Joe Stacey

Joe Stacey, Pasquale Minervini, Haim Dubossarsky et al.

Current Natural Language Inference (NLI) models achieve impressive results, sometimes outperforming humans when evaluating on in-distribution test sets. However, as these models are known to learn from annotation artefacts and dataset biases, it is unclear to what extent the models are learning the task of NLI instead of learning from shallow heuristics in their training data. We address this issue by introducing a logical reasoning framework for NLI, creating highly transparent model decisions that are based on logical rules. Unlike prior work, we show that improved interpretability can be achieved without decreasing the predictive accuracy. We almost fully retain performance on SNLI, while also identifying the exact hypothesis spans that are responsible for each model prediction. Using the e-SNLI human explanations, we verify that our model makes sensible decisions at a span level, despite not using any span labels during training. We can further improve model performance and span-level decisions by using the e-SNLI explanations during training. Finally, our model is more robust in a reduced data setting. When training with only 1,000 examples, out-of-distribution performance improves on the MNLI matched and mismatched validation sets by 13% and 16% relative to the baseline. Training with fewer observations yields further improvements, both in-distribution and out-of-distribution.

Joe Stacey, Hadas Orgad, Kentaro Inui et al.

Recent work has shown that the hidden states of large language models contain signals useful for uncertainty estimation and hallucination detection, motivating a growing interest in efficient probe-based approaches. Yet it remains unclear how robust existing methods are, and which probe designs provide uncertainty estimates that are reliable under distribution shift. We present a systematic study of supervised uncertainty probes across models, tasks, and OOD settings, training over 2,000 probes while varying the representation layer, feature type, and token aggregation strategy. Our evaluation highlights poor robustness in current methods, particularly in the case of long-form generations. We also find that probe robustness is driven less by architecture and more by the probe inputs. Middle-layer representations generalise more reliably than final-layer hidden states, and aggregating across response tokens is consistently more robust than relying on single-token features. These differences are often largely invisible in-distribution but become more important under distribution shift. Informed by our evaluation, we explore a simple hybrid back-off strategy for improving robustness, arguing that better evaluation is a prerequisite for building more robust probes.

Joe Stacey, Jianpeng Cheng, John Torr et al.

Spurred by recent advances in Large Language Models (LLMs), virtual assistants are poised to take a leap forward in terms of their dialogue capabilities. Yet a major bottleneck to achieving genuinely transformative task-oriented dialogue capabilities remains the scarcity of high quality data. Existing datasets, while impressive in scale, have limited domain coverage and contain few genuinely challenging conversational phenomena; those which are present are typically unlabelled, making it difficult to assess the strengths and weaknesses of models without time-consuming and costly human evaluation. Moreover, creating high quality dialogue data has until now required considerable human input, limiting both the scale of these datasets and the ability to rapidly bootstrap data for a new target domain. We aim to overcome these issues with LUCID, a modularised and highly automated LLM-driven data generation system that produces realistic, diverse and challenging dialogues. We use LUCID to generate a seed dataset of 4,277 conversations across 100 intents to demonstrate its capabilities, with a human review finding consistently high quality labels in the generated data.

Lisa Alazraki, Lihu Chen, Ana Brassard et al.

Large Language Models (LLMs) have achieved high accuracy on complex commonsense and mathematical problems that involve the composition of multiple reasoning steps. However, current compositional benchmarks testing these skills tend to focus on either commonsense or math reasoning, whereas LLM agents solving real-world tasks would require a combination of both. In this work, we introduce an Agentic Commonsense and Math benchmark (AgentCoMa), where each compositional task requires a commonsense reasoning step and a math reasoning step. We test it on 61 LLMs of different sizes, model families, and training strategies. We find that LLMs can usually solve both steps in isolation, yet their accuracy drops by ~30% on average when the two are combined. This is a substantially greater performance gap than the one we observe in prior compositional benchmarks that combine multiple steps of the same reasoning type. In contrast, non-expert human annotators can solve the compositional questions and the individual steps in AgentCoMa with similarly high accuracy. Furthermore, we conduct a series of interpretability studies to better understand the performance gap, examining neuron patterns, attention maps and membership inference. Our work underscores a substantial degree of model brittleness in the context of mixed-type compositional reasoning and offers a test bed for future improvement.

Alexandru Coca, Bo-Hsiang Tseng, Pete Boothroyd et al.

Programmable task-oriented dialogue (TOD) agents enable language models to follow structured dialogue policies, but their effectiveness hinges on accurate state tracking. We present PyTOD, an agent that generates executable code to track dialogue state and uses policy and execution feedback for efficient error correction. To this end, PyTOD employs a simple constrained decoding approach, using a language model instead of grammar rules to follow API schemata. This leads to state-of-the-art state tracking performance on the challenging SGD benchmark. Our experiments show that PyTOD surpasses strong baselines in both accuracy and robust user goal estimation as the dialogue progresses, demonstrating the effectiveness of execution-aware state tracking.

Joe Stacey, Pasquale Minervini, Haim Dubossarsky et al.

With recent advances, neural models can achieve human-level performance on various natural language tasks. However, there are no guarantees that any explanations from these models are faithful, i.e. that they reflect the inner workings of the model. Atomic inference overcomes this issue, providing interpretable and faithful model decisions. This approach involves making predictions for different components (or atoms) of an instance, before using interpretable and deterministic rules to derive the overall prediction based on the individual atom-level predictions. We investigate the effectiveness of using LLM-generated facts as atoms, decomposing Natural Language Inference premises into lists of facts. While directly using generated facts in atomic inference systems can result in worse performance, with 1) a multi-stage fact generation process, and 2) a training regime that incorporates the facts, our fact-based method outperforms other approaches.

Joe Stacey, Marek Rei

Knowledge distillation optimises a smaller student model to behave similarly to a larger teacher model, retaining some of the performance benefits. While this method can improve results on in-distribution examples, it does not necessarily generalise to out-of-distribution (OOD) settings. We investigate two complementary methods for improving the robustness of the resulting student models on OOD domains. The first approach augments the distillation with generated unlabelled examples that match the target distribution. The second method upsamples data points among the training set that are similar to the target distribution. When applied on the task of natural language inference (NLI), our experiments on MNLI show that distillation with these modifications outperforms previous robustness solutions. We also find that these methods improve performance on OOD domains even beyond the target domain.

Joe Stacey, Yonatan Belinkov, Marek Rei

Natural Language Inference (NLI) models are known to learn from biases and artefacts within their training data, impacting how well they generalise to other unseen datasets. Existing de-biasing approaches focus on preventing the models from learning these biases, which can result in restrictive models and lower performance. We instead investigate teaching the model how a human would approach the NLI task, in order to learn features that will generalise better to previously unseen examples. Using natural language explanations, we supervise the model's attention weights to encourage more attention to be paid to the words present in the explanations, significantly improving model performance. Our experiments show that the in-distribution improvements of this method are also accompanied by out-of-distribution improvements, with the supervised models learning from features that generalise better to other NLI datasets. Analysis of the model indicates that human explanations encourage increased attention on the important words, with more attention paid to words in the premise and less attention paid to punctuation and stop-words.

Joe Stacey, Pasquale Minervini, Haim Dubossarsky et al.

Natural Language Inference (NLI) datasets contain annotation artefacts resulting in spurious correlations between the natural language utterances and their respective entailment classes. These artefacts are exploited by neural networks even when only considering the hypothesis and ignoring the premise, leading to unwanted biases. Belinkov et al. (2019b) proposed tackling this problem via adversarial training, but this can lead to learned sentence representations that still suffer from the same biases. We show that the bias can be reduced in the sentence representations by using an ensemble of adversaries, encouraging the model to jointly decrease the accuracy of these different adversaries while fitting the data. This approach produces more robust NLI models, outperforming previous de-biasing efforts when generalised to 12 other datasets (Belinkov et al., 2019a; Mahabadi et al., 2020). In addition, we find that the optimal number of adversarial classifiers depends on the dimensionality of the sentence representations, with larger sentence representations being more difficult to de-bias while benefiting from using a greater number of adversaries.