Offline reinforcement learning (RL) offers a promising framework for training agents using pre-collected datasets without the need for further environment interaction. However, policies trained on offline data often struggle to generalise due to limited exposure to diverse states. The complexity of visual data introduces additional challenges such as noise, distractions, and spurious correlations, which can misguide the policy and increase the risk of overfitting if the training data is not sufficiently diverse. Indeed, this makes it challenging to leverage vision-based offline data in training robust agents that can generalize to unseen environments. To solve this problem, we propose a simple approach generating additional synthetic training data. We propose a two-step process, first augmenting the originally collected offline data to improve zero-shot generalization by introducing diversity, then using a diffusion model to generate additional data in latent space. We test our method across both continuous action spaces (Visual D4RL) and discrete action spaces (Procgen), demonstrating that it significantly improves generalization without requiring any algorithmic changes to existing model-free offline RL methods. We show that our method not only increases the diversity of the training data but also significantly reduces the generalization gap at test time while maintaining computational efficiency. We believe this approach could fuel additional progress in generating synthetic data to train more general agents in the future.

The centralized collection of search interaction logs for training ranking models raises significant privacy concerns. Federated Online Learning to Rank (FOLTR) offers a privacy-preserving alternative by enabling collaborative model training without sharing raw user data. However, benchmarks in FOLTR are largely based on random partitioning of classical learning-to-rank datasets, simulated user clicks, and the assumption of synchronous client participation. This oversimplifies real-world dynamics and undermines the realism of experimental results. We present AOL4FOLTR, a large-scale web search dataset with 2.6 million queries from 10,000 users. Our dataset addresses key limitations of existing benchmarks by including user identifiers, real click data, and query timestamps, enabling realistic user partitioning, behavior modeling, and asynchronous federated learning scenarios.

Speech foundation models (SFMs) are designed to serve as general-purpose representations for a wide range of speech-processing tasks. The last five years have seen an influx of increasingly successful self-supervised and supervised pre-trained models with impressive performance on various downstream tasks. Although the zoo of SFMs continues to grow, our understanding of the knowledge they acquire lags behind. This thesis presents a lightweight analysis framework using statistical tools and training-free tasks to investigate the acoustic and linguistic knowledge encoded in SFM layers. We conduct a comparative study across multiple SFMs and statistical tools. Our study also shows that the analytical insights have concrete implications for downstream task performance. The effectiveness of an SFM is ultimately determined by its performance on speech applications. Yet it remains unclear whether the benefits extend to spoken language understanding (SLU) tasks that require a deeper understanding than widely studied ones, such as speech recognition. The limited exploration of SLU is primarily due to a lack of relevant datasets. To alleviate that, this thesis contributes tasks, specifically spoken named entity recognition (NER) and named entity localization (NEL), to the Spoken Language Understanding Evaluation benchmark. We develop SFM-based approaches for NER and NEL, and find that end-to-end (E2E) models leveraging SFMs can surpass traditional cascaded (speech recognition followed by a text model) approaches. Further, we evaluate E2E SLU models across SFMs and adaptation strategies to assess the impact on task performance. Collectively, this thesis tackles previously unanswered questions about SFMs, providing tools and datasets to further our understanding and to enable the community to make informed design choices for future model development and adoption.

Recent advances in protein large language models, such as ProtTeX, represent both side-chain amino acids and backbone structure as discrete token sequences of residue length. While this design enables unified modeling of multimodal protein information, it suffers from two major limitations: (1) The concatenation of sequence and structure tokens approximately doubles the protein length and breaks the intrinsic residue-level alignment between modalities. (2) Constrained by the training corpus and limited context window, ProtTeX is typically trained on single-protein inputs, rendering it incompatible with in-context learning (ICL) and thus limiting its generalization capability. To address these issues, we propose ProtTeX-CC, a lightweight two-stage compression framework designed to enhance ProtTeX under few-shot settings. We first design a joint embedding compression mechanism that fuses sequence and structure representations at the residue level, effectively reducing the protein input length by half without sacrificing performance. Then we propose a self-compression module that aggregates each full demonstration into the latent space of the last few linguistic tokens, reducing the average demonstration length from 751 tokens to less than 16 tokens. Compared to the original ProtTeX, our self-compression approach achieves a compression ratio of approximately 93.68% in the total prompt length under the 16-shot setting. Without modifying the backbone model, ProtTeX-CC introduces only a small number of additional parameters through PEFT-based tuning in the joint embedding compression stage and a single trainable projection layer in the self-compression stage. Extensive experiments on protein function prediction show that ProtTeX-CC improves performance on the in-domain benchmark by 2%, and generalizes well to the out-of-domain dataset with a performance gain of 11%.

This paper introduces RLNVR (Reinforcement Learning from Non-Verified Rewards), a framework for training language models using noisy, real-world feedback signals without requiring explicit human verification. Traditional RLHF requires expensive, verified reward signals that are impractical in many real-world domains. RLNVR addresses this challenge through baseline normalization and semantic similarity-based reward transfer. We demonstrate RLNVR through Walter, a prototype system that optimizes social media content generation using actual engagement data from Bluesky. Our experimental results show significant improvements in content quality and training stability, with comprehensive evaluation planned for future work. Positioning: We present a practical framework that combines RLNVR with GSPO (Group Sequence Policy Optimization) and an optional UED (Unsupervised Environment Design) curriculum to improve stability and diversity under noisy, implicit rewards. To our knowledge, combining GSPO-style normalization with a UED-style curriculum for LLM content generation from implicit social engagement has not been previously documented in this applied setting; we frame this as an applied integration rather than a new algorithm.

While large language models (LLMs) have demonstrated remarkable performance across diverse tasks, they fundamentally lack self-awareness and frequently exhibit overconfidence, assigning high confidence scores to incorrect predictions. Accurate confidence estimation is therefore critical for enhancing the trustworthiness and reliability of LLM-generated outputs. However, existing approaches suffer from coarse-grained scoring mechanisms that fail to provide fine-grained, continuous confidence estimates throughout the generation process. To address these limitations, we introduce FineCE, a novel confidence estimation method that delivers accurate, fine-grained confidence scores during text generation. Specifically, we first develop a comprehensive pipeline for constructing training data that effectively captures the underlying probabilistic distribution of LLM responses, and then train a model to predict confidence scores for arbitrary text sequences in a supervised manner. Furthermore, we propose a Backward Confidence Integration (BCI) strategy that leverages information from the subsequent text to enhance confidence estimation for the current sequence during inference. We also introduce three strategies for identifying optimal positions to perform confidence estimation within the generation process. Extensive experiments on multiple benchmark datasets demonstrate that FineCE consistently outperforms existing classical confidence estimation methods. Our code and all baselines used in the paper are available on GitHub.

Bongard Problems (BPs) provide a challenging testbed for abstract visual reasoning (AVR), requiring models to identify visual concepts fromjust a few examples and describe them in natural language. Early BP benchmarks featured synthetic black-and-white drawings, which might not fully capture the complexity of real-world scenes. Subsequent BP datasets employed real-world images, albeit the represented concepts are identifiable from high-level image features, reducing the task complexity. Differently, the recently released Bongard-RWR dataset aimed at representing abstract concepts formulated in the original BPs using fine-grained real-world images. Its manual construction, however, limited the dataset size to just $60$ instances, constraining evaluation robustness. In this work, we introduce Bongard-RWR+, a BP dataset composed of $5\,400$ instances that represent original BP abstract concepts using real-world-like images generated via a vision language model (VLM) pipeline. Building on Bongard-RWR, we employ Pixtral-12B to describe manually curated images and generate new descriptions aligned with the underlying concepts, use Flux.1-dev to synthesize images from these descriptions, and manually verify that the generated images faithfully reflect the intended concepts. We evaluate state-of-the-art VLMs across diverse BP formulations, including binary and multiclass classification, as well as textual answer generation. Our findings reveal that while VLMs can recognize coarse-grained visual concepts, they consistently struggle with discerning fine-grained concepts, highlighting limitations in their reasoning capabilities.

The rise of generative AI tools like ChatGPT has significantly reshaped education, sparking debates about their impact on learning outcomes and academic integrity. While prior research highlights opportunities and risks, there remains a lack of quantitative analysis of student behavior when completing assignments. Understanding how these tools influence real-world academic practices, particularly assignment preparation, is a pressing and timely research priority. This study addresses this gap by analyzing survey responses from 388 university students, primarily from Russia, including a subset of international participants. Using the XGBoost algorithm, we modeled predictors of ChatGPT usage in academic assignments. Key predictive factors included learning habits, subject preferences, and student attitudes toward AI. Our binary classifier demonstrated strong predictive performance, achieving 80.1\% test accuracy, with 80.2\% sensitivity and 79.9\% specificity. The multiclass classifier achieved 64.5\% test accuracy, 64.6\% weighted precision, and 64.5\% recall, with similar training scores, indicating potential data scarcity challenges. The study reveals that frequent use of ChatGPT for learning new concepts correlates with potential overreliance, raising concerns about long-term academic independence. These findings suggest that while generative AI can enhance access to knowledge, unchecked reliance may erode critical thinking and originality. We propose discipline-specific guidelines and reimagined assessment strategies to balance innovation with academic rigor. These insights can guide educators and policymakers in ethically and effectively integrating AI into education.

The development of artificial intelligence (AI) agents--autonomous systems capable of perceiving their surroundings, reasoning about possible courses of action, and executing decisions--has evolved significantly in recent decades. Early AI agents, rooted in the symbolic reasoning systems of the 1950s and 1960s, relied on hand-crafted rules and logic-based methods, excelling in constrained domains but struggling with adaptability and uncertainty[1, 2]. The introduction of statistical learning and probabilistic reasoning in the 1980s and 1990s enhanced reliability, while the rise of reinforcement learning (RL) allowed agents to learn policies through trial-and-error interactions [3, 4, 5, 6]. The integration of deep neural networks with RL (DeepRL) led to breakthroughs such as superhuman performance in Atari games and Go [7, 8]. With growing computational power, recent advancements in statistical methods and machine learning, AI agents have incorporated advanced perception, natural language sequence modeling, and cognitive-inspired principles, enabling them to adapt, collaborate, and mirror aspects of human reasoning in dynamic environments [2, 9, 10, 11, 12, 13, 14]. Contemporary AI agents are increasingly deployed in high-stakes, real-world contexts: self-driving cars navigating congested urban environments [15, 16], autonomous laboratories accelerating scientific discovery [17, 18], virtual assistants managing complex user queries [19], and automated trading agents operating in financial markets [20].

Optimizing data mixtures for supervised fine-tuning (SFT) of large language models (LLMs) is critical for developing general-purpose models, yet this area remains underexplored. In this paper, we frame data mixing as an optimization problem and introduce a novel method designed to minimize validation loss. Our approach parametrizes the loss by modeling effective data transferred and leveraging scaling laws for fine-tuning. By experimenting with various small-scale data mixtures, we fit these parameters and derive the optimal weights. We provide both mathematical proofs and empirical results demonstrating that our algorithm achieves excellent overall and individual performance across all domains. Through controlled experiments, we show that models trained with our optimized weights perform on par with those using optimal weights determined via grid search, with per-domain loss only 0.66% higher than the best domain loss from grid search on average. Additionally, we show that reweighting popular SFT datasets using our method improves both validation loss and downstream performance. Finally, we discuss how our method can generalize to guide data selection for domain-specific models and provide insights into SFT.