TY - JOUR T1 - A Watermark for Large Language Models Y1 - 2023 A1 - John Kirchenbauer A1 - Jonas Geiping A1 - Yuxin Wen A1 - Jonathan Katz A1 - Ian Miers A1 - Tom Goldstein AB -

Potential harms of large language models can be mitigated by watermarking model output, i.e., embedding signals into generated text that are invisible to humans but algorithmically detectable from a short span of tokens. We propose a watermarking framework for proprietary language models. The watermark can be embedded with negligible impact on text quality, and can be detected using an efficient open-source algorithm without access to the language model API or parameters. The watermark works by selecting a randomized set of "green" tokens before a word is generated, and then softly promoting use of green tokens during sampling. We propose a statistical test for detecting the watermark with interpretable p-values, and derive an information-theoretic framework for analyzing the sensitivity of the watermark. We test the watermark using a multi-billion parameter model from the Open Pretrained Transformer (OPT) family, and discuss robustness and security.

UR - https://arxiv.org/abs/2301.10226 ER - TY - JOUR T1 - What happens to entropy production when conserved quantities fail to commute with each other Y1 - 2023 A1 - Twesh Upadhyaya A1 - William F. Braasch, Jr. A1 - Gabriel T. Landi A1 - Nicole Yunger Halpern AB -

We extend entropy production to a deeply quantum regime involving noncommuting conserved quantities. Consider a unitary transporting conserved quantities ("charges") between two systems initialized in thermal states. Three common formulae model the entropy produced. They respectively cast entropy as an extensive thermodynamic variable, as an information-theoretic uncertainty measure, and as a quantifier of irreversibility. Often, the charges are assumed to commute with each other (e.g., energy and particle number). Yet quantum charges can fail to commute. Noncommutation invites generalizations, which we posit and justify, of the three formulae. Charges' noncommutation, we find, breaks the formulae's equivalence. Furthermore, different formulae quantify different physical effects of charges' noncommutation on entropy production. For instance, entropy production can signal contextuality - true nonclassicality - by becoming nonreal. This work opens up stochastic thermodynamics to noncommuting - and so particularly quantum - charges.

UR - https://arxiv.org/abs/2305.15480 ER - TY - JOUR T1 - Where we are with quantum JF - Nature Physics Y1 - 2022 A1 - Yusuf Alnawakhtha A1 - Carl Miller AB -

A theoretical analysis shows how a person’s location in space could be verified by the transmission of single photons. A vital application of quantum networks may be within reach.

J1 - Nat. Phys. U5 - https://doi.org/10.1038/s41567-022-01597-w ER - TY - JOUR T1 - Why Bohr was (Mostly) Right Y1 - 2017 A1 - Jeffrey Bub AB -

After a discussion of the Frauchiger-Renner argument that no “singleworld” interpretation of quantum mechanics can be self-consistent, I propose a “Bohrian” alternative to many-worlds or QBism as the rational option.

UR - https://arxiv.org/abs/1711.01604 ER - TY - JOUR T1 - Wannier functions using a discrete variable representation for optical lattices JF - Physical Review A Y1 - 2016 A1 - Saurabh Paul A1 - Eite Tiesinga AB -

We propose a numerical method using the discrete variable representation (DVR) for constructing real-valued Wannier functions localized in a unit cell for both symmetric and asymmetric periodic potentials. We apply these results to finding Wannier functions for ultracold atoms trapped in laser-generated optical lattices. Following S. Kivelson [Phys. Rev. B 26, 4269 (1982)], for a symmetric lattice with inversion symmetry, we construct Wannier functions as eigenstates of the position operators xˆ, yˆ, and zˆ restricted to single-particle Bloch functions belonging to one or more bands. To ensure that the Wannier functions are real-valued, we numerically obtain the band structure and real-valued eigenstates using a uniform Fourier grid DVR. We then show, by a comparison of tunneling energies, that the Wannier functions are accurate for both inversion-symmetric and asymmetric potentials to better than 10 significant digits when using double-precision arithmetic. The calculations are performed for an optical lattice with double-wells per unit cell with tunable asymmetry along the x axis and a single sinusoidal potential along the perpendicular directions. Localized functions at the two potential minima within each unit cell are similarly constructed, but using a superposition of single-particle solutions from the two lowest bands. We finally use these localized basis functions to determine the two-body interaction energies in the Bose-Hubbard model and show the dependence of these energies on lattice asymmetry.

VL - 94 U4 - 033606 UR - http://journals.aps.org/pra/abstract/10.1103/PhysRevA.94.033606 CP - 3 U5 - http://dx.doi.org/10.1103/PhysRevA.94.033606 ER - TY - JOUR T1 - Whose Information? Information About What? JF - Quantum [Un]Speakables II: 50 Years of Bell’s Theorem Y1 - 2016 A1 - Jeffrey Bub A1 - Anton Zeilinger A1 - Reinhold Bertlmann ER - TY - JOUR T1 - When the asymptotic limit offers no advantage in the local-operations-and-classical-communication paradigm JF - Phys. Rev. A Y1 - 2014 A1 - Honghao Fu A1 - Debbie Leung A1 - Laura Mancinska AB -

We consider bipartite LOCC, the class of operations implementable by local quantum operations and classical communication between two parties. Surprisingly, there are operations that can be approximated to arbitrary precision but are impossible to implement exactly if only a finite number of messages are exchanged. This significantly complicates the analysis of what can or cannot be approximated with LOCC. Toward alleviating this problem, we exhibit two scenarios in which allowing vanishing error does not help. The first scenario is implementation of projective measurements with product measurement operators. The second scenario is the discrimination of unextendable product bases on two three-dimensional systems.

VL - 89 CP - 052310 U5 - https://doi.org/10.1103/PhysRevA.89.052310 ER - TY - JOUR T1 - Why the Tsirelson bound? JF - The Probable and the Improbable: The Meaning and Role of Probability in Physics Y1 - 2012 A1 - Jeffrey Bub AB - Wheeler's question 'why the quantum' has two aspects: why is the world quantum and not classical, and why is it quantum rather than superquantum, i.e., why the Tsirelson bound for quantum correlations? I discuss a remarkable answer to this question proposed by Pawlowski et al (2009), who provide an information-theoretic derivation of the Tsirelson bound from a principle they call 'information causality.' U4 - 167-185 UR - http://arxiv.org/abs/1208.3744v1 J1 - Published in Meir Hemmo and Yemima Ben-Menahem (eds.) U5 - 10.1007/978-3-642-21329-8_11 ER - TY - JOUR T1 - Wigner crystals of ions as quantum hard drives JF - Physical Review A Y1 - 2008 A1 - J. M. Taylor A1 - T. Calarco AB - Atomic systems in regular lattices are intriguing systems for implementing ideas in quantum simulation and information processing. Focusing on laser cooled ions forming Wigner crystals in Penning traps, we find a robust and simple approach to engineering non-trivial 2-body interactions sufficient for universal quantum computation. We then consider extensions of our approach to the fast generation of large cluster states, and a non-local architecture using an asymmetric entanglement generation procedure between a Penning trap system and well-established linear Paul trap designs. VL - 78 UR - http://arxiv.org/abs/0706.1951v1 CP - 6 J1 - Phys. Rev. A U5 - 10.1103/PhysRevA.78.062331 ER - TY - JOUR T1 - Weak Fourier-Schur sampling, the hidden subgroup problem, and the quantum collision problem Y1 - 2006 A1 - Andrew M. Childs A1 - Aram W. Harrow A1 - Pawel Wocjan AB - Schur duality decomposes many copies of a quantum state into subspaces labeled by partitions, a decomposition with applications throughout quantum information theory. Here we consider applying Schur duality to the problem of distinguishing coset states in the standard approach to the hidden subgroup problem. We observe that simply measuring the partition (a procedure we call weak Schur sampling) provides very little information about the hidden subgroup. Furthermore, we show that under quite general assumptions, even a combination of weak Fourier sampling and weak Schur sampling fails to identify the hidden subgroup. We also prove tight bounds on how many coset states are required to solve the hidden subgroup problem by weak Schur sampling, and we relate this question to a quantum version of the collision problem. UR - http://arxiv.org/abs/quant-ph/0609110v1 J1 - Proc. 24th Symposium on Theoretical Aspects of Computer Science (STACS 2007) U5 - 10.1007/978-3-540-70918-3_51 ER - TY - JOUR T1 - Why the quantum? Y1 - 2004 A1 - Jeffrey Bub AB - This paper is a commentary on the foundational significance of the Clifton-Bub-Halvorson theorem characterizing quantum theory in terms of three information-theoretic constraints (Foundations of Physics 33, 1561-1591 (2003); quant-ph/0211089). I argue that: (1) a quantum theory is best understood as a theory about the possibilities and impossibilities of information transfer, as opposed to a theory about the mechanics of nonclassical waves or particles, (2) given the information-theoretic constraints, any mechanical theory of quantum phenomena that includes an account of the measuring instruments that reveal these phenomena must be empirically equivalent to a quantum theory, and (3) assuming the information-theoretic constraints are in fact satisfied in our world, no mechanical theory of quantum phenomena that includes an account of measurement interactions can be acceptable, and the appropriate aim of physics at the fundamental level then becomes the representation and manipulation of information. UR - http://arxiv.org/abs/quant-ph/0402149v1 J1 - Studies in History and Philosophy of Modern Physics 35B ER -