Search

The capabilities and limitations of Large Language Models have been sketched out in great detail in recent years, providing an intriguing yet conflicting picture. On the one hand, LLMs demonstrate a general ability to solve problems. On the other hand, they show surprising reasoning gaps when compared to humans, casting doubt on the robustness of their generalisation strategies. The sheer volume of data used in the design of LLMs has precluded us from applying the method traditionally used to measure generalisation: train-test set separation. To overcome this, we study what kind of generalisation strategies LLMs employ when performing reasoning tasks by investigating the pretraining data they rely on. For two models of different sizes (7B and 35B) and 2.5B of their pretraining tokens, we identify what documents influence the model outputs for three simple mathematical reasoning tasks and contrast this to the data that are influential for answering factual questions. We find that, while the models rely on mostly distinct sets of data for each factual question, a document often has a similar influence across different reasoning questions within the same task, indicating the presence of procedural knowledge. We further find that the answers to factual questions often show up in the most influential data. However, for reasoning questions the answers usually do not show up as highly influential, nor do the answers to the intermediate reasoning steps. When we characterise the top ranked documents for the reasoning questions qualitatively, we confirm that the influential documents often contain procedural knowledge, like demonstrating how to obtain a solution using formulae or code. Our findings indicate that the approach to reasoning the models use is unlike retrieval, and more like a generalisable strategy that synthesises procedural knowledge from documents doing a similar form of reasoning.

Representation engineering methods have recently shown promise for enabling efficient steering of model behavior. However, evaluation pipelines for these methods have primarily relied on subjective demonstrations, instead of quantitative, objective metrics. We aim to take a step towards addressing this issue by advocating for four properties missing from current evaluations: (i) contexts sufficiently similar to downstream tasks should be used for assessing intervention quality; (ii) model likelihoods should be accounted for; (iii) evaluations should allow for standardized comparisons across different target behaviors; and (iv) baseline comparisons should be offered. We introduce an evaluation pipeline grounded in these criteria, offering both a quantitative and visual analysis of how effectively a given method works. We use this pipeline to evaluate two representation engineering methods on how effectively they can steer behaviors such as truthfulness and corrigibility, finding that some interventions are less effective than previously reported., Comment: Accepted to the NeurIPS 2024 - Workshop on Foundation Model Interventions

We consider the problem of finding spatial configurations of multiple objects in images, e.g., a mobile inspection robot is tasked to localize abandoned tools on the floor. We define the spatial configuration of objects by first-order logic in terms of relations and attributes. A neuro-symbolic program matches the logic formulas to probabilistic object proposals for the given image, provided by language-vision models by querying them for the symbols. This work is the first to combine neuro-symbolic programming (reasoning) and language-vision models (learning) to find spatial configurations of objects in images in an open world setting. We show the effectiveness by finding abandoned tools on floors and leaking pipes. We find that most prediction errors are due to biases in the language-vision model., Comment: 12 pages

Collecting and annotating real-world data for the development of object detection models is a time-consuming and expensive process. In the military domain in particular, data collection can also be dangerous or infeasible. Training models on synthetic data may provide a solution for cases where access to real-world training data is restricted. However, bridging the reality gap between synthetic and real data remains a challenge. Existing methods usually build on top of baseline Convolutional Neural Network (CNN) models that have been shown to perform well when trained on real data, but have limited ability to perform well when trained on synthetic data. For example, some architectures allow for fine-tuning with the expectation of large quantities of training data and are prone to overfitting on synthetic data. Related work usually ignores various best practices from object detection on real data, e.g. by training on synthetic data from a single environment with relatively little variation. In this paper we propose a methodology for improving the performance of a pre-trained object detector when training on synthetic data. Our approach focuses on extracting the salient information from synthetic data without forgetting useful features learned from pre-training on real images. Based on the state of the art, we incorporate data augmentation methods and a Transformer backbone. Besides reaching relatively strong performance without any specialized synthetic data transfer methods, we show that our methods improve the state of the art on synthetic data trained object detection for the RarePlanes and DGTA-VisDrone datasets, and reach near-perfect performance on an in-house vehicle detection dataset., Comment: Submitted to and presented at SPIE Defense + Commercial Sensing 2024, 13 pages, 4 figures, 3 tables

Common methods for aligning large language models (LLMs) with desired behaviour heavily rely on human-labelled data. However, as models grow increasingly sophisticated, they will surpass human expertise, and the role of human evaluation will evolve into non-experts overseeing experts. In anticipation of this, we ask: can weaker models assess the correctness of stronger models? We investigate this question in an analogous setting, where stronger models (experts) possess the necessary information to answer questions and weaker models (non-experts) lack this information. The method we evaluate is debate, where two LLM experts each argue for a different answer, and a non-expert selects the answer. We find that debate consistently helps both non-expert models and humans answer questions, achieving 76% and 88% accuracy respectively (naive baselines obtain 48% and 60%). Furthermore, optimising expert debaters for persuasiveness in an unsupervised manner improves non-expert ability to identify the truth in debates. Our results provide encouraging empirical evidence for the viability of aligning models with debate in the absence of ground truth., Comment: For code please check: https://github.com/ucl-dark/llm_debate

Despite widespread use of LLMs as conversational agents, evaluations of performance fail to capture a crucial aspect of communication: interpreting language in context -- incorporating its pragmatics. Humans interpret language using beliefs and prior knowledge about the world. For example, we intuitively understand the response "I wore gloves" to the question "Did you leave fingerprints?" as meaning "No". To investigate whether LLMs have the ability to make this type of inference, known as an implicature, we design a simple task and evaluate four categories of widely used state-of-the-art models. We find that, despite only evaluating on utterances that require a binary inference (yes or no), models in three of these categories perform close to random. However, LLMs instruction-tuned at the example-level perform significantly better. These results suggest that certain fine-tuning strategies are far better at inducing pragmatic understanding in models. We present our findings as the starting point for further research into evaluating how LLMs interpret language in context and to drive the development of more pragmatic and useful models of human discourse., Comment: Accepted as Spotlight at NeurIPS 2023

The first evaluation of an ultra-high granularity digital electromagnetic calorimeter prototype using 1.0-5.8 GeV/c electrons is presented. The $25\times10^6$ pixel detector consists of 24 layers of ALPIDE CMOS MAPS sensors, with a pitch of around 30~$\mu$m, and has a depth of almost 20 radiation lengths of tungsten absorber. Ultra-thin cables allow for a very compact design. The properties that are critical for physics studies are measured: electromagnetic shower response, energy resolution and linearity. The stochastic energy resolution is comparable with the state-of-the art resolution for a Si-W calorimeter, with data described well by a simulation model using GEANT and Allpix$^2$. The performance achieved makes this technology a good candidate for use in the ALICE FoCal upgrade, and in general demonstrates the strong potential for future applications in high-energy physics., Comment: 30 pages, 19 figures, submitted to JINST

A prototype of a new type of calorimeter has been designed and constructed, based on a silicon-tungsten sampling design using pixel sensors with digital readout. It makes use of the Alpide MAPS sensor developed for the ALICE ITS upgrade. A binary readout is possible due to the pixel size of $\approx 30 \times 30 \, \mu \mathrm{m}^2$. This prototype has been successfully tested with cosmic muons and with test beams at DESY and the CERN SPS. We report on performance results obtained at DESY, showing good energy resolution and linearity, and compare to detailed MC simulations. Also shown are preliminary results of the high-energy performance as measured at the SPS. The two-shower separation capabilities are discussed., Comment: Proceedings to PM2021 - The 15. PISA Meeting on Advanced Detectors, updated after referee review

Systematic generalization is the ability to combine known parts into novel meaning; an important aspect of efficient human learning, but a weakness of neural network learning. In this work, we investigate how two well-known modeling principles -- modularity and data augmentation -- affect systematic generalization of neural networks in grounded language learning. We analyze how large the vocabulary needs to be to achieve systematic generalization and how similar the augmented data needs to be to the problem at hand. Our findings show that even in the controlled setting of a synthetic benchmark, achieving systematic generalization remains very difficult. After training on an augmented dataset with almost forty times more adverbs than the original problem, a non-modular baseline is not able to systematically generalize to a novel combination of a known verb and adverb. When separating the task into cognitive processes like perception and navigation, a modular neural network is able to utilize the augmented data and generalize more systematically, achieving 70% and 40% exact match increase over state-of-the-art on two gSCAN tests that have not previously been improved. We hope that this work gives insight into the drivers of systematic generalization, and what we still need to improve for neural networks to learn more like humans do.

The nuclear medium effects are studied in the $\nu_\tau/\bar\nu_\tau$ interactions from nuclei in the deep inelastic scattering (DIS) region and applied to the $^{40}Ar$ nucleus to obtain the scattering cross sections in the energy region of the proposed DUNE experiment. The free nucleon structure functions ($F_{iN}(x,Q^2);~(i=1-5)$) have been calculated at the next-to-leading order (NLO) using Martin-Motylinski-Harland Lang-Thorne 2014 as well as The Coordinated Theoretical-Experimental Project on QCD parameterizations for parton distribution functions (PDFs) and including the effect of perturbative and nonperturbative QCD corrections \cite{Ansari:2020xne}. These free nucleon structure functions are then convoluted with the nucleon spectral function in the nucleus to obtain the nuclear structure functions ($F_{iA}(x,Q^2);~(i=1-5)$). The nucleon spectral function takes into account the Fermi motion and the binding energy of the nucleons as well as the nucleon correlations within the nucleus. These nuclear structure functions are then used to calculate the deep inelastic scattering cross sections. Moreover, the contribution of $\pi$ and $\rho$ mesons as well as the corrections due to the shadowing and antishadowing effects in the relevant kinematic region of the Bjorken variable $x$ are also included. The numerical results for the cross sections have been presented and compared with the results obtained in the phenomenological approach using nuclear structure functions from nCTEQnu., Comment: 26 pages and 20 figures and 1 table

Small and medium enterprises rely on detailed Web analytics to be informed about their market and competition. Focused crawlers meet this demand by crawling and indexing specific parts of the Web. Critically, a focused crawler must quickly find new pages that have not yet been indexed. Since a new page can be discovered only by following a new outlink, predicting new outlinks is very relevant in practice. In the literature, many feature designs have been proposed for predicting changes in the Web. In this work we provide a structured analysis of this problem, using new outlinks as our running prediction target. Specifically, we unify earlier feature designs in a taxonomic arrangement of features along two dimensions: static versus dynamic features, and features of a page versus features of the network around it. Within this taxonomy, complemented by our new (mainly, dynamic network) features, we identify best predictors for new outlinks. Our main conclusion is that most informative features are the recent history of new outlinks on a page itself, and of its content-related pages. Hence, we propose a new 'look back, look around' (LBLA) model, that uses only these features. With the obtained predictions, we design a number of scoring functions to guide a focused crawler to pages with most new outlinks, and compare their performance. The LBLA approach proved extremely effective, outperforming other models including those that use a most complete set of features. One of the learners we use, is the recent NGBoost method that assumes a Poisson distribution for the number of new outlinks on a page, and learns its parameters. This connects the two so far unrelated avenues in the literature: predictions based on features of a page, and those based on probabilistic modelling. All experiments were carried out on an original dataset, made available by a commercial focused crawler., Comment: 23 pages, 15 figures, 4 tables, uses arxiv.sty, added new title, heuristic features and their results added, figures 7, 14, and 15 updated, accepted version

Humans are good at compositional zero-shot reasoning; someone who has never seen a zebra before could nevertheless recognize one when we tell them it looks like a horse with black and white stripes. Machine learning systems, on the other hand, usually leverage spurious correlations in the training data, and while such correlations can help recognize objects in context, they hurt generalization. To be able to deal with underspecified datasets while still leveraging contextual clues during classification, we propose ProtoProp, a novel prototype propagation graph method. First we learn prototypical representations of objects (e.g., zebra) that are conditionally independent w.r.t. their attribute labels (e.g., stripes) and vice versa. Next we propagate the independent prototypes through a compositional graph, to learn compositional prototypes of novel attribute-object combinations that reflect the dependencies of the target distribution. The method does not rely on any external data, such as class hierarchy graphs or pretrained word embeddings. We evaluate our approach on AO-Clever, a synthetic and strongly visual dataset with clean labels, and UT-Zappos, a noisy real-world dataset of fine-grained shoe types. We show that in the generalized compositional zero-shot setting we outperform state-of-the-art results, and through ablations we show the importance of each part of the method and their contribution to the final results.

Humans easily interpret expressions that describe unfamiliar situations composed from familiar parts ("greet the pink brontosaurus by the ferris wheel"). Modern neural networks, by contrast, struggle to interpret novel compositions. In this paper, we introduce a new benchmark, gSCAN, for evaluating compositional generalization in situated language understanding. Going beyond a related benchmark that focused on syntactic aspects of generalization, gSCAN defines a language grounded in the states of a grid world, facilitating novel evaluations of acquiring linguistically motivated rules. For example, agents must understand how adjectives such as 'small' are interpreted relative to the current world state or how adverbs such as 'cautiously' combine with new verbs. We test a strong multi-modal baseline model and a state-of-the-art compositional method finding that, in most cases, they fail dramatically when generalization requires systematic compositional rules., Comment: accepted at NeurIPS 2020

We propose the Insertion-Deletion Transformer, a novel transformer-based neural architecture and training method for sequence generation. The model consists of two phases that are executed iteratively, 1) an insertion phase and 2) a deletion phase. The insertion phase parameterizes a distribution of insertions on the current output hypothesis, while the deletion phase parameterizes a distribution of deletions over the current output hypothesis. The training method is a principled and simple algorithm, where the deletion model obtains its signal directly on-policy from the insertion model output. We demonstrate the effectiveness of our Insertion-Deletion Transformer on synthetic translation tasks, obtaining significant BLEU score improvement over an insertion-only model., Comment: Accepted as an Extended Abstract at the Workshop of Neural Generation and Translation (WNGT 2019) at EMNLP 2019