Transformers have demonstrated remarkable performance in natural language processing and related domains, as they largely focus on sequential, autoregressive next-token prediction tasks. An emerging exploration in this direction is the Hierarchical Reasoning Model Wang et al. (2025a), which introduces a novel type of recurrent reasoning in the latent space of transformers, achieving remarkable performance on a wide range of 2D reasoning tasks. Despite the promising results, this line of models is still at an early stage and calls for in-depth investigation. In this work, we review this class of models, examine key design choices, test alternative variants and clarify common misconceptions. The Transformer architecture (V aswani et al., 2017) has become the foundation of modern large language models (LLMs), powering systems such as BERT (Devlin et al., 2019), PaLM (Chowdhery et al., 2022) and GPT series (Brown et al., 2020; Achiam et al., 2023).

Code editing plays a vital role in software engineering, requiring developers to adjust existing code according to natural language instructions while keeping functionality intact and avoiding unnecessary modifications. However, commit-based datasets commonly used for this task are often noisy, lack diversity, and fail to reflect the style of real-world edit instructions. To address this, we introduce OpenCodeEdit, an open-source pipeline that leverages multiple LLMs to synthesize realistic code-edit triplets. The pipeline produces both concise "lazy" instructions and more detailed "descriptive" ones, and applies filtering based on diffs and topics to guarantee data quality and variety. Using this process, we construct OCEDataFT, a curated dataset of 20K samples. Fine-tuning three advanced base models on OCEDataFT leads to significant performance boosts on the CanItEdit benchmark, with relative pass@1 improvements ranging from 4.50% to 20.79%. Notably, the resulting models achieve performance close to closed-source systems, narrowing the gap to GPT-4 to just 3.54%, without relying on proprietary resources or manual annotation.

Process Reward Models (PRMs) provide step-level supervision that improves the reliability of reasoning in large language models. While PRMs have been extensively studied in text-based domains, their extension to Vision Language Models (VLMs) remains limited. Existing Vision-Language PRMs (VL-PRMs) rely on Monte Carlo Tree Search (MCTS) for data construction, which can often produce noisy supervision signals and limit generalization across tasks. In this work, we aim to elucidate the design space of VL-PRMs by exploring diverse strategies for dataset construction, training, and test-time scaling. First, we introduce a hybrid data synthesis framework that combines MCTS with judgments from a strong VLM, producing more accurate step-level labels. Second, we propose perception-focused supervision, enabling our PRM to explicitly detect errors at the visual grounding stage of reasoning. Third, we systematically evaluate multiple test-time scaling strategies, showing that our PRMs can reliably guide VLMs toward more accurate solutions. Our experiments covering five diverse multimodal benchmarks (MMMU, PuzzleVQA, AlgoPuzzleVQA, MathVista, and MathVision) reveal several key insights: (i) VL-PRMs when used as Outcome Reward Models (ORMs) during test-time scaling (TTS) can outperform VL-PRM guided process step selection, (ii) smaller VL-PRMs can match or even surpass larger ones in detecting process errors, (iii) VL-PRMs uncover latent reasoning abilities in stronger VLM backbones, (iv) perception-level supervision leads to significant gains in test-time scaling, and (v) TTS performance of different policies improve on advanced math reasoning datasets despite not training VL-PRMs on such datasets. We hope our work will motivate further research and support the advancement of VLMs.

Large language models (LLMs) are increasingly used for decision-making tasks under uncertainty; however, their risk profiles and how they are influenced by prompting and alignment methods remain underexplored. Existing studies have primarily examined personality prompting or multi-agent interactions, leaving open the question of how post-training influences the risk behavior of LLMs. In this work, we propose a new pipeline for eliciting, steering, and modulating LLMs' risk profiles, drawing on tools from behavioral economics and finance. Using utility-theoretic models, we compare pre-trained, instruction-tuned, and RLHF-aligned LLMs, and find that while instruction-tuned models exhibit behaviors consistent with some standard utility formulations, pre-trained and RLHF-aligned models deviate more from any utility models fitted. We further evaluate modulation strategies, including prompt engineering, in-context learning, and post-training, and show that post-training provides the most stable and effective modulation of risk preference. Our findings provide insights into the risk profiles of different classes and stages of LLMs and demonstrate how post-training modulates these profiles, laying the groundwork for future research on behavioral alignment and risk-aware LLM design.

Post-training compression of large language models (LLMs) largely relies on low-rank weight approximation, which represents each column of a weight matrix in a shared low-dimensional subspace. While this is a computationally efficient strategy, the imposed structural constraint is rigid and can lead to a noticeable model accuracy drop. In this work, we propose CoSpaDi (Compression via Sparse Dictionary Learning), a novel training-free compression framework that replaces low-rank decomposition with a more flexible structured sparse factorization in which each weight matrix is represented with a dense dictionary and a column-sparse coefficient matrix. This formulation enables a union-of-subspaces representation: different columns of the original weight matrix are approximated in distinct subspaces spanned by adaptively selected dictionary atoms, offering greater expressiveness than a single invariant basis. Crucially, CoSpaDi leverages a small calibration dataset to optimize the factorization such that the output activations of compressed projection layers closely match those of the original ones, thereby minimizing functional reconstruction error rather than mere weight approximation. This data-aware strategy preserves better model fidelity without any fine-tuning under reasonable compression ratios. Moreover, the resulting structured sparsity allows efficient sparse-dense matrix multiplication and is compatible with post-training quantization for further memory and latency gains. We evaluate CoSpaDi across multiple Llama and Qwen models under per-layer and per-group settings at 20-50\% compression ratios, demonstrating consistent superiority over state-of-the-art data-aware low-rank methods both in accuracy and perplexity. Our results establish structured sparse dictionary learning as a powerful alternative to conventional low-rank approaches for efficient LLM deployment.

Autonomous cyber-physical systems like robots and self-driving cars could greatly benefit from using formal methods to reason reliably about their control decisions. However, before a problem can be solved it needs to be stated. This requires writing a formal physics model of the cyber-physical system, which is a complex task that traditionally requires human expertise and becomes a bottleneck. This paper experimentally studies whether Large Language Models (LLMs) can automate the formalization process. A 20 problem benchmark suite is designed drawing from undergraduate level physics kinematics problems. In each problem, the LLM is provided with a natural language description of the objects' motion and must produce a model in differential game logic (dGL). The model is (1) syntax checked and iteratively refined based on parser feedback, and (2) semantically evaluated by checking whether symbolically executing the dGL formula recovers the solution to the original physics problem. A success rate of 70% (best over 5 samples) is achieved. We analyze failing cases, identifying directions for future improvement. This provides a first quantitative baseline for LLM-based autoformalization from natural language to a hybrid games logic with continuous dynamics.

This pilot study explores the localisation capabilities of state-of-the-art multilingual AI models when translating figurative language, such as idioms and puns, from English into a diverse range of global languages. It expands on existing LLM translation research and industry benchmarks, which emphasise grammatical accuracy and token-level correctness, by focusing on cultural appropriateness and overall localisation quality - critical factors for real-world applications like marketing and e-commerce. To investigate these challenges, this project evaluated a sample of 87 LLM-generated translations of e-commerce marketing emails across 24 regional dialects of 20 languages. Human reviewers fluent in each target language provided quantitative ratings and qualitative feedback on faithfulness to the original's tone, meaning, and intended audience. Findings suggest that, while leading models generally produce grammatically correct translations, culturally nuanced language remains a clear area for improvement, often requiring substantial human refinement. Notably, even high-resource global languages, despite topping industry benchmark leaderboards, frequently mistranslated figurative expressions and wordplay. This work challenges the assumption that data volume is the most reliable predictor of machine translation quality and introduces cultural appropriateness as a key determinant of multilingual LLM performance - an area currently underexplored in existing academic and industry benchmarks. As a proof of concept, this pilot highlights limitations of current multilingual AI systems for real-world localisation use cases. Results of this pilot support the opportunity for expanded research at greater scale to deliver generalisable insights and inform deployment of reliable machine translation workflows in culturally diverse contexts.

Natural language generation (NLG) tasks are often subject to inherent variability; e.g. predicting the next word given a context has multiple valid responses, evident when asking multiple humans to complete the task. While having language models (LMs) that are aligned pluralistically, so that they are able to reproduce well the inherent diversity in perspectives of an entire population of interest is clearly beneficial, Ilia and Aziz (2024) show that LMs do not reproduce this type of linguistic variability well. They speculate this inability might stem from the lack of consistent training of LMs with data reflecting this type of inherent variability. As such, we investigate whether training LMs on multiple plausible word continuations per context can improve their ability to reproduce human linguistic variability for next-word prediction. We employ fine-tuning techniques for pre-trained and instruction-tuned models; and demonstrate their potential when fine-tuning GPT-2 and Mistral-7B-IT, using Provo Corpus. Our evaluation, which measures divergence among empirically estimated human and model next-word distributions across contexts before and after fine-tuning, shows that our multi-label fine-tuning improves the LMs' ability to reproduce linguistic variability; both for contexts that admit higher and lower variability.

Generative Engine Marketing (GEM) is an emerging ecosystem for monetizing generative engines, such as LLM-based chatbots, by seamlessly integrating relevant advertisements into their responses. At the core of GEM lies the generation and evaluation of ad-injected responses. However, existing benchmarks are not specifically designed for this purpose, which limits future research. To address this gap, we propose GEM-Bench, the first comprehensive benchmark for ad-injected response generation in GEM. GEM-Bench includes three curated datasets covering both chatbot and search scenarios, a metric ontology that captures multiple dimensions of user satisfaction and engagement, and several baseline solutions implemented within an extensible multi-agent framework. Our preliminary results indicate that, while simple prompt-based methods achieve reasonable engagement such as click-through rate, they often reduce user satisfaction. In contrast, approaches that insert ads based on pre-generated ad-free responses help mitigate this issue but introduce additional overhead. These findings highlight the need for future research on designing more effective and efficient solutions for generating ad-injected responses in GEM. The benchmark and all related resources are publicly available at https://gem-bench.org/.

This paper tackles \textbf{open-ended deep research (OEDR)}, a complex challenge where AI agents must synthesize vast web-scale information into insightful reports. Current approaches are plagued by dual-fold limitations: static research pipelines that decouple planning from evidence acquisition and monolithic generation paradigms that include redundant, irrelevant evidence, suffering from hallucination issues and low citation accuracy. To address these challenges, we introduce \textbf{WebWeaver}, a novel dual-agent framework that emulates the human research process. The planner operates in a dynamic cycle, iteratively interleaving evidence acquisition with outline optimization to produce a comprehensive, citation-grounded outline linking to a memory bank of evidence. The writer then executes a hierarchical retrieval and writing process, composing the report section by section. By performing targeted retrieval of only the necessary evidence from the memory bank via citations for each part, it effectively mitigates long-context issues and citation hallucinations. Our framework establishes a new state-of-the-art across major OEDR benchmarks, including DeepResearch Bench, DeepConsult, and DeepResearchGym. These results validate our human-centric, iterative methodology, demonstrating that adaptive planning and focused synthesis are crucial for producing comprehensive, trusted, and well-structured reports.