{"id":"6b6b15db-e82c-4abd-b0bc-92418ed84c41","arxiv_id":"2509.23212","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"A self-labeling linear classifier can bootstrap alignment with the latent structure of Gaussian data when regularization is strong enough, and label-exchange between agents can produce collective consensus.","lead":"This paper studies a simple classifying agent that labels fresh data with its own predictions and then trains on those labels, with no ground truth or reward. In simulations and statistical mechanics calculations, strong regularization makes this self-consistency loop lock onto the true structure of the data, and interacting agents can reach consensus through labels alone.","discovery_kind":"first_principles","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Fresh-stream assumption is load-bearing: finite or reused data changes the dynamics, and the paper's own data-constraint experiments show the predicted phase boundary is tied to the infinite fresh-stream idealization.","rationale":"The reader's weakest_assumption identifies the fresh-batch idealization as the key condition that makes the replica calculation tractable. The paper openly acknowledges this assumption and the SM's own experiments show that data reuse changes the dynamics, including a protocol (uniform splitting) where replicators fail. This makes the assumption genuinely load-bearing for the scope of the central claim: the 'genuine phase transition' is proven for an infinite stream, not for finite-data learning. However, the paper does not claim the transition survives all finite-data protocols; it presents the fresh-stream model as the minimal theoretical setting. The binary single-agent result is internally consistent, the replica prediction matches simulations for the stated protocol, and the phase boundary in Fig. 2(b) is not parameter-fitted. The other reader concerns (missing code/data, omitted error bars, unfair supervised comparison, unfinished C>2 replica) are real but secondary to the strongest_claim. Therefore the verdict should remain CONDITIONAL: the central theoretical result stands under its stated idealization, but the article should quantify the finite-data dependence before the mechanism is presented as broadly applicable.","tokens_in":23032,"tokens_out":23050,"duration_ms":256220,"concrete_test":"Using the SM's partial-resampling protocol (Protocol 3) on the C=2 Gaussian mixture with a fixed dataset of size D, sweep the resampling fraction X from 0% to 100% and measure the empirical phase boundary (critical lambda as a function of sigma) where the nontrivial fixed point appears. Repeat for N=500, 1000, 2000 to check finite-size convergence. If the critical lambda diverges or the boundary shifts by O(1) as X approaches 0, the infinite fresh-stream idealization is load-bearing; if the boundary converges to the analytical r=1 line for all X>0, the concern is substantially weakened.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The replica reduction to the scalar map m_{t+1}=f(m_t) in Eq. (5) and the stability line r=1 in Eq. (6) rely on drawing a fresh, independent batch at every timestep. The SM states this explicitly: 'Since we assume fresh data at each timestep, we conveniently do not have to deal with inter-step cross correlations.' If batches are finite or reused, the dynamics is no longer Markovian in m_t alone, and the SM's own Fig. S7 shows the outcome changes: under uniform splitting (0% resampling) the system typically fails to reach the structured state that the fresh-stream theory predicts for the same parameters. Alternating batches and partial resampling alter the alignment trajectories. Thus the central claim of a spontaneous learning phase transition is established only for an idealized infinite data supply; its domain of validity for finite data is not quantified. This is the least secure link between the model and the broader claim that self-consistency alone can bootstrap latent-structure extraction.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies a minimal unsupervised learning loop in which a linear classifier labels fresh samples from a symmetric two-component Gaussian mixture and then updates its weights by minimizing cross-entropy on its own labels with L2 regularization. In the thermodynamic limit and under a replica symmetric ansatz, the dynamics is reduced to a scalar recursion for the magnetization m_t, Eq. (5). The stability of the trivial fixed point, Eq. (6), yields a phase boundary r=1 in the (σ, λ) plane: above a critical regularization strength the agent spontaneously develops nonzero alignment with the latent centroid, despite never seeing ground-truth labels. The authors validate the boundary against simulations without fitted parameters, compare performance with supervised baselines, extend the phenomenology numerically to C>2 Gaussian mixture classes, and study a population of agents exchanging labels, finding consensus regimes with distinct O(M^2) and O(M) timescales.","tokens_in":23272,"tokens_out":3720,"duration_ms":35010,"significance":"If the central claim holds, the paper provides a clean statistical-mechanics example of a genuine learning phase transition in a self-supervised, teacher-free setting: self-consistency plus a simplicity bias is sufficient to bootstrap latent-structure extraction. The strength of the work is that the binary case is derived from the model definition via a replica calculation, with no fitted parameters in the phase boundary, and the theoretical transition line is checked against direct simulations. The data-constraint experiments in the SM, while highlighting an important limitation, are a useful and honest addition. The population part is more exploratory but connects the model to collective-learning and consensus phenomenology.","major_comments":[{"comment":"The scalar recursion and the stability line r=1 rely on drawing a fresh, independent batch at every timestep; the SM states this explicitly ('Since we assume fresh data at each timestep, we conveniently do not have to deal with inter-step cross correlations'). This is not a harmless technical convenience: SM Fig. S7 shows that under uniform splitting (0% resampling) the dynamics typically fails to reach the structured state that the fresh-stream theory predicts for the same parameters, and alternating/partial-resampling protocols change the trajectories. The phase transition is therefore established only for the infinite fresh-stream protocol. The paper should either quantify a finite-data/reuse regime in which the transition survives, or explicitly restrict the main claims to this idealization. As written, the abstract's 'without explicit guidance' and the general framing overstate the","section":"Theoretical analysis, Eq. (5)-(6), and SM I.D"},{"comment":"The phase boundary is derived under the Replica Symmetric (RS) ansatz. The paper justifies RS by 'the convex nature of the optimization problem', but the replicated free entropy involves the sign function and the iterated dynamics is not convex; RS is an assumption, not a consequence of per-step convexity. The excellent numerical agreement for m* and for the boundary is encouraging, but it does not by itself rule out replica-symmetry-breaking corrections to the free energy or to the stability condition. The authors should state more carefully that the boundary is obtained under RS and, if possible, provide a stability check of the RS saddle point or a direct test of RS order parameters.","section":"Theoretical analysis, replica computation (SM II.B.1)"},{"comment":"The abstract claims that 'interaction reshapes the learning phase boundary', but the population section does not compute a learning phase boundary. It reports the evolution of π_t and φ_t, the η=1/2 crossover, and O(M^2) vs O(M) timescales, but no boundary in (σ, λ, η) is derived or mapped. The claim that interaction reshapes the phase boundary is therefore unsupported by the presented evidence. Either compute the population-level boundary (even approximately) or soften the claim to describe the observed consensus/coordination crossover.","section":"Collective learning in interacting agents and Abstract"}],"minor_comments":[{"comment":"The caption says 'for increasing noise levels α = 0.15, 0.45, 0.75'; α is the load P/N, not a noise level. The variables should be labeled consistently, and the color code (λ? training time?) is unclear.","section":"SM I.A, Fig. S1 caption"},{"comment":"The average NMI is not self-averaging: footnote [38] states that low averages can be dominated by rare lucky initializations, and error bars are omitted from Fig. 2(c). Since this is the only quantitative evidence for the multiclass claim, the distribution or best-case curves should be shown alongside the average.","section":"Main text, Fig. 2(c) and footnote [38]"},{"comment":"The figure compares protocols but the marker for the replica solution is described only as starting 'from the experimental mean alignment at t=0'. Clarify whether the replica curve is the deterministic recursion or an average, and report variability across the 50 runs.","section":"SM I.D, Fig. S7"},{"comment":"The notation r = lim_{m_t -> 0} f'(m_t; α, σ, λ) is fine, but f itself is not given in closed form and is only defined implicitly through the replica saddle-point calculation. A sentence pointing to the explicit saddle-point equations in the SM would help reproducibility.","section":"Eq. (6)"}],"recommendation":"major_revision","confidential_remarks":"The central binary-case result is solid within the fresh-stream idealization, but the paper's broader framing and the population-phase-boundary claim go beyond what is demonstrated. The authors may want to reposition the population section as a numerical exploration rather than a phase-boundary result. The citation of the authors' own prior work (refs 36 and 51) is appropriate and not excessive."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The thing to know: this paper gives a minimal linear-classifier model where a self-consistency loop (predict on fresh data, train on your own labels) is reduced to a scalar recursion m_{t+1}=f(m_t). The phase boundary r=1, where the trivial fixed point destabilizes and a nonzero magnetization appears, is derived via a replica calculation and matches simulations without any fitted parameters. For the binary, single-agent case, the central claim—that a simplicity bias can bootstrap latent-structure extraction without labels—holds up. That is a real, transferable insight, and the connection to functional replicators is apt.\n\nThe paper does several things well. The replica computation is standard but carefully adapted to the turnover setting, and the RS ansatz is justified by the convex optimization at each step. The long-time correlation analysis showing that only the signal component persists is a nice touch. The population extension, where agents exchange only labels, is qualitatively interesting and shows a clear Moran-process-like consensus crossover. The writing is honest: the authors flag the unfair supervised baseline in their own supplement, and footnote 35 correctly notes Takahashi's toolbox could be adapted.\n\nThe soft spots are real but not fatal. The load-bearing assumption is the fresh independent batch at every timestep. The replica derivation explicitly avoids inter-step cross correlations, and the dynamics is Markovian in m_t only under that assumption. The paper's own Figure S7 shows that with finite or reused data—uniform splitting, alternating batches, partial resampling—the outcome changes; in some regimes the predicted structured state is not reached. That means the phase transition is established for an idealized infinite fresh-stream, and the domain of validity for finite data is left unquantified. This is the main caveat a referee should press on. The multiclass and population sections are mostly numerical, the C>2 replica computation is left unfinished, and no code or data are shipped. These are limitation-level issues, not errors. The multiclass NMI averages also mix rare successes and typical failures; the authors acknowledge this in a footnote but don't show error bars, which understates the multimodality.\n\nWho is this for? Stat-mech of learning people and anyone thinking about self-training, pseudo-labeling, or emergent consensus. The idealized setting means it won't directly answer practical questions, but as a minimal exactly solvable model it deserves attention. I'd send it to a serious referee; the binary case is solid enough for publication after the finite-data caveat is moved from a supplement aside to a clearly stated limitation, and ideally with some quantitative handle on how the phase boundary degrades under reuse.\n\nFor your reading group: worth a slot, especially if you discuss the fresh-stream assumption and how much it matters.","headline":"A clean analytically tractable phase transition for label-only self-consistent learning, held together by a fresh-data idealization that the paper states plainly but does not quantify.","tokens_in":23719,"tokens_out":1005,"would_cite":true,"duration_ms":11489,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["68T05","62H30","82B26"],"pacs":["05.20.-y","64.60.-i","07.05.Mh"],"model":"deepseek-v4-flash","headline":"A minimal agent that trains on its own predicted labels, with no ground truth or reward, can spontaneously learn the true latent structure of its environment, and this capacity appears as a sharp phase transition.","keywords":["functional replicators","self-consistency learning","unsupervised learning","phase transition","replica method","Gaussian mixture","collective learning","label exchange"],"falsifier":"Initialize agents at a small nonzero magnetization m0 in a parameter region where the replica calculation predicts r<1 (low λ, high σ) and confirm the magnetization relaxes to zero; and in a region where r>1 (high λ, low σ) confirm a small seed grows to m*>0. The replica prediction gives exact r in the thermodynamic limit, so a discrepancy between the measured slope of mt+1 vs mt at mt=0 and the replica r would settle the matter directly.","tokens_in":22947,"feed_emoji":"🔁","tokens_out":7533,"duration_ms":62325,"temperature":0.7,"pith_summary":"This paper claims that a classifying agent can extract genuine latent structure from data without any ground-truth labels, rewards, or teaching signal, purely by repeatedly labeling fresh data and then training on its own labels under a simplicity bias. The authors call the resulting self-sustaining labeling regimes 'functional replicators' and show, with replica-method statistical mechanics, that their onset is a phase transition: when the regularization is strong enough, the zero-magnetization fixed point becomes unstable and the agent aligns with the true data centroids. In a Gaussian-mixture environment a single agent reaches the same mutual information as a supervised classifier in most noise regimes, sometimes better. A population of such agents exchanging labels can reach consensus, and interaction can either enable or suppress spontaneous learning depending on a stubbornness parameter. If correct, this provides a minimal principled setting where learning emerges from self-consistency alone, without selection or external objectives.","feed_headline":"Self-made labels can unlock hidden structure in data","feed_subtitle":"A minimal agent that teaches itself from its own predictions spontaneously learns true clusters—no ground truth, no reward","key_machinery":"The central object is the magnetization mt = Wt·μ/||Wt||, the agent's alignment with the environment's centroid. The argument is carried by the scalar recursion mt+1 = f(mt; α, σ, λ), obtained via a replica calculation (replica-symmetric ansatz, β→∞) of the convex self-consistency loss with fresh data at each step; and by the stability condition r = lim_{mt→0} f'(mt), whose crossing of 1 marks the learning phase transition. The regularization λ acts as a simplicity bias that concentrates weights along informative directions, and the fresh-data assumption removes inter-step cross correlations, making the one-step replica computation exact in the thermodynamic limit. The population extension u","core_discovery":"On the paper's own terms: a linear classifier updated to be self-consistent with its own hard labels on a fresh batch of Gaussian-mixture data, plus an L2 regularization, has its whole high-dimensional dynamics reduced to a scalar recursion for the magnetization mt = (Wt·μ)/||Wt||. The paper shows that the trivial fixed point m*=0 loses stability when the derivative r = lim_{mt→0} f'(mt) crosses 1, which happens as the regularization λ grows relative to the noise σ; the stable fixed point then has m*>0, meaning the self-made labels are correlated with the environment's true centroids. Long-time correlations between agent generations equal (m*)², so the only persistent memory is the signal pr","pith_inferences":["Testable extension: for a fixed finite dataset, the phase boundary should renormalize as a function of the resampling rate, shrinking toward zero as the rate goes to zero; the paper's own 0% resampling result suggests this, and a quantitative finite-budget replica theory with temporal correlations would settle it.","An inference from the η=1/2 crossover: decentralized learning systems may have an optimal peer-distrust level where consensus is fastest even though individual learning is slowest—a possible design principle for swarm or federated learning that the paper does not state.","Because the only persistent channel is the signal projection, the model predicts that any self-consistency algorithm with strong regularization will discard all spurious directions; if a real self-distillation method shows persistent non-signal correlations, the model's minimal description is incomplete."],"forward_implications":["If correct, self-supervised bootstrap is a genuine phase transition: there is a sharp boundary in (σ, λ) separating agents that never align from agents that reach a stable informative labeling; small initial fluctuations get amplified.","The steady-state performance of the unsupervised turnover is comparable to supervised training (sometimes superior at intermediate noise and strong regularization), so self-consistency alone can be as good as ground truth in simple environments.","Functional replicators are the only long-time attractors in the binary case: the only persistent memory channel is the signal projection, since orthogonal components decorrelate exponentially.","In populations, label exchange alone (without access to other agents' weights) can drive consensus; interaction can enable learning in regimes where isolated agents would fail, but can also suppress it for high stubbornness.","Multiclass environments support information-bearing replicators, but full latent-structure recovery becomes harder as C grows; output overparameterization (K > C) increases the chance of full coverage."],"fun_headline_variants":["Agent learns true clusters from its own labels","Simplicity bias drives spontaneous learning","Self-consistent labels trigger a learning phase shift","Label exchange alone powers decentralized learning"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The load-bearing premise is that every update sees a fresh, independent batch of data—so that inter-step correlations vanish and the replica computation of the one-step map is exact in the thermodynamic limit; if data are finite or reused, the transition boundary changes or disappears, as the paper's own uniform-splitting example shows.","fun_headline_variants_meta":{"raw":{"variants":["Agent learns true clusters from its own labels","Simplicity bias drives spontaneous learning","Self-consistent labels trigger a learning phase shift","Label exchange alone powers decentralized learning"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000788,"raw_usage":{"total_tokens":3286,"prompt_tokens":692,"completion_tokens":2594,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":436,"completion_tokens_details":{"reasoning_tokens":2551}},"tokens_in":436,"tokens_out":2594,"duration_ms":16127,"temperature":1.0,"reasoning_tokens":2551,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-04T14:45:50.079936+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Initialize agents at a small nonzero magnetization m0 in a parameter region where the replica calculation predicts r<1 (low λ, high σ) and confirm the magnetization relaxes to zero; and in a region where r>1 (high λ, low σ) confirm a small seed grows to m*>0. The replica prediction gives exact r in the thermodynamic limit, so a discrepancy between the measured slope of mt+1 vs mt at mt=0 and the replica r would settle the matter directly.","supporting_citations":[],"review_version":1}