Self-supervised speech models have demonstrated the ability to learn rich acoustic representations. However, interpreting which specific phonological or acoustic features these models leverage within their highly polysemantic activations remains challenging. In this paper, we propose a straightforward and unsupervised probing method for model interpretability. We extract the activations from the final MLP layer of a pretrained HuBERT model and train a sparse autoencoder (SAE) using dictionary learning techniques to generate an over-complete set of latent representations. Analyzing these latent codes, we observe that a small subset of high-variance units consistently aligns with phonetic events, suggesting their potential utility as interpretable acoustic detectors. Our proposed method does not require labeled data beyond raw audio, providing a lightweight and accessible tool to gain insights into the internal workings of self-supervised speech models.

Large language models are powerful but costly. We ask whether meta-learning can make the pretraining of small language models not only faster but also more interpretable. We integrate first–order MAML with subset-masked LM pretraining, producing four LLama-style decoder-only models (11M–570M params), and evaluate on multilingual Universal NER. Compared with vanilla training, our hybrid setup (i) reaches the same loss up to 1.6× sooner, (ii) yields modest but consistent average gains on Universal NER at medium/large scales under equal compute (+2–3 percentage points), and (iii) and (iii) reveals phase-like learning dynamics: models first diversify their representations, then compress them in a pattern that aligns with improved episodic accuracy. These observations are correlational, not causal, and we do not claim generality beyond NER or across seeds. We also document a trade-off: perplexity on Paloma (a diverse language modeling benchmark spanning 18 domains) is worse at most scales. Code, checkpoints and analysis logs are released.

Transformer based models, specially large language models (LLMs) dominate the field of NLP with their mass adoption in tasks such as text generation, summarization and fake news detection. These models offer ease of deployment and reliability for most applications, however, they require significant amounts of computational power for training as well as inference. This poses challenges in their adoption in resource-constrained applications, specially in the open-source community where compute availability is usually scarce. This work proposes a graph-based approach for Environmental Claim Detection, exploring Graph Neural Networks (GNNs) and Hyperbolic Graph Neural Networks (HGNNs) as lightweight yet effective alternatives to transformer-based models. Re-framing the task as a graph classification problem, we transform claim sentences into dependency parsing graphs, utilizing a combination of word2vec & learnable part-of-speech (POS) tag embeddings for the node features and encoding syntactic dependencies in the edge relations. Our results show that our graph-based models, particularly HGNNs in the poincaré space (P-HGNNs), achieve performance superior to the state-of-the-art on environmental claim detection while using up to **30x fewer parameters**. We also demonstrate that HGNNs benefit vastly from explicitly modeling data in hierarchical (tree-like) structures, enabling them to significantly improve over their euclidean counterparts.

Continual learning (CL) presents a significant challenge for large pre-trained models, primarily due to catastrophic forgetting and the high computational cost of sequential knowledge updating. Parameter-Efficient Transfer Learning (PETL) methods offer reduced computational burdens but often struggle to effectively mitigate forgetting. This paper introduces Stacked Low-Rank Adaptation (SLoRA), a novel parameter-efficient approach that leverages the additive composition of task-specific, frozen low-rank adapters to enable modular continual learning with inherent support for explicit knowledge modification. SLoRA was evaluated on vision benchmarks, BERT-base, and the 1-billion-parameter Llama-3.2-1B model. Experiments demonstrated that SLoRA almost completely eliminated catastrophic forgetting, achieving a final average accuracy of 92.75% on Llama-3.2-1B while perfectly preserving prior task performance. Furthermore, SLoRA is computationally efficient, enabling up to a 15x training speed-up over full fine-tuning with 99.7% fewer trainable parameters per update. SLoRA offers a compelling balance of forgetting mitigation, parameter efficiency, and modularity, representing a promising direction for developing adaptable and efficient lifelong knowledgeable foundation models.

In this paper, we combine two-step knowledge distillation, structured pruning, truncation, and vocabulary trimming for extremely compressing multilingual encoder-only language models for low-resource languages. Our novel approach systematically combines existing techniques and takes them to the extreme, reducing layer depth, feed-forward hidden size, and intermediate layer embedding size to create significantly smaller monolingual models while retaining essential language-specific knowledge. We achieve compression rates of up to 92% while maintaining competitive performance, with average drops of 2–10% for moderate compression and 8–13% at maximum compression in four downstream tasks, including sentiment analysis, topic classification, named entity recognition, and part-of-speech tagging, across three low-resource languages. Notably, the performance degradation correlates with the amount of language-specific data in the teacher model, with larger datasets resulting in smaller performance losses. Additionally, we conduct ablation studies to identify the best practices for multilingual model compression using these techniques.

Traditional video-based tasks like soccer action spotting rely heavily on visual inputs, often requiring complex and computationally expensive models to process dense video data. We propose a shift from this video-centric approach to a text-based task, making it lightweight and scalable by utilizing Large Language Models (LLMs) instead of Vision-Language Models (VLMs). We posit that expert commentary, which provides rich descriptions and contextual cues contains sufficient information to reliably spot key actions in a match. To demonstrate this, we employ a system of three LLMs acting as judges specializing in outcome, excitement, and tactics for spotting actions in soccer matches. Our experiments show that this language-centric approach performs effectively in detecting critical match events coming close to state-of-the-art video-based spotters while using zero video processing compute and similar amount of time to process the entire match.

Reliable evaluation of Question Answering (QA) systems in low-resource Indic languages presents a significant challenge due to limited annotated datasets, linguistic diversity, and suitable evaluation metrics. Languages such as Sindhi, Manipuri, Dogri, Konkani, and Maithili are particularly underrepresented, creating difficulty in assessing Large Language Models (LLMs) on QA tasks. Existing metrics, including BLEU, ROUGE-L, and BERTScore, are effective in machine translation and high-resource settings; however, they often fail in low-resource QA due to score compression, zero-inflation, and poor scale alignment. To overcome this, LRMGS (Language-Robust Metric for Generative QA) is introduced to capture semantic and lexical agreement while preserving the score scale across languages. LRMGS is evaluated across 8 Indic languages and multiple LLMs, demonstrating consistently higher concordance with reference-based chrF++ scores, measured using the Concordance Correlation Coefficient (CCC). Experimental results indicate that LRMGS provides more accurate discrimination of system performance in very low-resource languages compared to existing metrics. This work establishes a robust and interpretable framework for evaluating QA systems in low-resource Indic languages, supporting more reliable multilingual model assessment.

Large language models show strong performance on knowledge intensive tasks such as fact checking and question answering, yet they often struggle with numerical reasoning. We present a systematic evaluation of state-of-the-art models for veracity prediction on numerical claims and evidence pairs using controlled perturbations, including label flipping probes, to test robustness. Our results indicate that even leading proprietary systems experience accuracy drops of up to 62% under certain perturbations. No model proves to be robust across all conditions. We further find that increasing context length generally reduces accuracy, but when extended context is enriched with perturbed demonstrations, most models substantially recover. These findings highlight critical limitations in numerical fact-checking and suggest that robustness remains an open challenge for current language models.

Theory of Mind (ToM) is the ability to understand others’ mental states, which is essential for human social interaction. Although recent studies suggest that large language models (LLMs) exhibit human-level ToM capabilities, the underlying mechanisms remain unclear. “Simulation Theory” posits that we infer others’ mental states by simulating their cognitive processes, which has been widely discussed in cognitive science. In this work, we propose a framework for investigating whether the ToM mechanism in LLMs is based on Simulation Theory by analyzing their internal representations. Following this framework, we successfully steered LLMs’ ToM reasoning through modeled perspective-taking and counterfactual interventions. Our results suggest that Simulation Theory may partially explain the ToM mechanism in state-of-the-art LLMs, indicating parallels between human and artificial social reasoning.

Large language models (LLMs) have the potential to be powerful instruments for psychotherapy. However, there is a shortage of practical tools to support their use in production. We develop a novel, iterative process of updating conversational context for tracking EIS (Emotional Intelligence Scale) instantaneously, and test Llama-70b. Through this, we show that (1) EIS varies more on psychotherapeutic (emotional support) conversations than control (emotionally unstimulating) conversations and (2) model responses can be systematically classified to identify consistent patterns. Thus, EIS is a valid indicator of empathetic model behavior. Rises in the EIS score correspond to prosocial behavior, and falls correspond to detached, unsocial behavior. These results suggest that psychometric questionnaires like EIS can provide a structured lens for observing empathetic stability of models and offer a foundation for future work on their role in psychotherapy.

Aspect-based summarization aims to generate summaries that highlight specific aspects of a text, enabling more personalized and targeted summaries. However, its application to books remains unexplored due to the difficulty of constructing reference summaries for long text. To address this challenge, we propose BookAsSumQA, a QA-based evaluation framework for aspect-based book summarization. BookAsSumQA automatically generates aspect-specific QA pairs from a narrative knowledge graph to evaluate summary quality based on its question-answering performance. Our experiments using BookAsSumQA revealed that while LLM-based approaches showed higher accuracy on shorter texts, RAG-based methods become more effective as document length increases, making them more efficient and practical for aspect-based book summarization.

Large language models (LLMs) excel across diverse natural language processing tasks but remain opaque and unreliable. This thesis investigates how LLM reasoning can be made both interpretable and reliable through systematic analysis of internal dynamics and targeted interventions. Unlike prior work that examines reasoning broadly, this research focuses on two representative domains: puzzle solving, where reasoning steps can be precisely tracked, and ontological inference, where hierarchical structures constrain valid reasoning. The central questions are: (1) How can systematic error patterns in domain specific reasoning be detected through layer wise probing and mitigated through targeted interventions? (2) How can probing frameworks and middle layer analyses reveal and enhance the computational mechanisms underlying inference? By combining probing methods, middle layer investigations, and probe guided interventions, the work aims to uncover interpretable reasoning patterns, identify systematic failure modes, and develop adaptive enhancement strategies. The expected outcome is a domain grounded framework that advances both theoretical understanding of neural reasoning and the design of practical, trustworthy AI systems.

Inference-time computation is a critical yet challenging paradigm for enhancing the reasoning performance of large language models (LLMs). While existing strategies improve reasoning stability and consistency, they suffer from notable limitations: self-correction often reinforces the model’s initial biases, and Multi-Agent Collaboration (MAC) often fails due to the lack of efficient coordination mechanisms, leading to collective errors. Although high-performing verifiers can detect reasoning errors, making them reliable requires substantial training. To address these challenges, we introduce a novel inference-time framework - **Adaptive Coopetition (AdCo)** - in which LLM agents utilize **an adaptive, UCB-based ‘coopetition’ mechanism**. At each round, agents leverage coarse verifier signals to determine whether to collaborate or compete, further iteratively refining their reasoning based on peer feedback. Without relying on high-performance verifiers, our adaptive strategy achieves significant performance gains on mathematical reasoning benchmarks, yielding **a 20% relative improvement** over baselines on the more challenging dataset. Our approach remains robust and consistent in terms of accuracy under different sample sizes and configurations. This adaptive, signal-guided ‘coopetition’ framework enhances reasoning robustness by leveraging bothmodel knowledge diversity and reasoning trace measure, while also promoting uncertainty-driven exploration, especially when participants have comparable capabilities. From this perspective, our work offers a fresh lens on inference-time computation and paves the way for more resilient multi-agent LLM systems.

We investigate whether Large Language Models (LLMs) exhibit human-like cognitive patterns under four established frameworks frompsychology: Thematic Apperception Test (TAT), Framing Bias, Moral Foundations Theory (MFT), and Cognitive Dissonance. We evaluated several proprietary and open-source models using structured prompts and automated scoring. Our findings reveal that these models often produce coherent narratives, show susceptibility to positive framing, exhibit moral judgments aligned with Liberty/Oppression concerns, and demonstrate self-contradictions tempered by extensive rationalization. Such behaviors mirror human cognitive tendencies yetare shaped by their training data and alignment methods. We discuss the implications for AI transparency, ethical deployment, and futurework that bridges cognitive psychology and AI safety.

Developing effective healthcare dialog systems requires controlling conversations to offer clear insight into the system’s understanding and to address the lack of patient-oriented conversational datasets. Moreover, evaluating these systems is equally challenging and requires user studies for robust evaluation. These challenges are even more pronounced when addressing the needs of minority populations with low health literacy and numeracy. This thesis proposal focuses on designing conversational architectures that deliver self-care information to African American patients with heart failure.Neuro-symbolic approaches provide a promising direction by integrating symbolic reasoning with the generative capabilities of Large Language Models (LLMs). In this proposal, we explore various approaches to creating a hybrid dialog model by combining the strengths of task-oriented dialog systems with the integration of neuro-symbolic rules into a Language Model (LM)/LLM-based dialog system, thereby controlling the dialog system. We propose a hybrid conversational system that uses schema graphs to control the flow of dialogue, while leveraging LLMs to generate responses grounded in these schemas. We will also conduct a user study to evaluate the system’s effectiveness.

Improving the performance of neural machine translation for low-resource languages is challenging due to the limited availability of parallel corpora. However, recently available Large Language Models (LLM) have demonstrated superior performance in various natural language processing tasks, including translation. In this work, we propose to incorporate an LLM into a Machine Translation (MT) model as a prior distribution to leverage its translation capabilities. The LLM acts as a teacher, instructing the student MT model about the target language. We conducted an experiment in four language pairs: English ⇔ German and English ⇔ Hindi. This resulted in improved BLEU and COMET scores in a low-resource setting.

Large language models have the potential to generate explanations for their own predictions in a variety of styles based on user instructions. Recent research has examined whether these self-explanations faithfully reflect the models’ actual behavior and has found that they often lack faithfulness. However, the question of how to improve faithfulness remains underexplored. Moreover, because different explanation styles have superficially distinct characteristics, it is unclear whether improvements observed in one style also arise when using other styles. This study analyzes the effects of training for faithful self-explanations and the extent to which these effects generalize, using three classification tasks and three explanation styles. We construct one-word constrained explanations that are likely to be faithful using a feature attribution method, and use these pseudo-faithful self-explanations for continual learning on instruction-tuned models. Our experiments demonstrate that training can improve self-explanation faithfulness across all classification tasks and explanation styles, and that these improvements also show signs of generalization to the multi-word settings and to unseen tasks. Furthermore, we find consistent cross-style generalization among three styles, suggesting that training may contribute to a broader improvement in faithful self-explanation ability.

While Large Language Models (LLMs) have shown remarkable performance in various Natural Language Processing (NLP) tasks, their effectiveness seem to be heavily biased toward high-resource languages. This proposal aims to address this gap by developing efficient training strategies for low-resource languages. We propose various techniques for efficient learning in simluated low-resource settings for English. We then plan to adapt these methods for low-resource languages. We plan to experiment with both natural language generation and understanding models. We evaluate the models on similar benchmarks as the BabyLM challenge for English. For other languages, we plan to use treebanks and translation techniques to create our own silver test set to evaluate the low-resource LMs.

We study automatic post-editing for patent translation, where accuracy and traceability are critical, and propose a two-step pipeline that combines a multilingual encoder for token-level error detection with an LLM for targeted correction. As no word-level annotations exist for Japanese–English patents, we create supervised data by injecting synthetic errors into parallel patent sentences and fine-tune mBERT, XLM-RoBERTa, and mDeBERTa as detectors. In the second stage, GPT-4o is prompted to revise translations either freely or under a restricted policy that allows edits only on detector-marked spans. For error detection, evaluation on synthetic errors shows that encoder-based detectors outperform LLMs in both F1 and MCC. For error correction, tests on synthetic, repetition, and omission datasets demonstrate statistically significant BLEU gains over LLM methods for synthetic and repetition errors, while omission errors remain challenging. Overall, pairing compact encoders with an LLM enables more accurate and controllable post-editing for key patent error types, reducing unnecessary rewrites via restricted edits. Future work will focus on strengthening omission modeling to better detect and correct missing content.

The relationship between tokenizer algorithm (e.g., Byte-Pair Encoding (BPE), Unigram), morphological alignment, tokenization quality (e.g., compression efficiency), and downstream performance remains largely unclear, particularly for languages with complex morphology. In this paper, we conduct a comprehensive evaluation of tokenizers using small-sized BERT models—from pre-training through fine-tuning—for Telugu (agglutinative), along with preliminary evaluation in Hindi (primarily fusional with some agglutination) and English (fusional). To evaluate morphological alignment of tokenizers in Telugu, we create a dataset containing gold morpheme segmentations of 600 derivational and 7000 inflectional word forms.Our experiments reveal two key findings for Telugu. First, the choice of tokenizer algorithm is the most significant factor influencing performance, with Unigram-based tokenizers consistently outperforming BPE across most settings. Second, while better morphological alignment shows a moderate, positive correlation with performance on text classification and structure prediction tasks, its impact is secondary to the tokenizer algorithm. Notably, hybrid approaches that use morphological information for pre-segmentation significantly boost the performance of BPE, though not Unigram. Our results further showcase the need for comprehensive intrinsic evaluation metrics for tokenizers that could explain downstream performance trends consistently.

Semantic role labeling (SRL) is a fundamental task in natural language processing that is crucial for achieving deep semantic understanding. Despite the success of large language models (LLMs) in several downstream NLP tasks, key tasks such as SRL remain a challenge for LLMs. Hence, in this study, we attempt to instantiate the efficacy of LLMs for the task of SRL via Question answering. Toward that goal, we investigate the effectiveness of five different LLMs (Llama, Mistral, Qwen, OpenChat, Gemini) using zero-shot and few-shot prompting. Our findings indicate that few-shot prompting enhances the performance of all models. Although Gemini outperformed others by a margin of 11%, Qwen and Llama are not too far behind. Additionally, we conduct a comprehensive error analysis to shed light on the cases where LLMs fail. This study offers valuable insights into the performance of LLMs for structured prediction and the effectiveness of simple prompting techniques in the Question-Answering framework for SRL.

Sarcasm is a specific form of ironic speech which can often be hard to understand for language models due to its nuanced nature. Recent improvements in the ability of such models to detect and generate sarcasm motivate us to try a new approach to help language models perceive sarcasm as a speech style, through a human cognitive perspective. In this work, we propose a multi-hop Chain of Thought (CoT) methodology to understand the context of an utterance that follows a dialogue and to perform bidirectional style transfer on that utterance, leveraging the Theory of Mind. We use small language models (SLMs) due to their cost-efficiency and fast response-time. The generated utterances are evaluated using both LLM-as-a-judge and human evaluation, suitable to the open-ended and stylistic nature of the generations. Along with these, we also evaluate scores of automated metrics such as DialogRPT, BLEU and SBERT; drawing valuable insights from them that support our evidence. Based on this, we find that our cognitive approach to sarcasm is an effective way for language models to stylistically understand and generate sarcasm with better authenticity.

While Large Language Models (LLMs) are known for their In-Context Learning (ICL) capabilities, there is no consensus on the underlying mechanisms. A key point of debate is whether ICL allows models to adapt to unseen tasks without parameter updates—that is, whether they can extrapolate. In this study, we address this question by constructing an arithmetic dataset based on the bivariate linear function z=ax+by to train a model and quantitatively evaluate its interpolation and extrapolation abilities through ICL. Our results show that while extrapolation was not achieved within our experimental design, tasks that were partially learned could be solved. We also found that the model acquires internal representations that can distinguish unseen tasks, and that greater task diversity in the training dataset improves ICL capabilities.

Large Language Models (LLMs) often hallucinate, generating nonsensical or false information that can be especially harmful in sensitive fields such as medicine or law. To study this phenomenon systematically, we introduce FalseCite, a curated dataset designed to capture and benchmark hallucinated responses induced by misleading or fabricated citations. Running GPT-4o-mini, Falcon-7B, and Mistral 7-B through FalseCite, we observed a noticeable increase in hallucination activity for false claims with deceptive citations, especially in GPT-4o-mini. Using the responses from FalseCite, we can also analyze the internal states of hallucinating models, visualizing and clustering the hidden state vectors. From this analysis, we noticed that the hidden state vectors, regardless of hallucination or non-hallucination, tend to trace out a distinct horn-like shape. Our work underscores FalseCite’s potential as a foundation for evaluating and mitigating hallucinations in future LLM research.

As neural networks are increasingly deployed in dynamic environments, they face the challenge of catastrophic forgetting, the tendency to overwrite previously learned knowledge when adapting to new tasks, resulting in severe performance degradation on earlier tasks. We propose Selective Forgetting-Aware Optimization (SFAO), a dynamic method that regulates gradient directions via cosine similarity and per-layer gating, enabling controlled forgetting while balancing plasticity and stability. SFAO selectively projects, accepts, or discards updates using a tunable mechanism with efficient Monte Carlo approximation. Experiments on standard continual learning benchmarks show that SFAO achieves competitive accuracy with markedly lower memory cost, a 90% reduction, and improved forgetting on MNIST datasets, making it suitable for resource-constrained scenarios.

Accurate classification in high-stakes domains requires not only correct predictions but transparent, traceable reasoning. We instantiate this need in clinical genomics and present VariantBench, a reproducible benchmark and scoring harness that evaluates both the final American College of Medical Genetics and Genomics/Association for Molecular Pathology (ACMG/AMP) labels and criterion-level reasoning fidelity for missense single-nucleotide variants (SNVs). Each case pairs a variant with deterministic, machine-readable evidence aligned to five commonly used criteria (PM2, PP3, PS1, BS1, BA1), enabling consistent evaluation of large language models (LLMs). Unlike prior work that reports only final labels, our framework scores the correctness and faithfulness of per-criterion justifications against numeric evidence. On a balanced 100-variant freeze, Gemini 2.5 Flash and GPT-4o outperform Claude 3 Opus on label accuracy and criterion detection, and both improve materially when the decisive PS1 cue is provided explicitly. Error analyses show models master population-frequency cues yet underuse high-impact rules unless evidence is unambiguous. VariantBench delivers a substrate to track such improvements and compare prompting, calibration, and aggregation strategies in genomics and other rule-governed, safety-critical settings.

This study investigated how well computational embeddings aligned with human semantic judgments in the processing of English compound words. We compared static word vectors (GloVe) and contextualized embeddings (BERT) against human ratings of lexeme meaning dominance (LMD) and semantic transparency (ST) drawn from a psycholinguistic dataset. Using measures of association strength (Edinburgh Associative Thesaurus), frequency (BNC), and predictability (LaDEC), we computed embedding-derived LMD and ST metrics and assessed their relationships with human judgments via Spearman’s correlation and regression analyses. Our results showed that BERT embeddings better captured compositional semantics than GloVe, and that predictability ratings were strong predictors of semantic transparency in both human and model data. These findings advanced computational psycholinguistics by clarifying the factors that drove compound word processing and offered insights into embedding-based semantic modeling.