{"id":"8e0f26ee-e557-4133-a9cb-4d183d7df2c2","arxiv_id":"2607.26179","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"A synthesis paper argues that LLMs and humans converge on five core principles of cognitive organization, casting them as 'cognitive cousins' rather than alien intelligences.","lead":"This paper argues that large language models are not 'alien' intelligences but share core organizing principles with human cognition — dual-process reasoning, production-like architecture, predictive learning, and reward-driven control. It assembles evidence across five dimensions to support a 'cognitive cousins' view that reframes the AI-human comparison.","discovery_kind":"review","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Production-system analogy in §4 is under-constrained; canalization in §5 is asserted and does not distinguish transformers from generic residual/sequence models.","rationale":"The reader's weakest_assumption correctly identifies architectural canalization as the load-bearing premise. My stress-test sharpens this: the premise fails not just because it is asserted without direct evidence, but because the production-system-like description of transformers in §4 is too under-constrained to make canalization a substantive or testable claim. The same residual-stream description applies to ResNets, LSTMs with gated updates, and state-space models, so without a precise formal criterion it cannot explain why transformers specifically should converge on human representations. The paper's own example of a fully connected network learning convolutional structure (Ingrosso & Goldt, ref 125) suggests task-driven convergence can occur independent of architecture, undermining the need for canalization over contravariance. This is a real soft spot in the central argument, but it does not warrant rejecting the paper as a perspective piece; the reader's CONDITIONAL verdict already captures the need to moderate the strong architectural claims. I also considered training-data leakage as a competing concern, but the paper's internal-representation evidence (e.g., neural alignment, reading-time prediction) is less exposed to that confound, and the engineering-history argument for prediction-based learning is independent of human annotations. Therefore the canalization/production-analogy issue remains the single most load-bearing concern, and the existing CONDITIONAL verdict should stand.","tokens_in":17264,"tokens_out":7270,"duration_ms":86272,"concrete_test":"Run a controlled architecture comparison on the same cognitive task suite: train (a) a transformer and (b) a deliberately non-production-like sequence model (e.g., an LSTM or state-space model) on identical data, then measure representational alignment to human neural/behavioral data using the methods of refs 100/104. If both show comparable alignment, production-system-like architecture is not doing causal work; canalization as proposed in §5 is unsupported and convergence is explained by contravariance/training data.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim depends on the architectural-canalization premise in §5: transformer architectures are production-system-like (persistent workspace + conditional operators) and therefore bias learning toward human-like solutions. But the §4 analogy is defined so loosely that it cannot do this work. Any residual network—ResNet, LSTM with gated state updates, state-space model—can be described as a workspace iteratively updated by context-sensitive operators. 'Condition matching' in attention/MLP blocks is just continuous query-key/value similarity, not a discrete production match; the paper's own caveat that transformers are 'graded, differentiable' versions strips the analogy of the properties (modularity, recombination, parallelism) that made production systems distinctive. Without a non-vacuous criterion for what counts as production-system-like, architectural canalization cannot explain why transformers (rather than any sequence model) align with human representations. The paper even cites Ingrosso & Goldt (ref 125), where a fully connected network converges to convolutional structure under task pressure—evidence that task constraints (contravariance) rather than shared architecture can produce convergence, undercutting canalization's necessity. If the analogy is vacuous, the 'shared computational family' thesis loses its architectural pillar.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper argues against the 'alien intelligence' framing of large language models. It claims that, despite differences in substrate, learning history, and environment, humans and contemporary LLM-based systems converge on several core principles of cognitive organization: dual-process inferential organization, production-system-like computational architecture, aligned representational structure, prediction-driven learning, and reinforcement-learning-like goal-directed control. The authors review a wide range of empirical and theoretical literature, propose contravariance and a new 'architectural canalization' mechanism to explain representational alignment, and conclude that humans and LLMs are 'cognitive cousins'—different members of a shared computational family. The paper is a literature synthesis rather than a new empirical study.","tokens_in":17530,"tokens_out":3907,"duration_ms":46210,"significance":"If the convergence thesis holds, the paper provides a valuable counterweight to the widely held assumption that LLM-human similarities are superficial or merely anthropomorphic. Its strengths are its breadth of relevant literature, its explicit acknowledgment of important differences (sample efficiency, limited RL infrastructure), and its framing of the debate in terms of mechanistic levels rather than behavioral mimicry. The paper also makes potentially testable claims, e.g., that architecture should modulate brain/behavioral alignment, and it honestly flags open empirical questions. However, the central explanatory mechanism—architectural canalization—is currently asserted rather than demonstrated, and it is load-bearing for the 'shared computational family' conclusion. The paper is best suited as a perspective/review that will inform future work, but the strength of its conclusions currently exceeds the support for its main mechanistic premise.","major_comments":[{"comment":"The production-system analogy is too under-constrained to support the canalization premise. In §4, the defining features are 'a persistent representational workspace [modified] by a large population of context-sensitive operators,' with attention heads and MLPs described as 'graded, differentiable' condition matching. As stated, this description applies equally to ResNets, gated RNNs, and state-space models, so it does not distinguish transformers within the class of neural sequence models. §5 then relies on this analogy: 'two systems built around similar production-system-like architectures... will tend to converge on similar solutions.' Without a non-vacuous criterion (e.g., modularity, recombinability, compositionality) and evidence that the residual-stream architecture specifically biases learning toward human-like representations, the canalization explanation cannot carry the weight","section":"§4–§5"},{"comment":"The 'architectural canalization' hypothesis is asserted with citations to Waddington and Ariew, but no mechanistic account is given for how the residual-stream architecture biases gradient descent toward particular representational solutions. Biological canalization concerns developmental buffering against perturbations; its transfer to transformer training requires a model of the learning dynamics—not just a loose analogy. The claim is central: it is one of two explanations for representational alignment, and the conclusion explicitly invokes 'learning that unfolds within a shared computational architecture.' Without support, the explanatory force reduces to contravariance, which is already established and does not require shared architecture. Please either supply a mechanistic model or direct evidence (e.g., comparing brain/behavioral alignment across transformers, state-space models,","section":"§5 (Architectural Canalization)"}],"minor_comments":[{"comment":"The section heading reads 'Reinforcement Learning and Reinforcement Learning and Agentive Control'—the duplicated phrase should be removed.","section":"§7 heading"},{"comment":"Minor grammatical issue: 'the RL infrastructure current deployed in artificial systems' should be 'currently deployed.'","section":"§7"},{"comment":"The Ingrosso & Goldt analogy is introduced as supporting the 'same destination despite different routes' claim, but the discussion in the same paragraph (and ref 126) emphasizes sample-efficiency differences. It would help to explicitly distinguish the two uses of this example: task-driven convergence of representations vs. efficiency of the learning route.","section":"§6"},{"comment":"The behavioral parallels section includes vision-language models and text-to-image models, not just LLMs. A brief clarification of the intended scope of 'LLM-based systems' at first mention would avoid ambiguity.","section":"§2"}],"recommendation":"major_revision","confidential_remarks":"This is a competent and well-cited synthesis. The main issue is that the strongest conclusion—'shared computational family'—rests on the canalization premise, which is currently an analogy rather than a demonstrated mechanism. The authors should be encouraged to either add evidence or soften the conclusion. No citation-pattern concerns; the self-citations are appropriate to the argument's grounding in their prior work."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague—\n\nThe paper is a well-written, well-cited position piece that makes a serious case for 'cognitive convergence' between LLMs and human cognition. It's not a new empirical result, and it doesn't pretend to be; its value is in organizing a large literature into five dimensions and in proposing a mechanism—'architectural canalization'—to explain representational alignment. That's a genuinely useful framing, and it deserves a real referee.\n\nThe strengths are real. The dual-process section is nicely tied to in-weight/in-context learning and chain-of-thought; the prediction-learning section makes good use of the BabyLM sample-efficiency numbers; the RL section is admirably honest about how sparse and brittle current artificial RL infrastructure is. The Locke point—distinguishing the process of furnishing from the arrangement of the furniture—is a good rhetorical move and conceptually sound.\n\nThe soft spots are equally real, and they're mostly concentrated in the middle of the paper. The production-system analogy in §4 is so loosely drawn that it's hard to falsify: any residual network with a persistent state and context-sensitive updates counts as 'production-system-like,' whether it's a ResNet, an LSTM, or a state-space model. That strips the analogy of the modularity/recombination properties that made production systems distinctive. The canalization hypothesis in §5 inherits this problem: it's asserted, not demonstrated. The paper itself cites Ingrosso & Goldt, where a fully connected network converges to convolutional structure purely through task pressure—that's evidence that contravariance alone can produce convergence, which undercuts the need for shared architecture. The paper doesn't engage with that tension.\n\nAlso, the treatment of divergences is selective. Sample efficiency and RL underdevelopment are acknowledged, but other well-known LLM-human differences (e.g., reversal curse, catastrophic forgetting, lack of core knowledge) are either absent or folded into 'underdevelopment.' That's a legitimate rhetorical move in a position paper, but it makes the strong conclusion—'shared computational family'—feel more secure than the evidence supports.\n\nMy overall take: this is a solid perspective piece, not a breakthrough. The central claim is plausible but overstated. The canalization mechanism needs to be either sharpened into something testable or demoted to a speculation. A good reviewer will catch this.\n\nRecommendation: send it to peer review, but expect revision. It's the kind of paper that generates useful discussion, and with a more disciplined mechanistic claim it could be a well-cited piece. I'd bring it to reading group; whether I cite it depends on how much of the strong claim survives.","headline":"A plausible, well-sourced synthesis arguing LLMs and humans converge on core cognitive principles, but its central mechanistic claim is asserted rather than demonstrated.","tokens_in":17973,"tokens_out":3258,"would_cite":true,"duration_ms":32026,"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":"LLMs and human minds are cognitive cousins, not alien intelligences.","keywords":["LLM cognition","cognitive convergence","dual-process theory","production systems","representational alignment","prediction-driven learning","reinforcement learning","architectural canalization"],"falsifier":"A concrete test: take a non-production-like architecture (e.g., a fully connected or convolutional network without a residual stream) and train it to match an LLM's performance on human-language tasks. If it shows equally strong alignment with human neural representations and behavioral error profiles under representational similarity analysis, then task demands alone, not shared architecture, explain the convergence, and the canalization claim is undermined. Alternatively, find a transformer variant whose residual stream is removed or radically altered but that still matches LLM behavior; if","tokens_in":17109,"feed_emoji":"🧠","tokens_out":5348,"duration_ms":47465,"temperature":0.7,"pith_summary":"The paper argues that the widespread view of LLMs as 'alien intelligences' is mistaken. Despite significant differences in physical substrate, training history, and environment, humans and contemporary LLM-based systems converge on five principles of cognitive organization with longstanding support in cognitive science: dual-process inference, production-system-like architecture, representational alignment, prediction-driven learning, and reinforcement-learning-like goal-directed control. The authors claim these are not superficial behavioral similarities but structural correspondences at the level of computational organization. To explain why such different systems converge, they propose two complementary principles: contravariance, which says harder tasks leave fewer viable solutions, and architectural canalization, which says a shared production-system-like architecture biases learning toward similar internal solutions. If the thesis is right, LLM–human parallels are evidence of genuinely shared computational principles, making humans and LLMs 'different members of a shared computational family.'","feed_headline":"LLMs and brains run on the same five cognitive principles","feed_subtitle":"Human and machine intelligence converge on shared inference, architecture, representation, learning, and control.","key_machinery":"The central mechanism is the transformer's residual stream: a persistent, high-dimensional vector at each token position that attention heads and MLP blocks read from and incrementally write back to, creating a graded condition-action system with modular, composable updates. The paper argues this is structurally analogous to the working memory and production rules of cognitive architectures. Two explanatory principles carry the convergence: contravariance (demanding tasks shrink the space of viable internal solutions) and architectural canalization (systems built on a shared production-system-like architecture will tend to converge on similar solutions even under different training regimes).","core_discovery":"The paper's central claim is that LLM-based systems and human cognition are 'importantly different members of a shared computational family.' The authors ground this in five correspondences: both exhibit dual-process organization, with fast compiled responses alongside slower serial reasoning that uses externalized scratchpads; both are organized like production systems, where a persistent workspace is modified by many context-sensitive conditional operators; internal representations align, as shown by LLM internal units encoding linguistic constructs that predict human reading times and brain responses, with model depth tracking the cortical hierarchy; both learn through prediction and erro","pith_inferences":["If architectural canalization holds, a testable prediction follows: two transformer-based systems trained on very different corpora (e.g., text-only vs. multimodal) should still develop similar human-aligned internal representations on shared tasks, whereas a non-production-like architecture trained to the same performance should not. This could be probed with representational similarity analysis ","The thesis implies that the space of feasible intelligent cognition is narrower than often assumed, which bears on AI safety: if genuinely alien forms of intelligence are rare, value alignment may be more tractable than if intelligence can take radically unconstrained forms.","A useful extension would be to formalize 'canalization' as a quantitative bias: measure the distribution of internal solutions across random initializations and training runs, and test whether production-system-like architectures show lower variance and stronger attraction to human-like solutions than equally powerful non-residual architectures.","The paper's treatment of RL suggests a specific empirical target: current frontier models should be evaluated for whether longer-horizon RL training induces functionally distinct controllers and metacontrol (e.g., arbitration between habitual and deliberative policies), which would further confirm the convergence."],"forward_implications":["If convergence is genuine, behavioral similarities—like comparable garden-path difficulty, serial-position effects, and conjunctive-search costs—should be treated as evidence of shared mechanisms, not as products of text mimicry.","The dual-process mapping predicts that LLM internal 'thinking token' counts should track human reaction times across many reasoning tasks, a pattern the paper reports with strong correlations.","Representational alignment implies that next-word prediction performance should continue to predict human neural and reading-time measures, and that layerwise activations should map onto cortical hierarchy, as current data show near the noise ceiling.","The sample-efficiency gap between LLMs and children is a difference in the learning process, not necessarily the learning target; adding structural inductive biases, world models, and curiosity-like exploration may reduce the gap while preserving cognitive convergence.","In the RL domain, the paper's claim implies that current artificial limitations—reward hacking, sparse rewards, absent homeostatic signals—are engineering gaps rather than evidence of a fundamentally alien agency, and may be narrowed by richer training experience and reward design."],"fun_headline_variants":["LLMs share five core cognitive principles with humans","Machine and human cognition converge on five shared principles","LLMs and human minds: same computational family, five shared principles","LLMs and brains: different members of the same cognitive family","Five cognitive principles unite LLM and human intelligence"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The argument's load-bearing premise is that transformer LLMs are genuinely production-system-like in their computational organization, so that a shared architecture—not merely shared task demands—biases both humans and LLMs toward the same internal solutions; if the residual stream is not truly analogous to human working memory plus production rules, the deep convergence thesis loses its mechanistic support.","fun_headline_variants_meta":{"raw":{"variants":["LLMs share five core cognitive principles with humans","Machine and human cognition converge on five shared principles","LLMs and human minds: same computational family, five shared principles","LLMs and brains: different members of the same cognitive family","Five cognitive principles unite LLM and human intelligence"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000474,"raw_usage":{"total_tokens":2136,"prompt_tokens":634,"completion_tokens":1502,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":378,"completion_tokens_details":{"reasoning_tokens":1424}},"tokens_in":378,"tokens_out":1502,"duration_ms":11003,"temperature":1.0,"reasoning_tokens":1424,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T00:32:33.982635+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A concrete test: take a non-production-like architecture (e.g., a fully connected or convolutional network without a residual stream) and train it to match an LLM's performance on human-language tasks. If it shows equally strong alignment with human neural representations and behavioral error profiles under representational similarity analysis, then task demands alone, not shared architecture, explain the convergence, and the canalization claim is undermined. Alternatively, find a transformer variant whose residual stream is removed or radically altered but that still matches LLM behavior; if","supporting_citations":[],"review_version":1}