PADS·Explainer

Pause-Aware Depth Scheduling

Using the natural silence before a person finishes speaking as free compute time for the expensive part of language-model inference — on hardware with no discrete GPU.

6 / 7
Go/No-Go falsification tests executed on the development host (one pending a root-only command)
1,050 ms
median real pause in the AMI corpus, versus a 164 ms extrapolated 7B decode step
13.5 → 20.5%
held-out early-exit acceptance at 8 of 16 layers after multi-exit training
+21.3 pp
held-out token-confidence agreement gain over the base model at 50% compute

The one-sentence bet

Conversational systems on CPU-only devices pay the full serial cost of every reasoning layer after the user finishes speaking, and the user waits for all of it. But human turns contain a short, predictable silence before they end — typically a few hundred milliseconds. PADS asks whether that silence can be used to begin full-depth computation early, using self-speculative early exit inside a single model, so that by the time the turn actually ends some or all of the reasoning cost is already paid for.

Why a single model

Routing between a small and a large model is a crowded, solved problem, and it cannot hand off KV-cache state between differently sized models — caches are tied to a model's own hidden dimension and layer count. PADS never switches models. It switches depth within one network: early layers draft, later layers verify, and the pause buys time to start the later layers sooner.

How it works

Partial utterance (streaming) Turn-taking predictor → p_end Dialogue-act classifier → p_deep Trigger policy — conservative AND gate Stay shallow (early exit) Trigger deep — precompute during pause End of turn: confirm or discard response generation
The PADS turn loop. A wrong prediction discards the speculative branch; because verification uses the same network, the final output distribution is unchanged — the correctness guarantee inherited from standard speculative decoding. Full diagram set on Architecture.

Two tiers, so a negative result is still a result

Safe core measured

A systematic characterisation of self-speculative early exit on CPU-bound hardware, plus a dialogue-act-driven exit policy compared against the standard token-confidence policy at matched compute. Independently publishable.

Stretch layer in progress

PADS proper: using the pause to start full-depth computation before the turn ends. Needs a turn-taking signal and real pause timing. If it shows no gain, it is written up as a limitation — the safe core stands on its own.

Status at a glance

PhaseWhat it coversState
Foundations & Go/No-GoLiterature gap, pre-registered falsification suite, environment, hardware characterisationdone
Phase 1 — safe coreLayerSkip reproduction, multi-exit training, real dialogue-act classifier, held-out acceptancedone
Phase 2 — exit policiesToken-confidence vs dialogue-act routing at matched compute, with uncertaintydone
Phase 3 — baselinesPlain decoding, two-model speculative, cascading, LayerSkip configuration reproduced locallyrunning
Phase 4 — pause-aware end-to-endTurn-taking proxy, real pause distributions, PADS vs safe coreplanned
Phase 5 — target hardware & manuscriptDell Latitude 5490 validation, energy/thermal protocol, IEEE manuscriptplanned
Read this before quoting any number

Every measurement on this site was taken on the development host (a laptop with a discrete GPU used for training), not on the Dell Latitude 5490 that the project's claims are about. Development-host numbers are labelled as such throughout. Wall-clock speedup for PADS proper is not measured, audio turn-taking is not implemented, and energy measurement is blocked on root access.

Explore

Architecture

The full system, training, evaluation, and deployment pipelines in diagrams.

Technology

Every technique and tool used, explained: early exit, LayerSkip, LoRA, GGUF, policies, statistics.

Experiments

What was measured, what it showed, and what remains unmeasured.

Glossary

Definitions for every technical term used on this site.