Long‑horizon reasoning is prone to early mistakes that corrupt later decisions. This work introduces QART (Quantum‑Augmented Reasoning Transformer), a quantum‑classical hybrid architecture whose backbone consists of a language model, quantum encoding, CIM‑based QUBO optimization, and quantum decoding.
Semantic information may be drawn from hidden representations or model‑generated text; the exact encoding and optimization pipelines are proprietary. Under explicit assumptions the authors derive a conditional asymptotic reliability separation compared with single‑trajectory autoregressive LLMs.
For a common family of tasks where optimality aligns with acceptance criteria, the autoregressive acceptance probability tends to zero as the cumulative conditional risk of irreversible errors diverges. In contrast, QART’s probability of recovering the optimal‑path remains bounded away from zero provided that conditional probabilities for path coverage, semantic fidelity, spectral certification, dynamical reachability, and faithful readout stay uniformly positive under a prescribed resource schedule.
Empirical evaluation on six long‑horizon benchmarks, using DeepSeek V4 Flash, GLM‑5.3, and GPT‑5.5 xhigh as backbones within a Codex agent environment, shows QART outperforming 14 of 15 backbone‑benchmark pairs. Gains reach 84.0% on SciCode, 47.6% on τ³‑Bench, and 44.4% on Terminal‑Bench 4.0, while the DeepSeek V4 Flash configuration regresses by 7.8% on DeepSWE. These results do not directly confirm the asymptotic separation.
The authors formulate several quantum scaling laws as conditional hypotheses. Claiming a quantum advantage requires demonstrating a CIM advantage over strong classical solvers and transferring that advantage to end‑to‑end reasoning after accounting for all system overheads.
In summary, QART outlines a plausible quantum‑accelerated path for long‑horizon reasoning, yet its practical benefit hinges on further advances in quantum hardware and optimization algorithms. Review