arXiv subjects

Paul M. Riechers

Publications and source records attributed to Paul M. Riechers.

At least 19 recordsLinked to original sources

arXiv2026

Large Language Models Develop Belief State Geometry In-Context

Large language models (LLMs) trained on next-token prediction exhibit remarkable in-context learning (ICL) abilities, yet the representations that support ICL remain poorly understood. We consider such representations in a controlled setting: prompting LLMs with data emitted from hidden Markov models (HMMs) and probing for the corresponding belief state -- the posterior distribution over the HMM's hidden states given the observed token history. Across six open-source LLMs prompted with data from 40 HMMs selected for non-trivial belief structure, we find that belief states are linearly decodable from residual stream activations, with peak probe $R^2$-values from 0.83-0.99 across HMM and LLM combinations, ranging from early to late layers. To establish functional relevance, we intervene directly on the probe-identified subspace via patching and steering, resulting in downstream prediction quality on the order of the untampered model, while controls degrade performance substantially. Together, these results provide representation-level evidence that ICL in open-source LLMs approximates optimal Bayesian prediction over a context-inferred generative model. More broadly, our findings extend prior results linking input-distribution structure to activation geometry: from toy networks trained explicitly on HMM data to production-scale LLMs.

cs.LG↗

arXiv2026

Time-ordered free energy in correlated quantum systems: An agentic approach

How much work can an agent extract from a temporal sequence of quantum states when it can only operate online under causal constraints---deciding which energy extraction method to use with knowledge of what it has observed before? Here, we study this problem in the context of quantum state sequences that are potentially non-Markovian---generated by some underlying hidden Markov machine that the agent cannot observe. Using techniques from dynamic programming and computational mechanics, we present a method to identify the provably optimal agent strategy, with time complexity that scales linearly with sequence length. This motivates us to introduce the maximum work such agents can extract---time-ordered free energy(TOFE)---as a fundamental measure of free energy available in a temporally correlated quantum system subject to causal considerations.

quant-ph↗

arXiv2026

Thermodynamic Advantage of Quantum Time-Reversal

Classical computations inherently require energy dissipation that increases significantly as the reliability of the computation improves. This dissipation arises when transitions between memory states are not balanced by their time-reversed counterparts. While classical memories exhibit a discrete set of possible time-reversal symmetries, quantum memory offers a continuum. This continuum enables the design of quantum memories that minimize irreversibility. As a result, quantum memory reduces energy dissipation several orders of magnitude below classical memory.

cond-mat.stat-mech↗

arXiv2026

Transformers learn factored representations

Transformers pretrained via next token prediction learn to factor their world into parts, representing these factors in orthogonal subspaces of the residual stream. We formalize two representational hypotheses: (1) a representation in the product space of all factors, whose dimension grows exponentially with the number of parts, or (2) a factored representation in orthogonal subspaces, whose dimension grows linearly. The factored representation is lossless when factors are conditionally independent, but sacrifices predictive fidelity otherwise, creating a tradeoff between dimensional efficiency and accuracy. We derive precise predictions about the geometric structure of activations for each, including the number of subspaces, their dimensionality, and the arrangement of context embeddings within them. We test between these hypotheses on transformers trained on synthetic processes with known latent structure. Models learn factored representations when factors are conditionally independent, and continue to favor them early in training even when noise or hidden dependencies undermine conditional independence, reflecting an inductive bias toward factoring at the cost of fidelity. This provides a principled explanation for why transformers decompose the world into parts, and suggests that interpretable low dimensional structure may persist even in models trained on complex data.

cs.LG↗

arXiv2025

Constrained belief updates explain geometric structures in transformer representations

What computational structures emerge in transformers trained on next-token prediction? In this work, we provide evidence that transformers implement constrained Bayesian belief updating -- a parallelized version of partial Bayesian inference shaped by architectural constraints. We integrate the model-agnostic theory of optimal prediction with mechanistic interpretability to analyze transformers trained on a tractable family of hidden Markov models that generate rich geometric patterns in neural activations. Our primary analysis focuses on single-layer transformers, revealing how the first attention layer implements these constrained updates, with extensions to multi-layer architectures demonstrating how subsequent layers refine these representations. We find that attention carries out an algorithm with a natural interpretation in the probability simplex, and create representations with distinctive geometric structure. We show how both the algorithmic behavior and the underlying geometry of these representations can be theoretically predicted in detail -- including the attention pattern, OV-vectors, and embedding vectors -- by modifying the equations for optimal future token predictions to account for the architectural constraints of attention. Our approach provides a principled lens on how architectural constraints shape the implementation of optimal prediction, revealing why transformers develop specific intermediate geometric structures.

cs.LG↗

arXiv2025

Identifiability and minimality bounds of quantum and post-quantum models of classical stochastic processes

To make sense of the world around us, we develop models, constructed to enable us to replicate, describe, and explain the behaviours we see. Focusing on the broad case of sequences of correlated random variables, i.e., classical stochastic processes, we tackle the question of determining whether or not two different models produce the same observable behavior. This is the problem of identifiability. Curiously, the physics of the model need not correspond to the physics of the observations; recent work has shown that it is even advantageous -- in terms of memory and thermal efficiency -- to employ quantum models to generate classical stochastic processes. We resolve the identifiability problem in this regime, providing a means to compare any two models of a classical process, be the models classical, quantum, or `post-quantum', by mapping them to a canonical `generalized' hidden Markov model. Further, this enables us to place (sometimes tight) bounds on the minimal dimension required of a quantum model to generate a given classical stochastic process.

quant-ph↗

arXiv2025

Next-token pretraining implies in-context learning

We argue that in-context learning (ICL) predictably arises from standard self-supervised next-token pretraining, rather than being an exotic emergent property. This work establishes the foundational principles of this emergence by focusing on in-distribution ICL, demonstrating how models necessarily adapt to context when trained on token sequences, especially from non-ergodic sources. Our information-theoretic framework precisely predicts these in-distribution ICL dynamics (i.e., context-dependent loss reduction). We verify this with experiments using synthetic datasets of differing types of correlational structure, reproducing characteristic phenomena like phase transitions in training loss for induction head formation and power-law scaling of in-context loss. We further show that a model's in-context performance on any task is mathematically coupled to the ensemble of tasks seen in pretraining, offering a fundamental explanation, grounded in architecture- and modality-independent principles, for such inference-time learning.

cs.LG↗

arXiv2025

Neural networks leverage nominally quantum and post-quantum representations

We show that deep neural networks, including transformers and RNNs, pretrained as usual on next-token prediction, intrinsically discover and represent beliefs over 'quantum' and 'post-quantum' low-dimensional generative models of their training data -- as if performing iterative Bayesian updates over the latent state of this world model during inference as they observe more context. Notably, neural nets easily find these representation whereas there is no finite classical circuit that would do the job. The corresponding geometric relationships among neural activations induced by different input sequences are found to be largely independent of neural-network architecture. Each point in this geometry corresponds to a history-induced probability density over all possible futures, and the relative displacement of these points reflects the difference in mechanism and magnitude for how these distinct pasts affect the future.

cs.LG↗

arXiv2025

Energetic advantages for quantum agents in online execution of complex strategies

Agents often execute complex strategies -- adapting their response to each input stimulus depending on past observations and actions. Here, we derive the minimal energetic cost for classical agents to execute a given strategy, highlighting that they must dissipate a certain amount of heat with each decision beyond Landauer's limit. We then prove that quantum agents can reduce this dissipation below classical limits. We establish the necessary and sufficient conditions for a strategy to guarantee quantum agents have energetic advantage, and illustrate settings where this advantage grows without bound. Our results establish a fundamental energetic advantage for agents utilizing quantum processing to enact complex adaptive behaviour.

quant-ph↗

arXiv2025

Transformers represent belief state geometry in their residual stream

What computational structure are we building into large language models when we train them on next-token prediction? Here, we present evidence that this structure is given by the meta-dynamics of belief updating over hidden states of the data-generating process. Leveraging the theory of optimal prediction, we anticipate and then find that belief states are linearly represented in the residual stream of transformers, even in cases where the predicted belief state geometry has highly nontrivial fractal structure. We investigate cases where the belief state geometry is represented in the final residual stream or distributed across the residual streams of multiple layers, providing a framework to explain these observations. Furthermore we demonstrate that the inferred belief states contain information about the entire future, beyond the local next-token prediction that the transformers are explicitly trained on. Our work provides a general framework connecting the structure of training data to the geometric structure of activations inside transformers.

cs.LG↗

arXiv2024

Geometry and Dynamics of LayerNorm

A technical note aiming to offer deeper intuition for the LayerNorm function common in deep neural networks. LayerNorm is defined relative to a distinguished 'neural' basis, but it does more than just normalize the corresponding vector elements. Rather, it implements a composition -- of linear projection, nonlinear scaling, and then affine transformation -- on input activation vectors. We develop both a new mathematical expression and geometric intuition, to make the net effect more transparent. We emphasize that, when LayerNorm acts on an N-dimensional vector space, all outcomes of LayerNorm lie within the intersection of an (N-1)-dimensional hyperplane and the interior of an N-dimensional hyperellipsoid. This intersection is the interior of an (N-1)-dimensional hyperellipsoid, and typical inputs are mapped near its surface. We find the direction and length of the principal axes of this (N-1)-dimensional hyperellipsoid via the eigen-decomposition of a simply constructed matrix.

cs.LG↗

arXiv2023

Engines for predictive work extraction from memoryful quantum stochastic processes

Quantum information-processing techniques enable work extraction from a system's inherently quantum features, in addition to the classical free energy it contains. Meanwhile, the science of computational mechanics affords tools for the predictive modeling of non-Markovian classical and quantum stochastic processes. We combine tools from these two sciences to develop a technique for predictive work extraction from non-Markovian stochastic processes with quantum outputs. We demonstrate that this technique can extract more work than non-predictive quantum work extraction protocols, on one hand, and predictive work extraction without quantum information processing, on the other. We discover a phase transition in the efficacy of memory for work extraction from quantum processes, which is without classical precedent. Our work opens up the prospect of machines that harness environmental free energy in an essentially quantum, essentially time-varying form.

quant-ph↗

arXiv2023

Ultimate limit on learning non-Markovian behavior: Fisher information rate and excess information

We address the fundamental limits of learning unknown parameters of any stochastic process from time-series data, and discover exact closed-form expressions for how optimal inference scales with observation length. Given a parametrized class of candidate models, the Fisher information of observed sequence probabilities lower-bounds the variance in model estimation from finite data. As sequence-length increases, the minimal variance scales as the square inverse of the length -- with constant coefficient given by the information rate. We discover a simple closed-form expression for this information rate, even in the case of infinite Markov order. We furthermore obtain the exact analytic lower bound on model variance from the observation-induced metadynamic among belief states. We discover ephemeral, exponential, and more general modes of convergence to the asymptotic information rate. Surprisingly, this myopic information rate converges to the asymptotic Fisher information rate with exactly the same relaxation timescales that appear in the myopic entropy rate as it converges to the Shannon entropy rate for the process. We illustrate these results with a sequence of examples that highlight qualitatively distinct features of stochastic processes that shape optimal learning.

cs.LG↗

arXiv2023

Thermodynamically ideal quantum-state inputs to any device

We investigate and ascertain the ideal inputs to any finite-time thermodynamic process. We demonstrate that the expectation values of entropy flow, heat, and work can all be determined via Hermitian observables of the initial state. These Hermitian operators encapsulate the breadth of behavior and the ideal inputs for common thermodynamic objectives. We show how to construct these Hermitian operators from measurements of thermodynamic output from a finite number of effectively arbitrary inputs. Behavior of a small number of test inputs thus determines the full range of thermodynamic behavior from all inputs. For any process, entropy flow, heat, and work can all be extremized by pure input states -- eigenstates of the respective operators. In contrast, the input states that minimize entropy production or maximize the change in free energy are non-pure mixed states obtained from the operators as the solution of a convex optimization problem. To attain these, we provide an easily implementable gradient descent method on the manifold of density matrices, where an analytic solution yields a valid direction of descent at each iterative step. Ideal inputs within a limited domain, and their associated thermodynamic operators, are obtained with less effort. This allows analysis of ideal thermodynamic inputs within quantum subspaces of infinite-dimensional quantum systems; it also allows analysis of ideal inputs in the classical limit. Our examples illustrate the diversity of 'ideal' inputs: Distinct initial states minimize entropy production, extremize the change in free energy, and maximize work extraction.

quant-ph↗

arXiv2023

Complexity-calibrated Benchmarks for Machine Learning Reveal When Next-Generation Reservoir Computer Predictions Succeed and Mislead

Recurrent neural networks are used to forecast time series in finance, climate, language, and from many other domains. Reservoir computers are a particularly easily trainable form of recurrent neural network. Recently, a "next-generation" reservoir computer was introduced in which the memory trace involves only a finite number of previous symbols. We explore the inherent limitations of finite-past memory traces in this intriguing proposal. A lower bound from Fano's inequality shows that, on highly non-Markovian processes generated by large probabilistic state machines, next-generation reservoir computers with reasonably long memory traces have an error probability that is at least ~ 60% higher than the minimal attainable error probability in predicting the next observation. More generally, it appears that popular recurrent neural networks fall far short of optimally predicting such complex processes. These results highlight the need for a new generation of optimized recurrent neural network architectures. Alongside this finding, we present concentration-of-measure results for randomly-generated but complex processes. One conclusion is that large probabilistic state machines -- specifically, large $ε$-machines -- are key to generating challenging and structurally-unbiased stimuli for ground-truthing recurrent neural network architectures.

cs.LG↗

arXiv2021

Time Symmetries of Memory Determine Thermodynamic Efficiency

While Landauer's Principle sets a lower bound for the work required for a computation, that work is recoverable for efficient computations. However, practical physical computers, such as modern digital computers or biochemical systems, are subject to constraints that make them inefficient -- irreversibly dissipating significant energy. Recent results show that the dissipation in such systems is bounded by the nonreciprocity of the embedded computation. We investigate the consequences of this bound for different types of memory, showing that different memory devices are better suited for different computations. This correspondence comes from the time-reversal symmetries of the memory, which depend on whether information is stored positionally or magnetically. This establishes that the time symmetries of the memory device play an essential roll in determining energetics. The energetic consequences of time symmetries are particularly pronounced in nearly deterministic computations, where the cost of computing diverges as minus log of the error rate. We identify the coefficient of that divergence as the dissipation divergence. We find that the dissipation divergence may be zero for a computation when implemented in one type of memory while it's maximal when implemented with another. Given flexibility in the type of memory, the dissipation divergence is bounded below by the average state compression of the computation. Moreover, we show how to explicitly construct the memory to achieve this minimal dissipation. As a result, we find that logically reversible computations are indeed thermodynamically efficient, but logical irreversibility comes at a much higher cost than previously anticipated.

cond-mat.stat-mech↗

arXiv2021

The impossibility of Landauer's bound for almost every quantum state

The thermodynamic cost of resetting an arbitrary initial state to a particular desired state is lower bounded by Landauer's bound. However, here we demonstrate that this lower bound is necessarily unachievable for nearly every initial state, for any reliable reset mechanism. Since local heating threatens rapid decoherence, this issue is of substantial importance beyond mere energy efficiency. For the case of qubit reset, we find the minimally dissipative state analytically for any reliable reset protocol, in terms of the entropy-flow vector introduced here. This allows us to verify a recent theorem about initial-state dependence of entropy production for any finite-time transformation, as it pertains to quantum state preparation.

quant-ph↗

arXiv2021

Initial-State Dependence of Thermodynamic Dissipation for any Quantum Process

New exact results about the nonequilibrium thermodynamics of open quantum systems at arbitrary timescales are obtained by considering all possible variations of initial conditions of a system, its environment, and correlations between them. First we obtain a new quantum-information theoretic equality for entropy production, valid for an arbitrary initial joint state of system and environment. For any finite-time process with a fixed initial environment, we then show that the contraction of the system's distinction -- relative to the minimally dissipative state -- exactly quantifies its thermodynamic dissipation. The quantum component of this dissipation is the change in coherence relative to the minimally dissipative state. Implications for quantum state preparation and local control are explored. For nonunitary processes -- like the preparation of any particular quantum state -- we find that mismatched expectations lead to divergent dissipation as the actual initial state becomes orthogonal to the anticipated one.

cond-mat.stat-mech↗