CoDraft+: Adaptive Multi-Model Collaboration through a Shared Canonical Workspace for Resource-Constrained Edge Devices

Deploying monolithic large language models on offline edge hardware (<8 GB RAM) triggers fatal memory exhaustion, so CoDraft+ intercepts each incoming query via an Adaptive Collaboration Controller that bypasses simple prompts (C(q) < 0.25) on a single-turn fast path while routing complex queries into structured multi-model collaboration. Rather than exchanging bloated dialogue transcripts, Drafting Specialist Model A initializes a four-section Shared Canonical Workspace (W_t = <D_t, L_t, H_t>) whose raw JSON outputs are ingested through a deterministic 4-Tier Resilient Patcher achieving a 0.00% crash rate across 1,500+ transactions. Patching Specialist Model B then audits the evolving draft in-place, scoring extracted atomic claims via conditional log-likelihood against thresholds tau_sup = -2.1 and tau_contra = -5.8 nats inside a Grounded Evidence Ledger that cuts hallucinations by 68.2% (24.5% down to 7.8%) and lifts claim accuracy from 68.2% to 89.4%. Execution halts automatically once normalized Levenshtein edit churn drops below Delta_t < 0.15 (triggering early stopping on 66.7% of queries), claim pressure reaches V_t = 0, or turn/patience budgets (T_max = 3, P_max = 2) expire. Evaluated strictly on genuine FP32 CPU forward passes across 11 datasets and 3 architectural families (Pythia / GPT-NeoX, BabyLM / GPT-BERT, and Stentor2 / Llama) with zero synthetic data:• Set 1 Superadditivity (Pythia): Pairing Pythia-14M + Pythia-31M (44.56M active parameters) achieves 8.25% ROUGE-L (+79.7% relative improvement over the 1.58x Pythia-70M monolith at 4.59% and +44.5% over isolated parallel ensembling at 5.71%) while using 1,234.1 MB peak RSS RAM (16 MB less than Pythia-70M's 1,250.1 MB).• Boundary Laws (Sets 2 & 3): Documents 512-token context-horizon collapse in un-tuned BabyLM models (establishing N_ctx >= 1,024 tokens) and asymmetric capability dilution when Stentor2-12M collaborates unweighted with Stentor2-30M.• Systems Profiling: ~100x latency acceleration via native KV caching (120s down to 0.85–1.77s/turn) and 6.46x scaling across 12 AVX2/oneDNN CPU threads (126,800 MFLOPS). Open-Source Code & Evaluation Suite: https://github.com/satyamsrj/codraft-plus

Authors

Institutions

Publication Details

Journal
Zenodo (CERN European Organization for Nuclear Research)
Published
2026-11-17
DOI
https://doi.org/10.5281/zenodo.24322588
Primary Topic
Topic Modeling
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

CoDraft+: Adaptive Multi-Model Collaboration through a Shared Canonical Workspace for Resource-Constrained Edge Devices

Satyam Saroj
Zenodo (CERN European Organization for Nuclear Research)
Topic Modeling
preprint

CoDraft+: Adaptive Multi-Model Collaboration through a Shared Canonical Workspace for Resource-Constrained Edge Devices

Satyam Saroj
preprint en

Abstract

Deploying monolithic large language models on offline edge hardware (<8 GB RAM) triggers fatal memory exhaustion, so CoDraft+ intercepts each incoming query via an Adaptive Collaboration Controller that bypasses simple prompts (C(q) < 0.25) on a single-turn fast path while routing complex queries into structured multi-model collaboration. Rather than exchanging bloated dialogue transcripts, Drafting Specialist Model A initializes a four-section Shared Canonical Workspace (W_t = <D_t, L_t, H_t>) whose raw JSON outputs are ingested through a deterministic 4-Tier Resilient Patcher achieving a 0.00% crash rate across 1,500+ transactions. Patching Specialist Model B then audits the evolving draft in-place, scoring extracted atomic claims via conditional log-likelihood against thresholds tau_sup = -2.1 and tau_contra = -5.8 nats inside a Grounded Evidence Ledger that cuts hallucinations by 68.2% (24.5% down to 7.8%) and lifts claim accuracy from 68.2% to 89.4%. Execution halts automatically once normalized Levenshtein edit churn drops below Delta_t < 0.15 (triggering early stopping on 66.7% of queries), claim pressure reaches V_t = 0, or turn/patience budgets (T_max = 3, P_max = 2) expire. Evaluated strictly on genuine FP32 CPU forward passes across 11 datasets and 3 architectural families (Pythia / GPT-NeoX, BabyLM / GPT-BERT, and Stentor2 / Llama) with zero synthetic data:• Set 1 Superadditivity (Pythia): Pairing Pythia-14M + Pythia-31M (44.56M active parameters) achieves 8.25% ROUGE-L (+79.7% relative improvement over the 1.58x Pythia-70M monolith at 4.59% and +44.5% over isolated parallel ensembling at 5.71%) while using 1,234.1 MB peak RSS RAM (16 MB less than Pythia-70M's 1,250.1 MB).• Boundary Laws (Sets 2 & 3): Documents 512-token context-horizon collapse in un-tuned BabyLM models (establishing N_ctx >= 1,024 tokens) and asymmetric capability dilution when Stentor2-12M collaborates unweighted with Stentor2-30M.• Systems Profiling: ~100x latency acceleration via native KV caching (120s down to 0.85–1.77s/turn) and 6.46x scaling across 12 AVX2/oneDNN CPU threads (126,800 MFLOPS). Open-Source Code & Evaluation Suite: https://github.com/satyamsrj/codraft-plus

Zenodo (CERN European Organization for Nuclear Research)
Ajay Kumar Garg Engineering College (IN)
Topic Modeling
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon