DeepSeek Experiments

Current DeepSeek RL launch and qualification entry: Miles baseline.

Lineages and controls

Phase	Trainable weights	Comparison
Frozen-backbone math adaptation	Added branches	Ordinary versus simplicial attention
Full-weight continuation	Text backbone and adapters	Each arm continues its own matched adapter checkpoint
Scratch studies	Randomly initialized backbone and branches	Ordinary, simplicial, linear and linearized-simplicial candidates
Math RL	Declared policy parameter set	Outcome rewards from matched SFT parents

The final matched supervised pair is step 4537/4537, with 756,364,650 supervised tokens each: 307 adapter updates followed by 4,230 full-weight updates.

Mechanism

Adapters enter after attention expands into the four residual streams and before the FFN read. The simplicial branch uses exact joint softmax over 32×512 causal key windows. The ordinary control uses the 512-token window.

The adapters match geometry and shared initialization, not parameter count: normal has 146.84M parameters; simplicial has 167.82M.

Results and boundaries

Normal slightly wins final held-out math CE and has higher recorded throughput. The earlier 91M-token simplicial advantage reversed later.

The original scratch pair stopped at 240.875M tokens of a planned 10B; it cannot establish the converged ranking. Its first-option-biased MMLU predictions are not evidence of strong knowledge.

Native sparse attention already uses TileLang. Serving and incremental-cache work require separate full-model qualification.

Records: docs/DEEPSEEK_V41_FULL_COMPARISON.md, docs/DEEPSEEK_V41_NORMAL_CONTROL.md, docs/DEEPSEEK_V41_DEBUG_20260922.md.

Next: Math RL · Evaluation