{"author":{"name":"Md Muntaqim Meherab, Noor Islam S. Mohammad, Faiza Feroz","slug":"md-muntaqim-meherab-noor-islam-s-mohammad-faiza-feroz","article_count":1,"latest_published_at":"2026-04-27T13:07:09.804+00:00","profile_url":"https://vff.ai/authors/md-muntaqim-meherab-noor-islam-s-mohammad-faiza-feroz","api_url":"https://vff.ai/api/authors/md-muntaqim-meherab-noor-islam-s-mohammad-faiza-feroz"},"articles":[{"slug":"causal-concept-graphs-in-llm-latent-space-for-stepwise-reasoning","title":"Mapping Causal Reasoning in LLMs with Sparse Concept Graphs","url":"https://vff.ai/article/2026/04/27/causal-concept-graphs-in-llm-latent-space-for-stepwise-reasoning","content_type":"research_summary","summary":"Researchers propose Causal Concept Graphs (CCG), a method that maps how concepts interact during multi-step reasoning in language models by combining sparse autoencoders with differentiable graph learning. The approach identifies sparse, interpretable latent features and learns directed causal dependencies between them, then validates these connections using a new Causal Fidelity Score metric. On three reasoning benchmarks with GPT-2 Medium, CCG significantly outperforms existing tracing and ranking methods, producing sparse, domain-specific graphs that remain stable across runs.","published_at":"2026-04-27T13:07:09.804+00:00","updated_at":"2026-05-07T02:19:52.375108+00:00","source":{"url":"https://arxiv.org/abs/2603.10377","name":"ArXiv (cs.AI)"},"featured_image":{"url":"https://www.wolframphysics.org/technical-introduction/the-updating-process-in-our-models/the-role-of-causal-graphs/img/the-role-of-causal-graphs_io_2.en.png","alt":null},"categories":[{"name":"Research","slug":"research"},{"name":"AI Safety & Alignment","slug":"ai-safety-alignment"},{"name":"LLMs","slug":"llms"}]}]}