Constrained Stochastic Shortest Path Problems (CSSPs) model problems with probabilistic effects, where a primary cost is min-imised subject to constraints over secondary costs, e.g., minimise time subject to monetary budget. Current heuristic search algorithms for CSSPs solve a sequence of increasingly larger CSSPs as linear programs until an optimal solution for the original CSSP is found. In this paper, we introduce a novel algorithm CARL, which solves a series of unconstrained Stochastic Shortest Path Problems (SSPs) with efficient heuristic search algorithms. These SSP subproblems are constructed with scalarisations that project the CSSP's vector of primary and secondary costs onto a scalar cost. CARL finds a maximising scalarisation using an optimisation algorithm similar to the subgradient method which, together with the solution to its associated SSP, yields a set of policies that are combined into an optimal policy for the CSSP . Our experiments show that CARL solves 50% more problems than the state-of-the-art on existing benchmarks.

Large Reasoning Models (LRMs) demonstrate remarkable capabilities on complex tasks, exhibiting emergent, human-like thinking patterns. Despite their advances, we identify a fundamental limitation: current LRMs lack a dedicated meta-level cognitive system-an essential faculty in human cognition that enables "thinking about thinking". This absence leaves their emergent abilities uncontrollable (non-adaptive reasoning), unreliable (intermediate error), and inflexible (lack of a clear methodology). To address this gap, we introduce Meta-R1, a systematic and generic framework that endows LRMs with explicit metacognitive capabilities. Drawing on principles from cognitive science, Meta-R1 decomposes the reasoning process into distinct object-level and meta-level components, orchestrating proactive planning, online regulation, and adaptive early stopping within a cascaded framework. Experiments on three challenging benchmarks and against eight competitive baselines demonstrate that Meta-R1 is: (I) high-performing, surpassing state-of-the-art methods by up to 27.3%; (II) token-efficient, reducing token consumption to 15.7% ~ 32.7% and improving efficiency by up to 14.8% when compared to its vanilla counterparts; and (III) transferable, maintaining robust performance across datasets and model backbones.

This thesis develops computational methods in similarity-preserving hashing, classification, and cancer genomics. Standard space partitioning-based hashing relies on Binary Search Trees (BSTs), but their exponential growth and sparsity hinder efficiency. To overcome this, we introduce Compressed BST of Inverted hash tables (ComBI), which enables fast approximate nearest-neighbor search with reduced memory. On datasets of up to one billion samples, ComBI achieves 0.90 precision with 4X-296X speed-ups over Multi-Index Hashing, and also outperforms Cellfishing.jl on single-cell RNA-seq searches with 2X-13X gains. Building on hashing structures, we propose Guided Random Forest (GRAF), a tree-based ensemble classifier that integrates global and local partitioning, bridging decision trees and boosting while reducing generalization error. Across 115 datasets, GRAF delivers competitive or superior accuracy, and its unsupervised variant (uGRAF) supports guided hashing and importance sampling. We show that GRAF and ComBI can be used to estimate per-sample classifiability, which enables scalable prediction of cancer patient survival. To address challenges in interpreting mutations, we introduce Continuous Representation of Codon Switches (CRCS), a deep learning framework that embeds genetic changes into numerical vectors. CRCS allows identification of somatic mutations without matched normals, discovery of driver genes, and scoring of tumor mutations, with survival prediction validated in bladder, liver, and brain cancers. Together, these methods provide efficient, scalable, and interpretable tools for large-scale data analysis and biomedical applications.

We present a novel goal-conditioned recurrent state space (GC-RSSM) model capable of learning latent dynamics of pick-and-place garment manipulation. Our proposed method LaGarNet matches the state-of-the-art performance of mesh-based methods, marking the first successful application of state-space models on complex garments. LaGarNet trains on a coverage-alignment reward and a dataset collected through a general procedure supported by a random policy and a diffusion policy learned from few human demonstrations; it substantially reduces the inductive biases introduced in the previous similar methods. We demonstrate that a single-policy LaGarNet achieves flattening on four different types of garments in both real-world and simulation settings.

Advances in large language models (LLMs) significantly enhance reasoning capabilities but their deployment is restricted in resource-constrained scenarios. Knowledge distillation addresses this by transferring knowledge from powerful teacher models to compact and transparent students. However, effectively capturing the teacher's comprehensive reasoning is challenging due to conventional token-level supervision's limited scope. Using multiple reasoning paths per query alleviates this problem, but treating each path identically is suboptimal as paths vary widely in quality and suitability across tasks and models. We propose Quality-filtered Routing with Cooperative Distillation (QR-Distill), combining path quality filtering, conditional routing, and cooperative peer teaching. First, quality filtering retains only correct reasoning paths scored by an LLM-based evaluation. Second, conditional routing dynamically assigns paths tailored to each student's current learning state. Finally, cooperative peer teaching enables students to mutually distill diverse insights, addressing knowledge gaps and biases toward specific reasoning styles. Experiments demonstrate QR-Distill's superiority over traditional single- and multi-path distillation methods. Ablation studies further highlight the importance of each component including quality filtering, conditional routing, and peer teaching in effective knowledge transfer. Our code is available at https://github.com/LzyFischer/Distill.

Data visualizations like charts are fundamental tools for quantitative analysis and decision-making across fields, requiring accurate interpretation and mathematical reasoning. The emergence of Multimodal Large Language Models (MLLMs) offers promising capabilities for automated visual data analysis, such as processing charts, answering questions, and generating summaries. However, they provide no visibility into which parts of the visual data informed their conclusions; this black-box nature poses significant challenges to real-world trust and adoption. In this paper, we take the first major step towards evaluating and enhancing the capabilities of MLLMs to attribute their reasoning process by highlighting the specific regions in charts and graphs that justify model answers. To this end, we contribute RADAR, a semi-automatic approach to obtain a benchmark dataset comprising 17,819 diverse samples with charts, questions, reasoning steps, and attribution annotations. We also introduce a method that provides attribution for chart-based mathematical reasoning. Experimental results demonstrate that our reasoning-guided approach improves attribution accuracy by 15% compared to baseline methods, and enhanced attribution capabilities translate to stronger answer generation, achieving an average BERTScore of $\sim$ 0.90, indicating high alignment with ground truth responses. This advancement represents a significant step toward more interpretable and trustworthy chart analysis systems, enabling users to verify and understand model decisions through reasoning and attribution.

Reasoning is a core capability of large language models, yet understanding how they learn and perform multi-step reasoning remains an open problem. In this study, we explore how different architectures and training methods affect model multi-step reasoning capabilities within a cellular automata framework. By training on state sequences generated with random Boolean functions for random initial conditions to exclude memorization, we demonstrate that most neural architectures learn to abstract the underlying rules. While models achieve high accuracy in next-state prediction, their performance declines sharply if multi-step reasoning is required. We confirm that increasing model depth plays a crucial role for sequential computations. We demonstrate that an extension of the effective model depth with recurrence, memory, and test-time compute scaling substantially enhances reasoning capabilities.

Traditional neural embeddings represent concepts as points, excelling at similarity but struggling with higher-level reasoning and asymmetric relationships. We introduce a novel paradigm: embedding concepts as linear subspaces. This framework inherently models generality via subspace dimensionality and hierarchy through subspace inclusion. It naturally supports set-theoretic operations like intersection (conjunction), linear sum (disjunction) and orthogonal complements (negations), aligning with classical formal semantics. To enable differentiable learning, we propose a smooth relaxation of orthogonal projection operators, allowing for the learning of both subspace orientation and dimension. Our method achieves state-of-the-art results in reconstruction and link prediction on WordNet. Furthermore, on natural language inference benchmarks, our subspace embeddings surpass bi-encoder baselines, offering an interpretable formulation of entailment that is both geometrically grounded and amenable to logical operations.

Interpreting figurative language such as sarcasm across multi-modal inputs presents unique challenges, often requiring task-specific fine-tuning and extensive reasoning steps. However, current Chain-of-Thought approaches do not efficiently leverage the same cognitive processes that enable humans to identify sarcasm. We present IRONIC, an in-context learning framework that leverages Multi-modal Coherence Relations to analyze referential, analogical and pragmatic image-text linkages. Our experiments show that IRONIC achieves state-of-the-art performance on zero-shot Multi-modal Sarcasm Detection across different baselines. This demonstrates the need for incorporating linguistic and cognitive insights into the design of multi-modal reasoning strategies. Our code is available at: https://github.com/aashish2000/IRONIC

Legal judgment prediction (LJP) has become increasingly important in the legal field. In this paper, we identify that existing large language models (LLMs) have significant problems of insufficient reasoning due to a lack of legal knowledge. Therefore, we introduce GLARE, an agentic legal reasoning framework that dynamically acquires key legal knowledge by invoking different modules, thereby improving the breadth and depth of reasoning. Experiments conducted on the real-world dataset verify the effectiveness of our method. Furthermore, the reasoning chain generated during the analysis process can increase interpretability and provide the possibility for practical applications.