Abstract

This thesis investigates the dynamics of answer formation in Large Language Models (LLMs) along
two dimensions: vertically across network layers, and horizontally across sequential reasoning tokens.
To evaluate vertical evolution, we apply layer-wise linear probes to open-weight models on a controlled character-counting task. The probing results show that early and middle layers reliably encode
sub-token character counts. Late-stage network components, specifically the penultimate and final multilayer perceptrons, actively suppress this information and default to degenerate prediction strategies. The
models thus fail at utilization, despite having the required representational capacity.
For horizontal evolution, we introduce Forced Answer Completion (FAC) to extract a model's intermediate answer commitment at each step of its chain-of-thought trace. Across multiple-choice, numeric,
and search tasks, FAC shows that models decide their final answer long before the reasoning trace concludes. After filtering out Transient Answer Flips (TAFs) with a causal smoothing operator, we find
that nearly half of all reasoning tokens are generated after the model's decision has stabilized. These
late-stage tokens function primarily as post-hoc justification.
These findings enable direct interventions in LLM inference. Vertical tracking isolates suppressed
knowledge, providing targets for structural adjustments. Horizontal tracking supports lightweight earlystopping gates by training linear probes on intermediate hidden states, we predict when an answer has
stabilized and halt generation. Testing on the Qwen3 architecture demonstrates that this early-stopping
method saves hundreds of tokens per query with a marginal drop in accuracy

Abstract

Large Language Models (LLMs) have demonstrated remarkable capabilities in natural language processing, yet their application to Machine Translation (MT), particularly for lowresource languages, remains an active area of research. This thesis investigates the adaptation
of LLMs for machine translation involving English and Indian languages, addressing challenges
in assessment, fine-tuning, alignment, evaluation, and inference efficiency.
We begin by assessing the translation capabilities of popular LLMs (LLaMA, Mistral, Bloom)
for 22 Indian languages under zero-shot and in-context learning settings. Our findings reveal
that while raw LLMs exhibit limited translation quality, parameter-efficient fine-tuning using
LoRA significantly improves performance, with two-stage fine-tuning achieving results competitive with specialized MT systems.
Building on these insights, we investigate fine-tuning strategies for extremely low-resource
Indic languages through participation in the WMT 2024 shared task. Our experiments across
Assamese, Mizo, Khasi, and Manipuri demonstrate that fine-tuned LLMs substantially outperform models trained from scratch, and that multilingual approaches consistently surpass
monolingual ones.
We then explore data-efficient alignment strategies using preference optimization. Our Progressive Perturbation approach with Kahneman-Tversky Optimization (KTO) demonstrates
that targeted perturbations of training data can enhance translation quality without requiring
extensive parallel corpora.
To address the challenge of evaluating MT for low-resource languages, we conduct a comprehensive analysis of statistical and neural evaluation metrics, examining their stability, consistency, and correlation with human judgments.
Finally, we tackle inference efficiency through DistillBeam, a multi-trajectory knowledge
distillation framework for speculative decoding. By aggregating supervision from multiple
high-probability teacher trajectories, DistillBeam achieves wall-clock speedups of 35-65% over
autoregressive decoding across 20 languages.
This thesis contributes practical methodologies and empirical insights for deploying LLMbased translation systems for Indian languages, balancing translation quality with computational efficiency.

Abstract

The Internet of Things (IoT) has revolutionized various fields by enabling the collection of data
in real-time and interacting remotely with physical systems. In the education sector, one of the most
impactful uses of IoT is remote labs, which enable students to perform experiments virtually. This
has been particularly impactful during the COVID-19 pandemic or for those who are located remotely.
Though remote labs are becoming extremely popular for physics and electronics due to their relatively
simpler requirements, chemistry has not yet found a suitable place due to the complexity involved. In
addition, existing remote labs are not focused on creating a deep level of reflection during learning.
This thesis presents a two-pronged approach to bridging the gap between traditional and remote
laboratory learning by focusing on chemistry education and pedagogical enhancement. First, it details
the design and implementation of a fully retrofitted, IoT-enabled titration system. The proposed system
has been implemented on the RLabs platform at IIIT Hyderabad and provides a facility for users to
conduct acid-base titration experiments virtually using a web-based platform. It consists of hardware
components that help in precise measurement of fluids, visual feedback mechanisms, and automatic
cleaning facilities. This helps users conduct repeated experiments without any physical interference.
The experimental setup has been tested for accuracy, usability, and cost-effectiveness, demonstrating its
potential as a scalable solution for remote chemistry education.
The second part of this work addresses the pedagogical shortcomings of existing remote labs, which
are mostly limited in their instruction and lack active engagement with learners during the experiment.
To fill this gap, a large language model (LLM)-powered assistant is integrated into the RLabs framework. Grounded in a dual-theoretical foundation of Kolb's Experiential Learning Cycle and Bloom's
Taxonomy, the assistant facilitates active experimentation, conceptual reflection, and assessment across
multiple cognitive levels. It includes context-based tutoring, structured viva examination, and experiment completion verification, mimicking the role of a human lab instructor. The system was evaluated
through user studies that benchmarked candidate language models and quantified the impact of staged
LLM-assisted guidance, showing clear improvements in learner performance across all six levels of
Bloom's Taxonomy.
These contributions show a marked advancement in the development of remote laboratory systems
by bringing together robust experimentation via IoT and intelligent tutoring. The contributions can be
seen as a step towards the development of more holistic and capable remote laboratory systems that can
effectively mimic in-person learning while being accessible to a wide audience.

Abstract

Speech prominence refers to the relative emphasis placed on certain units in speech to express the
communicative intent. This prominence is realized at two different levels. One, at the syllable level,
where one syllable within a word is made more prominent than the rest, referred to as lexical or syllablelevel prominence. Lexical prominence is generally produced correctly by native speakers and is commonly incorporated into speech synthesis through pronunciation lexicons, but non-native speakers frequently misplace it due to the influence of their native language, motivating the need for automatic
lexical stress detection. Two, at the word level, where certain words within an utterance are emphasized more than others, commonly referred to as sentence or word-level prominence. Unlike lexical
prominence, automatic detection and natural generation of sentence prominence remain challenging
problems for both native and non-native speech. Despite their importance, existing prominence modeling approaches continue to face several limitations. First, most methods rely on handcrafted prosodic
features, such as pitch, energy, and duration, which may not fully capture prominence-related information. Second, prominence modeling at both syllable and word levels typically depends on accurate
boundary annotations, obtained either through costly manual labeling or error-prone forced alignment.
Third, training these systems requires expensive prominence labels, limiting scalability to low resource
language backgrounds.
In this thesis, these limitations are addressed progressively, moving from supervised to unsupervised learning and from boundary-dependent to boundary-independent modeling across both linguistic
levels. At the syllable level, it is first demonstrated that self-supervised speech representations provide a substantially stronger foundation for prominence modeling than handcrafted features, with sequential modeling further improving performance by capturing inter-syllable dependencies. The dependence on boundary annotations is then removed through hierarchical temporal compression, introducing
boundary-independent frameworks that match and often exceed boundary-dependent counterparts. The
reliance on prominence labels is further addressed through Post-Net and Post-Net2.0, which leverage
pseudo-label generation and linguistic constraints to enable effective unsupervised prominence detection. A fully unsupervised and boundary-independent framework, SRAMA, is proposed, combining
Adaptive Local Monotonic Attention (AMA) with iterative deep clustering, approaching supervised
performance without requiring boundaries, transcriptions, or prominence labels. Finally, the scope is
extended to the word level through ProTTS, which integrates prominence discovery directly into SOTA
TTS, namely, FastSpeech 2, using word-level prominence inferred from prosodic predictions to guide
synthesis -- producing more natural and expressive speech and bridging the two central applications of
this thesis: automatic speech prominence detection and expressive speech synthesis.

Abstract

Despite negation being one of the core elements in any language, it remains a challenging
phenomenon for modern Large Language Models (LLMs). Recently, there have been growing efforts to evaluate how models handle negation. However, the existing probing datasets
are mostly English-centric, leaving morphologically rich languages largely unevaluated. To facilitate evaluation for Indian languages, especially Telugu, which expresses negation not only
through separate words but also through morphemes attached directly to words, we present
the NEGTEL benchmark. As part of this work, we first compile and verify an inventory
of Telugu negative morphemes and introduce NEGSEED, a manually curated set of example
sentences covering the full range of morphological negation patterns in Telugu. Building on
this resource, the NEGTEL benchmark is a test suite that contains five tasks: Negation Detection, Translation, Paraphrase Detection, Sentiment Classification, and Polarity Flipping. These
tasks are designed to evaluate models at increasing levels of linguistic and logical complexity,
starting from whether models can detect the presence of negation, to whether they can preserve
it during translation, distinguish negated sentence pairs, correctly classify sentiment in the presence of negation, and actively manipulate negation by flipping the polarity of a sentence. The
test suite is designed grounded in detailed linguistic analysis and includes annotations of different negation types. We evaluate a diverse set of multilingual LLMs on this benchmark, and
our evaluation reveals that most models struggle significantly with Telugu negation across all
tasks. A recurring finding is that standard automated metrics remain high even when models
fail semantically, highlighting the need for human evaluation when assessing negation-related
capabilities.

Abstract

In multi-armed bandit settings with Markovian rewards, the Gittins index policy is well known to
maximize the expected discounted return and is therefore the optimal arm-selection strategy. In practical
applications, however, the transition dynamics of the underlying Markov chains are seldom available,
making it impossible to compute the Gittins indices analytically. This motivates the use of reinforcement
learning (RL) techniques that balance exploration and exploitation to infer these indices directly from
data.
In this work, we introduce two learning-based approaches for estimating the Gittins index: a tabular method, QGI, and a deep reinforcement learning counterpart, DGN. Both algorithms are built upon
the retirement-formulation perspective of the multi-armed bandit problem. Compared to existing RL
approaches for Gittins index estimation, our methods offer substantially lower computational overhead,
require less memory due to a smaller Q-table in QGI and a reduced replay buffer in DGN, and exhibit faster and more stable empirical convergence to the true index values. These properties make the
proposed algorithms particularly effective for large state spaces and attractive alternatives to current
techniques. As an illustrative application, we show how these methods can be used to learn the optimal
scheduling policy that minimizes mean flow time in a batch-arrival job scheduling system with unknown
service-time distributions.

Abstract

Music's ability to communicate emotions is universally acknowledged, yet the relative contribution
of its constituent components, namely instrumentals, vocals, and lyrics, to perceived emotion remains
insufficiently understood. This thesis systematically investigates this question through a series of computational and perceptual experiments within the continuous valence-arousal (VA) emotion space.
We first establish that a substantial discrepancy exists between lyrics-based emotion annotations and
acoustically derived emotion estimates. Across three datasets spanning English, Hindi, and Telugu, we
find that lyrics-based labels frequently disagree with the emotion conveyed by the music itself, and that
this mismatch is not confined to a single language or cultural context.
To determine where the emotional signal resides, we construct a unified corpus of over 1,500 Englishlanguage music excerpts with continuous human-annotated VA values. Using source separation, we
isolate instrumental and vocal audio stems; lyrics are transcribed and human-verified. Deep learning
models are employed to predict VA values for each component independently. The results reveal a
consistent hierarchy: instrumental features exhibit the strongest alignment with human emotional annotations, followed by vocals, with lyrics trailing substantially. A perceptual validation study with human
participants confirms that these predictions are perceptually grounded.
Quadrant-wise analysis reveals a notable exception: in the low-valence, low-arousal region of the
VA space, all three modalities converge, with lyrics achieving their highest alignment with perceived
emotion. This pattern is unique to this region, suggesting that lyrics play a particularly important role
in conveying sadness-related emotions. Cross-corpus validation against independently derived emotion
sources, including Spotify audio features and crowd-sourced social tags, confirms that both the modality
hierarchy and this convergence pattern generalize beyond the primary corpus.
These findings carry implications for both dataset design and system architecture in music emotion recognition. Datasets annotated from lyrics alone measure textual sentiment rather than perceived
musical emotion, and current multimodal approaches that uniformly weight lyrics are suboptimal. Our
results motivate adaptive, context-sensitive fusion frameworks in which lyrical information is selectively
leveraged where it meaningfully enhances affective interpretation.

Abstract

The growth of digital information has created two related problems for natural language
processing. The first is how to accurately answer questions using large, structured knowledge
bases. The second is how to automatically check the truthfulness of claims made in plain text.
This thesis presents two new frameworks that offer practical solutions to these problems. The
first framework improves Question Answering over Knowledge Graphs (KGQA) by focusing on
the quality of the graph representation. The second framework introduces a complete system
for fact verification that works across multiple languages and at a more precise, fact-level of
detail.
The first part of this thesis focuses on KGQA systems that use a combination of pre-trained
Language Models (LMs) and Graph Neural Networks (GNNs). The performance of these systems depends on a key step: creating a small, relevant "Working Graph" (WG) from a much
larger Knowledge Graph. We find that this WG often has two main problems: it can be incomplete, leaving out key information from the user's question, and it is often noisy, containing
many irrelevant nodes that confuse the GNN model.
To solve this, we propose GrapeQA, a framework that directly improves the WG before
the main reasoning step. As detailed in Chapter 3, GrapeQA uses two main components:
(1) Prominent Entities for Graph Augmentation (PEGA), which adds new nodes to
the graph to represent important noun phrases from the question, and (2) QA ContextAware Node Pruning (CANP), which removes clusters of nodes that are not relevant to
the question's context. We tested GrapeQA on standard benchmarks, including OpenBookQA
and CommonsenseQA. The results show that our method provides significant and consistent
improvements over strong baseline models, proving that fixing the working graph is an effective
way to get better answers.
The second part of this thesis deals with automated fact verification. As discussed in Chapter
4, existing systems have two major limitations that limit their practical use: (1) they are mostly
monolingual and built for English, which does not work for global information platforms; and
(2) they operate at the sentence level, which is not detailed enough for complex sentences
that contain several individual facts.
To address these limitations, we developed XFactVer, a new dataset and an end-to-end
pipeline for cross-lingual, fact-level verification. First, we created and released a new, manually
vii
viii
labeled dataset covering English and five Indian languages to serve as a benchmark for this task.
Second, we built a three-stage pipeline that (1) uses a multilingual sequence-to-sequence model
(mT5) to extract individual facts from a sentence, (2) retrieves the best evidence sentence
from reference documents using a cross-lingual embedding model (SBERT), and (3) uses a
cross-lingual classifier (XLM-R) to label the fact as "Supported" or "Citation Needed." Our
experiments show that the pipeline achieves strong and consistent performance across all tested
languages, confirming that our approach is effective for building more detailed and globally
useful fact-checking systems.
In conclusion, this thesis contributes two complete and effective frameworks that advance
how computer systems process and validate information. By introducing a better method for
building graph representations for KGQA and a more precise, multilingual pipeline for fact
verification, this work helps in the development of more accurate and reliable information
systems.

Abstract

Recent advances in large language models (LLMs), vision-language models (VLMs), and text-tovideo (T2V) systems have sparked growing excitement about the future of education. Imagine a learning experience where complex ideas can be explained through synchronized text, visuals, narration,
and video; where educational content can be generated on demand; and where learning becomes more
interactive, accessible, and personalized. While these possibilities are increasingly within reach, an important question remains: do current multimodal AI systems truly understand and connect information
across different modalities in ways that meaningfully support learning? This thesis explores that question through the development of benchmarks and evaluation frameworks for multimodal educational
AI.
A familiar challenge for learners is following presentation slides while simultaneously listening to a
speaker. Slides are often dense, explanations move quickly, and learners can easily lose track of where
to focus their attention. To address this, we explore a method that automatically highlights slide regions
relevant to the ongoing narration, helping guide learners toward the most important visual information
in real time. Supporting this effort, we introduce a dataset designed to evaluate how effectively models
can align spoken language with slide content. The dataset consists of 14 presentation videos and 150
slides, where each slide is paired with its corresponding audio segment and transcript. It captures a wide
range of visual and textual elements--including figures, equations, tables, and diverse layouts--making
it a challenging and realistic benchmark for multimodal alignment.
We then turn to physics education, where visual intuition is often essential for understanding abstract
concepts, yet creating high-quality educational animations remains time-consuming and labor-intensive.
Motivated by recent advances in T2V generation, we investigate whether these models can create videos
that are not only visually realistic, but also pedagogically meaningful. To this end, we introduce PhyEduVideo, a benchmark designed to evaluate the ability of T2V models to generate simple, accurate, and
educational visualizations for physics concepts. Each concept is decomposed into fine-grained teaching
points and paired with carefully crafted prompts intended to elicit clear visual explanations. Beyond
visual quality and motion smoothness, we evaluate whether the generated videos actually convey the
underlying physics correctly. Our findings reveal a striking gap: while current models often produce
visually appealing outputs, their conceptual understanding of physics remains inconsistent.
Finally, we address a broader challenge in scientific communication: knowledge is often scattered
across research papers, slides, presentation videos, and explanatory videos. Although these modalities
frequently convey the same core ideas, they provide complementary perspectives and rarely contain exvi
vii
plicit links between one another. To bridge this gap, we introduce the Multimodal Conference Dataset
(MCD), the first benchmark to integrate all of these modalities from the same set of research works. Using MCD, we evaluate both embedding-based approaches and vision-language models on their ability
to discover fine-grained correspondences across formats. Our experiments show that while VLMs are
generally robust, they still struggle with precise alignment, whereas embedding-based methods effectively capture text-visual relationships but tend to isolate equations and symbolic content into separate
regions of the embedding space. These findings expose important limitations in current multimodal
systems while also highlighting promising directions for future research in multimodal scientific understanding.
Together, these contributions move toward educational AI systems that are not only visually impressive, but also accurate, aligned, and pedagogically meaningful--ultimately enabling richer, more
connected, and more effective learning experiences.

Abstract

Breast cancer is among the most prevalent and clinically heterogeneous malignancies
worldwide, and its characterization at multiple biological scales such as radiological
phenotype and genomic architecture represents one of the defining computational challenges
at the interface of information science and biomedicine. Despite significant algorithmic
progress in both medical image analysis and genomic data processing, reliable computational
characterization of breast cancer is constrained by a common and underappreciated
bottleneck, which is the absence of rigorous, standardized benchmarking infrastructure that
enables reproducible evaluation and fair comparison of methods. This thesis addresses two
distinct but complementary computational problems in breast cancer research: (i) the
development and evaluation of benchmark datasets and deep learning frameworks for
mammographic image analysis, and (ii) the systematic benchmarking of long-read sequencing
tools for structural variant detection in germline and somatic cancer contexts. Both problems
are unified by the same underlying bottleneck, which is the absence of rigorous evaluation
infrastructure. The contributions presented here address this bottleneck at both scales. The
first contribution is Mammo-Bench, a large-scale unified mammography benchmark dataset
comprising 19,731 images from 6,500 patients across six geographically diverse publicly
available resources, standardized through a comprehensive preprocessing pipeline and
annotated for five clinical tasks including cancer status, breast density, BI-RADS score,
abnormality type, and molecular subtype. Mammo-Bench is, to the best of our knowledge, the
largest open-access multi-task mammography benchmark currently available and establishes
a reproducible evaluation standard for mammographic computer-aided detection and
diagnosis systems. The second contribution is MammoInsight, a unified multi-task deep
learning framework for mammographic analysis, trained on an extended benchmark of 65,602
images. Built on the MedImageInsight foundation model backbone, MammoInsight
simultaneously performs five clinical classification tasks and lesion segmentation in a single
forward pass, incorporating SAM-guided region-of-interest pooling, cross-view attention
fusion, and ordinal regression losses for clinically ordered tasks. Evaluated against four large
foundation model baselines, MammoInsight achieves state-of-the-art performance on cancer
status prediction, breast density classification, and BI-RADS scoring, and attains a
segmentation Dice score of 0.804, substantially outperforming a fine-tuned SAM-Med2D
baseline (0.734) despite using only 367 million parameters compared to baselines up to 74
times larger. The third contribution is a systematic multi-context evaluation of long-read
structural variant (SV) callers across three experimental paradigms, which include simulated
chromosome 17 data at four coverage levels, germline benchmarking against the GIAB
HG002 Tier 1 truth set, and paired tumor-normal somatic analysis of HER2+ (HCC1954) and
triple-negative breast cancer (HCC1937) cell lines on both ONT PromethION R10 and PacBio
Revio platforms. Six germline callers and four dedicated somatic callers are evaluated using
a novel three-tier benchmarking framework that accounts for systematic differences in VCF
output representation across tools. Key findings include caller-specific coverage thresholds
for reliable SV detection (F1 > 0.85 achievable at 5x for Sniffles2, Kled, and SVDF), nearceiling germline performance at high coverage, and a consistent, substantial performance gap
between dedicated somatic callers and germline callers adapted through normal subtraction in
the cancer context. The thesis establishes benchmarking infrastructure and validates
computational frameworks across the imaging and genomic modalities of breast cancer
characterization and provides guidance for tool selection and evaluation design in both
research and clinical genomics applications.