Boosting is a widely used machine learning approach based on the idea of aggregating weak learning rules.

This document contains additional details, in-depth discussions, and more results that did not fit into the main paper.

With advances in generative AI, decision-making agents can now dynamically create new actions during online learning, but action generation typically incurs costs that must be balanced against potential benefits. We study an online learning problem where an agent can generate new actions at any time step by paying a one-time cost, with these actions becoming permanently available for future use. The challenge lies in learning the optimal sequence of two-fold decisions: which action to take and when to generate new ones, further complicated by the triangular tradeoffs among exploitation, exploration and $\textit{creation}$. To solve this problem, we propose a doubly-optimistic algorithm that employs Lower Confidence Bounds (LCB) for action selection and Upper Confidence Bounds (UCB) for action generation. Empirical evaluation on healthcare question-answering datasets demonstrates that our approach achieves favorable generation-quality tradeoffs compared to baseline strategies. From theoretical perspectives, we prove that our algorithm achieves the optimal regret of $O(T^{\frac{d}{d+2}}d^{\frac{d}{d+2}} + d\sqrt{T\log T})$, providing the first sublinear regret bound for online learning with expanding action spaces.

Recent text-only models demonstrate remarkable mathematical reasoning capabilities. Extending these to visual domains requires vision-language models to translate images into text descriptions. However, current models, trained to produce captions for human readers, often omit the precise details that reasoning systems require. This creates an interface mismatch: reasoners often fail not due to reasoning limitations but because they lack access to critical visual information. We propose Adaptive-Clarification Reinforcement Learning (AC-RL), which teaches vision models what information reasoners need through interaction. Our key insight is that clarification requests during training reveal information gaps; by penalizing success that requires clarification, we create pressure for comprehensive initial captions that enable the reasoner to solve the problem in a single pass. AC-RL improves average accuracy by 4.4 points over pretrained baselines across seven visual mathematical reasoning benchmarks, and analysis shows it would cut clarification requests by up to 39% if those were allowed. By treating clarification as a form of implicit supervision, AC-RL demonstrates that vision-language interfaces can be effectively learned through interaction alone, without requiring explicit annotations. Recent advances in reinforcement learning have produced text-based reasoning models with remarkable mathematical capabilities (Guo et al., 2025a; Shao et al., 2024). While these reasoning capabilities are impressive, extending them to visual domains requires careful consideration of how visual and linguistic information should interface.

Existing Reward Models (RMs), typically trained on general preference data, struggle in Retrieval Augmented Generation (RAG) settings, which require judging responses for faithfulness to retrieved context, relevance to the user query, appropriate refusals when context is insufficient, completeness and conciseness of information. To address the lack of publicly available RAG-centric preference datasets and specialised RMs, we introduce RAGferee, a methodology that repurposes question-answering (QA) datasets into preference pairs that prioritise groundedness over stylistic features, enabling the training of contextual RMs better suited to judging RAG responses. Using RAGferee, we curate a small preference dataset of 4K samples and fine-tune RMs ranging from 7B to 24B parameters. Our RAG-centric RMs achieve state-of-the-art performance on ContextualJudgeBench, surpassing existing 70B+ RMs trained on much larger (up to 2.4M samples) general corpora, with an absolute improvement of +15.5%.

Sparsity-aware training is an effective approach for transforming large language models (LLMs) into hardware-friendly sparse patterns, thereby reducing latency and memory consumption during inference. In this paper, we propose Continuous Adaptive Sparse Trainer (CAST), a fully continuous and differentiable sparsity-aware training framework for semi-structured (or "N:M") sparse models. Unlike previous approaches that optimize sparsity patterns and weights separately, CAST enables seamless joint optimization during training, while progressively transforming the model into the desired sparsity format. Specifically, CAST introduces three key components: 1) AdamS, a sparsity-aware optimizer that leverages adaptive L1 decay to promote uniform sparsification across all parameters; 2) Weight Scaling, a module designed to mitigate the magnitude reduction caused by decay while preserving desired sparsity patterns; 3) Knowledge Distillation, which employs the dense model as a self-teacher to enhance training efficiency. We evaluate CAST under 2:4 sparsity patterns across multiple model families, ranging from 125M to 13B parameters. Our results demonstrate significant improvements over previous state-of-the-art methods in both perplexity and zero-shot accuracy with minimal training resources. Notably, on LLaMA2-7B, our 2:4 sparse model achieves a negligible perplexity increase of 0.09 and a 0.36% gain in zero-shot accuracy compared to the dense model using only 2% of the original pretraining tokens. Additionally, we establish an accurate and robust empirical scaling law to predict sparse model performance given adequate training resources. Finally, we demonstrate the practical applicability of our sparse models by evaluating them under quantization and fine-tuning scenarios.

Reasoning quality in large language models depends not only on producing correct answers but also on generating valid intermediate steps. We study this through multiple-choice question answering (MCQA), which provides a controlled setting with fixed answer options. Our analysis shows that when questions are effectively unsolvable for a model, spurious chains of thought (CoTs) are more likely to appear, leading to false positives. By estimating the solvability of each question, we uncover an intermediate regime where learning is most effective. Building on this insight, we adapt outcome-supervised reward models and reinforcement learning with group-relative advantage to incorporate solvability into their objectives. Across experiments on math and multimodal datasets, these modifications consistently yield higher rates of process-correct reasoning and, in reinforcement learning, improved answer accuracy as well. Our results highlight solvability as a key factor for reducing hallucinations and increasing reliability in CoT reasoning. In many applications of CoT reasoning, the generated thought process is as important as the final answer. While some tasks provide gold-standard reasoning chains that can effectively be used for supervised training (Nye et al., 2021; Dziri et al., 2023; Hochlehnert et al., 2025), most datasets lack such annotations. For these cases, correct reasoning has to be incentivized by rewards on correct final answers (Wen et al., 2025). It is known that CoTs can lead to the correct answer, despite an incorrect explanation. Grattafiori et al. (2024) note that this often occurs for questions where only a small fraction of the generated answers is correct. In this work, we investigate this observation in controlled experiments on multiple datasets. To avoid confounding factors of noisy answer extraction and matching, we focus on multiple-choice question answering. This format is popular for evaluating models and widely used training sets like NuminaMath (LI et al., 2024) contain a large fraction of multiple-choice questions. The fixed number of answer options also allows us to explicitly model the solvability of a question.

Large language models (LLMs) deliver remarkable performance but are costly to deploy, motivating knowledge distillation (KD) for efficient inference. Existing KD objectives typically match student and teacher probabilities via softmax, which blurs valuable logit information. While direct logit distillation (DLD) mitigates softmax smoothing, it fails to account for logit shift invariance, thereby restricting the solution space. We propose Concrete Score Distillation (CSD), a discrete score-matching objective that overcomes both softmax-induced smoothing and restrictions on the optimal solution set. We resolve the training instability and quadratic complexity of discrete score-matching in autoregressive LLMs, and the resulting CSD objective aligns relative logit differences across all vocabulary pairs between student and teacher with flexible weighting. We provide both mode-seeking and mode-covering instances within our framework and evaluate CSD on task-agnostic instruction-following and task-specific distillation using GPT-2-1.5B, OpenLLaMA-7B, and GEMMA-7B-IT. Experiments show that CSD consistently surpasses recent KD objectives, achieves favorable fidelity-diversity trade-offs, and yields complementary gains when combined with on-policy techniques, demonstrating its scalability and effectiveness for LLM distillation.

Proximal Policy Optimization (PPO)-based reinforcement learning from human feedback (RLHF) is a widely adopted paradigm for aligning large language models (LLMs) with human preferences. However, its training pipeline suffers from substantial inefficiencies due to sequential multi-model dependencies (e.g., reward model depends on actor outputs) and long-tail response lengths, where a few long responses straggle the stage completion. We present OPPO, a novel, lightweight, and model-agnostic PPO-based RLHF framework that improves training efficiency by overlapping pipeline execution. OPPO introduces two novel techniques: (1) Intra-step overlap, which streams upstream model outputs (e.g., actor model) in right-sized chunks, enabling the downstream model (e.g., reward) to begin prefill while the upstream continues decoding; and (2) Inter-step overlap, which adaptively overcommits a few prompts and defers long generations to future steps, mitigating tail latency without discarding partial work. OPPO integrates easily with existing PPO implementations with a few lines of code change. Extensive evaluations show that OPPO accelerates PPO-based RLHF training by $1.8 \times-2.8 \times$ and improves GPU utilization by $1.4 \times-2.1 \times$ without compromising training convergence.

While large language models (LLMs) have demonstrated strong performance on factoid question answering, they are still prone to hallucination and untruthful responses, particularly when tasks demand information outside their parametric knowledge. Indeed, truthfulness requires more than accuracy -- models must also recognize uncertainty and abstain when unsure to avoid hallucinations. This presents a fundamental challenge for existing methods: approaches that optimize for accuracy often amplify hallucinations, while those that encourage abstention can become overly conservative, sacrificing correct answers. Both extremes ultimately compromise truthfulness. In this work, we present TruthRL, a general reinforcement learning (RL) framework that directly optimizes the truthfulness of LLMs. Specifically, we implement TruthRL using GRPO with a simple yet effective ternary reward that distinguishes correct answers, hallucinations, and abstentions. It incentivizes models to reduce hallucinations not only by providing correct responses, but also by enabling abstention when uncertain, thereby improving truthfulness. Extensive experiments across four knowledge-intensive benchmarks show that, compared to vanilla RL, TruthRL significantly reduces hallucinations by 28.9% and improves truthfulness by 21.1%, with consistent gains across various backbone models (e.g., Qwen, Llama) under both retrieval and non-retrieval setups. In-depth ablation study demonstrates that vanilla accuracy-driven methods, such as supervised fine-tuning or RL with a binary reward, struggle to balance factual correctness and uncertainty. In contrast, our proposed truthfulness-driven TruthRL achieves strong performance in both accuracy and truthfulness, underscoring the importance of learning objective design for developing truthful LLMs.