{"id":"40a350b9-79ca-4cdc-98a4-0a322b0ccac8","arxiv_id":"2411.18498","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":8,"one_line_summary":"Embodied agents with HKB oscillator brains reach consensus when internal, social, and environmental coupling are balanced.","lead":"This paper simulates tiny agents whose steering is controlled by coupled oscillators, and shows that groups of such agents can choose a common stimulus source only when internal brain coupling, social coupling, and environmental sensitivity are balanced. It offers a bridge between brain dynamics and collective behavior, with implications for neuro-AI and swarm robotics.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 'balance' conclusion in Fig. 4 is likely an artifact of the arbitrary ternary normalization; the three couplings are constrained to sum to 100, so the high-performance interior region may simply reflect that all three need to be non-negligible, not a genuine balance.","rationale":"The reader's weakest assumption was the missing social decay rate lambda_s. That is a legitimate reproducibility issue, but it does not directly threaten the central claim: even if lambda_s is unknown, the qualitative 'balance' could persist across reasonable values. The arbitrary ternary normalization, by contrast, shapes the very evidence for the balance claim. The paper presents Fig. 4 as the multi-agent demonstration of balance, but the sum-to-100 constraint means the three couplings are not independent; the 'balance' region is partly a consequence of the chosen parameterization. The single-agent results in Fig. 3 and Fig. S1 provide independent evidence for an intermediate optimal internal coupling when sensory sensitivity is fixed, so the balance idea may be correct. However, the multi-agent claim as stated is not supported by the reported scan alone. This is a concrete, addressable weakness: reporting the full 3D sweep would settle it. I therefore agree with the reader's CONDITIONAL verdict but identify a different primary concern; hence 'partial' agreement.","tokens_in":22869,"tokens_out":11503,"duration_ms":105640,"concrete_test":"Scan the full 3D grid c in {0,...,10}, S in {0,...,5}, av3,v4 in {0,...,1} without the sum-to-100 constraint, and map the high-performance set. If the high-performance set is an interior region in this unconstrained cube (i.e., performance drops when any single coupling is too strong or when several are simultaneously too strong), the balance claim survives. If the high-performance set instead touches the boundary of the cube or is determined by the arbitrary scaling, the Fig. 4 'balance' conclusion is an artifact of the normalization.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that collective decision success depends on a 'balance' of intra-agent, inter-agent, and agent-environment coupling rests primarily on the ternary plots in Fig. 4. In the multi-agent experiments (Sec. 5.2.2), the three parameters c, S, and av3,v4 are not scanned independently. The Fig. 4 caption states: 'the parameters fulfill the condition stimulussensitivity + socialsensitivity + internalcoupling = 100', with each raw parameter mapped to a [0,50] scale via different linear factors (c:0-10, S:0-5, av3,v4:0-1). This constraint makes the three couplings negatively correlated and confines the scan to a 2D manifold of the full 3D parameter space. The high-performance 'middle region' is therefore guaranteed to be a region where no single coupling dominates; the corners necessarily remove at least one coupling entirely. The authors do not report a full 3D parameter sweep or show that the high-performance set is an interior, bounded region when the sum constraint is removed. Moreover, the linear mapping of raw parameters to the [0,50] scale is never justified; a different scaling would shift the location of the 'balanced' region. Thus the central balance claim is scale-dependent and may be an artifact of the normalization rather than a property of the system.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents a multi-agent simulation study in which agents are controlled by HKB oscillator networks instead of simple behavioral rules. Each agent has two sensory and two motor oscillators; the motor phase difference controls heading, and in the multi-agent setting agents also emit and perceive a social stimulus that is added to the environmental stimulus. The authors report single-agent results linking gradient-ascent and binary-decision performance to internal coupling and stimulus sensitivity, and multi-agent results in which collective decision performance is displayed in ternary plots as a function of internal, environmental, and social coupling. The main claim is that successful collective decisions require a balance of intra-agent, inter-agent, and agent-environment coupling, and that environmental factors such as stimulus quality ratio and initial heading spread modulate decision difficulty.","tokens_in":23173,"tokens_out":3973,"duration_ms":38959,"significance":"If fully supported, the model would provide a useful bridge between coordination-dynamics neuroscience and embodied collective behavior, and the open-source implementation is a concrete strength. The paper honestly reports deterministic runs and provides a substantial set of behavioral and neural measures (PLV, KOP, wPLI). However, the central 'balance' conclusion is currently supported mainly by a constrained ternary parameterization whose normalization is arbitrary, and the social decay rate needed to reproduce the multi-agent results is never specified. The contribution is therefore valuable but needs additional analysis before the headline claim can be accepted.","major_comments":[{"comment":"The ternary normalization makes the central 'balance' claim scale-dependent. The caption states that stimulus sensitivity, social sensitivity, and internal coupling sum to 100, with heterogeneous linear maps (c in 0–10, S in 0–5, av3,v4 in 0–1 each mapped to a [0,50] scale). This constraint negatively correlates the three parameters and confines the scan to a 2D manifold. The bright interior region is therefore guaranteed to be a region where no single coupling is maximal, regardless of the system's actual behavior. Please report an unconstrained 3D sweep, or at minimum show that the qualitative location and boundedness of the high-performance region are invariant under different rescaling choices.","section":"§3.2.1, Fig. 4"},{"comment":"The social decay rate lambda_s is introduced in Eq. (6) but is never given a numerical value anywhere in the paper, and Eq. (3) instead uses the symbol lambda_a. This decay rate sets the spatial range of inter-agent influence and therefore directly shapes the consensus regions in Fig. 4 and all multi-agent outcomes. Without this value the model is not reproducible and the role of social coupling cannot be independently assessed. Please provide the value, reconcile the notation, and include a brief sensitivity analysis over lambda_s.","section":"§5.1, Eq. (6)"},{"comment":"It is unclear which internal couplings were varied in the multi-agent experiments. The methods text says 'internal coupling between the motor oscillators (avi,vj values from 0 to 1, in steps of 0.02, in Eq. 7)', while the Fig. 4 caption labels the axis specifically as av3,v4. The single-agent experiments varied all connections, so the reader cannot tell whether 'internal coupling' in the multi-agent balance claim refers only to the motor–motor link or to the full set of oscillator connections. Please clarify this and, if only av3,v4 was varied, discuss why the other internal couplings were held fixed.","section":"§5.2.2 and Fig. 4 caption"},{"comment":"The binary decision-making performance metric is misprinted: it reads min{D_source1(tend), D_source1(tend)} / D(t0). Taken literally, this metric ignores the second stimulus source and cannot measure a binary decision. Presumably the intended expression is min{D_source1(tend), D_source2(tend)}; please correct the equation and confirm that the implemented metric matches the corrected form, since this metric underlies the single-agent decision-making claims.","section":"§5.3, Eq. (10)"}],"minor_comments":[{"comment":"The social decay rate is written as lambda_a in Eq. (3) and lambda_s in Eq. (6); please use a single symbol consistently.","section":"§2.2, Eq. (3) and §5.1, Eq. (6)"},{"comment":"The supplementary text refers to 'Fig. 5A' and 'Fig. 6' of the main paper when the corresponding panels appear to be Fig. 4 and Fig. 5; please update all cross-references.","section":"Supplementary S3 and S4"},{"comment":"The caption contains run-together labels such as 'stimulussensitivity' and 'socialsensitivity'; please insert proper spacing.","section":"Fig. 4 caption"},{"comment":"The observation that high internal coupling leads to PLV near 1 and loss of gradient climbing is, to a large extent, a direct consequence of Eq. (8): a phase-locked motor oscillator pair gives a constant heading. The paper should acknowledge this structural entailment explicitly rather than presenting it only as an emergent empirical result.","section":"§3.1"}],"recommendation":"major_revision","confidential_remarks":"The manuscript has a solid empirical core and openly available code, but the headline 'balance' claim is not yet adequately supported because the ternary normalization and the unspecified lambda_s jointly determine the main result. I would not reject the paper, but the authors should be asked to provide an unconstrained parameter sweep or a clear invariance argument, and to specify the social decay rate, before publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Read this one before you cite it: the architecture is genuinely new and the paper is cleanly written, but the headline 'balance of intra-, inter-, and agent-environment coupling' is not actually established by the main figure. The four-node HKB controller with indirect social coupling through an emitted stimulus is a real extension of Aguilera et al. 2013, and the single-agent results showing an intermediate internal-coupling sweet spot are solid. The code is public, the equations are explicit, and the authors are honest that inter-agent synchrony cannot be shown to be causal.\n\nThe soft spot is Fig. 4. The three couplings are scanned under the constraint stimulus sensitivity + social sensitivity + internal coupling = 100, with each mapped to a [0,50] scale through different linear factors (c: 0-10, S: 0-5, a: 0-1). That makes the three negatively correlated, so the high-performing interior is guaranteed to be a region where no single coupling dominates. A 'balance' claim needs a full 3D sweep or at least a demonstration that the high-performance set is an interior, bounded region under independent variation. As is, the center of the ternary plot could just be the region where all three parameters are non-negligible. The scaling choice is arbitrary, and a different mapping would move the 'balanced' region.\n\nTwo smaller issues back this up. The social decay rate lambda_s in Eq. 6 is never given a value anywhere in the paper, so the spatial range of social influence is unspecified. And there is no non-neural baseline; without a controller without HKB dynamics, it's unclear what the oscillators add over a standard sensorimotor loop. Minor: Eq. 10 has a typo (source1 repeated), and supplementary figures refer to Fig. 6 that should be Fig. 5.\n\nTo be fair, the qualitative phenomenon is not purely an artifact. Fig. S2 is a one-dimensional scan of social influence at fixed c and internal coupling, and it shows performance rising then falling as S increases. That supports the claim that too much social coupling hurts. But it doesn't rescue the three-way 'balance' framing.\n\nVerdict: this deserves peer review. The novel combination and honest, reproducible effort justify referee time. The authors need to fix the missing parameter, run an unconstrained sweep, and add a baseline before the central claim is convincing. I'd take it to a reading group as a case study in how ternary normalization can manufacture a 'balanced' conclusion, and I'd cite it as a proof-of-concept for oscillator-controlled collective decisions, with a caveat on the balance claim.","headline":"A clean, reproducible proof-of-concept for oscillator-driven collective decisions, but the headline 'balance' claim rests on a constrained ternary scan that needs a robustness check.","tokens_in":23712,"tokens_out":3535,"would_cite":true,"duration_ms":31436,"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":"Oscillator-driven agents reach collective decisions without explicit communication or preferences.","keywords":["collective decision making","embodied cognition","Haken-Kelso-Bunz oscillators","neural dynamics","multi-agent systems","sensorimotor coordination","inter-brain synchrony","consensus"],"falsifier":"Set the social decay rate $\\lambda_s$ in Eq. (6) to near zero and to infinity while holding all other parameters fixed; the paper's claim predicts that collective performance collapses at both extremes, with a single intermediate optimum. A simulation or robot experiment that shows consensus persisting across the full range would falsify the balance claim.","tokens_in":22640,"feed_emoji":"🧠","tokens_out":6324,"duration_ms":57981,"temperature":0.7,"pith_summary":"This paper tries to show that collective decision making can emerge from the simplest kind of embodied neural dynamics, without any agent holding a preference or sending a deliberate signal. Ten agents, each controlled by four coupled phase oscillators, move up a stimulus gradient and, through mutual interference of their emitted stimulus, converge on the same food-like source. The central claim is that success requires a balance of three couplings: within the agent's own oscillator network, between agents, and between agent and environment. The authors argue that this balance is visible in neural coordination measures, and that environmental factors such as source quality and initial heading spread set how hard the collective decision is. If right, the work connects coordination dynamics of brains to self-organized multi-agent systems and gives neuro-AI a principled way to build social agents.","feed_headline":"Ten oscillator-driven agents choose one site together","feed_subtitle":"Coupled HKB brains make collective decisions when internal, social, and environmental ties stay balanced.","key_machinery":"The carrying object is the Haken–Kelso–Bunz (HKB) oscillator model, a system of coupled phase oscillators with both in-phase and anti-phase attraction. Each agent has four oscillator nodes—two sensory and two motor—and its steering angle is proportional to the phase difference between the two motor oscillators, closing a sensorimotor loop. Social coupling is implemented by having each agent emit the same exponential-decay stimulus it senses, so the perceived input in Eq. (6) is the environmental gradient plus the summed, distance-decayed emissions of other agents. The argument is carried by showing, in ternary parameter scans, that only an intermediate weighting of internal, environmental, and social influence produces consensus, and by linking that region to phase-locking measures (PLV, KOP, wPLI) that quantify integration and metastability inside and across agents.","core_discovery":"On the paper's own terms, the discovery is that a group of embodied agents whose only control system is a network of HKB phase oscillators can solve a collective binary decision: starting from the same point with different headings, ten agents choose one of two stimulus sources and gather there. The decision is not coded anywhere—there is no preferred-direction parameter, no communication channel, and no rule saying 'follow the majority.' Instead, heading emerges from the phase difference between two motor oscillators, and social influence enters only because each agent emits the same stimulus it senses, so other agents cannot tell social input from environmental input. In the single-agent setting, the same mechanism lets an agent choose between two sources based purely on how its neural dynamics coordinate with the environment. The results show a coherent parameter region, visible in ternary plots, in which internal coupling, environmental sensitivity, and social sensitivity are balanced; in that region movement alignment is high, alignment variability is low, and consensus is achieved. The paper also reports that collective performance falls when social influence is too strong, because agents become saturated with social information and lose adaptive contact with the environment.","pith_inferences":["Because social and environmental stimuli are indistinguishable in the model, a natural extension is to test whether agents that can label social input (for example, by giving it a distinct oscillation frequency) show a wider or more robust consensus region; the paper does not simulate this.","The balance-of-couplings picture suggests a testable prediction for robot swarms: a controller that tunes internal coupling online in response to social saturation should track the consensus region automatically.","The observed overshoot and deadlock regimes could be interpreted as emergent failure modes analogous to milling or group panic; a quantitative comparison with existing swarm or pedestrian models would be a direct next step.","If the spatial decay rate of social emission were measured rather than left unspecified, the model's quantitative predictions about consensus regions could be tested directly against physical multi-robot experiments."],"forward_implications":["A single agent needs intermediate internal coupling: too little and it is driven by raw sensory input, too much and it ignores the environment; only the metastable middle climbs the gradient.","Groups with divergent initial headings can reach consensus with no opinion parameters, purely through embodied sensorimotor coupling.","Increasing social influence improves consensus only up to a point; beyond it, agents overshoot, split, or deadlock, so collective behavior is not monotonic in sociability.","Environmental difficulty is set by the quality difference between sources and the spread of initial headings; decisions get nonlinearly harder as the second source brightens.","Consensus performance is accompanied by measurable movement alignment and by inter-agent neural phase covariance, connecting the model to inter-brain synchrony research."],"supporting_citations":[{"why":"Supplies the situated HKB agent architecture (sensory and motor oscillators in a sensorimotor loop) that this paper extends to four oscillator nodes and social coupling.","marker":"[Aguilera et al., 2013]"},{"why":"Introduces the HKB equations whose in-phase and anti-phase attraction give the agents bistable metastable dynamics.","marker":"[Haken et al., 1985]"},{"why":"Provides the N-component HKB phase update rule used in Eq. (1) and in the agent's controller.","marker":"[Zhang et al., 2019]"},{"why":"Defines the decision-versus-compromise problem for moving groups and the deterministic setup that the multi-agent experiments follow.","marker":"[Leonard et al., 2011]"},{"why":"Supplies the Kuramoto order parameter used to measure movement alignment and metastability in collective runs.","marker":"[Strogatz, 2000]"},{"why":"Supplies the weighted phase-lag index used to quantify intra-agent and inter-agent neural coordination.","marker":"[Vinck et al., 2011]"}],"fun_headline_variants":["Neural agents reach consensus with balanced ties","HKB oscillators drive collective decisions","No rules, just rhythm: agents choose together","Embodied neural agents pick one site as a team","Coupled brains, shared choice: agents agree"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that social influence can be modeled as an additive stimulus identical in kind to the environmental stimulus, with a spatial decay rate $\\lambda_s$ in Eq. (6) that the paper never fixes numerically.","fun_headline_variants_meta":{"raw":{"variants":["Neural agents reach consensus with balanced ties","HKB oscillators drive collective decisions","No rules, just rhythm: agents choose together","Embodied neural agents pick one site as a team","Coupled brains, shared choice: agents agree"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000195,"raw_usage":{"total_tokens":1370,"prompt_tokens":974,"completion_tokens":396,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":590,"completion_tokens_details":{"reasoning_tokens":327}},"tokens_in":590,"tokens_out":396,"duration_ms":4312,"temperature":1.0,"reasoning_tokens":327,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T11:07:59.398240+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Set the social decay rate $\\lambda_s$ in Eq. (6) to near zero and to infinity while holding all other parameters fixed; the paper's claim predicts that collective performance collapses at both extremes, with a single intermediate optimum. A simulation or robot experiment that shows consensus persisting across the full range would falsify the balance claim.","supporting_citations":[],"review_version":1}