2026 · arXiv

Measuring AI Reasoning: A Guide for Researchers

Munachiso Samuel Nwadike, Zangir Iklassov, Kareem Ali, Rifo Genadi, Kentaro Inui

A position-style guide arguing that AI reasoning should be evaluated through adaptive, multi-step search and process evidence, not final-answer accuracy alone. It gives researchers a cleaner lens for diagnosing reasoning traces and failure modes.

Sycophancy Hides Linearly in the Attention Heads

Rifo Genadi, Munachiso S. Nwadike, Nurdaulet Mukhituly, Hilal AlQuabeh, Tatsuya Hiraoka, Kentaro Inui

Finds that sycophancy-related signals are especially linearly accessible in attention-head activations. The work uses probing and steering to show that targeted attention-head interventions can reduce deference to incorrect user beliefs.

2025 · ACL Long Papers

RECALL: Library-Like Behavior in Language Models is Enhanced by Self-Referencing Causal Cycles

Munachiso S. Nwadike, Zangir Iklassov, Toluwani Aremu, Tatsuya Hiraoka, Benjamin Heinzerling, Velibor Bojkovic, Hilal AlQuabeh, Martin Takáč, Kentaro Inui

Introduces self-referencing causal cycles as a mechanism that helps autoregressive language models retrieve information in reverse or non-standard directions. The paper connects this mechanism to the reversal curse and proposes ReCall as a two-step retrieval process.

2025 · arXiv

Number Representations in LLMs: A Computational Parallel to Human Perception

Hilal AlQuabeh, Velibor Bojkovic, Munachiso S. Nwadike, Ahmed Oumar El-Shangiti, Tatsuya Hiraoka, Kentaro Inui

Studies whether language models internally encode numbers in a compressed, logarithmic-like way, paralleling human number perception. The results suggest that LLM number representations may be non-uniform rather than linearly spaced.