SlimeLLM: Attribute-Separated Reasoning Framework for Post-hoc Optimization of Large Language Models

This paper introduces SlimeLLM, a post-hoc reasoning framework that rethinks how large language models are used at inference time.

Rather than modifying or retraining LLMs, the framework operates entirely at the output level, decomposing monolithic embeddings into attribute-separated vector spaces and dynamically generating semantic structures from context. By treating LLM outputs as candidates instead of final answers, SlimeLLM enables drift detection, localized correction, and gradual convergence through reuse.

A central idea of this work is that many failures attributed to “hallucination” are not random errors, but structural mismatches: meanings, evaluations, and rhetorical roles collapsed into a single undifferentiated space. Separating these roles makes inconsistencies observable and correctable without increasing model scale.

The framework incorporates a candidate catalog indexed by attributes, allowing previously validated structures to be reused across similar queries. Over time, and across diverse users and languages, structurally equivalent queries tend to converge toward stable semantic cores, while local or transient errors fail to persist.

While the paper focuses on a minimal, model-agnostic implementation, it is informed by a broader line of work on non-sequential, order-independent structural indexing (referred to here as SlimeTree), which explores how meaning can be organized without fixed traversal order. That direction is orthogonal to the present contribution and is mentioned only to suggest that richer structural treatments are possible beyond the scope of this paper.

No patents are claimed on the methods described here. The intent of this publication is to provide a practical and conceptual entry point for researchers and practitioners interested in inference-time adaptation, explainability, and robustness in large language models.