Geometry of Knowledge Allows Extending Diversity Boundaries of Large Language Models

Starting from the hypothesis that knowledge in semantic space is organized along structured manifolds, we argue that this geometric structure renders the space explorable. By traversing it and using the resulting continuous representations to condition an LLM’s generation distribution, we can systematically expand the model’s reachable semantic range. We introduce a framework that requires no modification of LLM parameters and operationalizes this idea by constructing a conditioning distribution from a small set of diverse anchor generations. This distribution conditions LLM’s generation via an xRAG-style projector. Our experiments demonstrate that this manifold-based conditioning substantially increases generative diversity, with direct benefits for enhancing divergent thinking, a core facet of creativity, in language models.

💡 Research Summary

The paper “Geometry of Knowledge Allows Extending Diversity Boundaries of Large Language Models” proposes a novel way to increase the semantic diversity of large language model (LLM) generations without any fine‑tuning of the model parameters. The authors start from the hypothesis that knowledge in semantic space is organized along structured, multi‑component manifolds. Because these manifolds are continuous, they can be explored to reach regions of the space that are inaccessible to traditional prompt‑based or decoding‑based diversity techniques, which are limited to a finite set of contexts.

The core of the method is a continuous conditioning variable z that lives in the same embedding space used to represent text. For a given prompt x, an encoder E (e.g., the Mistral SRF embedding model) produces a dense representation e = E(x). The base LLM (Mistral‑7B‑Instruct) is first asked to generate a small set of diverse “anchor” responses. These responses are re‑encoded with E to obtain a discrete set Aₓ = {e₁,…,eₘ}. The authors then define a latent region Zₓ by interpolating between pairs of anchors (z = (1‑λ)eᵢ + λeⱼ) and optionally adding random perturbations or using meta‑heuristic search. This yields a continuous sub‑manifold that approximates the local geometry of the LLM’s semantic space.

To inject z into the generation process, the paper adopts the xRAG projector (Cheng et al., 2024), which maps the latent vector into the token‑embedding space of the LLM, forming a new context c = g(x, z). The LLM then samples from pθ(y | c). The overall output distribution is an integral over z: p(y | x) = ∫pθ(y | g(x, z)) qφ(z | E(x)) dz. By expanding the variance term in the law of total variance, the authors argue that the continuous latent variable dramatically increases the second term (variance of the conditional expectation), which is negligible in discrete prompt‑based settings.

Two evaluation tracks are presented. First, on NoveltyBench (2025), the method is measured with the “Distinct” metric (number of abstract equivalence classes) and a “Utility” metric that penalizes uninformative outputs. Across sampling budgets of 10‑30 generations, the manifold‑based approach discovers 20‑30 % more distinct classes and maintains higher utility than baselines such as temperature scaling, nucleus sampling, G2 (guided generation), and multi‑agent discussion. Second, the authors test divergent thinking using the Alternative Uses Test (AUT). Using an automatic originality scorer, the proposed method achieves the highest originality scores, approaching the upper bound of the scoring scale and outperforming all prior methods.

Key contributions are: (1) a theoretical analysis showing that prompt‑based diversity is bounded by the finite set of contexts; (2) a plug‑in latent‑conditioning framework that works with any frozen LLM; (3) an explanation of why VAE‑style latent spaces (with unimodal Gaussian priors) are mismatched to the clustered geometry of LLM embeddings; (4) empirical evidence of substantial gains in both semantic diversity and creative‑thinking benchmarks.

The paper’s strengths include its parameter‑free nature, clear theoretical grounding, and practical simplicity (anchor generation + linear interpolation). Limitations are noted: the quality of the manifold depends on the initial anchor set, linear interpolation may not capture highly non‑linear semantic transitions, and experiments are limited to a single LLM/embedding pair, leaving generalization to other models open. Future work suggested includes non‑linear flow‑based sampling, automated anchor selection via meta‑learning, multimodal extensions, and combining the conditioning with RL‑based diversity‑quality objectives.

Overall, the work demonstrates that treating the semantic space of LLMs as a traversable geometry offers a powerful, low‑cost avenue to push the boundaries of generative diversity and creativity, complementing existing fine‑tuning or decoding‑level techniques.