Keynotes | K&Ts | GACs | Talks | Posters | Search

Poster F in Poster Session F: Thursday, August 6, 1:45 – 3:30 pm, Kimmel Center, Shorin & Rosenthal Rooms

Representational Neighborhoods and Multi-Agent Optimization for Creative Concept Combination in Large Language Models

Hashmath Shaik¹, Manikya Venkata Gnaneswar Villuri¹, Alex Doboli¹; ¹State University of New York at Stony Brook

Presenter: Manikya Venkata Gnaneswar Villuri

Here is the abstract with TeX formatting: Conceptual combination (CC), merging distinct concepts into novel, coherent wholes, is central to creative cognition yet systematically fails in large language models (LLMs). Given the pair *slipper–bed*, state-of-the-art models produce generic bedroom descriptions while omitting "slipper" entirely; the genuinely creative blend, *a bed shaped like a slipper*, is never generated. We attribute this to **representational collapse**: internal circuits converge to high-probability default framings, rendering emergent, metaphorical, and relational combinations unreachable. We present a three-part framework. First, we formalize **Neighborhood Graphs**, where nodes are CC sentences and edges encode input perturbation (Δ_in), neural circuit divergence (Δ_NN), and output difference (Δ_out), as an explicit map of an LLM's concept combination space. Analysis across 47 concept pairs reveals small-world topology (clustering coefficient 0.64, path length 2.3) and neutral sets of ≈3.7 semantically equivalent sentences, echoing evolvability structures in biological fitness landscapes. Second, we develop **multi-scale metrics** bridging neural-circuit measurements to output quality: a composite MI exploration score Φ, built from Sparse Autoencoder (SAE) feature distances and cross-layer coherence, discriminates CC types that behavioral metrics conflate (metaphorical: Φ = 0.664; concept-2-dominant default: Φ = 0.621). Third, a closed-loop **three-agent system** (Generator, Behavioral Validator, Representational Probe) that optimizes J = α Q + βΦ outperforms behavioral-only optimization: quality Q rises from 86.1 to 88.4, novelty Φ from 0.631 to 0.648, and joint score J from 0.762 to 0.790. These results demonstrate that mechanistic interpretability tools can actively guide creative generation, connecting AI systems research to the cognitive neuroscience of conceptual combination.

Topic Area: Methods, Tools, Theory & Neural Coding