[GH-ISSUE #856] OpenAIBox — Mapping reasoning-specific dimensions in DeepSeek R1 vs standard LLMs #708

New issue

Closed

opened 2026-10-26 21:25:49 +00:00 by bsakaguchi · 2 comments

bsakaguchi commented

2026-10-26 21:25:49 +00:00

Owner

Originally created by @Tryboy869 on GitHub (Jun 10, 2026).
Original GitHub issue: https://github.com/deepseek-ai/DeepSeek-R1/issues/856

Hi DeepSeek team,

I'm Daouda Abdoul Anzize, creator of OpenAIBox — a Python tool for universal LLM introspection that maps the semantic role of each hidden dimension via forward hook tracing.

Background

Goodfire.ai showed that DeepSeek R1 is qualitatively different from non-reasoning LLMs. But SAEs describe features, not dimension roles across the full architecture.

OpenAIBox works differently: it contrasts activation patterns across semantic categories (causality, certainty, temporality, emotion...) and identifies which specific hidden dimensions carry each concept — across all layers.

What it found on SmolLM-360M

dim_696  →  syntax + causality + certainty + abstraction + time + emotion  (score=1.0)
dim_295  →  causality + temporality only  (specialist)

Separability scores:
  syntax_semantics  →  cos_sim=0.645  (VERY_SEPARABLE)
  causality         →  cos_sim=0.951  (SEPARABLE)
  certainty         →  cos_sim=0.987  (LOW_SEPARABILITY)
  temporality       →  cos_sim=0.992  (LOW_SEPARABILITY)

Why DeepSeek R1 is the most interesting target

R1 is trained with chain-of-thought. My hypothesis:

Dimensions encoding causality and temporality are stronger/more separable in R1 than in standard LLMs
R1 may have dedicated reasoning dimensions that don't exist in non-reasoning models
The separability of "certainty" (0.987 in SmolLM) may shift significantly in R1, explaining its calibration behavior

This is a direct mechanistic analysis of why R1 reasons differently — not just that it does.

3 lines to run

pip install openaibox

from openaibox import OpenAIBox
gr = OpenAIBox("deepseek-ai/DeepSeek-R1")
gr.discover()
gr.map_dimensions()
gr.export("deepseek_r1_graph.json")

Full graph.json shared immediately. Open to co-authorship.

Daouda Abdoul Anzize
PyPI: openaibox · GitHub: Tryboy869 · Twitter: @Nexusstudio100
Cotonou, Benin → Global Remote

Originally created by @Tryboy869 on GitHub (Jun 10, 2026). Original GitHub issue: https://github.com/deepseek-ai/DeepSeek-R1/issues/856 Hi DeepSeek team, I'm **Daouda Abdoul Anzize**, creator of [OpenAIBox](https://pypi.org/project/openaibox) — a Python tool for universal LLM introspection that maps the **semantic role of each hidden dimension** via forward hook tracing. ### Background Goodfire.ai showed that DeepSeek R1 is **qualitatively different** from non-reasoning LLMs. But SAEs describe *features*, not *dimension roles across the full architecture*. OpenAIBox works differently: it contrasts activation patterns across semantic categories (causality, certainty, temporality, emotion...) and identifies which specific hidden dimensions carry each concept — across all layers. ### What it found on SmolLM-360M ``` dim_696 → syntax + causality + certainty + abstraction + time + emotion (score=1.0) dim_295 → causality + temporality only (specialist) Separability scores: syntax_semantics → cos_sim=0.645 (VERY_SEPARABLE) causality → cos_sim=0.951 (SEPARABLE) certainty → cos_sim=0.987 (LOW_SEPARABILITY) temporality → cos_sim=0.992 (LOW_SEPARABILITY) ``` ### Why DeepSeek R1 is the most interesting target R1 is trained with chain-of-thought. My hypothesis: - Dimensions encoding **causality and temporality** are stronger/more separable in R1 than in standard LLMs - R1 may have **dedicated reasoning dimensions** that don't exist in non-reasoning models - The separability of "certainty" (0.987 in SmolLM) may shift significantly in R1, explaining its calibration behavior This is a direct mechanistic analysis of *why* R1 reasons differently — not just *that* it does. ### 3 lines to run ```python pip install openaibox from openaibox import OpenAIBox gr = OpenAIBox("deepseek-ai/DeepSeek-R1") gr.discover() gr.map_dimensions() gr.export("deepseek_r1_graph.json") ``` Full `graph.json` shared immediately. Open to co-authorship. --- **Daouda Abdoul Anzize** PyPI: [openaibox](https://pypi.org/project/openaibox) · GitHub: [Tryboy869](https://github.com/Tryboy869) · Twitter: [@Nexusstudio100](https://twitter.com/Nexusstudio100) Cotonou, Benin → Global Remote

bsakaguchi

2026-10-26 21:25:49 +00:00

closed this issue
added the
stale

closed-as-stale
labels

bsakaguchi commented

2026-10-26 21:25:50 +00:00

Author

Owner

@github-actions[bot] commented on GitHub (Jul 11, 2026):

This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. If you believe this issue is still relevant, please leave a comment to keep it open. Thank you for your contributions!

@github-actions[bot] commented on GitHub (Jul 11, 2026): This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. If you believe this issue is still relevant, please leave a comment to keep it open. Thank you for your contributions!

bsakaguchi commented

2026-10-26 21:25:50 +00:00

Author

Owner

@github-actions[bot] commented on GitHub (Jul 25, 2026):

false

@github-actions[bot] commented on GitHub (Jul 25, 2026): false