Research
RASP. The foundational breakthrough.
The science underneath: RASP keeps the part of the model that reasons and prunes the part that does not, measured in reasoning per unit of energy.
The setup
Why pruning normally breaks reasoning.
all cells active, full cost
error unbounded, reasoning degrades
principal block kept dense, residual pruned, error bounded by σ(r)
Sparsity is not the trick. Knowing what the matrix can afford to lose is.
The decomposition
Protect what reasons. Prune what doesn't. Move the dial.
Rank is the control knob. Kept dense scales with r.
Residual tail only. The principal part is untouched.
Applied post hoc. No retraining required.
Qualitative bound. The bar tightens as r grows. Reasoning held, training free.
The R/E metric
Reasoning per unit of energy. The unit the bill should be in.
reasoning per energy
- Q · reasoning delivered
- Benchmark score across the evaluation suites. Units: normalized accuracy, 0 to 1.
- E · energy consumed
- Joules, measured at the device via NVML.
Dense inference, same open weights, same hardware. R/E is 1.00 by construction.
(QRASP / ERASP) ÷ (Qdense / Edense) = 3.51, so +251%.
σ(r) denotes the residual tail: the standard deviation of the residual after the principal part of a weight matrix is kept. The RASP output error is bounded by σ(r), not by the full weight. Rank is the control knob.
The measured results
Three models. One reduction.
Full compute on every token.
Same weights, RASP at 50% sparsity. +251% R/E.
Per token, at the same reasoning quality.
Device level, measured with NVML.
How we measure
methods plate · disclosed in full
- models tested
- Llama-3 8B · Gemma-2 9B · DeepSeek-R1 7B
- sparsity setting
- 50%
- evaluation suites
- four canonical reasoning suites
- energy measurement
- NVML, device level
- baseline
- dense inference · same weights · same hardware
Methods questions welcome: research@prizmal.ai
Ahead of academic sparsity baselines: +7.93 WANDA [1], +7.33 GRIFFIN [2], +2.10 TEAL [3]. Lineage of per-token decoding research: speculative decoding [4]. Enterprise spend signals: Deloitte [5], Gartner [6].
The live loop
RASP is the foundation. The loop makes it live.
A disagreement-aware controller reads the model from inside as it runs and tunes how hard it works, block by block. This is not compression. It is governing when the full model runs.
The patents
Five families. One path. The moat is on the mechanism, not the weights.
Five USPTO patent families across one inference path. Filed Dec 2025, PCT in progress.
Provenance and access
From the bench to your traffic.
References
- [1]Sun et al., A Simple and Effective Pruning Approach for Large Language Models (WANDA), 2023.
- [2]Dong et al., GRIFFIN: Mixing Memory-Efficient Linear Recurrences with Local Attention, 2024.
- [3]Liu et al., TEAL: Training-Free Activation Sparsity in Large Language Models, 2024.
- [4]Leviathan, Kalman and Matias, Fast Inference from Transformers via Speculative Decoding, ICML 2023.
- [5]Deloitte, State of Generative AI in the Enterprise (annual series).
- [6]Gartner, Forecast: Generative AI, Worldwide.
- [7]RASP: Residual-Aware Sparsification (Prizmal). Internal technical report, external publication forthcoming.
Run the science on your traffic
The R/E Audit measures reasoning per unit of energy on your real workload, read only and reversible.