ForceDream Research OS · FD-2026-007
Cross-Model Cognitive Task Combiner: Ensemble Architectures for Heterogeneous LLM Outputs
ForceDream Research Team, Cognitive Architecture Division2026-01-30v1.13 pages
InferenceL1L4L3
WORM Access Seal · L828
fd2026007a1e9b3c
CMCTC achieves 9.4% hallucination reduction and 7.2% factual grounding improvement across 12,000 diverse tasks by aggregating outputs from multiple heterogeneous LLM providers using confidence-weighted voting calibrated to task type. Mean additional latency: 34ms for soft-weighted synthesis mode.
1. Motivation
Single-model inference pipelines exhibit systematic failure modes: hallucination rates of 3-8% on factual tasks, distributional shift under prompt variation, and task-type miscalibration. Scaling alone does not eliminate these failure modes. CMCTC addresses them through ensemble aggregation.
2. Confidence-Weighted Voting
For each task, CMCTC queries k providers (typically k=3) and assigns each output a confidence score using a lightweight calibration model. The final output is produced by a confidence-weighted voting function optimising over discrete or continuous output spaces.
📄
Unlock the full paper
Enter your name and email to read all 5 sections and receive the PDF. Free. WORM-sealed. New papers delivered automatically.
✓ Free access✓ WORM-sealed✓ No spam✓ Auto-delivered
3. Three Operating Modes
Hard-majority: highest weighted vote count selected, used for classification. Soft-weighted: confidence-weighted combination, used for generation. Adversarial: contradiction detection against WORM-sealed Atlas records, triggering re-query with explicit contradiction context, used for factual recall.
4. Evaluation
Across 12,000 diverse tasks: hallucination rate 9.4% reduction (p<0.001), factual grounding 7.2% improvement vs Atlas ground truth, task success rate 4.1% improvement, additional latency (soft-weighted) 34ms, throughput impact 2.1% reduction.
5. Conclusions
CMCTC demonstrates reliable quality improvements across diverse task types. The adversarial mode provides grounding in WORM-sealed facts, critical for compliance and fraud detection where hallucination is a regulatory risk.
Live API Endpoints
POST /v1/inference/racePOST /v1/inference/routeGET /v1/models/listPOST /v1/inference/ensembleCitation
ForceDream Research Team (2026). Cross-Model Cognitive Task Combiner. ForceDream Intelligence OS Research Series, FD-2026-007. https://forcedream.com/research/cross-model-cognitive-task-combiner-llm-ensemble-architecture
Build on ForceDream
Free API key. 78% earnings on every call. WORM-sealed.
Get free API key →