h t t p s : / / w w w . f r o n t i e r s i n . o r g / j o u r n a l s / p s y c h o l o g y / a r t i c l e s / 1 0 . 3 3 8 9 / f p s y g . 2 0 2 6 . 1 8 4 9 0 8 3 / f u l l

Ingest: Conceptual blending in humans and language models

Type: types/ingest-report.md

Classification

This is a peer-reviewed conceptual-analysis article in Frontiers in Psychology (vol. 17, 2026; received April, published July 2026; three named reviewers). It reports no experiment. Its only computational artifact is a supplementary Python toy model, a vector symbolic architecture (VSA) of one example, which the author calls "a model-building hypothesis". The capture is the full article text without figures or the supplementary script. Author: Spencer K. Lynn, a single author at Charles River Analytics (from the correspondence address). He has prior work on Bayesian predictive-processing cognitive architectures (Scruff, 2019) and on embodied generative agents (AAAI symposium, 2023), which this paper builds on. The author reports using Gemini and ChatGPT for literature review and Copilot to generate the toy code.

Summary

The paper asks what decides which elements of two source concepts are inherited into a blend. Conceptual blending (Fauconnier and Turner's conceptual integration theory) is the construction of a new concept from a selective combination of source features; "house boat" and "boat house" take different features from the same two sources. Lynn's answer is that a goal selects the features. In humans the goal is "foregrounding" features relevant to it: in the running example, the goal of understanding a never-seen Osprey tiltrotor selects propeller orientation and wings from "helicopter" and "airplane" and ignores window count. He proposes predictive processing as a candidate mechanism. A goal or context node is a parent in a generative hierarchy. It raises the precision of goal-relevant features, roles and affordances of the co-active source concepts and lowers the precision of irrelevant ones. A blend node integrates the selected features into a hypothesis tested against observations, simulations, linguistic context or action possibilities. Prediction error, generated when the blend fails the goal, revises which features are kept. Human goals are ultimately rooted in physiological regulation (allostasis); "goal" is used at neurocognitive, behavioral and toy-model levels, which the paper separates. For language models, Lynn proposes "proxy-goals": the purposiveness comes from instruction tuning, RLHF and DPO, prompt framing, the context window, in-context learning and decoding, not from an internal causal model of a goal. Blending arises as the model "composes statistically associated features, roles, and constraints" into a continuation; the outputs are "distributionally novel blends". The goal "is a proxy signal that steers token probabilities" and is "more like working-memory than a revisable, hierarchical controller". Lynn then claims three differences. First, LLMs hold linguistic traces of image-schematic and force-dynamic structure (containment, support, blockage and similar patterns) without the perception-action grounding that makes them operative constraints. Second, attention-weighted superposition means conflicting constraints produce interference during decoding: tokens "lexically consistent with all constraints while not satisfactorily addressing any", which he calls "generic compromises". Third, in-context learning gives short-term adaptation but no lasting model change, and reflection needs prompted scaffolding. He qualifies these as differences of organization, not an all-or-none opposition. The proposed remedy is an "intermediate affordance layer" in which a goal becomes an affordance specification, then a candidate, a causal test and a revision. Every mechanism claim is conceptual.

Contribution assessment (our opinion)

The problem is real and old: conceptual integration theory describes selective projection well but has said less about what drives the selection. The paper's main addition is a clearly stated candidate answer: the goal is a precision-setting parent node, with error-driven revision of the blend. That answer connects blending theory to predictive-processing vocabulary, and it makes "goal-relevance" an explicit selection variable rather than an after-the-fact description. The human mechanism is a reasonable synthesis of existing predictive-processing accounts of concepts (Michel) and goals (Barrett). The author hedges it properly as one candidate within a contested program. It is not tested: the VSA toy shows only that the selection can be implemented, and the paper gives no observation that would distinguish it from rival accounts of selective projection. The LLM side is weaker as a mechanism claim. "Proxy-goal" mostly relabels a widely accepted point, that a model's apparent purposiveness comes from training, alignment and the prompt. The superposition-to-"generic compromise" step cites only toy-model superposition work (Elhage et al.) and offers no measurement of compromise outputs. The most testable contribution is that generic-compromise prediction, together with the claim that an affordance layer would reduce it. Neither is tested here. Originality relative to Sato (2025) on prompt-induced blending in LLMs, which the paper cites, cannot be judged from this capture.

Quotes

No source quotes have been retained yet.

Connections Found

For the workshop's question, this source is the anchor statement of goal-selected projection as the productive mode of blending. It is the first artifact in the KB to treat conceptual blending at all. It enters the KB as a conceptual comparison partner, not as evidence: every connection below is a comparison, because the paper offers no data.

It compares with LLM generation can hide a relaxed goal where human writing exposes a stall on what LLM output looks like when a set of constraints cannot all be met. That note conjectures an argmax over plausibility that silently drops one conjunct. Lynn conjectures superposition interference that addresses every constraint weakly. The two rival mechanisms predict different signatures: one conjunct missing, against all conjuncts thinly served. Neither has been measured.

It compares with Human analogies can motivate functions without determining component boundaries on inferring an engineered architecture from a human cognitive mechanism. Lynn's "intermediate affordance layer" moves directly from a human predictive-processing hierarchy to a recommended AI component. The note's test separates the function that may transfer (goal-conditioned feature selection with error-driven revision) from the component boundary that does not follow.

It compares with Psychology-to-agent transfer needs per-principle failure-mode testing on goal-gated selection. That note judged goal-relevant filtering of memory retrieval mostly decorative for agents. Lynn's goal gates which features enter a constructed concept, not which memories are retrieved, so the earlier verdict does not obviously cover it.

It compares with Knowledge storage does not imply contextual activation on available against operative knowledge. Lynn's claim that LLMs hold "linguistic traces" of image schemas without making them "operative constraints" draws the same distinction. He locates the missing gate in embodied grounding; the note locates it in contextual activation.

It compares with World models assess explanatory-reach through action-conditioned prediction on closed act-test-revise loops. Lynn's "goal to affordance specification to candidate realization to causal test to revision" applies that loop to concept construction.

It compares with LLM learning phases fall between human learning modes on mapping LLM adaptation onto human learning. Lynn's third difference treats in-context learning as short-term adaptation without lasting change. The note argues the phases do not map one-to-one onto human modes, which bounds how far that contrast can be read as a dichotomy.

Among sources, it compares with the seed-programmed autonomous general learning ingest as another architecture proposal for goal-driven causal model construction and revision; Thórisson's has implementation cases, Lynn's has a toy script. It compares with the interpolation, extrapolation, hyperpolation ingest, whose vocabulary for novelty relative to a training distribution could make Lynn's "distributionally novel blends" more precise.

Lynn's interference claim should not be added to the empirical irrelevant-context studies already in the KB as one more instance of a common mechanism. It concerns conflicting task constraints, not unlicensed context content, and it is not measured.

Learning Claims (our opinion)

The mechanism on its own terms. In Lynn's human account, a blend is a hypothesis built under a goal and revised by error. The goal node sets precision over source features. The blend node predicts observations. Prediction error flows upward and changes which features are retained, suppressed or reweighted, cycle by cycle, until the blend is stable. In the Osprey example, "helicopter" and "airplane" each fail as whole categories, but selected features of each jointly explain the observed flight. For LLMs, Lynn claims adaptation is transient: in-context learning can approximate Bayesian updating, but there is no "lasting model change or explicit causal roll-out", and reflection depends on prompted protocols rather than an intrinsic uncertainty-aware controller.

Mapping onto Commonplace concepts. Judged against theory builder, the human blending loop as Lynn describes it meets consumption (the blend governs categorization and action) and iteration (error results shape the next cycle). It is close to criticism in form (the blend is tested against observations), but the error signal is precision-weighted belief update, not criticism aimed at what a stated unit says. Its localized content is at most the toy model's labeled VSA vectors; in the brain the paper claims no stated units. So it maps onto the gradient-like error elimination the definition contrasts with criticism, not onto theory building. That mapping is ours; Lynn does not use these terms. Persistence of the human blend is claimed across encounters (blends become "entrenched") but not argued. For the LLM side, Lynn's "no lasting model change" matches the KB's within-episode persistence grade for in-context adaptation. His view that the remedy is an internal controller differs from Commonplace's route, which keeps revision in external, addressable artifacts. Lynn does not consider that alternative.

What it adds or questions. Under learning inside a fixed decomposition inherits its mistakes, Lynn's account fixes the decomposition outside the learner: the source concepts, the goal and the image-schema vocabulary are supplied, and only precision over features is updated. The Osprey blend improves within that space; the account does not say how a wrong source pair or a missing image schema gets corrected. The paper's learning claims are a plausible mechanism at best. No improvement is observed and no causal evidence is offered, so they neither support nor challenge the KB's account. They add one useful distinction: selection by goal relevance and correction by error are separate operations, and the second is what keeps the first honest.

Extractable Value

  1. The role of goals in Lynn's theory, in his terms -- the goal answers the "central puzzle" of what motivates the inheritance of specific source elements. It foregrounds goal-relevant features and lets the blend omit goal-irrelevant ones. In the predictive-processing implementation, the goal is a parent node that raises precision of expected-relevant features and lowers precision of irrelevant or misleading ones, and the blend is then tested; prediction error, generated "when the candidate blend fails to satisfy the goal", revises the selection. Goals are disambiguated by level: allostatic regulation, behavioral task or desired state, and high-level prediction in the toy model. For the workshop, the key point is that selection and test are coupled in one loop: goal relevance chooses features, and observation decides whether the combination holds. [quick-win]
  2. The proposed LLM mechanism, in his terms -- language models produce "distributionally novel blends" under "proxy-goals" supplied by instruction tuning, RLHF/DPO, prompt framing, the context window, in-context learning and decoding. Blending is the composition of statistically associated features through attention-weighted superposition. The goal is externalized ("residing in the context window, the instructions, and human preferences"), acts "more like working-memory than a revisable, hierarchical controller", and does "not encode explanations of why a particular blend should exist or how it would work". The paper explicitly says LLMs have structured representations; the claimed difference is the source and grounding of purposiveness and whether the control loop is closed. [quick-win]
  3. A tension for the workshop's watched failure -- the workshop watches two useful ideas being fused because both serve one goal. Lynn presents goal-selected fusion as blending working correctly. On his account, what separates a productive blend from an unsupported relation is not the selection but the subsequent test against observation. His LLM account gives the model the selection without that internal test. Read that way (our interpretation, not Lynn's claim), an unsupported relation would be a goal-selected blend that no error signal has checked. This is a conjecture for the workshop to weigh, not evidence. [experiment]
  4. Two rival signatures for unsatisfiable constraints -- Lynn's "generic compromise" (all constraints thinly addressed) and the KB's relaxation note (one constraint silently dropped) predict observably different outputs. A small comparison on prompts with deliberately conflicting constraints could discriminate them. Both are unmeasured conjectures at present. [experiment]
  5. A fresh case for the function-versus-boundary test -- the affordance-layer recommendation is a direct human-to-AI component inference. The transferable function is goal-conditioned selection plus error-driven revision; whether it needs a dedicated affordance layer, or can live in external artifacts or tool loops, is not argued. [just-a-reference]
  6. Pointer to the direct LLM-blending source -- Lynn cites Sato (2025), "The way we prompt: conceptual blending, neural dynamics, and prompt-induced transitions in LLMs" (arXiv 2505.10948), as evidence that LLMs blend. It is not ingested and may carry the empirical content this paper lacks. [just-a-reference]

Limitations (our opinion)

Nothing in the paper tests its mechanisms. The human account is one candidate implementation inside predictive processing, a program the author himself calls broad and contested. The paper offers no observation that predictive processing explains and a simpler selective-projection account does not. The Osprey walkthrough is an illustration whose outcome is built into its setup: the goal node is defined so that the right features win. The toy VSA script was not in the capture and is described as an in-principle demonstration.

The LLM mechanism is hard to vary in only one place, the generic-compromise claim; elsewhere "proxy-goal" absorbs any source of LLM purposiveness (training, alignment, prompt, tools, decoding), so it rules out little. The superposition-to-compromise step rests on toy-model superposition work, not on evidence about decoding under conflicting constraints. It may also be explained more simply: an aligned model hedges or averages under conflicting instructions without any representational interference.

The three differences are stated as a contrast with an idealized human. The claim that "humans detect and repair incoherence proactively" ignores the large human literature on unnoticed incoherence. The comparison is therefore asymmetric: LLM brittleness is described from engineering observations, human competence from theory. The paper does not discuss how goal-driven selection fails in humans, for example motivated or goal-congruent projection that stabilizes because no disconfirming observation arrives. That is exactly the failure the occasion concerns, so the paper says nothing direct about when goal-selected blends go wrong.

The capture omits the four figures and the supplementary code; the text describes them sufficiently for this analysis.

Update the E20 entry in the epistemic-relevance-and-conceptual-blending workshop's evidence register with Lynn's two-part answer (goals set precision over source features and are coupled to an error test in humans; LLM goals are externalized proxy-goals with no internal error-driven revision), marked as conceptual and untested. Record the tension in Extractable Value item 3 as an open question for the workshop's review point.

Relevant Notes: