QLASS European Use Glass & Light to Create Quantum Chips
QLASS in Quantum Computing
“Quantum Glass-based Photonic Integrated Circuits” (QLASS) is quantum computing. This European cooperative research initiative aims to build quantum computers using glass and light. Giulia Acconcia's Fondazione Politecnico di Milano brings together French, Italian, and German SMEs and researchers.
Key QLASS Quantum Computing Project Features:
The ambitious goal of a quantum photonic integrated circuit is to generate 3D waveguides inside glass using femtosecond laser writing (FLW).
Ephos, an Italian startup developing QLASS, is unique in making glass quantum photonic chips. Glass transmits light well and confines photons, preventing absorption and data loss.
Performance: Ephos' glass chips can reconfigure up to 200 optical modes to dynamically change light propagation. Glass has extremely low interface losses (less than 5%), making it ideal for scalable, modular circuit connectivity.
Technological Components: The project uses circuits with 200 cryogenic-detector channels, 1000 phase shifters, high-performance single-photon sources, and superconducting nanowire detectors (SNSPDs) for reconfigurable state manipulation.
End Goal: The ultimate goal is to construct a quantum photonics platform for Variational Quantum Algorithms (VQAs), the most promising near-term quantum advantage approach.
Real-World Applications: QLASS researchers are developing novel pharmaceuticals, lithium-ion batteries (needed to store renewable energy and power vehicles), and materials. These usage match Giulia Acconcia's interest in green technologies.
Collaboration: The following people helped with this pan-European project:
Ephos, an Italian business, laser-writes glass chips.
Pixel Photonics (Germany): High-sensitivity laser enhancement. German firm Schott AG makes quality glass substrates.
Italian Polytechnic University of Milan team led by Giulia Acconcia makes high-performance electronics.
Sapiensza University (Rome, Italy): This experimental quantum optics centre produces single photons and plans to build a photonic quantum device by 2026.
The French National University develops open-source quantum software.
Université de Montpellier (France) and the National Centre for Scientific Research will model and test innovative energy-storage options.
Europe's Digital Decade and Chips Act targets of a domestic quantum-chip industry by 2030 and the continent's first quantum-accelerated supercomputer will be supported by 2025. It is sponsored by the €1 billion, 10-year EU Quantum Technologies Flagship.
QLASS in Language Agents
A research study on arxiv.org explains a second meaning of “Q-guided Language Agent Stepwise Search” (QLASS). This novel QLASS improves open-source language agents' inference ability for complex interactive tasks.
QLASS Language Agent Challenges:
Data Scarcity & Costly Annotations: Training language agents for complex tasks often requires human annotations of intermediate exchanges, which limits scalability and is expensive.
Sub-optimal Policies: Many existing approaches use outcome reward models with one final reward. These models cannot reliably offer feedback for every step in long, complex trajectories, which may lead to ineffective or suboptimal responses.
The wide action space in linguistic agent challenges may render traditional reinforcement learning methods like Q-learning and direct exploration ineffective.
How Language Agents QLASS Addresses These Challenges: The innovative QLASS method produces intermediate annotations for linguistic agents using estimated Q-values, providing crucial model inference guidance.
Explore trees are used to formalise self-generated exploratory routes in process reward modelling using Q-Value estimation.
Unlike outcome-based incentives, the Bellman equation estimates Q-values for every intermediate state-action pair in the exploration tree. This provides more exact reasoning control and shows how current decisions have long-term benefits.
Tree trimming reduces computing load by stopping generation on branches with zero-outcome rewards and limiting trajectory extension in the beginning.
The projected Q-values from the exploration trees are used to train a Q-network (QNet) under supervision. This QNet uses pre-trained Large Language Models (LLMs) with a value head to forecast Q-values. This strategy avoids employing online Q-learning in linguistic situations due to its instability and high exploration costs.
Q-Guided Generation:
The trained QNet guides agent inference.
The agent samples numerous actions at each stage and does the one with the highest Q-value (as predicted by QNet). This ensures smarter decisions throughout. Parsing task descriptions with perturbation-augmented generation increases action diversity for WebShop.
Language Agent Performance and Robustness:
QLASS consistently outperforms all open-sourced baselines in WebShop, SciWorld, and ALFWorld agent contexts. It even outperforms proprietary closed-source models like GPT-4 on several benchmarks.
Effectiveness: QLASS outperforms “Best-of-N” sampling with fewer completion tokens across search budgets.
Limited Supervision: QLASS performs well even when behaviour cloning lowers almost half of the annotated data, showing its resiliency in the absence of expert data.
Effective Decision Making: A case study shows how QLASS prevents repetitive, unnecessary operations like opening and closing a refrigerator by understanding the stepwise value. Llama-2-13B and other base LLMs have confirmed its architecture-independent robustness.