arXiv subjects

Milica Gašić

Publications and source records attributed to Milica Gašić.

At least 19 recordsLinked to original sources

arXiv2026

Post-Training Large Language Models via Reinforcement Learning from Self-Feedback

Large Language Models (LLMs) often produce plausible but poorly-calibrated answers, limiting their reliability on reasoning-intensive tasks. Recent research suggests that Chain-of-Thought (CoT) reasoning paths are inherent in pre-trained LLMs and can be elicited by simply altering the decoding process, where the presence of a CoT path correlates with higher answer confidence. Building on these insights, we present Reinforcement Learning from Self-Feedback (RLSF), a post-training stage that utilises the model's intrinsic confidence as a self-generated reward. By generating multiple CoT decoding beams from a frozen LLM, we compute the confidence of each final answer span and rank the resulting traces accordingly to create synthetic preferences. These preferences are subsequently utilised to fine-tune the policy through standard preference optimisation, requiring no human labels, gold answers, or externally curated rewards. RLSF simultaneously (i) refines the model's probability estimates--restoring well-behaved calibration--and (ii) strengthens step-by-step reasoning, yielding improved performance on arithmetic reasoning and multiple-choice question answering. By converting a model's own uncertainty into structured self-feedback, RLSF affirms reinforcement learning on intrinsic model behaviour as a principled and data-efficient component of the LLM post-training pipeline. Our results demonstrate that leveraging these inherent reasoning capabilities provides a robust path for enhancing model reliability without manual prompt engineering or external supervision.

cs.CL↗

arXiv2026

All Learning Has an Emotional Basis, So Does Task-Oriented Dialogue

Task-oriented dialogue (ToD) systems aim to help users accomplish goals through natural language interaction. Beyond task success, effective ToD systems must also maintain positive emotional interaction and accurately convey information in inherently noisy and ambiguous conversational environments. Recent advances in large language models (LLMs) have substantially improved conversational fluency and contextual understanding. However, although emotion has been incorporated into existing ToD systems, its integration is typically confined to limited roles, restricting its influence on overall dialogue behaviour and user experience. To address this, we propose a novel end-to-end framework that combines fully lexicalised representations with an LLM backbone for improved semantic accuracy and employs online reinforcement learning with short-term emotional and long-term task-success rewards for joint optimisation of emotional interaction and task success. We further develop a challenging and realistic simulation environment that enables systematic evaluation across diverse ToD set-ups. Experiments show that incorporating coarse affective signals improves task success and sentiment rating in aggregate. Furthermore, integrating emotional interaction throughout the dialogue system using our proposed framework results in more pronounced improvements in both task success and user emotional experience.

cs.CL↗

arXiv2026

Prompt reinforcing for long-term planning of large language models

Large language models (LLMs) have achieved remarkable success in a wide range of natural language processing tasks and can be adapted through prompting. However, they remain suboptimal in multi-turn interactions, often relying on incorrect early assumptions and failing to track user goals over time, which makes such tasks particularly challenging. Prior works in dialogue systems have shown that long-term planning is essential for handling interactive tasks. In this work, we propose a prompt optimisation framework inspired by reinforcement learning, which enables such planning to take place by only modifying the task instruction prompt of the LLM-based agent. By generating turn-by-turn feedback and leveraging experience replay for prompt rewriting, our proposed method shows significant improvement in multi-turn tasks such as text-to-SQL and task-oriented dialogue. Moreover, it generalises across different LLM-based agents and can leverage diverse LLMs as meta-prompting agents. This warrants future research in reinforcement learning-inspired parameter-free optimisation methods.

cs.CL↗

arXiv2026

Exploration Through Introspection: A Self-Aware Reward Model

Understanding how artificial agents model internal mental states is central to advancing Theory of Mind in AI. Evidence points to a unified system for self- and other-awareness. We explore this self-awareness by having reinforcement learning agents infer their own internal states in gridworld environments. Specifically, we introduce an introspective exploration component that is inspired by biological pain as a learning signal by utilizing a hidden Markov model to infer "pain-belief" from online observations. This signal is integrated into a subjective reward function to study how self-awareness affects the agent's learning abilities. Further, we use this computational framework to investigate the difference in performance between normal and chronic pain perception models. Results show that introspective agents in general significantly outperform standard baseline agents and can replicate complex human-like behaviors.

cs.AI↗

arXiv2025

Less is More: Local Intrinsic Dimensions of Contextual Language Models

Understanding the internal mechanisms of large language models (LLMs) remains a challenging and complex endeavor. Even fundamental questions, such as how fine-tuning affects model behavior, often require extensive empirical evaluation. In this paper, we introduce a novel perspective based on the geometric properties of contextual latent embeddings to study the effects of training and fine-tuning. To that end, we measure the local dimensions of a contextual language model's latent space and analyze their shifts during training and fine-tuning. We show that the local dimensions provide insights into the model's training dynamics and generalization ability. Specifically, the mean of the local dimensions predicts when the model's training capabilities are exhausted, as exemplified in a dialogue state tracking task, overfitting, as demonstrated in an emotion recognition task, and grokking, as illustrated with an arithmetic task. Furthermore, our experiments suggest a practical heuristic: reductions in the mean local dimension tend to accompany and predict subsequent performance gains. Through this exploration, we aim to provide practitioners with a deeper understanding of the implications of fine-tuning on embedding spaces, facilitating informed decisions when configuring models for specific applications. The results of this work contribute to the ongoing discourse on the interpretability, adaptability, and generalizability of LLMs by bridging the gap between intrinsic model mechanisms and geometric properties in the respective embeddings.

cs.CL↗

arXiv2025

Dialogue Ontology Relation Extraction via Constrained Chain-of-Thought Decoding

State-of-the-art task-oriented dialogue systems typically rely on task-specific ontologies for fulfilling user queries. The majority of task-oriented dialogue data, such as customer service recordings, comes without ontology and annotation. Such ontologies are normally built manually, limiting the application of specialised systems. Dialogue ontology construction is an approach for automating that process and typically consists of two steps: term extraction and relation extraction. In this work, we focus on relation extraction in a transfer learning set-up. To improve the generalisation, we propose an extension to the decoding mechanism of large language models. We adapt Chain-of-Thought (CoT) decoding, recently developed for reasoning problems, to generative relation extraction. Here, we generate multiple branches in the decoding space and select the relations based on a confidence threshold. By constraining the decoding to ontology terms and relations, we aim to decrease the risk of hallucination. We conduct extensive experimentation on two widely used datasets and find improvements in performance on target ontology for source fine-tuned and one-shot prompted large language models.

cs.CL↗

arXiv2025

A Confidence-based Acquisition Model for Self-supervised Active Learning and Label Correction

Supervised neural approaches are hindered by their dependence on large, meticulously annotated datasets, a requirement that is particularly cumbersome for sequential tasks. The quality of annotations tends to deteriorate with the transition from expert-based to crowd-sourced labelling. To address these challenges, we present CAMEL (Confidence-based Acquisition Model for Efficient self-supervised active Learning), a pool-based active learning framework tailored to sequential multi-output problems. CAMEL possesses two core features: (1) it requires expert annotators to label only a fraction of a chosen sequence, and (2) it facilitates self-supervision for the remainder of the sequence. By deploying a label correction mechanism, CAMEL can also be utilised for data cleaning. We evaluate CAMEL on two sequential tasks, with a special emphasis on dialogue belief tracking, a task plagued by the constraints of limited and noisy datasets. Our experiments demonstrate that CAMEL significantly outperforms the baselines in terms of efficiency. Furthermore, the data corrections suggested by our method contribute to an overall improvement in the quality of the resulting datasets.

cs.CL↗

arXiv2024

Learning from Noisy Labels via Self-Taught On-the-Fly Meta Loss Rescaling

Correct labels are indispensable for training effective machine learning models. However, creating high-quality labels is expensive, and even professionally labeled data contains errors and ambiguities. Filtering and denoising can be applied to curate labeled data prior to training, at the cost of additional processing and loss of information. An alternative is on-the-fly sample reweighting during the training process to decrease the negative impact of incorrect or ambiguous labels, but this typically requires clean seed data. In this work we propose unsupervised on-the-fly meta loss rescaling to reweight training samples. Crucially, we rely only on features provided by the model being trained, to learn a rescaling function in real time without knowledge of the true clean data distribution. We achieve this via a novel meta learning setup that samples validation data for the meta update directly from the noisy training corpus by employing the rescaling function being trained. Our proposed method consistently improves performance across various NLP tasks with minimal computational overhead. Further, we are among the first to attempt on-the-fly training data reweighting on the challenging task of dialogue modeling, where noisy and ambiguous labels are common. Our strategy is robust in the face of noisy and clean data, handles class imbalance, and prevents overfitting to noisy labels. Our self-taught loss rescaling improves as the model trains, showing the ability to keep learning from the model's own signals. As training progresses, the impact of correctly labeled data is scaled up, while the impact of wrongly labeled data is suppressed.

cs.CL↗

arXiv2024

Local Topology Measures of Contextual Language Model Latent Spaces With Applications to Dialogue Term Extraction

A common approach for sequence tagging tasks based on contextual word representations is to train a machine learning classifier directly on these embedding vectors. This approach has two shortcomings. First, such methods consider single input sequences in isolation and are unable to put an individual embedding vector in relation to vectors outside the current local context of use. Second, the high performance of these models relies on fine-tuning the embedding model in conjunction with the classifier, which may not always be feasible due to the size or inaccessibility of the underlying feature-generation model. It is thus desirable, given a collection of embedding vectors of a corpus, i.e., a datastore, to find features of each vector that describe its relation to other, similar vectors in the datastore. With this in mind, we introduce complexity measures of the local topology of the latent space of a contextual language model with respect to a given datastore. The effectiveness of our features is demonstrated through their application to dialogue term extraction. Our work continues a line of research that explores the manifold hypothesis for word embeddings, demonstrating that local structure in the space carved out by word embeddings can be exploited to infer semantic properties.

cs.CL↗

arXiv2024

Affect Recognition in Conversations Using Large Language Models

Affect recognition, encompassing emotions, moods, and feelings, plays a pivotal role in human communication. In the realm of conversational artificial intelligence, the ability to discern and respond to human affective cues is a critical factor for creating engaging and empathetic interactions. This study investigates the capacity of large language models (LLMs) to recognise human affect in conversations, with a focus on both open-domain chit-chat dialogues and task-oriented dialogues. Leveraging three diverse datasets, namely IEMOCAP (Busso et al., 2008), EmoWOZ (Feng et al., 2022), and DAIC-WOZ (Gratch et al., 2014), covering a spectrum of dialogues from casual conversations to clinical interviews, we evaluate and compare LLMs' performance in affect recognition. Our investigation explores the zero-shot and few-shot capabilities of LLMs through in-context learning as well as their model capacities through task-specific fine-tuning. Additionally, this study takes into account the potential impact of automatic speech recognition errors on LLM predictions. With this work, we aim to shed light on the extent to which LLMs can replicate human-like affect recognition capabilities in conversations.

cs.CL↗

arXiv2024

Infusing Emotions into Task-oriented Dialogue Systems: Understanding, Management, and Generation

Emotions are indispensable in human communication, but are often overlooked in task-oriented dialogue (ToD) modelling, where the task success is the primary focus. While existing works have explored user emotions or similar concepts in some ToD tasks, none has so far included emotion modelling into a fully-fledged ToD system nor conducted interaction with human or simulated users. In this work, we incorporate emotion into the complete ToD processing loop, involving understanding, management, and generation. To this end, we extend the EmoWOZ dataset (Feng et al., 2022) with system affective behaviour labels. Through interactive experimentation involving both simulated and human users, we demonstrate that our proposed framework significantly enhances the user's emotional experience as well as the task success.

cs.CL↗

arXiv2023

Speech-based Slot Filling using Large Language Models

Recently, advancements in large language models (LLMs) have shown an unprecedented ability across various language tasks. This paper investigates the potential application of LLMs to slot filling with noisy ASR transcriptions, via both in-context learning and task-specific fine-tuning. Dedicated prompt designs and fine-tuning approaches are proposed to improve the robustness of LLMs for slot filling with noisy ASR transcriptions. Moreover, a linearised knowledge injection (LKI) scheme is also proposed to integrate dynamic external knowledge into LLMs. Experiments were performed on SLURP to quantify the performance of LLMs, including GPT-3.5-turbo, GPT-4, LLaMA-13B and Vicuna-13B (v1.1 and v1.5) with different ASR error rates. The use of the proposed fine-tuning together with the LKI scheme for LLaMA-13B achieved an 8.3% absolute SLU-F1 improvement compared to the strong Flan-T5-base baseline system on a limited data setup.

cs.CL↗

arXiv2023

ConvLab-3: A Flexible Dialogue System Toolkit Based on a Unified Data Format

Task-oriented dialogue (TOD) systems function as digital assistants, guiding users through various tasks such as booking flights or finding restaurants. Existing toolkits for building TOD systems often fall short of in delivering comprehensive arrays of data, models, and experimental environments with a user-friendly experience. We introduce ConvLab-3: a multifaceted dialogue system toolkit crafted to bridge this gap. Our unified data format simplifies the integration of diverse datasets and models, significantly reducing complexity and cost for studying generalization and transfer. Enhanced with robust reinforcement learning (RL) tools, featuring a streamlined training process, in-depth evaluation tools, and a selection of user simulators, ConvLab-3 supports the rapid development and evaluation of robust dialogue policies. Through an extensive study, we demonstrate the efficacy of transfer learning and RL and showcase that ConvLab-3 is not only a powerful tool for seasoned researchers but also an accessible platform for newcomers.

cs.CL↗

arXiv2023

From Chatter to Matter: Addressing Critical Steps of Emotion Recognition Learning in Task-oriented Dialogue

Emotion recognition in conversations (ERC) is a crucial task for building human-like conversational agents. While substantial efforts have been devoted to ERC for chit-chat dialogues, the task-oriented counterpart is largely left unattended. Directly applying chit-chat ERC models to task-oriented dialogues (ToDs) results in suboptimal performance as these models overlook key features such as the correlation between emotions and task completion in ToDs. In this paper, we propose a framework that turns a chit-chat ERC model into a task-oriented one, addressing three critical aspects: data, features and objective. First, we devise two ways of augmenting rare emotions to improve ERC performance. Second, we use dialogue states as auxiliary features to incorporate key information from the goal of the user. Lastly, we leverage a multi-aspect emotion definition in ToDs to devise a multi-task learning objective and a novel emotion-distance weighted loss function. Our framework yields significant improvements for a range of chit-chat ERC models on EmoWOZ, a large-scale dataset for user emotion in ToDs. We further investigate the generalisability of the best resulting model to predict user satisfaction in different ToD datasets. A comparison with supervised baselines shows a strong zero-shot capability, highlighting the potential usage of our framework in wider scenarios.

cs.CL↗

arXiv2023

ChatGPT for Zero-shot Dialogue State Tracking: A Solution or an Opportunity?

Recent research on dialogue state tracking (DST) focuses on methods that allow few- and zero-shot transfer to new domains or schemas. However, performance gains heavily depend on aggressive data augmentation and fine-tuning of ever larger language model based architectures. In contrast, general purpose language models, trained on large amounts of diverse data, hold the promise of solving any kind of task without task-specific training. We present preliminary experimental results on the ChatGPT research preview, showing that ChatGPT achieves state-of-the-art performance in zero-shot DST. Despite our findings, we argue that properties inherent to general purpose models limit their ability to replace specialized systems. We further theorize that the in-context learning capabilities of such models will likely become powerful tools to support the development of dedicated and dynamic dialogue state trackers.

cs.CL↗

arXiv2023

EmoUS: Simulating User Emotions in Task-Oriented Dialogues

Existing user simulators (USs) for task-oriented dialogue systems only model user behaviour on semantic and natural language levels without considering the user persona and emotions. Optimising dialogue systems with generic user policies, which cannot model diverse user behaviour driven by different emotional states, may result in a high drop-off rate when deployed in the real world. Thus, we present EmoUS, a user simulator that learns to simulate user emotions alongside user behaviour. EmoUS generates user emotions, semantic actions, and natural language responses based on the user goal, the dialogue history, and the user persona. By analysing what kind of system behaviour elicits what kind of user emotions, we show that EmoUS can be used as a probe to evaluate a variety of dialogue systems and in particular their effect on the user's emotional state. Developing such methods is important in the age of large language model chat-bots and rising ethical concerns.

cs.CL↗

arXiv2022

Dynamic Dialogue Policy for Continual Reinforcement Learning

Continual learning is one of the key components of human learning and a necessary requirement of artificial intelligence. As dialogue can potentially span infinitely many topics and tasks, a task-oriented dialogue system must have the capability to continually learn, dynamically adapting to new challenges while preserving the knowledge it already acquired. Despite the importance, continual reinforcement learning of the dialogue policy has remained largely unaddressed. The lack of a framework with training protocols, baseline models and suitable metrics, has so far hindered research in this direction. In this work we fill precisely this gap, enabling research in dialogue policy optimisation to go from static to dynamic learning. We provide a continual learning algorithm, baseline architectures and metrics for assessing continual learning models. Moreover, we propose the dynamic dialogue policy transformer (DDPT), a novel dynamic architecture that can integrate new knowledge seamlessly, is capable of handling large state spaces and obtains significant zero-shot performance when being exposed to unseen domains, without any growth in network parameter size.

cs.CL↗

arXiv2022

Dialogue Evaluation with Offline Reinforcement Learning

Task-oriented dialogue systems aim to fulfill user goals through natural language interactions. They are ideally evaluated with human users, which however is unattainable to do at every iteration of the development phase. Simulated users could be an alternative, however their development is nontrivial. Therefore, researchers resort to offline metrics on existing human-human corpora, which are more practical and easily reproducible. They are unfortunately limited in reflecting real performance of dialogue systems. BLEU for instance is poorly correlated with human judgment, and existing corpus-based metrics such as success rate overlook dialogue context mismatches. There is still a need for a reliable metric for task-oriented systems with good generalization and strong correlation with human judgements. In this paper, we propose the use of offline reinforcement learning for dialogue evaluation based on a static corpus. Such an evaluator is typically called a critic and utilized for policy optimization. We go one step further and show that offline RL critics can be trained on a static corpus for any dialogue system as external evaluators, allowing dialogue performance comparisons across various types of systems. This approach has the benefit of being corpus- and model-independent, while attaining strong correlation with human judgements, which we confirm via an interactive user trial.

cs.CL↗