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.

+251% R/E vs dense99% reasoning kept at 50% sparsitytraining free

The setup

Why pruning normally breaks reasoning.

Dense

all cells active, full cost

Naive pruning

error unbounded, reasoning degrades

RASP

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
r / 12

Rank is the control knob. Kept dense scales with r.

Error bound
≤ σ(r+1)

Residual tail only. The principal part is untouched.

Training
none

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.

R/E = Q / E

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.
baseline

Dense inference, same open weights, same hardware. R/E is 1.00 by construction.

worked ratio

(QRASP / ERASP) ÷ (Qdense / Edense) = 3.51, so +251%.

notation

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.

Open, served dense
1.00x

Full compute on every token.

Open, PrizmalRun
3.51x

Same weights, RASP at 50% sparsity. +251% R/E.

Compute
-60%

Per token, at the same reasoning quality.

Energy
-25.1%

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.

AGREE · STAY SHALLOWDISAGREE · FULL DEPTHTOKENDRAFTproposesVERIFIERchecksAGREEMENT?NEXT TOKEN

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.

01
Isocline Pruning
Per-token compute pruned at zero divergence from the full answer.
02
Draft-Guided Router
The draft's live signals set how much of the large model each token engages.
03
Disagreement Control Loop
Relaxes pruning the instant draft and verifier diverge.
04
Layer Sensitivity Caps
Per-layer limits that bound error as it propagates through depth.
05
Masked Matrix Multiply
Hardware-aware kernels that turn pruned compute into speed on the metal.

Provenance and access

From the bench to your traffic.

The bench
Research rooted at Mila. The researchers who wrote RASP build PrizmalRun.
The proof on your traffic
The R/E Audit runs these mechanisms against your real workload.

References

  1. [1]Sun et al., A Simple and Effective Pruning Approach for Large Language Models (WANDA), 2023.
  2. [2]Dong et al., GRIFFIN: Mixing Memory-Efficient Linear Recurrences with Local Attention, 2024.
  3. [3]Liu et al., TEAL: Training-Free Activation Sparsity in Large Language Models, 2024.
  4. [4]Leviathan, Kalman and Matias, Fast Inference from Transformers via Speculative Decoding, ICML 2023.
  5. [5]Deloitte, State of Generative AI in the Enterprise (annual series).
  6. [6]Gartner, Forecast: Generative AI, Worldwide.
  7. [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.