| C. Troussas, A. Krouska, Ph. Mylonas, C. Sgouropoulou, I. Voyiatzis |
| You Only Think Once: Unifying Retrieval, Reasoning, Generation, and Verification in a Single Forward Pass |
| 11th South-East Europe Design Automation, Computer Engineering, Computer Networks and Social Media Conference (SEEDA-CECNSM 2026), Thessaloniki, Greece, September 18-20, 2026 |
|
ABSTRACT
|
| Current state-of-the-art complex knowledge task AI systems often rely on a sequential chain of retrieval, reasoning, generation and verification modules in a multi-stage pipeline. While this design has achieved strong performance, it is prone to compounded error from each pipeline stage, suffers from high latency and inability of inter-stage end-to-end optimization. In this paper, we propose UniCog (Unified Cognitive Architecture), a system performing retrieval, reasoning, generation and self-verification simultaneously in a single transformer forward pass. We introduce three key architectural innovations for UniCog: (1) shared transformer backbone with parallelized cognitive task heads and task-specific lightweight adapters; (2) cross-cognitive attention mechanism which facilitates information exchange between task heads at inference time; and (3) joint training objective that simultaneously optimizes all cognitive abilities through multi-task learning with dynamic loss balancing. We evaluate UniCog on 6 diverse benchmarks including retrieval (Natural Questions), reasoning (GSM8K), code generation (HumanEval), factual verification (TruthfulQA), knowledge (MMLU) and science reasoning (ARC-Challenge). Our system UniCog-7B achieves an average score of 0.891, outperforming its equivalent size pipeline system (0.878) with 3.3 times reduction in inference latency (47ms vs 156ms). Scaling experiments from 1.3B to 13B parameters prove that UniCog continuously outperforms its pipeline counterparts across all scales. Our work sets a new paradigm for cognitive AI: UniCog reveals that complex cognitive abilities need not be sequenced into pipeline modules.
|
| 18 September, 2026 |
| C. Troussas, A. Krouska, Ph. Mylonas, C. Sgouropoulou, I. Voyiatzis, "You Only Think Once: Unifying Retrieval, Reasoning, Generation, and Verification in a Single Forward Pass", 11th South-East Europe Design Automation, Computer Engineering, Computer Networks and Social Media Conference (SEEDA-CECNSM 2026), Thessaloniki, Greece, September 18-20, 2026 |
[
BibTex] [
Print] [
 Back] |
