Posts by Collection

Portfolio item number 1

Short description of portfolio item number 1
Read more

Portfolio item number 2

Short description of portfolio item number 2
Read more

Query-Conditioned Active Causal World Models for Language Agents

M. Emre Bilgin

conference 2026

tldr LLM agents should not answer causal questions immediately when the necessary evidence is missing. We propose a query-conditioned causal agent that first asks: Do I have enough evidence to answer this specific causal question? If not, it selects the most useful next action—observe, intervene, retrieve, ask, or simulate—to reduce uncertainty about the target answer. The goal is not to learn the whole causal graph, but to gather just enough causal evidence to answer well.

PDF Website GitHub arXiv

Information Decomposition Shapes Human Reasoning

M. Emre Bilgin

conference 2026

tldr This study examines whether people, when making decisions under uncertainty, are sensitive not only to the question “How much evidence is there?” but also to how the evidence is organized. Initial pilot findings suggest that redundant information is utilized more directly, while synergistic information is more difficult to process but becomes actionable once connections are established. In short: What matters in reasoning is not just the amount of information, but its architecture.

PDF Website GitHub arXiv

A Partially Shared Latent Neural Space for Deductive Reasoning

M. Emre Bilgin

conference 2026

tldr Most reasoning fMRI studies ask which brain regions activate. This project asks a different question: can trial-wise brain activity during reasoning be explained by a small number of latent neural factors? We compare syllogistic and transitive reasoning, estimate ROI-level activity for each trial, and use PCA/factor analysis to search for stable low-dimensional structure. Functional connectivity is used as a complementary check: do these latent factors reflect coordinated network activity?

PDF Website GitHub arXiv

Beyond Final Answers: Information-Architecture Blindness in LLM Belief Updating

M. Emre Bilgin

conference 2026

tldr Final-answer accuracy can hide important failures in LLM reasoning. We introduce a human-grounded, Bayesian-referenced benchmark showing that many LLMs fail to update their beliefs at the right time and for the right kind of evidence. We call this failure mode information-architecture blindness: weak sensitivity to whether evidence is carried by individual cues or by relations between cues.

PDF Website GitHub arXiv