Pith. sign in

REVIEW 3 major objections 5 minor 13 references

From Basic Affordances to Symbolic Thought: A Computational Phylogenesis of Biological Intelligence

T0 review · 3 major / 5 minor · reviewed 2026-08-05 · deepseek-v4-flash

Pith's one-line read The paper claims that two distinct capacities for hierarchical integration—multi-place predicates and structure mapping—are the minimal extra ingredients, beyond dynamic binding, that make symbolic thought possible, and that neither can sub

desk verdict First orthogonal ablation of LISA's relational capacities; careful but single-run simulations, and the 'minimal requirements' claim reaches past the evidence. read the letter →

arxiv 2508.15082 v1 pith:PZEUOI7E submitted 2025-08-20 cs.NE cs.AIcs.LGq-bio.NC

classification cs.NEcs.AIcs.LGq-bio.NC
keywords symbolicthoughtdynamicbindingmulti-placepredicatesstructuremappingrelationalreasoningcognitivearchitecturegeneralizationphylogenesis
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper asks what extra neural machinery, beyond dynamic binding, turns a merely intelligent brain into a symbolic one. It argues that two forms of hierarchical integration—combining role-bindings into multi-place relations, and mapping one relational structure onto another—are jointly necessary and minimally sufficient for basic symbolic thought. To test this, the authors stripped the LISA relational-reasoning architecture of each capacity in turn, yielding four architectures, and ran them on four task types specially built to be unsolvable by featural similarity. The pattern of results is exactly as the hypothesis predicts: tasks requiring a capacity are solvable only by architectures that have it, and neither capacity substitutes for the other. If right, this locates the evolutionary leap to symbolic cognition in the coincidence of two independent innovations.

What carries the argument

The central object is the LISA neural architecture and its 'LISAese' representational scheme, in which propositions are built from token units (objects, predicates, role-bindings, propositions) bound dynamically by synchrony of neural firing. The two capacities under test are: (1) multi-place predicates, which hierarchically compose several role-bindings under a single proposition unit, conferring recursion and unbounded compositionality; and (2) structure mapping, implemented as incremental Hebbian learning of mapping connections between driver and recipient structures, with a one-to-one mapping constraint. Ablating relation formation is done by flattening two-place predicates into single-p

What would settle it

A concrete counterexample would be an architecture with dynamic binding and structure mapping but only single-place predicates that solves the coffee-maker task (binding the affordance to the water tank) after a second exposure or a small amount of additional training, or a nonhuman species that performs one-trial relational transfer across novel objects while evidence shows it never composes multiple role-bindings into a single structure. Either observation would break the claimed joint necessity.

Watch

Extended reading notes

Core claim

Using the LISA model of analogical reasoning as a base, the paper defines four cognitive architectures—dynamic-binding-only (DBO), relations-only (RO), mapping-only (MO), and relations-and-mapping (R&M)—by ablating the capacity to form multi-place predicates (replacing each two-place relation by two single-place predicates with identical semantic features) and/or the capacity to compute structure mappings (setting the mapping-connection learning rate µ to zero). Across 17 simulations spanning four task types, only the architectures possessing the relevant capacity succeeded on tasks that required it: all four architectures solved the DBO affordance task; only RO and R&M solved the relations

Load-bearing premise

The entire argument rests on assuming that LISA's specific algorithmic implementations—Hebbian structure mapping with the one-to-one constraint, and proposition-unit hierarchical binding—faithfully capture the corresponding cognitive capacities in real brains, so that removing them in the simulation is equivalent to their absence in biology.

Editorial extensions

If this is right

  • An animal with dynamic binding alone can still generalize basic affordances to novel objects and environments, explaining how cats, bees, and other non-symbolic species navigate new situations.
  • Possessing relational knowledge in memory does not help an organism that cannot perceive relations in the current stimulus, a finding the paper turns into a novel prediction about patients with Balint's syndrome.
  • Modern machine-learning systems that rely on entangled, conjunctive representations may need enormous data to reach generalization that a symbol-capable architecture reaches in one example; the paper suggests biologically plausible architecture principles for more sample-efficient AI.
  • If symbolic thought requires two (and maybe three, with CWSG) independent innovations, its late and rare appearance in evolution is expected: the capacities must co-occur in the same organism.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • An extension of the paper's logic: the two capacities may have separable neural correlates, so one could look for dissociations between relational perception and cross-domain mapping in brain-damaged patients.
  • A testable extension would apply the same ablation logic to transformer models, selectively degrading attention (soft dynamic binding) versus cross-attention or positional encoding (soft mapping), to see whether the same non-compensatory pattern appears.
  • The account suggests a sharp prediction about phylogeny: species in the 'RO' or 'MO' cells should show a qualitative gap in task repertoire, not a graded one, when probed with one-shot relational transfer.
  • If the two mutations are truly independent, symbolic cognition should have arisen multiple times convergently wherever both capacities co-occur; the paper's own timeline argument implies a strong constraint on that convergence.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The paper proposes that two forms of hierarchical integration—multi-place predicates and structure mapping—are minimal requirements, on top of dynamic binding, for basic symbolic thought. The authors take the LISA model of relational reasoning and ablate these two capacities orthogonally, yielding four architectures (DBO, RO, MO, R&M). They also construct four tasks (DBO, RO, MO, R&M) that cross whether the task requires multi-place predicates and/or structure mapping. In 17 simulations, all architectures succeed on the DBO task; only relation-capable architectures (RO, R&M) succeed on the RO task; only mapping-capable architectures (MO, R&M) succeed on the MO task; and only the fully capable R&M architecture succeeds on the R&M task. The authors interpret this pattern as evidence that multi-place predicates and structure mapping are independent, non-compensable capacities, jointly necessary for basic symbolic thought.

Significance. If the conclusion holds, the paper offers a computationally explicit dissociation between two hypothesized building blocks of symbolic cognition, with implications for comparative psychology, evolutionary accounts of symbolic thought, and bio-inspired AI. The study has notable strengths: the code is publicly available; the tasks are designed with abstract, featurally uninformative stimuli; the 2x2 design is systematic; and it yields at least one concrete, falsifiable prediction (Balint's patients should fail allocentric relation-stacking judgments). The simulations are also carefully controlled so that object semantics never disambiguate the correct inference.

major comments (3)
  1. [Relations Only Task (RO) and Figure 4] The categorical claim that structure mapping cannot compensate for missing multi-place predicates rests on a single parameterization: the MO architecture run with the default mapping-learning rate μ=0.9 and only two presentations of the Perception propositions ('double time'). Since LISA's mapping connections are learned incrementally and the one-to-one constraint propagates previously discovered correspondences, the outcome could depend on μ, the number of presentation cycles, or the firing order. The paper reports no sweep over μ, no variation in cycle count, and no multiple random seeds (the model includes noise). A stronger case for 'neither can compensate' requires either a parameter robustness analysis (e.g., μ from 0.1 to 0.9, 1–10 cycles, several seeds) or an analytical argument that the initial featural bias cannot be overcome by any setting of the mapping mechanism. Without thi
  2. [Taking Away LISA's Abilities / Simulations] The necessary-condition conclusion is stated at a general level—minimal requirements for biological symbolic thought—but the evidence comes from one implementation family: LISA's proposition/role-binding units for multi-place predicates and its Hebbian mapping connections for structure mapping. Alternative neural implementations (e.g., tensor-product binding, complex-valued distributed codes, graph-matching approaches) might realize the abstract capacities differently, and it is not shown that the same dissociation would hold there. The authors should either implement the same task battery in a second, architecturally distinct model, or explicitly state the computational-level assumptions under which LISA's implementation is representative. As written, the simulations establish a dissociation within LISA but leave open whether it is a property of the specific algorithmic implementation r
  3. [Four Kinds of Tasks] The tasks are defined in terms of the very capacities under test: they are constructed by 'orthogonally crossing whether a task requires multi-place predicates, structure mapping, or both.' Consequently, the mapping from task to required capacity is assumed rather than independently established, and the broad success/failure pattern is partly analytic. The informative part is the feature-control and the cross-dissociation (MO fails RO, RO fails MO), but the paper should more sharply separate the analytic claim (a task that truly requires hierarchical structure can only be solved by an architecture with such structure) from the empirical claim (LISA's mechanisms implement that structure in a way relevant to biological cognition). This would make the sense in which the results support 'minimal requirements' clearer and avoid the appearance of circularity.
minor comments (5)
  1. [Table 2.1] In the 'Semantic coding of objects' block, the rows list 'Distractor2' twice; the second should presumably read 'Distractor3'.
  2. [References] Several typographical errors: 'Claton & Emery' should be 'Clayton & Emery'; 'Faulkenhainer' in text and one reference should be 'Falkenhainer'; 'Langrage' in the Deacon reference should be 'Language'; 'Franics' should be 'Francis'.
  3. [Figure 3] The text states that all four architectures performed the DBO task 'with equal speed and accuracy,' but no quantitative measure is reported. Please provide iteration counts or a defined convergence criterion so the claim is verifiable.
  4. [Mapping Only Task (MO)] The text says each proposition in Perception fires once, yet Figure 5 shows mapping connections being learned between iterations 270 and 380. Please clarify what 'once' means (e.g., one presentation per proposition, with multiple oscillation cycles per presentation) so the temporal dynamics are consistent with the description.
  5. [Figure 8] The LLM-generated image examples are anecdotal and not systematically scored. They are framed as illustrative in the General Discussion, but consider labeling them explicitly as demonstrations rather than evidence, and perhaps move them to supplementary material if the journal prefers a stricter separation.

Circularity Check

3 steps flagged · score 6.0 of 10

The 'minimal requirements' conclusion is largely entailed by the self-definitional task-architecture matrix: tasks are defined as requiring the very capacities whose necessity is claimed, with all featural alternatives removed.

  1. self definitional [Abstract; Simulations, 'Task Overview']
    "The simulations were as generic as possible, in that no task could be performed based on any diagnostic features, depending instead on the capacity for multi-place predicates and structure mapping. ... The Relations Only (RO) task requires the capacity to represent multi-place relations and dynamically bind them to their arguments but does not require structure mapping. ... The Mapping Only (MO) task requires dynamic binding and structure mapping but does not require multi-place relations."

    The architectures are defined by the presence/absence of multi-place predicates and structure mapping, and the tasks are defined as 'requiring' exactly those same capacities. The observed pattern—only architectures with capacity X succeed on the task defined as requiring X—is therefore a direct consequence of the design matrix. The paper's central conclusion that multi-place predicates and structure mapping are minimal requirements for symbolic thought restates these operational definitions rather than providing independent evidence for them.

  2. self definitional [Simulations, 'Relations Only Task', 'Logic of the Simulations' and 'Results and Discussion']
    "the semantic features of the predicates, relational roles, and objects were all configured so that, based only on the semantic overlap between the relational roles and/or single-place predicates, the Critical object in Perception (the Toy) would map to the Distractor object in Memory (a Fork), resulting in an incorrect inference ... But any cognitive architecture that can exploit the hierarchical structure of the multi-place smaller-than () relation should map the Critical object (the Toy) in Perception to the Target object (the Mouse) in Memory ... The MO architecture did no better firing eac"

    The stimulus was deliberately constructed so that all featural cues favor the wrong mapping; the only route to the correct answer is the hierarchical proposition unit that defines multi-place predicates. An architecture lacking that unit (MO) is therefore guaranteed to fail, independent of its mapping ability. The claimed non-compensability of structure mapping for missing multi-place predicates is not discovered from the simulation but is built into the task design, since no alternative cue or parameter regime was explored.

1 more flagged steps
  1. self definitional [Simulations, 'Mapping Only Task', 'Logic of the Simulation' and 'Results and Discussion']
    "A cognitive architecture that can learn structure mappings would therefore learn that the Critical object corresponds to the Target object ... However, to a cognitive architecture that can learn structure mappings, the Critical object is already known to correspond to the Target object in Memory (based on the prior mapping of vision1 (Critical) to memory1 (Target)). Given the 1:1 mapping constraint on structure mapping, this prior mapping will tend to favor the mapping of vision2 (Critical) to memory3 (Target). ... the only architectures that solved this task were those capable of structure ma"

    The task logic specifies that success requires using previously learned mapping connections to override featural similarity. Architectures without mapping (RO) have no mechanism to carry the first correspondence forward, so their failure is entailed by the stated logic. The conclusion that multi-place predicates cannot compensate for missing structure mapping is therefore a restatement of the rule that the task can only be solved via mapping connections, not an independent demonstration of non-compensability.

full rationale

The central derivation chain is: (i) define architectures by removing multi-place predicates and/or structure mapping from LISA; (ii) define tasks as 'requiring' exactly those capacities, with featural cues configured to be misleading; (iii) observe that architectures lacking a capacity fail the task defined over that capacity; (iv) conclude the capacity is a minimal requirement for symbolic thought. Step (iv) is already contained in step (ii): a task that is defined as requiring multi-place relations and that has all alternative cues removed will, by construction, be unsolvable by an architecture defined as lacking multi-place relations. The non-compensability results (MO fails RO; RO fails MO) likewise follow from the engineered stimulus statistics rather than from a search over possible mapping or relational strategies; the paper reports a single parameterization (e.g., µ=0.9, 'double time') with no sweep. This is the self-definitional pattern: the 'prediction' is an unfolding of the design matrix. That said, the simulations do provide existence proofs that LISA's specific mechanisms can implement the target capacities, and the Balint's-syndrome prediction is not entailed by the task definitions, so the circularity is partial rather than total. No load-bearing self-citation or imported uniqueness theorem was found; the main issue is the self-definitional alignment of tasks and architectures.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The central claims rest on LISA's specific mechanisms and hand-crafted stimulus semantics. There are no fitted parameters to behavioral data, but the architecture and tasks are constructed by the authors so that the predicted pattern is strongly favored by design. The paper does not introduce new physical or ontological entities; it reconfigures an existing model.

free parameters (4)
  • mu (mapping connection learning rate) = 0.9 (default), 0 (ablated)
    Controls whether structure mapping connections are learned. Set from prior LISA work, not fitted to data in this paper. Central to the manipulation of mapping capacity.
  • Hand-coded semantic feature sets = Various (e.g., vision1: [v1...v6])
    The overlap among semantic features of predicates and objects was hand-designed so that features would not disambiguate tasks, and in some cases to bias toward distractors. These choices determine which mappings are favored and thus the success or failure of each architecture.
  • LISA integration time constants = Not specified in paper (inherited from Hummel & Holyoak, 1997)
    Leaky integrator time scales for object, predicate, role-binding, and proposition units. Fixed by the prior model, not varied here.
  • Firing order and iteration counts (e.g., double time for MO task) = e.g., MO task fired each proposition twice
    Experimental design choices that affect the time traces and whether mappings learned during the first firing can influence the second firing.
assumptions (5)
  • domain assumption Dynamic binding (by synchrony of firing) is necessary for symbolic thought.
    The paper takes this as established by prior evidence and does not test it; it is the foundation on which the 'beyond dynamic binding' hypothesis is built (Beyond Dynamic Binding section).
  • domain assumption LISA's implementation of dynamic binding, multi-place predicates, and structure mapping is representative of the corresponding neural and cognitive capacities.
    The entire ablation logic assumes that removing these mechanisms in LISA models their absence in biological brains (Simulations, 'Taking Away LISA's Abilities').
  • domain assumption Replacing a two-place relation by two single-place predicates with identical semantic features preserves everything except hierarchical integration.
    Used to construct the DBO and MO architectures; if the replacement changes other computational properties, the ablation is confounded (Simulations, first section).
  • domain assumption Binding the Affordance semantic unit in synchrony with the Critical object is a valid measure of making the correct inference.
    Success is operationalized as synchronous firing of these units in the Simulation Format section; this is a weaker inference than full symbolic inference, and the paper acknowledges it is inadequate for complex symbolic thought.
  • standard math A representational system with n elements and multi-place predicates plus recursion can express unbounded structures, whereas single-place binding is limited to the finite powerset 2^n.
    Mathematical claim in the 'Beyond Dynamic Binding' section used to argue that multi-place predicates are qualitatively necessary for open-ended compositionality.

how reviews work

0 comments
Cite this review

Pith. "Pith review of From Basic Affordances to Symbolic Thought: A Computational Phylogenesis of Biological Intelligence." pith.science (2026). https://pith.science/paper/PZEUOI7E

@misc{pith2026250815082,
  author       = {Pith},
  title        = {Pith review of: From Basic Affordances to Symbolic Thought: A Computational Phylogenesis of Biological Intelligence},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/PZEUOI7E}},
  note         = {Machine review of arXiv:2508.15082}
}
read the original abstract

What is it about human brains that allows us to reason symbolically whereas most other animals cannot? There is evidence that dynamic binding, the ability to combine neurons into groups on the fly, is necessary for symbolic thought, but there is also evidence that it is not sufficient. We propose that two kinds of hierarchical integration (integration of multiple role-bindings into multiplace predicates, and integration of multiple correspondences into structure mappings) are minimal requirements, on top of basic dynamic binding, to realize symbolic thought. We tested this hypothesis in a systematic collection of 17 simulations that explored the ability of cognitive architectures with and without the capacity for multi-place predicates and structure mapping to perform various kinds of tasks. The simulations were as generic as possible, in that no task could be performed based on any diagnostic features, depending instead on the capacity for multi-place predicates and structure mapping. The results are consistent with the hypothesis that, along with dynamic binding, multi-place predicates and structure mapping are minimal requirements for basic symbolic thought. These results inform our understanding of how human brains give rise to symbolic thought and speak to the differences between biological intelligence, which tends to generalize broadly from very few training examples, and modern approaches to machine learning, which typically require millions or billions of training examples. The results we report also have important implications for bio-inspired artificial intelligence.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

13 extracted references · 8 canonical work pages

  1. [4]

    double time

    to give the architecture the opportunity to solve the task (during the second firing of the propositions) based on any mappings it discovered (during the first firing of the propositions). As before, those architectures with the capacity for structure mapping (MO and R&M) had µ (the mapping connection learning rate) set to the default of 0.9, whereas thos...

  2. [11]

    L., Bowers, J., Costa, R

    Montero, M. L., Bowers, J., Costa, R. P., Ludwig, C. J. H., Malhotra, G. (2022), Lost in latent space: Examining failures of disentangled models at combinatorial generalization. Advances in Neural Information Processing Systems, 35, 10136-10149. Moreno, A. M., de Souza, D. D. G., & Reinhard, J. (2012). A comparative study of relational learning capacity i...

  3. [12]

    Minds, Brains, and Programs

    Searle, J. (1980). “Minds, Brains, and Programs.” Behavioral and Brain Sciences 3, 417-424. Shastri, L., & Ajjanagadde, V. (1993). From simple associations to systematic reasoning: A connectionist representation of rules, variables and dynamic bindings using temporal synchrony. Behavioral and Brain Sciences, 16(3), 417-494. https://doi.org/10.1017/S014052...

  4. [43]

    Doumas, L. A. A., Puebla, G., Martin, A. E. & Hummel, J. E. (2022). A theory of relation learning and cross-domain generalization. Psychological Review. https://doi.org/10.1037/rev0000346 Esser, P., Rombach, R., & Ommer, B. (2021). Taming transformers for high-resolution image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pat...

  5. [1657]

    Skinner, B. F. (1957). Verbal behavior. Appleton-Century-Crofts. https://doi.org/10.1037/11256-000 Smith, L. & Yu, C. (2008). Infants rapidly learn word-referent mappings via cross-situational statistics. Cognition, 106, 1558-1568. Stopfer, M. (2007). Olfactory processing: massive convergence onto sparse codes. Current biology, 17(10), R363-R364. Strong, ...

  6. [1983]

    Can you create a photorealistic image in a 1:1 aspect ratio of a new daytime viewpoint of a tent erected on the wing of a plane which is parked on the wing of seagull?

    subjects used a story of a general sending troops to capture a fortress in the center of a city from multiple directions (the source analog) to infer that a doctor could use radiation converging from multiple directions to destroy a tumor (the target analog). In this analogy, the subjects are representing the radiation problem in terms of the story about ...

  7. [1997]

    I can walk on that surface

    and make much more complex inferences using CWSG. Relations Only Task (RO): Open-Ended (World-to-World) Relations It is one thing to generalize a relation between the world and oneself such as, “I can walk on that surface”, a relation that can be compressed into a single-place predicate. But it is a different matter to generalize, or even represent, a rel...

  8. [2003]

    Balint’s

    and relation discovery in DORA (Doumas et al., 2008, 2022). We intentionally designed the R&M task to not require these capacities so that we could use the same metric of performance, namely the ability to dynamically bind the affordance to the Critical object in Perception, for all four architectures. This metric, because it is within reach of the DBO, R...

Show all 13 references
  1. [2008]

    can make their way in the world but generally cannot reason about the relations between objects exterior to themselves. However, to manipulate the world as people do, it is necessary to understand the relations between various objects in the world and be able to reason about t...

  2. [2010]

    Garcia, J., W. G. Hankins & K. W. Rusiniak. (1974). Behavioral regulation of the milieu interne in man and rat. Science, 184: 824–831. Gentner, D. (1983). Structure-mapping: A theoretical framework for analogy. Cognitive Science, 7, 155-170. Gibson, J. J. (1977). The theory of...

  3. [2014]

    I can walk on this

    demonstrated that architectures capable of CWSG can also perform more complex symbolic tasks than even the R&M task explored here. To keep the mechanism of inference constant across our simulations, the simulations reported here were designed to work using an inference mechani...

  4. [2021]

    accidental

    in order to try to expand the convex hull. LLMs as models of human cognition must be trained on a body of text that it would take a human thousands of years to read, even assuming said human did nothing but read the entire time. A human, by contrast, learns to converse intelli...

  5. [2025]

    Can you create a photorealistic image in a 1:1 aspect ratio of a new daytime viewpoint of a tent erected on the wing of a plane which is parked on the wing of seagull?

    showed that they struggled to represent seagulls, planes, and tents independently of each other and in the correct relations. As shown in Figure 8, the networks hybridized the airplanes and seagulls, put objects in the wrong relationships to each other or on the wrong parts of...

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.