Any piece of knowledge is usually expressed in one or a handful of natural languages on the web or in any large corpus. Large Language Models (LLMs) act as a bridge by acquiring knowledge from a source language and making it accessible when queried from target languages. Prior research has pointed to a cross-lingual gap, viz., a drop in accuracy when the knowledge is queried in a target language compared to when the query is in the source language. Existing research has rationalized divergence in latent representations in source and target languages as the source of cross-lingual gap. In this work, we take an alternative view and hypothesize that the variance of responses in the target language is the main cause of this gap. We present extensive experimental evidence which support proposed formulation and hypothesis. We then reinforce our hypothesis through multiple inference-time interventions that control the variance and reduce the cross-lingual gap. We demonstrate a simple prompt instruction to reduce the response variance, which improved target accuracy by 20-25% across different models. Large Language Models (LLMs) have revolutionized information access. Central to LLM's mission is to assimilate knowledge universally and make it available generally without any barriers. State-of-art LLMs are multilingual: Gemini supports over 40 languages (Gemini, 2025), GPT -5 supports at least 12 languages (GPT, 2025) (with no official number of supported languages) and open-source models like Gemma-3 support over 100 spoken languages (Gemma, 2025). Because pretraining data cannot contain duplicate information for every language, cross-lingual generalization is a necessary capability for LLMs. However, LLMs are known to have disparity in recalling knowledge across languages (Jiang et al., 2020; Kassner et al., 2021; Qi et al., 2023; Chua et al., 2024a; Goldman et al., 2025). Our objective is to understand the causes of poor transfer of knowledge encoded in parameters across languages. We, therefore, evaluate models on knowledge-intensive tasks in a closed-book QA setting, i.e., without access to such tools as grounding in search. Cross-lingual gaps are quantified through disparity on parallel datasets that alter language-specific surface form of the prompts.

Multimodal retrieval, which seeks to retrieve relevant content across modalities such as text or image, supports applications from AI search to contents production. Despite the success of separate-encoder approaches like CLIP align modality-specific embeddings with contrastive learning, recent multimodal large language models (MLLMs) enable a unified encoder that directly processes composed inputs. While flexible and advanced, we identify that unified encoders trained with conventional contrastive learning are prone to learn modality shortcut, leading to poor robustness under distribution shifts. We propose a modality composition awareness framework to mitigate this issue. Concretely, a preference loss enforces multimodal embeddings to outperform their unimodal counterparts, while a composition regularization objective aligns multimodal embeddings with prototypes composed from its unimodal parts. These objectives explicitly model structural relationships between the composed representation and its unimodal counterparts. Experiments on various benchmarks show gains in out-of-distribution retrieval, highlighting modality composition awareness as a effective principle for robust composed multimodal retrieval when utilizing MLLMs as the unified encoder. Multimodal retrieval, which aims to retrieve semantically relevant contents across multiple modalities such as text, image and audio, is a fundamental task in various information fields. The core ability of multimodal retrieval is to represent multimodal inputs in a shared and comparable embedding space. A prevailing approach to this problem is to adopt unimodal encoders and align the encoded embeddings through contrastive learning (CL).

Subword tokenization methods like Byte Pair Encoding (BPE) are widely used in large language models due to their balance of vocabulary compactness and representational power. However, they suffer from inefficiencies in representing rare words and require large embedding matrices. Character-level models address these issues but introduce performance bottlenecks, particularly in Transformer-based architectures. Recent hierarchical models attempt to merge the benefits of both paradigms by grouping characters into patches, but existing patching strategies either rely on whitespace-limiting applicability to certain languages, or require auxiliary models that introduce new dependencies. In this paper, we propose a dynamic character grouping method that leverages the structure of existing BPE tokenization without requiring additional models. By appending explicit end-of-patch markers to BPE tokens and introducing a second-level BPE compression stage to control patch granularity, our method offers efficient, flexible, and language-agnostic representations. Empirical results demonstrate that our approach matches or exceeds the performance of dynamic entropy- and whitespace-based patching strategies, while maintaining a compact vocabulary.

Reinforcement learning (RL) has been pivotal in enhancing the reasoning capabilities of large language models (LLMs), but it often suffers from limited exploration and entropy collapse, where models exploit a narrow set of solutions, leading to a loss of sampling diversity and subsequently preventing RL from further improving performance. This issue is exacerbated in parallel sampling methods, where multiple outputs are drawn from the same distribution, potentially causing the model to converge to similar solutions. We propose SESA, a novel SEquential SAmpling framework that mitigates this challenge by generating diverse solution sketches sequentially before expanding them into full reasoning paths. This approach ensures broader exploration by conditioning each new output on previous ones, promoting diversity throughout the process and preventing policy collapse. Our experiments on a synthetic task show that sequential sampling consistently outperforms traditional RL methods in terms of path diversity and recovery from collapse. Further evaluations on real-world tasks demonstrate that SESA improves both the exploration of valid strategies and the overall performance of LLMs. On three agent benchmarks, SESA lifts success rates by $+0.25$, $+0.42$, and $+0.07$ absolute over the base model (up to an additional $211\%$ relative improvement over baseline RL), underscoring its exploration advantage. This work introduces a structured approach to exploration, paving the way for more effective and diverse reasoning in RL-trained LLMs. Our code is released at https://github.com/MuLabPKU/sesa.

Large Language Models (LLMs) can generate code, but can they generate fast code? In this paper, we study this question using a dataset of 65 real-world tasks mined from open-source Java programs. We specifically select tasks where developers achieved significant speedups, and employ an automated pipeline to generate patches for these issues using two leading LLMs under four prompt variations. By rigorously benchmarking the results against the baseline and human-authored solutions, we demonstrate that LLM-generated code indeed improves performance over the baseline in most cases. However, patches proposed by human developers outperform LLM fixes by a statistically significant margin, indicating that LLMs often fall short of finding truly optimal solutions. We further find that LLM solutions are semantically identical or similar to the developer optimization idea in approximately two-thirds of cases, whereas they propose a more original idea in the remaining one-third. However, these original ideas only occasionally yield substantial performance gains.

Source code clones pose risks ranging from intellectual property violations to unintended vulnerabilities. Effective and efficient scalable clone detection, especially for diverged clones, remains challenging. Large language models (LLMs) have recently been applied to clone detection tasks. However, the rapid emergence of LLMs raises questions about optimal model selection and potential LLM-ensemble efficacy. This paper addresses the first question by identifying 76 LLMs and filtering them down to suitable candidates for large-scale clone detection. The candidates were evaluated on two public industrial datasets, BigCloneBench, and a commercial large-scale dataset. No uniformly 'best-LLM' emerged, though CodeT5+110M, CuBERT and SPTCode were top-performers. Analysis of LLM-candidates suggested that smaller embedding sizes, smaller tokenizer vocabularies and tailored datasets are advantageous. On commercial large-scale dataset a top-performing CodeT5+110M achieved 39.71\% precision: twice the precision of previously used CodeBERT. To address the second question, this paper explores ensembling of the selected LLMs: effort-effective approach to improving effectiveness. Results suggest the importance of score normalization and favoring ensembling methods like maximum or sum over averaging. Also, findings indicate that ensembling approach can be statistically significant and effective on larger datasets: the best-performing ensemble achieved even higher precision of 46.91\% over individual LLM on the commercial large-scale code.

Mobile agents rely on Large Language Models (LLMs) to plan and execute tasks on smartphone user interfaces (UIs). While cloud-based LLMs achieve high task accuracy, they require uploading the full UI state at every step, exposing unnecessary and often irrelevant information. In contrast, local LLMs avoid UI uploads but suffer from limited capacity, resulting in lower task success rates. We propose $\textbf{CORE}$, a $\textbf{CO}$llaborative framework that combines the strengths of cloud and local LLMs to $\textbf{R}$educe UI $\textbf{E}$xposure, while maintaining task accuracy for mobile agents. CORE comprises three key components: (1) $\textbf{Layout-aware block partitioning}$, which groups semantically related UI elements based on the XML screen hierarchy; (2) $\textbf{Co-planning}$, where local and cloud LLMs collaboratively identify the current sub-task; and (3) $\textbf{Co-decision-making}$, where the local LLM ranks relevant UI blocks, and the cloud LLM selects specific UI elements within the top-ranked block. CORE further introduces a multi-round accumulation mechanism to mitigate local misjudgment or limited context. Experiments across diverse mobile apps and tasks show that CORE reduces UI exposure by up to 55.6% while maintaining task success rates slightly below cloud-only agents, effectively mitigating unnecessary privacy exposure to the cloud. The code is available at https://github.com/Entropy-Fighter/CORE.

Test-time scaling seeks to improve the reasoning performance of large language models (LLMs) by adding computational resources. A prevalent approach within the field is sampling-based test-time scaling methods, which enhance reasoning by generating multiple reasoning paths for a given input during inference. However, despite its practical success, the theoretical foundations remain underexplored. In this paper, we provide the first theoretical framework for analyzing sampling-based test-time scaling methods, grounded in the perspective of confidence estimation. Based on the framework, we analyze two dominant paradigms: self-consistency and perplexity, and reveal key limitations: self-consistency suffers from high estimation error while perplexity exhibits substantial modeling error and possible degradation of the estimation error convergence. To address these limitations, we introduce RPC, a hybrid method that leverages our theoretical insights through two key components: Perplexity Consistency and Reasoning Pruning. Perplexity Consistency combines the strengths of self-consistency and perplexity, boosting the convergence rate of estimation error from linear to exponential while preserving model error. Reasoning Pruning prevents degradation by eliminating low-probability reasoning paths. Both theoretical analysis and empirical results across seven benchmark datasets demonstrate that RPC has a strong potential for reducing reasoning error. Notably, RPC achieves reasoning performance comparable to self-consistency while not only enhancing confidence reliability but also reducing sampling costs by 50%. The code and resources are available at https://wnjxyk.github.io/RPC.

Long-form video reasoning remains a major challenge for Video Large Language Models (Video LLMs), as static uniform frame sampling leads to information dilution and obscures critical evidence. Furthermore, existing pixel-space video reasoning agents, which are designed to actively interact with the video to acquire new visual information, remain suboptimal due to their lack of rigorous reward mechanisms to enforce evidence purity and their inability to perform temporal information supplementation beyond pre-sampled frames. T o address this critical gap, we propose a novel evidence-prioritized adaptive framework built upon our core philosophy: "Select Less, Reason More. " Our core contribution is the evidence-aware reinforcement learning (EARL) framework, which transforms the model into an active interrogator of evidence. EARL is precisely engineered to dynamically select the most relevant frames and, crucially, to perform localized re-sampling around the selected key frames to access fine-grained temporal detail. Extensive experiments on five demanding video reasoning benchmarks demonstrate that our EARLtrained model achieves new state-of-the-art among open-source Video LLMs, simultaneously learning an effective and high-purity visual evidence selection policy.

This study presents a controllable abstract summary generation method for large language models based on prompt engineering. To address the issues of summary quality and controllability in traditional methods, we design a multi-stage prompt generation framework. This framework generates summaries with varying levels of abstraction by performing semantic analysis, topic modeling, and noise control on the input text. The experiment uses the CNN/Daily Mail dataset and provides a detailed analysis of different prompt lengths, data noise, and text types. The experimental results show that prompt length has a significant impact on the quality of generated summaries. Both very short and very long prompt tokens result in a decrease in summary quality. Data noise also negatively affects the summary generation process. As noise levels increase, the ROUGE-L score gradually decreases. Furthermore, different text types have varying effects on the model's ability to generate summaries. The model performs best when handling news texts, while its performance is worse when processing academic articles. This research provides new insights into improving summary generation using large language models, particularly in how controlling prompt strategies and optimizing text preprocessing can enhance summary accuracy and controllability.