Concept Blending

Draft — This content has not been approved for publication.

Responsible: Ben Goertzel

Papers: Hyperon for AGI⇒ASI Whitepaper (2025), §6.5; Goertzel (2025-02), A Formalization of Concept Blending: An Information-Theoretic Approach to Quantifying and Creating Emergent and Coherent Concept Blends; Goertzel (2026-01), Compositional Theorems for Quantale-Based Blending, Quantale Weakness Based Conceptual Blending, and Paraconsistent Interference in Conceptual Blending.

Status: Multi-paradigm. Per Non-clustered HAA cluster pilot Source 4 (closed 2026-05-07): [CONCEPT-BLENDING-PARTIAL-IMPLEMENTATION]. Information-theoretic + GA/CMA-ES + category-theoretic colimit/pushout + PeTTa-port branches are [IMPLEMENTATION-BACKED-CORE] in iCog-Labs-Dev/conceptBlending at HEAD 9675e5e. Quantale weakness, paraconsistent interference, and HoTT-adjacent material are [PAPER-ONLY-FORMAL-EXTENSION] at this snapshot. The TransWeave / MORK-native / WILLIAM / Symbolic-Heads downstream paths described below are whitepaper / roadmap claims, not currently-wired implementation.

Concept

Concept blending is a primary heuristic for creativity — creating novel ideas by merging properties from two or more source concepts so that the result shows useful emergent structure. In Hyperon, blends are not bag-of-words tricks but metagraph constructs that can be tested by the same evaluators that rate patterns, proofs, or programs, making blending a first-class, auditable operation.

How It Works (intuitive)

Blending takes two or more source concepts (represented as subgraphs in AtomSpace), identifies structural alignments between them, and constructs a merged graph that preserves key constraints from each source while introducing novel combinations. The whitepaper gives the intuitive example of "sky-garden" = aerial platform + botanical garden — a blend that inherits structural properties from both sources.

Paper Stack (2025-2026 Hyperon-era, Goertzel)

PaperYearApparatusVerdict
A Formalization of Concept Blending: An Information-Theoretic Approach to Quantifying and Creating Emergent and Coherent Concept Blends2025-02Information-theoretic + emergence/coherence quantification[IMPLEMENTATION-BACKED-CORE]
Compositional Theorems for Quantale-Based Blending: How Weakness Interoperates with Categorical Structure2026-01Quantale weakness + categorical compositionality[PAPER-ONLY-FORMAL-EXTENSION]
Quantale Weakness Based Conceptual Blending: Integrating Information-Theoretic and Category-Theoretic Approaches2026-01Integration of quantale + info-theoretic + cat-theoretic[PAPER-ONLY-FORMAL-EXTENSION]
Paraconsistent Interference in Conceptual Blending: When Properties Resonate or Cancel2026-01Paraconsistent logic + interference dynamics[PAPER-ONLY-FORMAL-EXTENSION]

Author attribution note: Ben Goertzel sole-author on all four 2025-2026 Hyperon-era papers in the local archive; targeted searches for "Eskridge" / "eskridge" returned zero hits. Older OpenCog-era Goertzel/Eskridge concept-blending work may exist outside the local Mattermost archive (per V4-1 cluster-pilot finding); flagged for future retrieval.

Implementation Backing (cluster-pilot extraction, Source 4 close 2026-05-07)

The iCog-Labs-Dev/conceptBlending repo at committed HEAD 9675e5e contains four implementation branches. Per V4-2 methodology guard: extract against committed HEAD only; the local working tree had hundreds of staged file deletions (likely a reorganization in flight) that are NOT a valid filesystem source surface.

BranchTagMaps to paperImplementation evidence
an_infotheoretic_approach/[INFO-THEORETIC-IBC]Blending (2025-02) information-theoreticinfo-theoretic.metta:111 (mu-hqblend); GA-imp.metta:166,:190; cma-es-imp.metta:249,:272 (GA + CMA-ES iteration loops)
peTTa_version/[PETTA-PORT-EXPERIMENTAL]Info-theoretic without GA + multi-objective CMA-ESpeTTa_version/info-theoretic.metta:116 (mu-hqblend)
a_categorytheoretic_approach/[CATEGORY-THEORETIC-COLIMIT]Whitepaper categorical framing + colimit-based blendinga_categorytheoretic_approach/libs/colimit.py:266 (compute_colimit); algspec_builder.metta; generalization_builder.metta; morphism_finder.metta; LLM + concept-AtomSpace prompt scaffolding
a_naive_approach/[NAIVE-PROTOTYPE]Prototype (graph matching + GPT)Mirror/simplex/single-scope/double-scope blending network LLM prompts; not the categorical implementation

[PAPER-ONLY-FORMAL-EXTENSION] applies to: quantale-weakness-class structures (no quantale or weakness implementation in conceptBlending HEAD); paraconsistent interference (zero paraconsistent / interference hits at git grep -i); and HoTT-adjacent identity-aware semantics (zero HoTT / homotopy hits). These formalisms are paper-rigorous but not exercised by code in this repo at the audited snapshot.

What This Card Is Not (disambiguation)

  • NOT one unified algorithm. Four parallel implementation branches in conceptBlending, plus three paper-only formal extensions. The whitepaper "concept blending" is the conceptual umbrella; the actual implementation surface is multi-paradigm.
  • NOT TransWeave-implemented. The TransWeave-based "transport path in concept space" / "braid/near-commutation rules" framing is a whitepaper roadmap claim, not currently exercised by conceptBlending HEAD code.
  • NOT MORK-native. The whitepaper's "content-addressed blending with shared sub-parts deduplicating automatically" + "transport metadata reuse" framing is roadmap; depends on the AtomSpace Integration Phase 4+/5+ work (per AtomSpace).
  • NOT current PLN/WILLIAM/Symbolic-Heads/MOSES integration. The "Role in PRIMUS" downstream paths below describe whitepaper-projected uses; they are not currently wired in Hyperon code paths. Concept-blending repo HEAD has zero references to semantic_parsing / RelEx / Link Grammar / pattern-mining-side miners.
  • NOT the same as PGM-HoTT identity semantics. Goertzel's HoTT-semantics papers (Identity-Aware Commonsense Reasoning via Paraconsistent Geodesic Modal Homotopy Type Theory + Transweaving Temporal and Counterfactual Identities) are adjacent cognitive-semantics work deferred to a future PRIMUS / HoTT-Semantics cluster pilot per V4-A.Q4. Boundary-only mention here.

Roadmap claims (qualify before citing as implementation)

The following are whitepaper / roadmap descriptions, not currently-wired implementation. Treat as design intent, not extracted code surface:

  • TransWeave Formalization: blending as a transport path in concept space; choose maps that minimize effort while preserving each source's constraints; braid/near-commutation rules let blending interleave with pattern mining, inference, or program search.
  • MORK-native implementation: sources, alignments, and resulting blends are content-addressed; shared sub-parts deduplicate automatically; transport metadata stored alongside.
  • Role in PRIMUS ambient cognitive loop: blended concepts feed PLN (composite hypotheses), pattern mining (revealing patterns not visible in either source), MOSES (blended program fragments), and goal management (novel subgoals from creative combination).

Key References

  • Goertzel, B. (2025). Hyperon for AGI⇒ASI Whitepaper, §6.5: Concept Blending
  • Goertzel, B. (2025-02). A Formalization of Concept Blending: An Information-Theoretic Approach to Quantifying and Creating Emergent and Coherent Concept Blends. RawData.
  • Goertzel, B. (2026-01). Compositional Theorems for Quantale-Based Blending: How Weakness Interoperates with Categorical Structure.
  • Goertzel, B. (2026-01). Quantale Weakness Based Conceptual Blending: Integrating Information-Theoretic and Category-Theoretic Approaches.
  • Goertzel, B. (2026-01). Paraconsistent Interference in Conceptual Blending: When Properties Resonate or Cancel.
  • See also: Goertzel HoTT-semantics papers (deferred to future PRIMUS / HoTT-Semantics cluster pilot per V4-A.Q4): Identity-Aware Commonsense Reasoning via Paraconsistent Geodesic Modal Homotopy Type Theory; Transweaving Temporal and Counterfactual Identities (Practical Examples with Branching Contexts).

Cluster-pilot extraction archive: scripts/archive/non_clustered_haa_pilot/source4_concept_blending_semantic_parsing/ (findings_codex.txt + findings_gemini.txt + findings_reconciled_crossmodel.txt). V4-1..V4-11 carry-forwards locked.



Discussion