Search

Mechanistic interpretability is an emerging diagnostic approach for neural models that has gained traction in broader natural language processing domains. This paradigm aims to provide attribution to components of neural systems where causal relationships between hidden layers and output were previously uninterpretable. As the use of neural models in IR for retrieval and evaluation becomes ubiquitous, we need to ensure that we can interpret why a model produces a given output for both transparency and the betterment of systems. This work comprises a flexible framework for diagnostic analysis and intervention within these highly parametric neural systems specifically tailored for IR tasks and architectures. In providing such a framework, we look to facilitate further research in interpretable IR with a broader scope for practical interventions derived from mechanistic interpretability. We provide preliminary analysis and look to demonstrate our framework through an axiomatic lens to show its applications and ease of use for those IR practitioners inexperienced in this emerging paradigm., Comment: 5 pages, 2 figures, Accepted to ECIR 2025 as a Demo Paper

For a long time, the study of thermal effects at three-dimensional (3D) short-ranged wetting transitions considered only the effect of interfacial fluctuations. We show that an entropic Casimir contribution, missed in previous treatments, produces significant effects when it is included; in particular, mean-field predictions are no longer obtained when interfacial fluctuations are ignored. The Casimir term arises from the many different microscopic configurations that correspond to a given interfacial one. By employing a coarse-graining procedure, starting from a microscopic Landau-Ginzburg-Wilson Hamiltonian, we identify the interfacial model for 3D wetting and the exact form of the Casimir term. The Casimir contribution does not alter the Nakanishi-Fisher surface phase diagram; it significantly increases the adsorption near a first-order wetting transition and completely changes the predicted critical singularities of tricritical wetting, including the nonuniversality occurring in 3D arising from interfacial fluctuations. We illustrate how the Casimir term leads to a reappraisal of the critical singularities at wetting transitions., Comment: 6 pages, 1 figure; Proceedings of the "Fifth International Symposium on the Casimir Effect", Piran (Slovenia). Submitted to Int. J. Mod. Phys. A

Recent studies of wetting in a two-component square-gradient model of interfaces in a fluid mixture, showing three-phase bulk coexistence, have revealed some highly surprising features. Numerical results show that the density profile paths, which form a tricuspid shape in the density plane, have curious geometric properties, while conjectures for the analytical form of the surface tensions imply that nonwetting may persist up to the critical end points, contrary to the usual expectation of critical point wetting. Here, we solve the model exactly and show that the profile paths are conformally invariant quartic algebraic curves that change genus at the wetting transition. Being harmonic, the profile paths can be represented by an analytic function in the complex plane which then conformally maps the paths onto straight lines. Using this, we derive the conjectured form of the surface tensions and explain the geometrical properties of the tricuspid and its relation to the Neumann triangle for the contact angles. The exact solution confirms that critical point wetting is absent in this square-gradient model., Comment: 6 pages, 3 figures, Physical Review Letters (accepted)

Neural approaches to ranking based on pre-trained language models are highly effective in ad-hoc search. However, the computational expense of these models can limit their application. As such, a process known as knowledge distillation is frequently applied to allow a smaller, efficient model to learn from an effective but expensive model. A key example of this is the distillation of expensive API-based commercial Large Language Models into smaller production-ready models. However, due to the opacity of training data and processes of most commercial models, one cannot ensure that a chosen test collection has not been observed previously, creating the potential for inadvertent data contamination. We, therefore, investigate the effect of a contaminated teacher model in a distillation setting. We evaluate several distillation techniques to assess the degree to which contamination occurs during distillation. By simulating a ``worst-case'' setting where the degree of contamination is known, we find that contamination occurs even when the test data represents a small fraction of the teacher's training samples. We, therefore, encourage caution when training using black-box teacher models where data provenance is ambiguous., Comment: 4 pages

In Minkowski spacetime quantum field theory, each stationary motion is associated with an effective, energy-dependent notion of temperature, which generalises the familiar Unruh temperature of uniform linear acceleration. Motivated by current experimental interest in circular motion, we analyse the effective temperature for drifted Rindler motion, generated by a boost and a spacelike translation (drift), and the way in which drifted Rindler motion can be smoothly (and in fact real analytically) deformed to circular motion through a third type of motion known as parator. For an Unruh-DeWitt detector coupled linearly to a massless scalar field in 2+1 and 3+1 spacetime dimensions, we establish analytic results in the limits of large gap, small gap and large drift speed. For fixed proper acceleration, the drifted Rindler temperature remains bounded in the large gap limit, in contrast to the circular motion temperature, which can be arbitrarily large in this limit. Finally, in 2+1 dimensions, we trace the vanishing of the circular motion temperature in the small gap limit to the weak decay of the Wightman function, and we show that, among all types of stationary motion in all dimensions, this phenomenon is unique to 2+1 dimensions and therein to circular and parator motion., Comment: 32 pages, 1 table. v2: Presentational clarifications, including new title. Real analyticity of the deformation mentioned

A supervised ranking model, despite its advantage of being effective, usually involves complex processing - typically multiple stages of task-specific pre-training and fine-tuning. This has motivated researchers to explore simpler pipelines leveraging large language models (LLMs) that are capable of working in a zero-shot manner. However, since zero-shot inference does not make use of a training set of pairs of queries and their relevant documents, its performance is mostly worse than that of supervised models, which are trained on such example pairs. Motivated by the existing findings that training examples generally improve zero-shot performance, in our work, we explore if this also applies to ranking models. More specifically, given a query and a pair of documents, the preference prediction task is improved by augmenting examples of preferences for similar queries from a training set. Our proposed pairwise few-shot ranker demonstrates consistent improvements over the zero-shot baseline on both in-domain (TREC DL) and out-domain (BEIR subset) retrieval benchmarks. Our method also achieves a close performance to that of a supervised model without requiring any complex training pipeline., Comment: Accepted to EMNLP 2024

Tunable biphoton quantum entanglement generated from nonlinear processes is highly desirable for cutting-edge quantum technologies, yet its tunability is substantially constrained by the symmetry of material nonlinear tensors. Here, we overcome this constraint by introducing symmetry-breaking in nonlinear polarization to generate optically tunable biphoton entanglement at picosecond speeds. Asymmetric optical responses have made breakthroughs in classical applications like non-reciprocal light transmission. We now experimentally demonstrate the nonlinear asymmetry response for biphoton entanglement using a semiconductor metasurface incorporating [110] InGaP nano-resonators with structural asymmetry. We realize continuous tuning of polarization entanglement from near-unentangled states to a Bell state. This tunability can also extend to produce tailored hyperentanglement. Furthermore, our nanoscale entanglement source features an ultra-high coincidence-to-accidental ratio of $\approx7\times10^4$, outperforming existing semiconductor flat optics by two orders of magnitude. Introducing asymmetric nonlinear response in quantum metasurfaces opens new directions for tailoring on-demand quantum states and beyond., Comment: 27 pages, 4 figures

Quantum computers can perform certain operations exponentially faster than classical computers, but designing quantum circuits is challenging. To that end, researchers used evolutionary algorithms to produce probabilistic quantum circuits that give the correct output more often than not for any input. They can be executed multiple times, with the outputs combined using a classical method (such as voting) to produce the final output, effectively creating a homogeneous ensemble of circuits (i.e., all identical). Inspired by n-version programming and ensemble learning, we developed a tool that uses an evolutionary algorithm to generate heterogeneous ensembles of circuits (i.e., all different), named QuEEn. We used it to evolve ensembles to solve the Iris classification problem. When using ideal simulation, we found the performance of heterogeneous ensembles to be greater than that of homogeneous ensembles to a statistically significant degree. When using noisy simulation, we still observed a statistically significant improvement in the majority of cases. Our results indicate that evolving heterogeneous ensembles is an effective strategy for improving the reliability of quantum circuits. This is particularly relevant in the current NISQ era of quantum computing where computers do not yet have good tolerance to quantum noise.

Nguyen has shown that on averaging over $a=1,...,q$ the 3-fold divisor function has exponent of distribution 2/3, following \cite {banks}. We follow [2] which leads to stronger bounds., Comment: To be published in the Ramanujan journal

A deep conjecture of Montgomery and Soundararajan on the distribution of prime numbers in short intervals of length $h$ says that the third moment is bounded by $\ll h^{\frac {3}{2}-c}$ for some $c>0$. There is in the literature some conditional evidence towards this conjecture whilst in the first article to this series we gave the first instance of unconditional evidence in the form of a bound corresponding to $\ll h^{7/5+o(1)}$. In this article we push the exponent down to $\ll h^{1+o(1)}$ which more or less is expected to be best possible.

Ramsey et al. report on the characteristics and temporal distribution of an interesting vibrational signal that they term the whooping signal, primarily based upon a long-term study of vibrations recorded by accelerometers placed inside two honey bee (Apis mellifera) colonies, one in France and one in the United Kingdom. The goal of the study, the long-term automated recording and analysis of honey bee vibrational signaling, is worthwhile. However, we believe that some of the conclusions drawn by the authors are not well supported, given the evidence.

The preservation of the corticospinal tract (CST) is key to good motor recovery after stroke. The gold standard method of assessing the CST with imaging is diffusion tensor tractography. However, this is not available for most intracerebral haemorrhage (ICH) patients. Non-contrast CT scans are routinely available in most ICH diagnostic pipelines, but delineating white matter from a CT scan is challenging. We utilise nnU-Net, trained on paired diagnostic CT scans and high-directional diffusion tractography maps, to segment the CST from diagnostic CT scans alone, and we show our model reproduces diffusion based tractography maps of the CST with a Dice similarity coefficient of 57%. Surgical haematoma evacuation is sometimes performed after ICH, but published clinical trials to date show that whilst surgery reduces mortality, there is no evidence of improved functional recovery. Restricting surgery to patients with an intact CST may reveal a subset of patients for whom haematoma evacuation improves functional outcome. We investigated the clinical utility of our model in the MISTIE III clinical trial dataset. We found that our model's CST integrity measure significantly predicted outcome after ICH in the acute and chronic time frames, therefore providing a prognostic marker for patients to whom advanced diffusion tensor imaging is unavailable. This will allow for future probing of subgroups who may benefit from surgery., Comment: Accepted to Miccai Switch Workshop

Detecting flying animals (e.g., birds, bats, and insects) using weather radar helps gain insights into animal movement and migration patterns, aids in management efforts (such as biosecurity) and enhances our understanding of the ecosystem.The conventional approach to detecting animals in weather radar involves thresholding: defining and applying thresholds for the radar variables, based on expert opinion. More recently, Deep Learning approaches have been shown to provide improved performance in detection. However, obtaining sufficient labelled weather radar data for flying animals to build learning-based models is time-consuming and labor-intensive. To address the challenge of data labelling, we propose a self-supervised learning method for detecting animal movement. In our proposed method, we pre-train our model on a large dataset with noisy labels produced by a threshold approach. The key advantage is that the pre-trained dataset size is limited only by the number of radar images available. We then fine-tune the model on a small human-labelled dataset. Our experiments on Australian weather radar data for waterbird segmentation show that the proposed method outperforms the current state-of-the art approach by 43.53% in the dice co-efficient statistic.

Energy performance certificate (EPC) and thermal infrared (TIR) images both play a key role in the energy performance mapping of the urban building stock. In this paper, we developed parametric building archetypes using an EPC database and conducted temperature clustering on TIR images acquired through drones and satellite datasets. We evaluated 1725 EPCs of existing building stock in Cambridge, UK to generate energy consumption profiles. Drone-based TIR images of individual buildings in two Cambridge University colleges were processed using a machine learning pipeline for thermal anomaly detection and investigated the influence of two specific factors that affect the reliability of TIR for energy management applications: ground sample distance (GSD) and angle of view (AOV). The EPC-level results suggest that the construction year of the buildings influences their energy consumption. For example, modern buildings were over 30% more energy-efficient than older ones. Parallelly, older buildings were found to show almost double the energy savings potential through retrofitting than newly constructed buildings. TIR imaging results showed that thermal anomalies can only be properly identified in images with a GSD of 1m/pixel or less. A GSD of 1-6m/pixel can detect hot areas of building surfaces. We found that a GSD > 6m/pixel cannot characterise individual buildings but does help identify urban heat island effects. Additional sensitivity analysis showed that building thermal anomaly detection is more sensitive to AOV than to GSD. Our study informs newer approaches to building energy diagnostics using thermography and supports decision-making for retrofitting at a large scale.

Large Language Models (LLMs) have significantly impacted many facets of natural language processing and information retrieval. Unlike previous encoder-based approaches, the enlarged context window of these generative models allows for ranking multiple documents at once, commonly called list-wise ranking. However, there are still limits to the number of documents that can be ranked in a single inference of the model, leading to the broad adoption of a sliding window approach to identify the k most relevant items in a ranked list. We argue that the sliding window approach is not well-suited for list-wise re-ranking because it (1) cannot be parallelized in its current form, (2) leads to redundant computational steps repeatedly re-scoring the best set of documents as it works its way up the initial ranking, and (3) prioritizes the lowest-ranked documents for scoring rather than the highest-ranked documents by taking a bottom-up approach. Motivated by these shortcomings and an initial study that shows list-wise rankers are biased towards relevant documents at the start of their context window, we propose a novel algorithm that partitions a ranking to depth k and processes documents top-down. Unlike sliding window approaches, our algorithm is inherently parallelizable due to the use of a pivot element, which can be compared to documents down to an arbitrary depth concurrently. In doing so, we reduce the number of expected inference calls by around 33% when ranking at depth 100 while matching the performance of prior approaches across multiple strong re-rankers., Comment: 16 pages, 3 figures, 2 tables

This paper describes our participation in the TREC 2023 Deep Learning Track. We submitted runs that apply generative relevance feedback from a large language model in both a zero-shot and pseudo-relevance feedback setting over two sparse retrieval approaches, namely BM25 and SPLADE. We couple this first stage with adaptive re-ranking over a BM25 corpus graph scored using a monoELECTRA cross-encoder. We investigate the efficacy of these generative approaches for different query types in first-stage retrieval. In re-ranking, we investigate operating points of adaptive re-ranking with different first stages to find the point in graph traversal where the first stage no longer has an effect on the performance of the overall retrieval pipeline. We find some performance gains from the application of generative query reformulation. However, our strongest run in terms of P@10 and nDCG@10 applied both adaptive re-ranking and generative pseudo-relevance feedback, namely uogtr_b_grf_e_gb., Comment: 5 pages, 5 figures, TREC Deep Learning 2023 Notebook

With the increasing ability of large language models (LLMs), in-context learning (ICL) has evolved as a new paradigm for natural language processing (NLP), where instead of fine-tuning the parameters of an LLM specific to a downstream task with labeled examples, a small number of such examples is appended to a prompt instruction for controlling the decoder's generation process. ICL, thus, is conceptually similar to a non-parametric approach, such as $k$-NN, where the prediction for each instance essentially depends on the local topology, i.e., on a localised set of similar instances and their labels (called few-shot examples). This suggests that a test instance in ICL is analogous to a query in IR, and similar examples in ICL retrieved from a training set relate to a set of documents retrieved from a collection in IR. While standard unsupervised ranking models can be used to retrieve these few-shot examples from a training set, the effectiveness of the examples can potentially be improved by re-defining the notion of relevance specific to its utility for the downstream task, i.e., considering an example to be relevant if including it in the prompt instruction leads to a correct prediction. With this task-specific notion of relevance, it is possible to train a supervised ranking model (e.g., a bi-encoder or cross-encoder), which potentially learns to optimally select the few-shot examples. We believe that the recent advances in neural rankers can potentially find a use case for this task of optimally choosing examples for more effective downstream ICL predictions., Comment: 9 Pages, 3 Figures, Accepted as Perspective paper to SIGIR 2024

In recent years, neural ranking models (NRMs) have been shown to substantially outperform their lexical counterparts in text retrieval. In traditional search pipelines, a combination of features leads to well-defined behaviour. However, as neural approaches become increasingly prevalent as the final scoring component of engines or as standalone systems, their robustness to malicious text and, more generally, semantic perturbation needs to be better understood. We posit that the transformer attention mechanism can induce exploitable defects through positional bias in search models, leading to an attack that could generalise beyond a single query or topic. We demonstrate such defects by showing that non-relevant text--such as promotional content--can be easily injected into a document without adversely affecting its position in search results. Unlike previous gradient-based attacks, we demonstrate these biases in a query-agnostic fashion. In doing so, without the knowledge of topicality, we can still reduce the negative effects of non-relevant content injection by controlling injection position. Our experiments are conducted with simulated on-topic promotional text automatically generated by prompting LLMs with topical context from target documents. We find that contextualisation of a non-relevant text further reduces negative effects whilst likely circumventing existing content filtering mechanisms. In contrast, lexical models are found to be more resilient to such content injection attacks. We then investigate a simple yet effective compensation for the weaknesses of the NRMs in search, validating our hypotheses regarding transformer bias., Comment: 8 pages, 4 main figures, 7 appendix pages, 2 appendix figures, Accepted to ACL 2024 Findings

We use the Petrow-Young [10] subconvexity bound for Dirichlet $L$-functions to show that $d_4(n)$ has exponent of distribution $4/7$ when we allow an average over $a$ mod $q$, thereby giving an equidistribution result for $d_4(n)$ which goes past the $1/2$ barrier for the first time.

We give a relatively simple proof that \[ \int _0^1\left |\sum _{n\leq x}d(n)e(n\alpha )\right |d\alpha \asymp \sqrt x.\]

Modern sequence-to-sequence relevance models like monoT5 can effectively capture complex textual interactions between queries and documents through cross-encoding. However, the use of natural language tokens in prompts, such as Query, Document, and Relevant for monoT5, opens an attack vector for malicious documents to manipulate their relevance score through prompt injection, e.g., by adding target words such as true. Since such possibilities have not yet been considered in retrieval evaluation, we analyze the impact of query-independent prompt injection via manually constructed templates and LLM-based rewriting of documents on several existing relevance models. Our experiments on the TREC Deep Learning track show that adversarial documents can easily manipulate different sequence-to-sequence relevance models, while BM25 (as a typical lexical model) is not affected. Remarkably, the attacks also affect encoder-only relevance models (which do not rely on natural language prompt tokens), albeit to a lesser extent., Comment: 13 pages, 3 figures, Accepted at ECIR 2024 as a Full Paper