Hi HKUDS team, congrats on the NeurIPS 2025 Spotlight!
I'm Daouda Abdoul Anzize, creator of OpenAIBox — a tool that maps the semantic role of each hidden dimension in any LLM via PyTorch forward hook tracing.
Why this is relevant to AI-Researcher
AI-Researcher generates novel hypotheses autonomously. OpenAIBox can help validate those hypotheses at the mechanistic level by showing which model dimensions are activated during specific reasoning tasks.
Example: if AI-Researcher generates the hypothesis "this model struggles with temporal reasoning", OpenAIBox can verify whether the temporality dimensions (low separability = 0.992 in SmolLM) are indeed poorly specialized.
What OpenAIBox produces
pip install openaibox
from openaibox import OpenAIBox
gr = OpenAIBox("HuggingFaceTB/SmolLM-360M")
gr.discover().map_dimensions().export("graph.json")
gr.print_summary()Output example:
dim_696 → syntax, causality, certainty, abstraction, time, emotion (score=1.0)
dim_295 → causality + temporality (pure specialist)
Injection points: INPUT → MEMORY → DECISION → OUTPUT
Potential collaboration
OpenAIBox produces the mechanistic map. AI-Researcher generates the hypotheses. Together: hypothesis generation + mechanistic validation in a single pipeline.
Happy to discuss how both tools could work together. Open to co-authorship.
Daouda Abdoul Anzize
PyPI: openaibox · GitHub: Tryboy869 · Twitter: @Nexusstudio100
Cotonou, Benin → Global Remote
Hi HKUDS team, congrats on the NeurIPS 2025 Spotlight!
I'm Daouda Abdoul Anzize, creator of OpenAIBox — a tool that maps the semantic role of each hidden dimension in any LLM via PyTorch forward hook tracing.
Why this is relevant to AI-Researcher
AI-Researcher generates novel hypotheses autonomously. OpenAIBox can help validate those hypotheses at the mechanistic level by showing which model dimensions are activated during specific reasoning tasks.
Example: if AI-Researcher generates the hypothesis "this model struggles with temporal reasoning", OpenAIBox can verify whether the temporality dimensions (low separability = 0.992 in SmolLM) are indeed poorly specialized.
What OpenAIBox produces
Output example:
Potential collaboration
OpenAIBox produces the mechanistic map. AI-Researcher generates the hypotheses. Together: hypothesis generation + mechanistic validation in a single pipeline.
Happy to discuss how both tools could work together. Open to co-authorship.
Daouda Abdoul Anzize
PyPI: openaibox · GitHub: Tryboy869 · Twitter: @Nexusstudio100
Cotonou, Benin → Global Remote