analiza aiqbench
This study identifies distinct, low-rank activation subspaces within LLMs dedicated specifically to logical inference, demonstrating a functional dissociation from other cognitive capacities like arithmetic or working memory. By ablating these specific regions, researchers can selectively disable logic without affecting general knowledge retention, suggesting that abstract reasoning in neural networks may be localized similarly to specialized brain networks. This finding provides critical insight into the internal architecture of large language models and the modularity of their reasoning capabilities.
Logical subspace in LLMs
arXiv:2609.32907v1 Announce Type: cross Abstract: Recent work has identified a human brain network specialized for abstract formal reasoning (Kean et al., 2025). Does the same hold true in language models? To answer this question, we introduce the minimal viable subspace (MVS) method, which searches for the lowest-rank activation subspace at a layer that preserves task performance when everything outside that subspace is ablated. Using MVS, we demonstrate low-rank subspaces supporting logical inference on Gemma and Qwen models. Furthermore, these subspaces exhibit a clear dissociation from model capacities on other tasks, such that retaining these late logic subspaces preserves inference while impairing factual knowledge, working memory, cognitive control, and arithmetic. Conversely, ablating them reduces logical inference accuracy to chance while largely sparing these other capacities. Our results suggest a functionally localizable core machinery for logic akin to that in the human brain.
Abstract:Recent work has identified a human brain network specialized for abstract formal reasoning (Kean et al., 2025). Does the same hold true in language models? To answer this question, we introduce the minimal viable subspace (MVS) method, which searches for the lowest-rank activation subspace at a layer that preserves task performance when everything outside that subspace is ablated. Using MVS, we demonstrate low-rank subspaces supporting logical inference on Gemma and Qwen models. Furthermore, these subspaces exhibit a clear dissociation from model capacities on other tasks, such that retaining these late logic subspaces preserves inference while impairing factual knowledge, working memory, cognitive control, and arithmetic. Conversely, ablating them reduces logical inference accuracy to chance while largely sparing these other capacities. Our results suggest a functionally localizable core machinery for logic akin to that in the human brain.
Artificial Intelligence (cs.AI); Computation and Language (cs.CL)
arXiv:2609.32907 [cs.AI]
(or arXiv:2609.32907v1 [cs.AI] for this version)
https://doi.org/10.48550/arXiv.2609.32907 Focus to learn more arXiv-issued DOI via DataCite (pending registration)
arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.
Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.
Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.
Text integral sub licența deschisă a publicației · 234 cuvinte
Subiecte conexe
Hallucinations as Orthogonal Noise: Inference-Time Manifold Alignment via Dynamic Contextual Orthogonalization
arXiv · 3 iun. 2026
Last Layer Logits to Logic: Empowering LLMs with Logic-Consistent Structured Knowledge Reasoning
arXiv · 5 oct. 2026
Semiotic logical hexagon theory for LLM logical reasoning
arXiv · 28 iul. 2026
Functional Subspace, where language models can use vector algebra to solve problems
arXiv · 1 oct. 2026