{"id":"ad29d374-d73e-4ad1-958a-7e26b8128242","arxiv_id":"2507.00951","paper_version":3,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":2.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"A broad survey arguing that AGI requires modular, memory-augmented, embodied architectures rather than scaled-up token prediction, with a brief proposal to decompose intelligence into five components.","lead":"This paper reviews how artificial general intelligence (AGI) might move beyond next-token prediction by integrating memory, reasoning, embodiment, and modular agents. It argues that scale alone will not produce human-like intelligence and outlines a roadmap of cognitive and societal considerations.","discovery_kind":"review","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central 'architectural, not parametric' necessity claim is asserted rather than demonstrated; Figure 4 and Table 1 analogies do not rule out non-brain-like monolithic architectures, so the paper overstates what it has shown.","rationale":"The reader correctly identifies the anthropocentric blueprint as the weakest assumption. I refine this: it is not just that non-brain-like alternatives are unconsidered; the paper's central modal claim requires that token-level prediction is structurally incapable, which is an empirical and complexity claim. No formal lower bound or controlled comparison is given. The paper's own internal text contains unfinished markers (Section 5.2.2's 'Section X'; Section 10's placeholder; Figure 8's 'Merge our proposal'), which independently corroborate that the roadmap is incomplete. I credit the survey for useful synthesis, including the generalization theory tables and governance frameworks; those are not what fail. The concern is about the strength of the conclusion, not the value of the collection. Since the reader's conditional verdict already asks for corrections and the survey remains informative, I would not change the verdict; I agree only partially with the reader's formulation because my attack targets the modal 'must' rather than merely the absence of non-brain-like alternatives.","tokens_in":49767,"tokens_out":6804,"duration_ms":191657,"concrete_test":"Run a controlled architectural comparison using benchmarks the paper itself lists in Table 4: with matched training compute, data, and inference budget, evaluate (A) a monolithic next-token-prediction transformer augmented only with RLVR and test-time search (no modular agent decomposition, persistent external memory, or multi-agent coordination) versus (B) a modular brain-inspired agent system on held-out splits of ARC, BIG-Bench, MineDojo, and AgentBench. If system A performs at or above system B, the claim that the limitations are 'architectural, not parametric' is falsified. If such a run is infeasible, a weaker scoping check is to repeat Figure 4 with 50+ diverse proportional-reasoning images and 10+ prompts per model; prompt- or item-sensitive failures would show the Figure 4 result is anecdotal rather than evidence of an architectural ceiling.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's most consequential step is the modal jump in Section 1: current models are 'fundamentally limited' and the missing capabilities are 'architectural, not parametric,' therefore AGI requires integrating modular memory, reasoning, and agency. This necessity claim is load-bearing because the abstract, introduction, and Section 9.2 all state it as the reason scaling and post-training are insufficient. The support offered is (i) one anecdotal Figure 4 showing four LMMs mis-estimating a Christmas-tree height, and (ii) analogical brain-region-to-network mappings in Table 1 and Section 3. The example shows that those particular models, with those prompts, cannot do that visual-proportion task; it does not show that no next-token-predicting architecture with sufficient compute, data, or test-time search could. Table 1 maps regions to network families but includes no controlled comparison with non-brain-inspired designs; the mapping is descriptive, not causal. Section 5.2.2 contains a dangling '(further discussed in Section X)', and Section 10 with Figure 8's 'Merge our proposal' is a placeholder, so the paper's own engineering proposal is never specified. Thus the central thesis is a plausible research hypothesis, not an established result; the 'must' and 'cannot' language overstates the evidence.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper is a cross-disciplinary position/review paper arguing that AGI cannot be reached by scaling token-prediction models alone and that true intelligence requires the integration of modular memory, reasoning, agency, world models, and cognitive scaffolding. It surveys historical AI, brain-inspired computing, generalization theory, reinforcement learning and alignment, agentic architectures, VLMs, recent advances such as Large Concept Models and Large Reasoning Models, benchmarks, and societal/ethical governance. The paper claims to offer a unified framework and roadmap for AGI that synthesizes computational architectures, cognitive neuroscience, and societal alignment, and it uses an anecdotal visual-reasoning example (Figure 4) and brain-region-to-network mappings (Table 1) to support the claim that current models are fundamentally limited by token-level prediction.","tokens_in":50004,"tokens_out":5342,"duration_ms":64133,"significance":"If the central thesis were established, the paper would be a valuable roadmap for redirecting AGI research toward architectural innovation rather than pure scaling. The survey is broad and timely, and it collects useful material in one place: agent communication protocols (MCP, ACP, A2A, ANP), test-time adaptation, Agentic RAG, self-play frameworks such as AZR, and recent model families (LCMs, LRMs, MoE). The paper includes algorithmic sketches and comparative tables that could serve as a convenient entry point for researchers. However, the paper provides no new experiments, machine-checked proofs, parameter-free derivations, or falsifiable predictions; its main claim is supported primarily by analogy and a single anecdotal evaluation. The contribution is therefore taxonomic and synthetic rather than demonstrative, and its central necessity claim is not established.","major_comments":[{"comment":"There are two load-bearing problems with Figure 4. First, Section 3.1.4 states that Figure 4 illustrates a cognitive pipeline using a soccer player example, but the actual Figure 4 shows the Christmas-tree VLM evaluation; the figure and its caption do not match the surrounding text. Second, the caption claims that 'all models... fail to estimate the Christmas tree height correctly,' yet GPT-4o's answer of 195 cm differs from the ground truth of 190 cm by exactly 5 cm, which is the rounding increment requested in the prompt ('to the nearest 5 cm'). GPT-4o's answer is therefore correct under the stated output convention. Since this figure is the only concrete empirical evidence in Section 1 for the paper's central claim that token-prediction models are 'fundamentally limited,' the misreporting directly undermines that claim. Please correct the figure/caption and either provide multiple controlled trials or weaken the claim to 'these four models, on this prompt, were unreliable.'","section":"Figure 4 and Section 3.1.4"},{"comment":"The paper asserts that limitations of current LLMs are 'architectural, not parametric' and that scaling alone 'cannot achieve AGI.' This modal claim is load-bearing because it motivates the entire paper. However, no formal argument or controlled comparison is provided; the cited references [14, 15] concern data minimization and AI-native memory, not scaling limits. Moreover, the paper itself later surveys Decision Transformers, Large Concept Models, Large Reasoning Models, Tree-of-Thoughts, ReAct, and test-time compute—all of which operate within or alongside autoregressive token-prediction frameworks and improve reasoning. The necessity claim is a plausible research hypothesis, but the text states it as an established result. Please rephrase the claim as a hypothesis ('we argue' / 'the evidence suggests') and add a discussion of whether monolithic architectures with test-time search or other non-brain-like designs could achieve the same capabilities.","section":"Section 1 ('Why Further Scaling Will Not Lead to AGI?')"},{"comment":"Section 10, titled 'Our Proposal,' contains only a single paragraph enumerating five intelligence components (spatial, geometrical, sequential, multimodal, social), and Figure 8 includes the placeholder text 'Merge our proposal.' No architecture, training objective, evaluation protocol, or relationship to the surveyed cognitive modules is specified. The paper's own engineering roadmap is therefore never actually presented. Either expand Section 10 with a concrete design and validation plan, or remove the 'proposal' framing and label this section as 'outlook' or 'research agenda.'","section":"Section 10 and Figure 8"},{"comment":"The paragraph under 'Solving Inductive Bias Technique' contains the dangling reference '(further discussed in Section X)' with no target section, indicating an incomplete draft. The paragraph is also too short to justify the heading; it mentions uncertainty estimation and adaptive regularization without connecting them to 'solving inductive bias.' Please fix the cross-reference and either expand the content or revise the heading.","section":"Section 5.2.2"},{"comment":"The brain-region-to-neural-network mappings in Table 1 are descriptive analogies, not evidence that AGI must be organized in this way. The paper uses these mappings (e.g., the 'Key Insight' in Section 3.2.3) to conclude that AGI 'may benefit from' and later 'must' integrate modular, brain-like components. There is no controlled comparison with non-brain-inspired architectures, and the possibility of a monolithic architecture with sufficient compute, data, or test-time search is not considered. This is an anthropocentric modeling choice, not a derivation. Please frame the brain-inspired mappings as heuristics or hypotheses, and explicitly acknowledge that alternative architectural routes to general intelligence are not ruled out.","section":"Table 1 and Section 3.2.3"}],"minor_comments":[{"comment":"The phrase 'lack grounded agency' is grammatically awkward; consider 'lack grounding and agency.'","section":"Abstract"},{"comment":"The agent communication protocols MCP, ACP, A2A, and ANP are referenced as 'Source Link' without proper citations or URLs; please provide full references.","section":"Section 8.1.1"},{"comment":"The term 'techical' in Section 7.3 should be 'technical.'","section":"Section 9.4"},{"comment":"The word 'componensts' in Section 10 should be 'components.'","section":"Section 10"},{"comment":"The word 'empathatically' in Section 9.3 should be 'empathetically.'","section":"Section 9.3"},{"comment":"The glossary entry for ARC appears truncated ('Benchmark that eva...'); the table should be completed.","section":"Table A1"}],"recommendation":"major_revision","confidential_remarks":"The manuscript relies heavily on self-citations, including unpublished arXiv preprints (e.g., refs. [73], [74], [154], [165], [185], [232], [273]) to ground substantive claims. For a survey, I would advise the authors to use independent sources for contested claims and to ensure self-citations are disclosed and not overused. The paper also claims to be 'the first review' to evaluate AGI through three integrated lenses; this novelty claim is unverified and should be softened. The scope is more that of a position/survey paper than an experimental paper, so the fit depends on whether the target journal publishes such broad reviews. After the factual and structural issues are repaired, the paper could be acceptable as a roadmap, but in its current state the central evidence is misreported and the proposal section is unfinished."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThis is a sprawling survey claiming that token-level next-token prediction is architecturally incapable of AGI and that true intelligence requires integrating modular memory, reasoning, and agency. The thesis is plausible, but the paper argues for it by assertion more than evidence. The one empirical demonstration, Figure 4, is wrong: it claims all four LMMs fail the Christmas-tree height task, but GPT-4o's 195 cm answer is exactly correct to the nearest 5 cm of the 190 cm ground truth. That undercuts the anecdote.\n\nWhat the paper does well: it assembles a broad, current map of AGI-related research—memory taxonomies, world models, agentic RAG, LCMs, LRMs, MoE, benchmarks, governance—and organizes it with useful tables, notably the brain-region-to-network mapping. For someone entering the field, it is a passable orientation. The discussion of compression as a bridge between memorization and reasoning is thoughtful and connects to real theoretical work.\n\nSoft spots, in order of severity. First, the central \"architectural, not parametric\" necessity claim is load-bearing but unsupported. The paper offers the flawed Figure 4 and brain analogies that are descriptive, not causal. There is no controlled comparison against non-brain-like monolithic architectures, so the \"must\" and \"cannot\" language overstates what is shown. Second, the paper contains unfinished artifacts: Section 5.2.2 refers to a nonexistent \"(further discussed in Section X)\", and Section 10 is labeled \"Merge our proposal\" in Figure 8, while the actual proposal is a five-item decomposition with no specification. Third, the novelty claim—\"first review through three integrated lenses\"—is a packaging claim, not a contribution. The survey is a synthesis, not a new result. The heavy self-citation is mostly for background and is not itself a problem.\n\nWho is this for? A reader who wants a broad, current overview of AGI-related ideas without deep technical detail. It is not for someone looking for a rigorous argument that scaling is insufficient. I think it deserves a serious referee because a corrected and tightened version could be a useful reference, but as submitted it needs major revision: fix Figure 4, remove or resolve the placeholder, specify the proposal, and rewrite the necessity claim as a hypothesis with clearly marked evidence. I would not cite it in its current form.","headline":"A broad, useful AGI survey that overreaches its central claim; the key empirical exhibit is wrong and the proposal is a placeholder.","tokens_in":50622,"tokens_out":2456,"would_cite":false,"duration_ms":28287,"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 reaching artificial general intelligence requires integrating memory, reasoning, and grounded agency—scale and token prediction alone will not get there.","keywords":["artificial general intelligence","token-level prediction","agentic AI","brain-inspired intelligence","memory and reasoning","world models","information compression","value alignment"],"falsifier":"A concrete test: benchmark a large next-token predictor with no external memory, no explicit world model, and no multi-agent coordination against an equally sized agentic system with those components on long-horizon open-ended tasks; if the monolithic predictor matches or beats the agentic system, the paper's integration claim is falsified. A weaker check is to ablate memory and action modules while holding parameter count fixed and show no improvement in out-of-distribution generalization.","tokens_in":49574,"feed_emoji":"🧠","tokens_out":6071,"duration_ms":67543,"temperature":0.7,"pith_summary":"This paper is a cross-disciplinary synthesis arguing that current AI systems, however fluent, remain anchored to next-token prediction and lack grounded agency, persistent memory, and goal-directed behavior. It claims that true general intelligence arises not from scale alone but from the orchestration of modular, interactive, and self-improving components, with compression acting as the bridge from memorization to reasoning. The paper maps human brain functions to computational modules, reviews agentic retrieval-augmented frameworks, world models, reinforcement learning, and generalization theory, and concludes that architectural change, not further parameter scaling, is the necessary route to AGI. A careful reader should care because this reframes the AGI agenda from model size to cognitive scaffolding and points to concrete design principles for memory, action, and modularity.","feed_headline":"AGI needs memory, reasoning, and agency","feed_subtitle":"The review's claim: AGI will come from modular, self-improving systems with memory and action, not from scaling tokens.","key_machinery":"The load-bearing mechanism is architectural integration: a modular cognitive scaffold in which specialized components—perception, reasoning, memory, action, and world modeling—interact through coordination loops. The paper concretizes this through agentic retrieval–planning–tool-use loops that combine retrieval, planning, and dynamic tool use, and through the mapping of brain regions to neural network models that serves as a design blueprint. It also treats information compression as the bridge between memorization and reasoning: high compression of training data into abstract principles is what enables flexible generalization. This mechanism carries the argument because it is what the paper says scaling lacks: inductive biases for structured reasoning, persistent memory, and self-models.","core_discovery":"On its own terms, the paper's central claim is that token-level prediction is a foundation, not a destination: current models 'remain fundamentally limited by their reliance on token-level prediction and lack of grounded agency.' The argument is that intelligence requires the integration of memory and reasoning—an orchestration of modular, interactive, and self-improving components—in which compression enables adaptive behavior. The paper presents a cognitive blueprint in which brain-region specialization, memory hierarchies, and perception-action loops map onto AI modules, and it argues that agentic architectures with retrieval, planning, tool use, persistent memory, and world models are the concrete instantiation of that blueprint. It concludes that reaching AGI means building systems that are grounded, embodied, memory-augmented, and value-aligned, not merely larger.","pith_inferences":["If the integration thesis is correct, an ablation test follows: holding parameter count fixed, systems with external memory and action loops should outperform monolithic predictors on long-horizon, open-ended tasks—this comparison is implied but not run in the paper.","The compression-as-reasoning view can be pushed to a testable continuum: models that compress training data more (by minimum-description-length-style measures) should show stronger out-of-distribution reasoning, which could be measured directly.","The brain-region mapping suggests a falsifiable architectural recipe: adding modules that mimic episodic replay and planning-like prefrontal loops should improve continual learning more than equal-parameter generic additions.","A societal consequence the paper leaves implicit: if AGI depends on agency and memory, then evaluation should shift from single-turn benchmark accuracy to sustained autonomy and goal-directed behavior over time."],"forward_implications":["Larger parameter counts and better post-training will hit diminishing returns; further AGI progress depends on adding persistent memory, world models, and goal-directed agency to architectures.","Agentic retrieval–planning–tool-use systems become a primary design route rather than a peripheral add-on to language models.","Vision-language models should be treated as perception modules embedded in agentic systems, not as standalone general intelligences.","Generalization efforts should focus on compression, test-time adaptation, and training-free methods as core capabilities, not merely regularization tricks.","AGI development must proceed together with societal governance, cognitive-debt mitigation, and energy-aware design, because the same integration that enables intelligence also raises systemic risks."],"supporting_citations":[{"why":"Supplies the chain-of-thought decomposition the paper adopts as a core reasoning technique.","marker":"[20]"},{"why":"Provides multi-path search with lookahead and self-evaluation, used as the tree-of-thought reasoning example.","marker":"[21]"},{"why":"Interleaves reasoning with environment actions, the basis of the agentic retrieval–planning–tool-use loop.","marker":"[18]"},{"why":"Reframes reinforcement learning as trajectory-level sequence modeling, the paper's example of going beyond next-token generation.","marker":"[16]"},{"why":"Maps brain functions to AI modules, the blueprint from which the paper derives its design principles.","marker":"[94]"},{"why":"Introduces a society-of-mind architecture of communicating specialized agents, grounding multi-agent coordination.","marker":"[75]"},{"why":"Establishes the information-bottleneck compression principle the paper uses to link compression to generalization.","marker":"[187]"},{"why":"Introduces retrieval-augmented generation, the central mechanism for persistent external memory and factual grounding.","marker":"[216]"}],"fun_headline_variants":["Scaling tokens won't yield AGI; memory and agency will","AGI needs memory reasoning agency, not just scale","Beyond token prediction: AGI requires memory and action","Token-scale alone won't build AGI; grounded agency will","AGI arises from memory and reasoning, not token counts"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the human cognitive blueprint—specialized brain regions, memory hierarchies, and perception–action loops—is the necessary template for general intelligence, so a system without these components cannot reach AGI; if a radically non-brain-like architecture succeeded, the paper's conclusion would not follow.","fun_headline_variants_meta":{"raw":{"variants":["Scaling tokens won't yield AGI; memory and agency will","AGI needs memory reasoning agency, not just scale","Beyond token prediction: AGI requires memory and action","Token-scale alone won't build AGI; grounded agency will","AGI arises from memory and reasoning, not token counts"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000378,"raw_usage":{"total_tokens":2043,"prompt_tokens":1007,"completion_tokens":1036,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":623,"completion_tokens_details":{"reasoning_tokens":954}},"tokens_in":623,"tokens_out":1036,"duration_ms":9987,"temperature":1.0,"reasoning_tokens":954,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T21:00:27.588263+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A concrete test: benchmark a large next-token predictor with no external memory, no explicit world model, and no multi-agent coordination against an equally sized agentic system with those components on long-horizon open-ended tasks; if the monolithic predictor matches or beats the agentic system, the paper's integration claim is falsified. A weaker check is to ablate memory and action modules while holding parameter count fixed and show no improvement in out-of-distribution generalization.","supporting_citations":[],"review_version":1}