{"id":"f05ed4ee-3641-49ce-b876-e9025f0bad1b","arxiv_id":"2505.10569","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":4,"one_line_summary":"The paper proposes active inference as the basis for future 6G network AI and introduces a conceptual 'cyberfungi' generative model that uses diffusion models to hallucinate small-world networks for social contagion.","lead":"An invited optics journal paper argues that active inference, a neuroscience theory, should replace today's data-hungry artificial intelligence in future optical and 6G networks. It sketches a 'cyberfungi' agent that mimics underground fungal networks to spread social norms through small-world network effects.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The cyberfungi demonstration does not implement active inference: no derivation links DDPM denoising to free-energy minimization, so the central claim remains unverified.","rationale":"The reader's weakest_assumption isolates the same unsupported mapping between active inference and DDPM diffusion/denoising. My stress-test sharpens this: the Section 6.C simulation is not merely missing a rigorous derivation; it omits the defining components of active inference, namely variational and expected free energy, POMDP belief updating, and policy selection. The reported small-world contagion effect is a known network-topology result, and Fig. 8 even shows that the forward diffusion process alone achieves the minimum contagion time, which undermines the claim that the cyberfungi backward process provides the active-inference benefit. This is not an internal inconsistency in the exposition of active inference in Sections 3 and 4, which read as a fair tutorial, but it means the central claim of the paper is asserted rather than demonstrated. The paper is transparent about being a vision piece and points to a research agenda; with code and data release, baselines, and softened wording it can serve as a useful roadmap. Therefore the existing CONDITIONAL verdict remains appropriate; no change is required.","tokens_in":19487,"tokens_out":4294,"duration_ms":45448,"concrete_test":"Re-run the Fig. 8 simulation with the trained DDPM replaced by a direct Watts-Strogatz small-world network generator (matching average degree and rewiring probability) and with the same contagion and token-reward rules. If the contagion time and actuated-token counts are statistically indistinguishable from Fig. 8, then the DDPM component, and hence the asserted active-inference mapping, is not load-bearing for the reported result.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 4 defines active inference via a POMDP generative model and free-energy minimization: variational free energy F (Eq. 7) and expected free energy G (Eq. 13) drive belief updating and policy selection. Section 6.C replaces this machinery with OpenAI's improved DDPM, asserting that forward diffusion biomimics hyphal branching and backward denoising biomimics fusion/homing. No derivation shows that DDPM denoising is Bayesian belief updating under the free-energy principle; the paper calls the mapping 'adapt and extend' but supplies no formal identity. The only quantitative result (Fig. 8) measures social contagion time and token counts on a 64-node ring lattice that is diffused and then hallucinated by the DDPM. Fast contagion on small-world, scale-free topologies is a known property of those topologies and does not test active inference. The same figure shows that the forward diffusion process also reaches the minimum contagion time after about 500 steps, so the DDPM's backward process is not uniquely responsible for the claimed benefit. Moreover, the DDPM is trained on 1.3 million synthetic small-world, scale-free networks, making the hallucination outcome circular. Since the generative model contains no free-energy term, no belief update, and no policy selection, the central claim that active inference resolves training, learning, and explainability challenges and grounds the 6G world brain is unsupported by the demonstration.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper argues that active inference, premised on the free-energy principle, can overcome the training, learning, and explainability limitations of today's AI and should become the foundation of future optical and 6G network intelligence. It provides an introductory review of the active inference framework in Section 4, including the Markov blanket, POMDP generative models, variational free energy F (Eq. 7), and expected free energy G (Eq. 13). The paper then proposes a 'cyberfungi' generative model based on OpenAI's improved DDPM, using forward diffusion to mimic hyphal branching and backward denoising to mimic fusion/homing, and applies it to a 64-node ring lattice. The simulation in Fig. 8 reports social contagion times and actuated teleological tokens under different spreading probabilities. The conclusions assert that active inference 'not only resolves the open key AI challenges' but also advocates an active AI that views networks as living organisms, leading to a '6G world brain'.","tokens_in":19774,"tokens_out":5130,"duration_ms":54078,"significance":"If the central mapping between diffusion models and active inference were formally established, the paper would open a genuinely novel research direction connecting generative AI, active inference, and network science. The paper's accessible exposition of active inference mathematics in Section 4 is a strength, as it fills a gap for optical networking researchers, and the literature survey correctly identifies the absence of active inference in optical network research. However, the paper offers no machine-checked proofs, no reproducible code, and no parameter-free derivation; the math presented is standard textbook material. The specific contribution—the cyberfungi demonstration—is not a valid test of active inference, and the claimed resolution of AI challenges is asserted rather than demonstrated. As a vision or tutorial paper the work has some value, but the current manuscript overstates the evidence for its central thesis.","major_comments":[{"comment":"The paper does not provide a formal link between the DDPM forward/backward processes and the variational free energy F (Eq. 7) or expected free energy G (Eq. 13) defined in Section 4. The assertion that diffusion denoising corresponds to Bayesian belief updating under free-energy minimization is not derived. Since the cyberfungi model contains no variational distribution q(s,π), no belief update, and no policy selection, the simulation in Fig. 8 does not implement active inference. Consequently, the abstract and Section 7's claim that active inference 'resolves the open key AI challenges' is unsupported by the demonstration.","section":"Section 6.C, Fig. 8"},{"comment":"The reported benefit of the backward hallucination process is confounded. The text states that the forward diffusion process reaches the minimum contagion time of 2 after roughly 500 steps for all values of r; the backward process then maintains this minimum. Thus the low contagion time is not uniquely attributable to the denoising or hallucination process. In addition, Fig. 8 shows no error bars or multiple runs, and the only baseline is the regular ring lattice. The training set of 1,304,483 synthetic small-world, scale-free networks (mentioned in Section 6.C) makes the hallucination outcome circular: the DDPM is trained to produce exactly the topologies whose contagion properties are then measured. This does not validate the proposed cyberfungi model.","section":"Section 6.C, Fig. 8"},{"comment":"The paper claims that active inference overcomes the training, learning, and explainability limitations of today's AI, but no mechanism is specified. The 'no big data' claim in Section 3.C is asserted rather than derived, and the paper does not explain how the proposed generative model performs continual learning or yields explainable decisions. Section 4's mathematics is standard active inference, but the mapping to optical networks and 6G is left at the level of analogy (e.g., Markov blanket as an 'interface'). Without a concrete architecture or a formal equivalence, the central thesis remains an unverified research program rather than a demonstrated result.","section":"Section 3.C and Section 7"}],"minor_comments":[{"comment":"There is a duplicated word 'to to' in the phrases 'used to to biomimic' (two occurrences); this should be corrected.","section":"Section 6.C"},{"comment":"The word 'schezophrenia' is a misspelling of 'schizophrenia'.","section":"Section 2"},{"comment":"The word 'environmnet' appears to be a typo for 'environment'.","section":"Section 3.B"},{"comment":"The caption describes entropy but does not define the right-hand y-axis, which conveys the number of actuated teleological tokens; the caption should be expanded to explain both axes.","section":"Fig. 8 caption"},{"comment":"The paper correctly notes that popular media accounts of mycorrhizal networks are not scientifically supported, but it then continues to use the 'wood-wide web' framing as a basis for the cyberfungi model. The authors should either temper the use of this metaphor or cite the primary literature more carefully.","section":"Section 5.A"},{"comment":"The categorical phrasing 'the key to true AI' and 'not only resolves the open key AI challenges' goes beyond what the evidence supports; a more measured framing as a research vision would align better with the paper's actual contribution.","section":"Title and Abstract"}],"recommendation":"major_revision","confidential_remarks":"This is an invited paper, and its primary value appears to be as a perspective or tutorial for the optical networking community. The current demonstration, however, is not a valid validation of active inference, and the central claims are overstated. I would encourage the editor to consider whether the journal's standards permit a position paper; if so, the author should substantially revise the abstract and conclusions, clearly label Fig. 8 as illustrative rather than confirmatory, and add the missing baselines and statistical details. The paper also leans heavily on Friston's roadmap and the author's own prior INTERBEING framework; the novelty relative to optical networking itself is modest."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know two things about this one. First, it is a genuinely useful tutorial: Section 4 gives a clear, correct walkthrough of active inference—POMDPs, variational free energy F, expected free energy G—that an optics researcher could actually learn from. Second, the load-bearing claim in the abstract and conclusions—that active inference 'resolves' the open AI challenges of training, learning, and explainability and grounds a 6G world brain—is not supported by the evidence in the paper. The only quantitative result, Figure 8, is a toy simulation that doesn't test active inference at all.\n\nWhat's new here is the cyberfungi metaphor: using a DDPM to hallucinate small-world scale-free networks, with the forward diffusion standing in for hyphal branching and the backward denoising for homing and fusion. That's a creative analogy, but it is just an analogy. The paper provides no derivation linking DDPM denoising to Bayesian belief updating under free-energy minimization. Without that link, the demonstration is not an implementation of active inference; it's a diffusion model generating networks that look like the ones it was trained on.\n\nThe stress-test note is on target. The DDPM was trained on 1.3 million synthetic small-world, scale-free networks, so the hallucinated networks naturally have those properties, and fast social contagion on such topologies is known. Worse, Figure 8 shows the forward diffusion process also reaches the minimum contagion time of 2 after about 500 steps, so the backward denoising step is not uniquely responsible for the claimed benefit. There are no error bars, no baselines, and no code or data released. These are fixable, but as it stands the figure doesn't validate the cyberfungi model.\n\nTo the author's credit, the paper is transparent about being a vision piece. It flags the popular-media overstatement of mycorrhizal 'wood-wide web' research and cites the relevant active-inference tutorial material. The Section 4 equations are standard and correct. The self-citations are not a problem—the author's prior work is relevant. But the jump from 'this is an interesting biomimetic framing' to 'this is the key to true AI' is too large.\n\nWho should read this? Optical and 6G network researchers who want an accessible entry point into active inference. It's a good perspective piece, not a research result. For peer review, I'd let it through as an invited paper but condition acceptance on the author softening the overclaims, adding a proper baseline or at least discussing the forward-diffusion result, and either providing code/data or explicitly labeling the simulation as illustrative. It deserves review, but it needs revision before I'd want it cited as evidence for the 6G world brain.","headline":"This is a readable vision piece that introduces active inference to optical network researchers, but its central claim that active inference resolves AI's open challenges is asserted rather than demonstrated, and the toy simulation is too weak to carry the weight.","tokens_in":20305,"tokens_out":1946,"would_cite":false,"duration_ms":21652,"reading_group":"maybe","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 active inference—acting to minimize surprise under the free-energy principle—should replace today's training-heavy AI and become the foundation of a 6G 'world brain' built on mycorrhizal-network-inspired generative…","keywords":["active inference","free-energy principle","6G networks","optical networks","mycorrhizal networks","generative model","diffusion models","world brain"],"falsifier":"Run the paper's social-contagion simulation on a generic small-world network of the same 64 nodes and degree, with no DDPM hallucination and no tokens; if it also spreads the norm in about two time units, the cyberfungi generative model is not what produces the fast contagion. Separately, if the DDPM's reverse process cannot be written as variational free-energy minimization over the agent's inferred state space, the claimed equivalence between denoising and active inference fails.","tokens_in":19231,"feed_emoji":"🍄","tokens_out":9217,"duration_ms":89043,"temperature":0.7,"pith_summary":"Active inference, a biomimetic framework in which an agent acts to minimize the surprise of its sensory states, is presented as the key to true AI: a replacement for today's training-hungry, black-box models that resolves the open AI challenges of training, learning, and explainability. The paper delivers a self-contained introduction to the framework's mathematics—Markov blankets, generative models, variational and expected free energy—aimed at optical and 6G network researchers, and argues that future networks should be viewed as living organisms rather than static machines. Its longer-range proposal is a '6G world brain': a techno-social superorganism in which humans and digital 'cyberfungi' agents are coupled through a generative model that hallucinates scale-free networks with small-world properties, spreads desirable social norms by token rewards, and rewires collective intelligence. The supporting demonstration is a simulation in which a diffusion model's forward and backward processes stand in for fungal branching and fusion, and the hallucinated small-world networks achieve near-instant social contagion.","feed_headline":"Active inference, not bigger models, is the key to true AI","feed_subtitle":"A diffusion-model 'cyberfungi' agent grows small-world networks and rewards social norms, sketching a 6G world brain.","key_machinery":"The load-bearing mechanism is the pairing of active inference's free-energy minimization with a denoising diffusion generative model. Active inference supplies the normative loop: a Markov blanket $b = (u, y)$ separates an agent from its environment, Bayes's theorem updates beliefs about hidden states, variational free energy $F$ bounds surprise and makes inference tractable, and expected free energy $G$ scores future policies against preferred observations. The paper then maps the diffusion model's forward process to the branching and exploration of hyphae and the backward denoising process to their fusion and homing, so the model hallucinates scale-free small-world networks that serve as the substrate for social contagion. Purpose-driven tokens act as the action channel that closes the loop: they reward pairs of humans who spread a new social norm, actuating the cyberfungi's 'spores' and rewiring the human-AI network. The argument is carried by this analogy—network structure from denoising, network function from token-rewarded contagion—held together by the claim that both are expressions of entropy minimization.","core_discovery":"The paper claims that active inference is not merely an AI technique but the key to true AI, because it replaces supervised training with continuous Bayesian belief updating and explains behavior as self-evidencing under the free-energy principle. To make this concrete for networking, it proposes a metamorphic generative model called cyberfungi, built by adapting a denoising diffusion probabilistic model whose forward diffusion process mimics the branching, exploratory growth of fungal hyphae and whose backward denoising process mimics their fusion and homing. The model hallucinates scale-free networks with small-world properties; humans connect to these networks, and when one person adopts a desirable new social norm, it spreads to directly linked humans with probability $r$, with each successful pair receiving teleological tokens (cybernetic spores) as reciprocal rewards. Simulation results show that the hallucinated small-world networks spread a social norm through all 64 nodes in about two time units across all tested $r$, and that the backward process yields up to almost 2,500 actuated tokens for $r=1.0$, which the paper presents as evidence that cyberfungi act as brokers of cyborganic entanglement for the 6G world brain.","pith_inferences":["The diffusion-to-active-inference mapping is asserted by analogy; a testable extension would be to prove that the DDPM denoising objective is a free-energy bound, for instance by identifying the reverse process with variational inference over the same latent states the agent infers.","The fast-contagion result may owe nothing to the fungal metaphor: replacing the hallucinated networks with any small-world network of the same size and degree distribution should reproduce the two-step spreading time, isolating topology as the cause.","The token-reward mechanism could be tested on its own terms by comparing cyberfungi incentives with standard influence-maximization strategies on identical networks; if outcomes match, active inference is not the operative ingredient.","The world-brain vision assumes human values can be encoded as preferred observations in a generative model; a stress test would be whether conflicting preferences across humans produce free-energy gradients that pull the collective apart rather than toward homeostasis."],"forward_implications":["If active inference is the right foundation, AI-native 6G networks would learn continuously from interaction rather than from large offline datasets, dissolving the training bottleneck the paper identifies.","Optical network management could be reorganized around Markov-blanket interfaces that keep humans in the loop, giving the network a principled channel for explanation and human-AI co-creation.","A cyberfungi-type generative model would make the network itself an embodied, enactive agent that spreads social norms and tokens, effectively treating collective behavior as a network resource to be cultivated.","Because hallucinated small-world, scale-free networks spread norms in roughly two time units regardless of spreading probability, the 6G world brain would propagate desirable behaviors quickly and at low reward cost.","Under active inference, network control becomes homeostasis: the 6G brain would act to keep its sensory states within preferred ranges, rather than optimizing a fixed task objective."],"supporting_citations":[{"why":"Supplies the canonical active-inference framework, including the claim that the generative model is the central design challenge.","marker":"[3]"},{"why":"Foundational statement of the free-energy principle that the paper uses as the first-principles basis for active inference.","marker":"[17]"},{"why":"Sets out the AI vision and R&D roadmap, including mycorrhizal networks as a model for shared intelligence, that the 6G world brain proposal extends.","marker":"[20]"},{"why":"Provides the step-by-step mathematical and coding treatment of active inference the paper leans on for its introductory sections.","marker":"[23]"},{"why":"Introduces the Markov blanket concept, which the paper adopts as the interface between active-inference agents and their environment.","marker":"[24]"},{"why":"Supports the claim that any Markov-blanketed system can engage in active inference, used to justify extending the framework to fungi and networks.","marker":"[25]"},{"why":"Supplies the metamorphic-agent and 'knit your own Markov blanket' ideas that ground the human-cyberfungi cocoon narrative.","marker":"[26]"},{"why":"Provides the experimental finding of carbon transfer between trees via common mycorrhizal networks that grounds the wood-wide-web analogy.","marker":"[32]"},{"why":"Flags overinterpretation and open questions about mycorrhizal-network structure and function, which the paper uses to position its own cyberfungi model as a research opportunity.","marker":"[33]"},{"why":"Earlier work on Internet–human-being symbiosis that the proposed human-cyberfungi mutualism directly builds on.","marker":"[37]"}],"fun_headline_variants":["Active inference: key to true AI and 6G world brain","Cyberfungi agents build small-world networks for a 6G brain","True AI arrives via active inference, not scale","From AI to active inference: wiring the 6G world brain"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole proposal stands or falls on treating a diffusion model's denoising as equivalent to an active-inference agent's free-energy-minimizing belief update, an equivalence the paper asserts by analogy rather than proves.","fun_headline_variants_meta":{"raw":{"variants":["Active inference: key to true AI and 6G world brain","Cyberfungi agents build small-world networks for a 6G brain","True AI arrives via active inference, not scale","From AI to active inference: wiring the 6G world brain"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000671,"raw_usage":{"total_tokens":3145,"prompt_tokens":1124,"completion_tokens":2021,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":740,"completion_tokens_details":{"reasoning_tokens":1949}},"tokens_in":740,"tokens_out":2021,"duration_ms":14969,"temperature":1.0,"reasoning_tokens":1949,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T05:11:59.629035+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the paper's social-contagion simulation on a generic small-world network of the same 64 nodes and degree, with no DDPM hallucination and no tokens; if it also spreads the norm in about two time units, the cyberfungi generative model is not what produces the fast contagion. Separately, if the DDPM's reverse process cannot be written as variational free-energy minimization over the agent's inferred state space, the claimed equivalence between denoising and active inference fails.","supporting_citations":[{"cited_title":"The free-energy principle: a unified brain theory?","cited_arxiv_id":null,"evidence_quote":"Foundational statement of the free-energy principle that the paper uses as the first-principles basis for active inference."},{"cited_title":"Designing ecosystems of intelligence from first principles,","cited_arxiv_id":null,"evidence_quote":"Sets out the AI vision and R&D roadmap, including mycorrhizal networks as a model for shared intelligence, that the 6G world brain proposal extends."},{"cited_title":"A step-by-step tutorial on active inference and its applications to empirical data,","cited_arxiv_id":null,"evidence_quote":"Provides the step-by-step mathematical and coding treatment of active inference the paper leans on for its introductory sections."},{"cited_title":"Pearl, Probabilistic Reasoning in Intelligent Systems: Networks of Plausible Inference (Morgan Kaufmann Publishers Inc., 1988)","cited_arxiv_id":null,"evidence_quote":"Introduces the Markov blanket concept, which the paper adopts as the interface between active-inference agents and their environment."},{"cited_title":"The Markov blankets of life: autonomy, active inference and the free energy principle,","cited_arxiv_id":null,"evidence_quote":"Supports the claim that any Markov-blanketed system can engage in active inference, used to justify extending the framework to fungi and networks."},{"cited_title":"Clark, Philosophy and Predictive Processing, 3 (MIND Group, Frank- furt am Main, Germany, 2017), chap","cited_arxiv_id":null,"evidence_quote":"Supplies the metamorphic-agent and 'knit your own Markov blanket' ideas that ground the human-cyberfungi cocoon narrative."},{"cited_title":"Net transfer of carbon between ectomycorrhizal tree species in the field,","cited_arxiv_id":null,"evidence_quote":"Provides the experimental finding of carbon transfer between trees via common mycorrhizal networks that grounds the wood-wide-web analogy."},{"cited_title":"Positive citation bias and overinterpreted results lead to misinformation on common mycorrhizal networks in forests,","cited_arxiv_id":null,"evidence_quote":"Flags overinterpretation and open questions about mycorrhizal-network structure and function, which the paper uses to position its own cyberfungi model as a research opportunity."},{"cited_title":"INTERBEING: On the Symbiosis Between INTERnet and Human BEING,","cited_arxiv_id":null,"evidence_quote":"Earlier work on Internet–human-being symbiosis that the proposed human-cyberfungi mutualism directly builds on."}],"review_version":1}