Abstract
We propose a novel inference-time and training-time paradigm for improving the epistemic reliability of large language models (LLMs) and tool-augmented AI systems, termed Forced Epistemic Reset (FER). Unlike conventional self-consistency or critique-based methods, FER introduces a structured adversarial self-invalidation cycle in which a model is periodically required to challenge, falsify, and independently re-derive its own outputs using heterogeneous reasoning paths. The system explicitly treats intermediate conclusions as provisional hypotheses subject to internal falsification pressure. We formalize the architecture, training objectives, reset triggering mechanisms, and multi-path recomputation strategy. Empirical evaluation protocols are also outlined, focusing on contradiction recovery, robustness under reasoning perturbation, and cross-path convergence stability. We argue that FER represents a shift from passive verification to active epistemic adversariality in neural reasoning systems.
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.
Recommended Citation
Manuel-Devadoss, Johny and Prasad, Deepthi, "An Adversarial Self-Invalidation Framework for Robust Reasoning", Technical Disclosure Commons, ()
https://www.tdcommons.org/dpubs_series/10034