Solo researcher from Bangladesh presenting at NeurIPS 2026

My work is mainly about reasoning efficiency, so the link to AI risk is indirect, but it's real in three ways.

First, it shows that reasoning models often abandon a correct line of reasoning and drift into unfocused elaboration. In my error analysis, 76% of failures were this kind of premature termination. Understanding when models stop reasoning well matters for trusting their outputs.

Second, I found that GSM8K accuracy does not predict accuracy on graduate-level questions (r = −0.027 across 18 models), and 12 of 13 models I profiled collapsed on GPQA. Evaluations that rely on easy benchmarks can hide real failures.

Third, my Q-score measures a model's reasoning behaviour from its outputs alone, without access to internals. That kind of external check can help evaluate closed models.

Going to NeurIPS lets me get feedback from people working on evaluation and reasoning reliability, and build connections so I can keep doing this work.