{"id":"27738336-f2f5-4302-bce0-ad892136d42d","arxiv_id":"2507.04575","paper_version":1,"verdict":"REJECT","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":0,"one_line_summary":"A conceptual framework in which untrained modular LLMs are developed through simulated life and token-based chemical signaling, with the goal of enabling empirical study of consciousness emergence via Integrated Information Theory.","lead":"This paper proposes LILITH, an architecture of modular language models that act as brain regions and communicate through learned token protocols, meant to mimic chemical signaling. It suggests training these modules from scratch through simulated life experiences and evolutionary selection, aiming to study how consciousness might emerge.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"IIT's Φ is not well-defined on LILITH because no formal mapping from LLM modules and token streams to IIT mechanisms, states, and transitions is supplied; the central empirical claim therefore rests on an undefined quantity.","rationale":"The reader's weakest assumption concerned IIT's validity as a measure of consciousness. I partially agree, but I would sharpen the objection: IIT validity is a field-level dispute, whereas the more internal and decisive gap is that the paper supplies no formalism connecting LILITH to IIT quantities. The strongest claim has two conjuncts: (1) IIT tracks consciousness, and (2) LILITH is a system on which IIT can be computed. The paper gives no derivation, implementation, or simulation supporting (2), and it explicitly says its goal is to put the idea forward. That honesty is commendable, and the writing is clear, but neither supports the advertised 'direct empirical investigation' claim. The suggested concrete test would settle whether the architecture has a well-defined IIT evaluation at all; if it does not, the central claim fails regardless of the IIT validity debate. Since the reader already recommended REJECT with high confidence, my analysis does not change that verdict; it only replaces a somewhat external assumption with a specific, testable formal gap.","tokens_in":3428,"tokens_out":7185,"duration_ms":80347,"concrete_test":"Implement the smallest faithful instance of LILITH: two finite-state modules (stand-ins for LLMs) communicating over a finite token alphabet, with deterministic transitions, in a toy lifetime environment. Compute Φ with the standard IIT 3.0 algorithm (e.g., PyPhi) at the module level and at the whole-system level. Repeat under two changes: (i) split each token into a two-bit sequence (finer state space, same behavior), and (ii) compose the two modules into one input-output-equivalent monolithic automaton. If Φ changes, or if its relative ordering across levels changes, then the IIT measurement is not well-defined on the token-based architecture, and the central empirical claim would not land.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is that LILITH 'would enable direct empirical investigation of consciousness emergence using Integrated Information Theory metrics' (Implications for Consciousness Research). For that claim to hold, IIT's quantitative measure, Φ, must be defined on the proposed architecture. It is not. IIT requires a physical substrate with a definite set of elements, states, and transition probabilities; LILITH specifies only 'LLM modules' exchanging 'tokens.' The paper never states whether the IIT elements are attention heads, whole modules, token types, or something else, nor what the state space is at each level. During developmental training the modules' weights, prompts, and sampling distributions change, so the substrate is not fixed. The token-based protocol is described at the level of abstract symbols, but Φ is not invariant under arbitrary choices of how tokens are grouped or split; changing token granularity changes the state space and the causal structure. Finally, the Optimization section proposes evolutionary search over inter-region signaling, but no objective function connecting that search to Φ is given, so 'optimizing for consciousness emergence' is disconnected from the IIT measurement the paper advertises. This is not a disagreement with IIT's validity; it is an internal formal gap: even granting IIT, the proposed system is not specified enough to compute Φ.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes LILITH, a modular LLM architecture in which distinct modules correspond to brain regions, communicate through learned token-based 'chemical signaling', and are trained via simulated development rather than static pretraining. The authors claim this architecture would enable direct empirical investigation of consciousness emergence using Integrated Information Theory (IIT) metrics, and that optimizing for consciousness emergence rather than task performance could reveal multi-scale neural correlates. The paper is explicitly a conceptual proposal: no implementation, simulation, data, or mathematical derivation is provided. The central claims are stated as future possibilities, and the authors acknowledge substantial open challenges, particularly in optimization.","tokens_in":3684,"tokens_out":6383,"duration_ms":74264,"significance":"The idea of combining modular LLMs with learned inter-module communication to test IIT-style measures is a creative extension of recent work on modular and brain-inspired AI. The paper is clearly written and honest in its limitations, and it cites relevant foundational references. However, as it stands it offers no formal model, no algorithms beyond natural-language descriptions, and no experimental results. Its significance is therefore entirely promissory. If the mapping from LILITH modules and token streams to IIT's formal objects were supplied and a concrete optimization scheme specified, the proposed architecture could become a useful instrument for studying integrated information in modular networks; the current manuscript does not yet provide that instrument.","major_comments":[{"comment":"The sentence 'The system can be directly tested using Integrated Information Theory metrics' is unsupported. IIT applies to a system with a specified set of elements, a state space, and transition probabilities; the manuscript never defines what counts as an element (modules, attention heads, token types?), how token streams map to states, or how the time-varying developmental process provides the stationary transition structure that Φ requires. Without this mapping, the central claim that LILITH enables direct IIT-based investigation cannot be evaluated.","section":"Implications for Consciousness Research"},{"comment":"The proposed optimization scheme is disconnected from the IIT measurement. The paper mentions 'auto-encoding objectives' for regions and 'evolutionary algorithms applied to inter-region signaling', but it explicitly states that 'detailed investigation of these optimization frameworks remains an important direction for future research'. Consequently, the Abstract's phrase 'optimizing for consciousness emergence rather than task performance' has no operational meaning in this manuscript, and the link between the evolutionary search and any IIT-based quantity is missing.","section":"On the Optimization of such an Architecture"},{"comment":"If LILITH is trained to maximize IIT-based metrics, then any later increase in those metrics is partly by construction, not independent evidence of consciousness emergence. The paper does not state whether IIT is intended as a training objective or only as a post-hoc evaluation criterion. This ambiguity is load-bearing because the proposed evidential value of the framework depends on IIT being an independent measure of the phenomenon under study.","section":"Implications for Consciousness Research / Abstract"},{"comment":"The architecture is specified at an informal level: success criteria are said to 'be carefully defined' but are not, the 'breeding' step relies on LLM voting without details, and the token protocol is described only through an example. Because the manuscript's goal is to propose a research direction, this level of specification might be acceptable in a perspective piece; however, the strength of the empirical claims in the Abstract and Implications sections ('unprecedented insight', 'direct empirical investigation') is disproportionate to what the manuscript actually delivers.","section":"Developmental Training Framework"}],"minor_comments":[{"comment":"The running title reads 'DEVELOPMENT AL MODULAR LLMS'; this should be 'DEVELOPMENTAL MODULAR LLMS'.","section":"Header"},{"comment":"The heading 'Modular Brain RegionDesign' is missing a space, and 'T oken-Based' in the following heading contains an extra space.","section":"Section 2 heading"},{"comment":"Reference [9] lists 'and et al.' after the first authors, which is redundant; either list all authors or use 'et al.' alone.","section":"References"}],"recommendation":"reject","confidential_remarks":"The manuscript is more of a position statement than a research paper. Its central claim about enabling IIT-based empirical investigation is not backed by the formal machinery needed even to define Φ on the proposed system. The authors openly defer the optimization details, which are essential to the proposal. I see no internal inconsistency or bad faith, but the contribution is currently at the level of an extended abstract; the missing formalization and experiments would require a fundamentally new manuscript rather than a revision within the scope of this submission."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe Lilith preprint is a short, clearly written position paper with a genuinely new combination: untrained modular LLMs serving as brain regions, token-based communication standing in for chemical signaling, developmental training in a simulated life, and evolutionary selection. I have not seen that exact synthesis before, and the authors deserve credit for saying plainly that this is a conceptual sketch, not a completed system.\n\nWhat the paper does well: it identifies a real gap in current AI—static pretraining and fixed routing—and proposes a plausible-sounding alternative. The writing is straightforward and the references are appropriate.\n\nThe main problem is the central empirical claim. The paper says LILITH 'would enable direct empirical investigation of consciousness emergence using Integrated Information Theory metrics.' That claim is unsupported in a specific and fixable way: IIT's Φ is not defined for this architecture. IIT requires a substrate with definite elements, states, and transition probabilities. The paper does not say whether Φ's elements are attention heads, LLM modules, token types, or something else, nor what the state space is at each level. The token protocol is specified only at the level of abstract symbols, and Φ depends on how tokens are grouped. The evolutionary optimization is described without any objective function connecting it to Φ. So even taking IIT at face value, the proposed system is not specified enough to compute the advertised quantity. This is an internal formal gap, not a philosophical disagreement with IIT.\n\nThe circularity concern is secondary but real: if the system is trained to maximize Φ, then increases in Φ are partly by construction. That weakens the claim that LILITH would independently reveal consciousness emergence.\n\nThere is no implementation, data, or formal derivation, so as a research paper it does not yet substantiate its claims. But the authors are honest about that. The paper is coherent and could be a useful discussion piece for an interdisciplinary workshop or a venue that welcomes speculative frameworks. I would not send it to a standard technical journal for peer review in its current form—there isn't enough technical content to referee. If the authors add a precise mapping from the architecture to IIT's substrate and a concrete optimization objective, it would become a serious proposal worth revisiting.","headline":"A new but wholly unsupported architectural idea; the IIT evaluation rests on an undefined Φ, so treat it as a position paper, not a research result.","tokens_in":4163,"tokens_out":3871,"would_cite":false,"duration_ms":41376,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"LILITH proposes that modular LLMs with chemical-style signalling and developmental training could let researchers measure consciousness emergence with Integrated Information Theory.","keywords":["developmental training","modular language models","chemical signaling","consciousness emergence","Integrated Information Theory","token-based communication","evolutionary optimization","multi-region brain modelling"],"falsifier":"One could train a LILITH system and record its Integrated Information Theory score across development together with behavioural markers such as flexible response to novel inputs or goal-directed action; if the score rises while every behavioural marker stays flat, the claim that optimizing for IIT metrics tracks consciousness emergence would be contradicted.","tokens_in":3257,"feed_emoji":"🧠","tokens_out":6990,"duration_ms":65103,"temperature":0.7,"pith_summary":"This paper proposes LILITH, an architecture in which separate language-model modules play the roles of brain regions—thinking, memory, sensory, and regulatory—and communicate through token-based signals that stand in for chemical neurotransmitters. The authors' central conjecture is that modelling the brain at the level of interacting regions with chemical-style signalling, instead of only at the level of neurons, is a productive step toward understanding the emergence of consciousness. To that end, LILITH would skip pre-training entirely: untrained LLM architectures would live simulated life cycles, develop communication pathways through environmental interaction, and be selected across generations by evolutionary optimization. The paper's stated payoff is that such a system could be measured directly with Integrated Information Theory, which seeks to quantify consciousness from causal structure, enabling empirical study of when and where consciousness-like integrated information appears during development. The paper is a conceptual proposal and acknowledges substantial implementation challenges.","feed_headline":"Modular LLMs that signal like neurons may test consciousness","feed_subtitle":"A proposal to grow untrained language models through simulated lives and measure emerging awareness with IIT.","key_machinery":"The load-bearing mechanism is the token-based communication protocol between modular LLMs. Each module is a specialised brain region with restricted capabilities—a thinking region that can prompt itself, a memory region that alone can save items, a sensory region that alone receives external input, and a regulatory brain stem with preprompted control powers—and the tokens these modules send to one another are the analogue of neurotransmitter signals. The machinery does two jobs: it makes inter-region signalling observable and editable, since tokens can be logged and measured, and it is what the developmental and evolutionary process shapes, because the meaning and routing of tokens are not predefined but emerge over simulated lives. This is the part of the design that connects the biology-inspired vocabulary to testable Integrated Information Theory calculations.","core_discovery":"The central claim is that consciousness emergence can be investigated empirically by combining developmental training of modular language models with brain-inspired token-based communication. Concretely, the paper argues that distinct brain regions should be modelled as specialised LLM modules, that inter-region signals should be learned tokens rather than predefined routes, and that the whole system should start untrained and learn through simulated life experience. On the authors' account, this design would allow Integrated Information Theory metrics to be applied at three scales—individual modules, inter-agent communication, and the whole system—so a researcher could see at which scale integrated information is strongest and how it grows over developmental time. The discovery being proposed, in other words, is a method: a way to build and measure a developing artificial system whose design objective is consciousness emergence rather than task performance.","pith_inferences":["The authors leave implicit that the same training objective could be swapped: any computable theory of consciousness could replace Integrated Information Theory, making LILITH a general platform for comparing candidate theories rather than a test of IIT alone.","A natural extension would be an ablation study of the regulatory brain-stem module: removing it during development and observing whether token routing and inter-module signalling collapse would test whether the architecture's regulatory component plays the arousal-like role it is assigned.","Because the token vocabulary is emergent, one could measure the entropy and stability of token usage over developmental time; the authors do not, but a concrete prediction of their framework is that stable, differentiated signalling tokens should appear before any rise in integrated information."],"forward_implications":["If LILITH works as proposed, consciousness metrics can be tracked continuously during development, giving a time series of integrated information as communication pathways form.","The multi-scale design would let researchers compare integrated information at the level of single modules, inter-module signalling, and the whole system, and thereby locate the organizational level where integration is maximal.","Shifting the optimization target from task performance to consciousness measures would change the evaluation culture: systems would be judged by their developmental trajectory and sentience markers rather than by benchmark scores.","The emergent token protocols themselves become observable objects, so the framework would yield a record of how inter-region chemical-style signalling changes with experience."],"supporting_citations":[{"why":"It supplies the Integrated Information Theory metrics that LILITH would use to measure consciousness emergence.","marker":"[14]"},{"why":"It supports the developmental learning premise that intelligence emerges through lived experience rather than static training.","marker":"[9]"},{"why":"It supplies the argument that human-like intelligence requires learning and thought beyond pattern matching, motivating developmental training.","marker":"[2]"},{"why":"It defines the limitation of pretrained LLMs as stochastic parrots, the gap LILITH aims to escape.","marker":"[1]"},{"why":"It provides the network-science basis for treating modularity as vital to brain function.","marker":"[11]"},{"why":"It extends the modular-brain rationale to network neuroscience, grounding the multi-region architecture.","marker":"[12]"},{"why":"It supplies the chain-of-thought prompting method used by the thinking region.","marker":"[13]"},{"why":"It supplies the prior modular architecture that LILITH's modular design extends.","marker":"[10]"}],"fun_headline_variants":["Growing modular LLMs that signal like brain chemicals to test consciousness","Lilith: developmental LLM modules with neurotransmitter-inspired signals","LLM modules that evolve chemical-like signaling to probe consciousness emergence","Proposal: grow untrained LLMs with brain-style signaling to measure IIT","Developmental LLM with chemical signaling: a testbed for consciousness emergence"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The proposal depends on Integrated Information Theory being a valid measure of consciousness, so that optimizing a system for its metrics counts as studying consciousness emergence.","fun_headline_variants_meta":{"raw":{"variants":["Growing modular LLMs that signal like brain chemicals to test consciousness","Lilith: developmental LLM modules with neurotransmitter-inspired signals","LLM modules that evolve chemical-like signaling to probe consciousness emergence","Proposal: grow untrained LLMs with brain-style signaling to measure IIT","Developmental LLM with chemical signaling: a testbed for consciousness emergence"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000403,"raw_usage":{"total_tokens":2083,"prompt_tokens":908,"completion_tokens":1175,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":524,"completion_tokens_details":{"reasoning_tokens":1083}},"tokens_in":524,"tokens_out":1175,"duration_ms":11680,"temperature":1.0,"reasoning_tokens":1083,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T19:43:45.477228+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"One could train a LILITH system and record its Integrated Information Theory score across development together with behavioural markers such as flexible response to novel inputs or goal-directed action; if the score rises while every behavioural marker stays flat, the claim that optimizing for IIT metrics tracks consciousness emergence would be contradicted.","supporting_citations":[{"cited_title":"Integrated information theory: from consciousness to its physical substrate","cited_arxiv_id":null,"evidence_quote":"It supplies the Integrated Information Theory metrics that LILITH would use to measure consciousness emergence."},{"cited_title":"Elman, Elizabeth A","cited_arxiv_id":null,"evidence_quote":"It supports the developmental learning premise that intelligence emerges through lived experience rather than static training."},{"cited_title":"Bender, Timnit Gebru, Angelina McMillan-Major, and Shmargaret Shmitchell","cited_arxiv_id":null,"evidence_quote":"It defines the limitation of pretrained LLMs as stochastic parrots, the gap LILITH aims to escape."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"It provides the network-science basis for treating modularity as vital to brain function."},{"cited_title":"Bassett and Olaf Sporns","cited_arxiv_id":null,"evidence_quote":"It extends the modular-brain rationale to network neuroscience, grounding the multi-region architecture."}],"review_version":1}