arXiv subjects

Einar Urdshals

Publications and source records attributed to Einar Urdshals.

12 recordsLinked to original sources

arXiv2026

Interpreting Reinforcement Learning Agents with Susceptibilities

Susceptibilities are a technique for neural network interpretability that studies the response of posterior expectation values of observables to perturbations of the loss. We generalize this construction to the setting of the regret in deep reinforcement learning and investigate the utility of susceptibilities in a simple gridworld model that nevertheless exhibits non-trivial stagewise development. We argue that susceptibilities reveal internal features of the development of the model in parameter space that one cannot detect purely by studying the development of the learned policy. We validate these results with activation-steering, and discuss the framework's extension to RLHF post-training.

cs.LG↗

arXiv2026

Stagewise Reinforcement Learning and the Geometry of the Regret Landscape

Singular learning theory characterizes Bayesian learning as an evolving tradeoff between accuracy and complexity, with transitions between qualitatively different solutions as sample size increases. We extend this theory to reinforcement learning, proving that the concentration of a generalized posterior over policies is governed by the local learning coefficient (LLC), an invariant of the geometry of the regret function. This theory predicts that deep reinforcement learning with SGD should proceed from simple policies with high regret to complex policies with low regret. We verify this prediction empirically in a gridworld environment exhibiting stagewise policy development: phase transitions over training manifest as "opposing staircases" where regret decreases sharply while the LLC increases.

cs.LG↗

arXiv2025

Beyond Tokens in Language Models: Interpreting Activations through Text Genre Chunks

Understanding Large Language Models (LLMs) is key to ensure their safe and beneficial deployment. This task is complicated by the difficulty of interpretability of LLM structures, and the inability to have all their outputs human-evaluated. In this paper, we present the first step towards a predictive framework, where the genre of a text used to prompt an LLM, is predicted based on its activations. Using Mistral-7B and two datasets, we show that genre can be extracted with F1-scores of up to 98% and 71% using scikit-learn classifiers. Across both datasets, results consistently outperform the control task, providing a proof of concept that text genres can be inferred from LLMs with shallow learning models.

cs.CL↗

arXiv2025

Compressibility Measures Complexity: Minimum Description Length Meets Singular Learning Theory

We study neural network compressibility by using singular learning theory to extend the minimum description length (MDL) principle to singular models like neural networks. Through extensive experiments on the Pythia suite with quantization, factorization, and other compression techniques, we find that complexity estimates based on the local learning coefficient (LLC) are closely, and in some cases, linearly correlated with compressibility. Our results provide a path toward rigorously evaluating the limits of model compression.

stat.ML↗

arXiv2025

Electronic structure of liquid xenon in the context of light dark matter direct detection

We present a description of the electronic structure of xenon within the density-functional theory formalism with the goal of accurately modeling dark-matter-induced ionisation in liquid xenon detectors. We compare the calculated electronic structures of the atomic, liquid and crystalline solid phases, and find that the electronic charge density and its derivatives in momentum space are similar in the atom and the liquid, consistent with the weak interatomic van der Waals bonding. The only notable difference is a band broadening of the highest occupied $5p$ levels, reflected in the densities of states of the condensed phases, as a result of the inter-atomic interactions. We therefore use the calculated density of states of the liquid phase, combined with the standard literature approach for the isolated atom, to recompute ionisation rates and exclusion limit curves for the XENON10 and XENON1T experiments. We find that the broadening of the 5$p$ levels induced by the liquid phase is relevant only for dark matter masses below 6 MeV, where it increases the ionisation rate relative to that of the isolated atom. For most of the explored mass range, the energies of the discrete 4$d$ and 5$s$ levels have the strongest effect on the rate. Our findings suggest a simple scheme for calculating dark matter-electron scattering rates in liquid noble gas detectors, using the calculated values for the atom weighted by the density of states of the condensed phase.

hep-ph↗

arXiv2025

Structure Development in List-Sorting Transformers

We study how a one-layer attention-only transformer develops relevant structures while learning to sort lists of numbers. At the end of training, the model organizes its attention heads in two main modes that we refer to as vocabulary-splitting and copy-suppression. Both represent simpler modes than having multiple heads handle overlapping ranges of numbers. Interestingly, vocabulary-splitting is present regardless of whether we use weight decay, a common regularization technique thought to drive simplification, supporting the thesis that neural networks naturally prefer simpler solutions. We relate copy-suppression to a mechanism in GPT-2 and investigate its functional role in our model. Guided by insights from a developmental analysis of the model, we identify features in the training data that drive the model's final acquired solution. This provides a concrete example of how the training data shape the internal organization of transformers, paving the way for future studies that could help us better understand how LLMs develop their internal structures.

cs.LG↗

arXiv2024

Dark Matter-induced electron excitations in silicon and germanium with Deep Learning

We train a deep neural network (DNN) to output rates of dark matter (DM) induced electron excitations in silicon and germanium detectors. Our DNN provides a massive speedup of around $5$ orders of magnitude relative to existing methods (i.e. QEdark-EFT), allowing for extensive parameter scans in the event of an observed DM signal. The network is also lighter and simpler to use than alternative computational frameworks based on a direct calculation of the DM-induced excitation rate. The DNN can be downloaded $\href{https://github.com/urdshals/DEDD}{\text{here}}$.

hep-ph↗

arXiv2023

Direct searches for general dark matter-electron interactions with graphene detectors: Part I. Electronic structure calculations

We develop a formalism to describe electron ejections from graphene-like targets by dark matter (DM) scattering for general forms of scalar and spin 1/2 DM-electron interactions and compare their applicability and accuracy within the density functional theory (DFT) and tight binding (TB) approaches. This formalism allows for accurate prediction of the daily modulation signal expected from DM in upcoming direct detection experiments employing graphene sheets as the target material. A key result is that the physics of the graphene sheet and that of the DM and the ejected electron factorise, allowing for the rate of ejections from all forms of DM to be obtained with a single graphene response function. We perform a comparison between the TB and DFT approaches to modeling the initial state electronic wavefunction within this framework, with DFT emerging as the more self-consistent and reliable choice due to the challenges in the embedding of an appropriate atomic contribution into the TB approach.

hep-ph↗

arXiv2023

Direct searches for general dark matter-electron interactions with graphene detectors: Part II. Sensitivity studies

We use a formalism that describes electron ejections from graphene-like targets by dark matter (DM) scattering for general forms of scalar and spin 1/2 DM-electron interactions in combination with state-of-the-art density functional calculations to produce predictions and reach estimates for various possible carbon-based detector designs. Our results indicate the importance of a proper description of the target electronic structure. In addition, we find a strong dependence of the predicted observed signal for different DM candidate masses and interaction types on the detailed geometry and design of the detector. Combined with directional background vetoing, these dependencies will enable the identification of DM particle properties once a signal has been established.

hep-ph↗

arXiv2022

Dark matter - electron interactions in materials beyond the dark photon model

The search for sub-GeV dark matter (DM) particles via electronic transitions in underground detectors attracted much theoretical and experimental interest in the past few years. A still open question in this field is whether experimental results can in general be interpreted in a framework where the response of detector materials to an external DM probe is described by a single ionisation or crystal form factor, as expected for the so-called dark photon model. Here, ionisation and crystal form factors are examples of material response functions: interaction-specific integrals of the initial and final state electron wave functions. In this work, we address this question through a systematic classification of the material response functions induced by a wide range of models for spin-0, spin-1/2 and spin-1 DM. We find several examples for which an accurate description of the electronic transition rate at DM direct detection experiments requires material response functions that go beyond those expected for the dark photon model. This concretely illustrates the limitations of a framework that is entirely based on the standard ionisation and crystal form factors, and points towards the need for the general response-function-based formalism we pushed forward recently [1,2]. For the models that require non-standard atomic and crystal response functions, we use the response functions of [1,2] to calculate the DM-induced electronic transition rate in atomic and crystal detectors, and to present 90% confidence level exclusion limits on the strength of the DM-electron interaction from the null results reported by XENON10, XENON1T, EDELWEISS and SENSEI.

hep-ph↗

arXiv2022

Crystal responses to general dark matter-electron interactions

We develop a formalism to describe the scattering of dark matter (DM) particles by electrons bound in crystals for a general form of the underlying DM-electron interaction. Such a description is relevant for direct-detection experiments of DM particles lighter than a nucleon, which might be observed in operating DM experiments via electron excitations in semiconductor crystal detectors. Our formalism is based on an effective theory approach to general non-relativistic DM-electron interactions, including the anapole, and magnetic and electric dipole couplings, combined with crystal response functions defined in terms of electron wave function overlap integrals. Our main finding is that, for the usual simplification of the velocity integral, the rate of DM-induced electronic transitions in a semiconductor material depends on at most five independent crystal response functions, four of which were not known previously. We identify these crystal responses, and evaluate them using density functional theory for crystalline silicon and germanium, which are used in operating DM direct detection experiments. Our calculations allow us to set 90% confidence level limits on the strength of DM-electron interactions from data reported by the SENSEI and EDELWEISS experiments. The novel crystal response functions discovered in this work encode properties of crystalline solids that do not interact with conventional experimental probes, suggesting the use of the DM wind as a probe to reveal new kinds of hidden order in materials.

hep-ph↗

arXiv2020

On dark atoms, massive dark photons and millicharged sub-components

We present a simple model of two dark matter species with opposite millicharge that can form electrically neutral bound states via the exchange of a massive dark photon. If bound state formation is suppressed at low temperatures, a sub-dominant fraction of millicharged particles remains at late times, which can give rise to interesting features in the 21 cm absorption profile at cosmic dawn. The dominant neutral component, on the other hand, can have dipole interactions with ordinary matter, leading to non-standard signals in direct detection experiments. We identify the parameter regions predicting a percent-level ionisation fraction and study constraints from laboratory searches for dark matter scattering and dark photon decays.

hep-ph↗