: the mean-field/block replacement of the disordered interaction matrix, compounded by the quasi-equilibration that drops the fluctuation term in Eq. (D8). My reading of the manuscript confirms that this is the single point on which the central quantitative claim depends. The monodominant and block-model results are exact or well-controlled, but the extension to disordered matrices is explicitly 'uncontrolled' and shows its largest deviations precisely for rare states. I add a sharper technical observation: the derivation of Eq. (5) assumes positive dressed prefactors χ_α; after replacing D by μ, this requires μ > 1/B_α for every state, a condition the paper never verifies. If it fails, Eq. (5) is formally outside its derivation. This does not overturn the paper's main qualitative finding—biomass is a strong predictor of likelihood—nor the honest reporting of limitations. But it does mean the headline claim of predicting 'the entire landscape' using only biomass and diversity is not fully established. The reader's CONDITIONAL verdict already captures this. No verdict change is needed; the requested conditions (tighten the abstract, characterize the mean-field approximation, release code/data) remain appropriate. I propose a concrete computational check that would determine whether the uncontrolled fluctuation term is truly responsible for the rare-state overprediction or whether the overlap structure is the limiting factor.","tokens_in":22568,"tokens_out":9858,"duration_ms":106877,"concrete_test":"Repeat the Fig. 3 analysis on the same random matrix A, but for every observed state α record χ_α = μ − 1/B_α and the mean over its basin of the neglected fluctuation term F_α = (⟨A⟩_α − 1) Σ_{i∈α}(N_i − B_α/L_α)^2 / B_α from Eq. (D8), where ⟨A⟩_α is the average in-state interaction. Then compare predicted vs. observed likelihoods restricted to states with χ_α > 0 and small |F_α| against the full set. If the restricted subset shows substantially better agreement, the uncontrolled approximation is the cause of the rare-state overprediction and the 'entire landscape' claim must be weakened; if agreement is no better, the failure stems from the overlapping-block structure (Appendix E) rather than the dropped fluctuation term.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The quantitative prediction of the full likelihood landscape rests on the derivation of Eq. (5) from the block model in Appendix D. In the disordered case this requires three uncontrolled substitutions: replacing every inter-state inhibition D by the global mean μ, replacing each state's diversity L_α by a common average L, and dropping the intra-block fluctuation term in Eq. (D8) as 'rapidly dissipating.' The authors themselves call this an 'uncontrolled approximation' (Appendix G). This is not peripheral: the same regime where the approximation is known to fail—rare states with a dominant species and atypically strong interactions—is exactly where Fig. 3b shows order-of-magnitude overprediction. Moreover, the derivation of Eq. (5) presumes χ_α = D − A_eff_αα > 0; after the replacement D→μ, this becomes μ > 1/B_α, a condition not checked for any observed state. If a state with B_α < 1/μ exists, its dressed initial condition is negative and the Gaussian integral (5) is applied outside its domain of validity. Since the 'entire landscape' claim depends on Eq. (5) being a genuine derived consequence of the GLV dynamics, not a fitted effective model, this uncontrolled mean-field step is the most load-bearing weakness.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies the multi-stable phase of the random Generalized Lotka-Volterra (GLV) model with strong interactions, asking what determines the relative likelihood (basin volume fraction) of each stable state. Numerically, for a fixed random interaction matrix, state likelihoods span orders of magnitude and increase sharply with the state's equilibrium biomass. To explain this, the authors first solve exactly a monodominant model in which each stable state contains a single species: the winner is the species with the largest dressed initial condition χ_i N_i(0), and the likelihood is given by Eq. (4). They then treat each stable state as an effective block with effective self-inhibition equal to the inverse of the block's equilibrium biomass, obtaining the block-likelihood formula Eq. (5) as the probability that one Gaussian-dressed block initial condition is the maximum. Finally, they apply Eq. (5) to random disordered matrices by replacing the inter-state inhibition D with the global mean interaction μ and using a common block size equal to the mean state diversity. This predicts simulated likelihoods for most states but systematically overpredicts the rarest, low-biomass states.","tokens_in":22935,"tokens_out":4069,"duration_ms":43156,"significance":"If the random-matrix extension were fully justified, the paper would establish a striking and practically useful result: ecological basin sizes are controlled by macroscopic state properties (biomass, diversity, and global mean interaction) rather than by the detailed interaction matrix. The monodominant solution Eq. (4) and the block-model integral Eq. (5) are genuine analytic contributions, and the block-model simulations validate them in their intended setting. The paper also contains a clear numerical demonstration that the biomass-likelihood relationship is robust across system sizes, asymmetry, carrying-capacity heterogeneity, and initial-condition distributions. The weakness is that the step from perfect block models to disordered random matrices relies on approximations the authors themselves describe as 'uncontrolled' (Appendix G); since this step carries the paper's central claim about random ecosystems, the strength of the final conclusion is currently disproportionate to the evidence.","major_comments":[{"comment":"Equation (5) is derived for a block-structured matrix with a single inter-state inhibition D, and the dressed prefactor is χ_α = D − A^eff_αα. In the random-matrix application this becomes χ_α = μ − 1/B_α. The derivation requires χ_α > 0 for every block; otherwise the Gaussian 'dressed initial condition' has negative mean and the maximum-of-independent-Gaussians picture in Eq. (5) is no longer the correct basin-volume integral. The manuscript does not check whether B_α > 1/μ holds for the states to which Eq. (5) is applied, nor does it state the condition as a validity criterion. This is load-bearing because Eq. (5) is the quantitative prediction underlying the 'entire landscape' claim.","section":"Appendix G, Eq. (5)"},{"comment":"The mean-field approximation fails precisely in the regime the paper claims to predict: rare, low-likelihood states are overpredicted by orders of magnitude (the gray region in Fig. 3b, and the text notes these states can be 10^5 times less probable than the most likely state). The authors acknowledge this, but the abstract and main text nevertheless state that the model 'accurately predicts the entire landscape.' Since the overprediction is systematic in the tail of the likelihood distribution, the central claim needs to be qualified, and the manuscript should provide an error analysis or an a priori criterion for when the block/mean-field description applies.","section":"Fig. 3b and Appendix G"},{"comment":"The reduction from the exact block dynamics Eq. (D6) to the effective single-unit form Eq. (D8) drops the intra-block fluctuation term (A_αα − 1) Σ(N_i − N̄_α)^2 / B_α, asserting it 'rapidly dissipates.' No timescale separation is demonstrated, and for the block sizes used (L ≈ 4–6) the fluctuations are not obviously negligible at early times. In the random-matrix setting, where states overlap and the block structure is only approximate, this uncontrolled truncation is a second independent approximation. The good agreement in Fig. S12 shows the truncation is harmless in the clean block model, but it does not establish that it is harmless for the disordered matrices that the main text targets.","section":"Appendix D, Eq. (D8)"}],"minor_comments":[{"comment":"The phrase 'accurately predict the entire landscape' is stronger than what Fig. 3b and Appendix G support. Suggest revising to 'the likelihoods of the vast majority of states' or adding a quantified statement about the low-likelihood tail.","section":"Abstract and main text"},{"comment":"The hyperbolic fit log(p) ∝ (B* − B)^−1 is introduced in the main text, but B* is not defined in the caption or in the surrounding text. Since B* is a fitted parameter, its role should be clarified so readers do not mistake the fit for a parameter-free prediction.","section":"Fig. 1c caption"},{"comment":"Typos: 'biomass-likelihood relationsip' should be 'relationship'; 'Radaumethod' should be 'RADA method' or similar; 'uninvasible' should be 'uninvadable.'","section":"Appendix B"},{"comment":"The text refers to 'Fig. S12b–c' when discussing the block-model comparison; the figure has panels (a) and (b). Please correct the cross-reference.","section":"Appendix D"}],"recommendation":"major_revision","confidential_remarks":"The paper has a solid analytic core: the monodominant solution and the block-model likelihood integral are correct and well validated. The main issue is that the headline claim about disordered random ecosystems rests on an acknowledged 'uncontrolled' mean-field substitution that is not derived and is demonstrably inaccurate for the rare states that the abstract's 'entire landscape' language explicitly covers. This is fixable by reframing the random-matrix application as a validated effective model with stated validity conditions, and by quantifying the failure at low likelihoods. I would not reject the paper, but the current version overclaims."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's my read of arXiv:2511.06697.\n\nThe novel core is real. The monodominant likelihood formula (Eq. 4) is exact, and the block-model integral (Eq. 5) follows from a clean extreme-value argument. Importantly, the biomass–likelihood relation is derived, not fitted: in the monodominant model the winner is the species with the largest dressed initial condition, and p_i increases monotonically with χ_i = D − A_ii, with biomass B_i = 1/A_ii. That is a genuine explanatory result, and it generalizes Ref. [44] to a broader class. The paper is also honest: it reports the overprediction for rare states and labels the mean-field replacement “uncontrolled” in Appendix G. That honesty is a credit, and it defines the ceiling.\n\nThe main soft spot is the load-bearing mean-field step. For disordered matrices, Eq. (5) is applied after replacing the inter-state inhibition D by the global mean μ and the state-specific diversity L_α by the mean L. The derivation of Eq. (5) requires χ_α = D − A_eff_αα > 0; after the substitution this becomes μ > 1/B_α. The paper never checks this condition for the observed states. If any state has B_α < 1/μ, its dressed initial condition is negative and the Gaussian integral is outside its domain. That is a testable, specific gap. The same approximation is what causes the order-of-magnitude overprediction for rare states (Fig. 3b), which the authors acknowledge. So I would not call the “entire landscape” claim established; I'd call it a well-supported conjecture within the random GLV model class, with a clear failure mode.\n\nMinor issues: the hyperbolic fit log p ∝ (B* − B)^−1 uses a fitted B*, so it's not a parameter-free prediction; state-specific diversity is not predictive, contrary to the abstract's phrase “biomass and species diversity”; and no code or data are shipped, which makes numerical verification heavier than it should be. None of these break the core, but they matter for the packaging.\n\nWho's this for? People working on multistability in ecological models and statistical physics of Lotka–Volterra dynamics. The exact monodominant result and the block integral will be useful as tools. I'd send it to a serious referee. The referee should ask for a check of the χ > 0 condition, a quantitative characterization of when the mean-field replacement fails, and release of simulation code/data. With those, the paper would be stronger than most in this area.","headline":"A genuinely derived biomass–likelihood relation in random GLV, with an honest but uncontrolled mean-field step that should be flagged in the abstract and checked against the χ>0 condition.","tokens_in":23416,"tokens_out":1938,"would_cite":true,"duration_ms":17842,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["92D25","92D40","60G70"],"pacs":[],"model":"deepseek-v4-flash","headline":"In the random Generalized Lotka–Volterra model, the likelihood of reaching a stable ecosystem state rises sharply with that state's total biomass, and the entire landscape of state likelihoods can be predicted from macroscopic observables a","keywords":["generalized Lotka-Volterra","multistability","attractor basins","state likelihood","biomass","emergent self-inhibition","coarse-graining","random interaction matrices"],"falsifier":"Run a large number of simulations from random initial conditions on a fixed GLV matrix in the multistable regime, and compare each observed stable state's frequency with Eq. (5) using only its biomass and mean diversity; if two states with equal biomass systematically differ in likelihood, or if the rank order is reversed for a substantial fraction of states, the claim is falsified. The paper's own data already show the formula overpredicts the rarest states, so the falsifier should target mid- and high-likelihood states where the match is claimed to be good.","tokens_in":22383,"feed_emoji":"🌿","tokens_out":4694,"duration_ms":45242,"temperature":0.7,"pith_summary":"This paper asks why some stable states of a species-rich ecosystem are reached far more often than others. In the random Generalized Lotka–Volterra model, where the same species pool can settle into many alternative states, the authors show that a state's likelihood increases sharply with its total biomass, often by orders of magnitude. They explain this by coarse-graining each stable state into a single effective unit whose self-inhibition is the inverse of its biomass, and derive an analytic formula for state likelihoods that depends only on each state's biomass, the mean interaction strength, and average species diversity. The formula predicts simulated basin sizes for hundreds of states in disordered ecosystems without detailed knowledge of the interaction matrix. The result matters because biomass and diversity are measurable in real communities, so state probabilities could be estimated experimentally.","feed_headline":"A state's biomass predicts its odds of winning","feed_subtitle":"In random species-interaction models, only biomass and diversity are needed to predict the full landscape of stable states.","key_machinery":"The central device is the block coarse-graining of stable states. Each observed stable state is treated as an effective species with self-inhibition equal to the inverse of its equilibrium biomass, and with an initial abundance that is the sum over its member species, approximated as Gaussian by the central limit theorem. Dressed initial conditions then determine the winner, and Eq. (5) computes each state's likelihood as an extreme-value probability. This machinery converts a high-dimensional dynamical problem into a one-dimensional competition among a few effective units.","core_discovery":"The central claim is that basin size—the fraction of random initial conditions leading to a given stable state—is controlled by an emergent, state-level self-inhibition that equals the inverse of the state's equilibrium biomass. Because high-biomass states have low self-inhibition, they grow faster and outcompete other states, making them sharply more likely outcomes. This is made quantitative in a block model: each state behaves as a Gaussian-dressed initial abundance, and the probability that a state wins is the probability that its dressed initial abundance is the largest, giving Eq. (5). The same formula predicts the observed likelihoods of hundreds of stable states in random GLV matrice","pith_inferences":["If the biomass–likelihood relation generalizes beyond random matrices to the structured microbial communities grown from parallel enrichments, then simple cell-count and 16S data could rank alternative stable states without long time-series.","The paper's surprising finding that using each state's own diversity worsens predictions suggests that overlapping species renormalize effective block size; a testable extension would be to predict the optimal effective block size from overlap statistics between states.","The hyperbolic form of the biomass–likelihood relationship is shown to be only one possible shape; the curvature depends on the distribution of self-inhibition, so deviations in real data could be used to infer the underlying inhibition structure.","Outside the symmetric, strong-interaction setting, nonreciprocal interactions can lead to chaos, where the very notion of a stable state and its basin needs replacement; whether a biomass-biased occupation measure survives in chaotic regimes is an open question the paper does not resolve."],"forward_implications":["In the multistable region of the random GLV model, replicate communities started from random initial conditions should show a strongly skewed distribution of outcomes, with high-biomass states orders of magnitude more frequent than low-biomass ones.","State probabilities can be estimated from macroscopic measurements—total biomass and species richness per state—without knowing the species-level interaction matrix, which is usually unmeasurable.","The biomass–likelihood ordering survives moderate interaction asymmetry, variation in carrying capacities, changes in system size, and different initial-condition sampling schemes.","The analytic formula for monodominant and block-structured models is exact, giving a benchmark for studying more complex or overlapping state structures.","Because high biomass corresponds to low self-inhibition and faster growth, transient growth rates rather than detailed equilibration dynamics may be enough to rank outcomes."],"fun_headline_variants":["Self-inhibition predicts which ecosystem state wins","Biomass of a state determines its odds of occurring","High biomass states dominate in random ecosystems","Emergent self-inhibition shapes stable state odds","Inverse biomass predicts attractor likelihood in ecosystems"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The disordered interaction matrix is replaced by an effective block model with a single mean inter-state inhibition and a common average block size, while fluctuations within each state are assumed to dissipate quickly; the paper itself calls this an 'uncontrolled' mean-field approximation.","fun_headline_variants_meta":{"raw":{"variants":["Self-inhibition predicts which ecosystem state wins","Biomass of a state determines its odds of occurring","High biomass states dominate in random ecosystems","Emergent self-inhibition shapes stable state odds","Inverse biomass predicts attractor likelihood in ecosystems"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":9.6e-05,"raw_usage":{"total_tokens":834,"prompt_tokens":730,"completion_tokens":104,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":474,"completion_tokens_details":{"reasoning_tokens":44}},"tokens_in":474,"tokens_out":104,"duration_ms":1960,"temperature":1.0,"reasoning_tokens":44,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-03T23:13:55.473903+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run a large number of simulations from random initial conditions on a fixed GLV matrix in the multistable regime, and compare each observed stable state's frequency with Eq. (5) using only its biomass and mean diversity; if two states with equal biomass systematically differ in likelihood, or if the rank order is reversed for a substantial fraction of states, the claim is falsified. The paper's own data already show the formula overpredicts the rarest states, so the falsifier should target mid- and high-likelihood states where the match is claimed to be good.","supporting_citations":[],"review_version":1}