QLASS in Quantum Computing
The term “Quantum Glass-based Photonic Integrated Circuits” (QLASS) refers to the field of quantum computing. This project is a cooperative European research endeavor that seeks to create quantum computers using glass and light. SMEs and researchers from France, Italy, and Germany are brought together by the Fondazione Politecnico di Milano, which is run by Giulia Acconcia.
You can also read Free Space Optical Communication in Satellite-Based QKD
Key Aspects of the QLASS Quantum Computing Project:
- Objective:Using femtosecond laser writing (FLW) to create 3D waveguides inside glass is the ambitious objective of creating a quantum photonic integrated circuit (QPIC).
- Material Innovation: When it comes to creating quantum photonic chips from glass, Ephos, an Italian startup working on QLASS, is unique. Because glass is a great medium for light transmission and helps confine photons, it helps avoid absorption and the loss of important information.
- Performance Features: Ephos’ glass chips have up to 200 optical modes that may be reconfigured, enabling dynamic modification of the way light propagates within the device. Glass also has incredibly low interface losses (less than 5%), which is essential for scalable and modular systems that link several circuits.
- Technological Components: The project’s technological components include circuits that allow for reconfigurable state manipulation through the use of a large number of cryogenic-detector channels (200) and phase shifters (1000), high-performance single-photon sources, and superconducting nanowire single-photon detectors (SNSPDs).
- End Goal: The ultimate goal is to create a quantum photonics platform that can be used with Variational Quantum Algorithms (VQAs), the most promising method for near-term quantum advantage.
- Real-World Applications: The researchers at QLASS are working to find solutions to real-world issues like developing new medications, improved lithium-ion batteries (which are essential for storing renewable energy and energising transportation), and new materials. Giulia Acconcia’s interest in eco-friendly technologies is in line with these uses.
- Collaborative Effort: The following individuals have contributed to this really pan-European endeavour:
- Ephos, an Italian company, uses laser writing to create the glass chips.
- Pixel Photonics (Germany): Enhancing lasers with high sensitivity.
- Schott AG, a German company, provides premium glass substrates.
- The Polytechnic University of Milan (Italy) team led by Giulia Acconcia creates high-performance electronics.
- Sapiensza University (Rome, Italy): Known for its proficiency in experimental quantum optics, this institution manages the production of single photons and is expected to construct an operational photonic quantum device by 2026.
- The French National University is developing open-source software for quantum processes.
- Innovative energy-storage possibilities will be modelled and tested by the Université de Montpellier (France) and the National Centre for Scientific Research.
- By 2025, the effort will assist Europe’s Digital Decade and Chips Act goals of a domestic quantum-chip industry by 2030 and the continent’s first quantum-accelerated supercomputer. The €1 billion, ten-year EU Quantum Technologies Flagship sponsors it.
You can also read Cosmic Rays Quantum Computing For Superconducting CPUs
QLASS in Language Agents
The term “Q-guided Language Agent Stepwise Search” (QLASS) has a second meaning that is explained in a research paper from arxiv.org. For complicated interactive tasks in particular, this QLASS provides an innovative way to improve the inference skills of open-source language agents.
Challenges Addressed by QLASS (Language Agents):
- Data Scarcity & Costly Annotations: Many times, training linguistic agents for intricate tasks necessitates significant human annotations of intermediate exchanges, which restricts scalability and is costly.
- Sub-optimal Policies: A lot of current approaches rely on outcome reward models, which only offer one ultimate reward. These models are unable to accurately provide feedback for every step in lengthy, complex trajectories, which may result in ineffective or suboptimal actions.
- Inefficiency in Search Space: Traditional reinforcement learning techniques, such as Q-learning, and direct exploration may be rendered ineffective by the large action space in linguistic agent tasks.
You can also read Alpine Quantum Technologies Developed And Include PIAST-Q
How QLASS (Language Agents) Addresses These Challenges:
With the use of calculated Q-values, QLASS presents a novel approach that creates intermediate annotations for language agents, offering vital step-by-step direction during model inference.
- Process Reward Modeling with Q-Value Estimation:
- Exploration trees are used to formalise self-generated exploratory pathways.
- The Bellman equation is used to estimate Q-values for every intermediate state-action pair in the exploration tree, in contrast to straightforward outcome-based rewards. This gives more precise control over the paths of reasoning and illustrates how decisions made now result in longer-term benefits.
- Tree pruning is used to lessen the computational load by preventing generation on branches that result in zero-outcome rewards and restricting expansion for the initial phases of a trajectory.
- QNet Training: A Q-network (QNet) is trained under supervision using the predicted Q-values from the exploration trees. To predict Q-values, this QNet employs a pre-trained Large Language Models (LLMs) as its backbone, with an attached value head. Bypassing online Q-learning’s instability and large exploration costs, this method avoids using it directly in linguistic contexts.
- Q-Guided Generation Strategy:
- The agent’s inference is guided by the trained QNet.
- The agent samples multiple actions at each stage, and the action with the greatest Q-value (as predicted by QNet) is carried out. This guarantees better decision-making at every stage.
- By paraphrasing task specifications, perturbation-augmented generation is utilised to boost action diversity for tasks such as WebShop.
Performance and Robustness (Language Agents):
- Superior Performance: In a variety of agent contexts, including WebShop, SciWorld, and ALFWorld, QLASS continuously receives the highest scores among all open-sourced baselines. It even shows similar or better performance on some benchmarks than proprietary closed-source models such as GPT-4.
- Effectiveness: Across a range of search budgets, QLASS outperforms “Best-of-N” sampling, obtaining superior results with a considerable reduction of completion tokens.
- Limited Supervision: It is noteworthy that QLASS maintains its good performance even when behaviour cloning reduces over half of the annotated data, demonstrating its effectiveness and resilience in situations with a lack of expert data.
- Effective Decision Making: A case study demonstrates how QLASS helps the agent to make good decisions in fewer steps by preventing repetitive, pointless activities (like opening and closing a refrigerator repeatedly) by comprehending the stepwise value. Its robustness across architectures has also been confirmed on several base LLMs, such as Llama-2-13B.
You can also read What is PEA, How Does PEA Work in Quantum Noise Mitigation
