Standing in 2025, at a critical juncture in the pursuit of Artificial General Intelligence (AGI), reinforcement fine-tuning (RFT) has demonstrated significant potential in enhancing the reasoning capability of large language models (LLMs) and has led to the development of cutting-edge AI models such as OpenAI-o1 and DeepSeek-R1. Moreover, the efficient application of RFT to enhance the reasoning capability of multimodal large language models (MLLMs) has attracted widespread attention from the community. In this position paper, we argue that reinforcement fine-tuning powers the reasoning capability of multimodal large language models. To begin with, we provide a detailed introduction to the fundamental background knowledge that researchers interested in this field should be familiar with. Furthermore, we meticulously summarize the improvements of RFT in powering reasoning capability of MLLMs into five key points: diverse modalities, diverse tasks and domains, better training algorithms, abundant benchmarks and thriving engineering frameworks. Finally, we propose five promising directions for future research that the community might consider. We hope that this position paper will provide valuable insights to the community at this pivotal stage in the advancement toward AGI. Summary of works done on RFT for MLLMs is available at https://github.com/Sun-Haoyuan23/Awesome-RL-based-Reasoning-MLLMs.

Many existing models for clinical trial outcome prediction are optimized using task-specific loss functions on trial phase-specific data. While this scheme may boost prediction for common diseases and drugs, it can hinder learning of generalizable representations, leading to more false positives/negatives. To address this limitation, we introduce CLaDMoP, a new pre-training approach for clinical trial outcome prediction, alongside the Successful Clinical Trials dataset(SCT), specifically designed for this task. CLaDMoP leverages a Large Language Model-to encode trials' eligibility criteria-linked to a lightweight Drug-Molecule branch through a novel multi-level fusion technique. To efficiently fuse long embeddings across levels, we incorporate a grouping block, drastically reducing computational overhead. CLaDMoP avoids reliance on task-specific objectives by pre-training on a "pair matching" proxy task. Compared to established zero-shot and few-shot baselines, our method significantly improves both PR-AUC and ROC-AUC, especially for phase I and phase II trials. We further evaluate and perform ablation on CLaDMoP after Parameter-Efficient Fine-Tuning, comparing it to state-of-the-art supervised baselines, including MEXA-CTP, on the Trial Outcome Prediction(TOP) benchmark. CLaDMoP achieves up to 10.5% improvement in PR-AUC and 3.6% in ROC-AUC, while attaining comparable F1 score to MEXA-CTP, highlighting its potential for clinical trial outcome prediction. Code and SCT dataset can be downloaded from https://github.com/murai-lab/CLaDMoP.

Foundation models based on large language models (LLMs) have shown great success in handling various tasks and modalities. However, adapting these models for general-purpose audio-language tasks is challenging due to differences in acoustic environments and task variations. In this work, we introduce LiSTEN Learning Soft Token Embeddings for Neural Audio LLMs), a framework for adapting LLMs to speech and audio tasks. LiSTEN uses a dynamic prompt selection strategy with learnable key-value pairs, allowing the model to balance general and task-specific knowledge while avoiding overfitting in a multitask setting. Our approach reduces dependence on large-scale ASR or captioning datasets, achieves competitive performance with fewer trainable parameters, and simplifies training by using a single-stage process. Additionally, LiSTEN enhances interpretability by analyzing the diversity and overlap of selected prompts across different tasks.

Cross-capability transfer is a key challenge in large language model (LLM) research, with applications in multi-task integration, model compression, and continual learning. Recent works like FuseLLM and FuseChat have demonstrated the potential of transferring multiple model capabilities to lightweight models, enhancing adaptability and efficiency, which motivates our investigation into more efficient cross-capability transfer methods. However, existing approaches primarily focus on small, homogeneous models, limiting their applicability. For large, heterogeneous models, knowledge distillation with full-parameter fine-tuning often overlooks the student model's intrinsic capacity and risks catastrophic forgetting, while PEFT methods struggle to effectively absorb knowledge from source LLMs. To address these issues, we introduce GraftLLM, a novel method that stores source model capabilities in a target model with SkillPack format. This approach preserves general capabilities, reduces parameter conflicts, and supports forget-free continual learning and model fusion. We employ a module-aware adaptive compression strategy to compress parameter updates, ensuring efficient storage while maintaining task-specific knowledge. The resulting SkillPack serves as a compact and transferable knowledge carrier, ideal for heterogeneous model fusion and continual learning. Experiments across various scenarios demonstrate that GraftLLM outperforms existing techniques in knowledge transfer, knowledge fusion, and forget-free learning, providing a scalable and efficient solution for cross-capability transfer. The code is publicly available at: https://github.com/duguodong7/GraftLLM.

--Federated Learning (FL) facilitates the fine-tuning of Foundation Models (FMs) using distributed data sources, with Low-Rank Adaptation (LoRA) gaining popularity due to its low communication costs and strong performance. While recent work acknowledges the benefits of heterogeneous LoRA in FL and introduces flexible algorithms to support its implementation, our theoretical analysis reveals a critical gap: existing methods lack formal convergence guarantees due to parameter truncation and biased gradient updates. Specifically, adapting client-specific LoRA ranks necessitates truncating global parameters, which introduces inherent truncation errors and leads to subsequent inaccurate gradient updates that accumulate over training rounds, ultimately degrading performance. T o address the above issues, we propose FedHL, a simple yet effective Federated Learning framework tailored for Heterogeneous LoRA. By leveraging the full-rank global model as a calibrated aggregation basis, FedHL eliminates the direct truncation bias from initial alignment with client-specific ranks. Furthermore, we derive the theoretically optimal aggregation weights by minimizing the gradient drift term in the convergence upper bound. Our analysis shows that FedHL guarantees O (1 / T) convergence rate, and experiments on multiple real-world datasets demonstrate a 1-3% improvement over several state-of-the-art methods. Zihao Peng, Jiandian Zeng, and Guo Li are with the Institute of Artificial Intelligence and Future Networks, Beijing Normal University, Zhuhai 519087, China (e-mail: { pzh cs, liguo }@mail.bnu.edu.cn; Boyuan Li is with the School of Computer Science and Artificial Intelligence, Zhengzhou University, Zhengzhou 450001, China (e-mail: l202311841010602@gs.zzu.edu.cn). Shengbo Chen is with the School of Software, Nanchang University, Nanchang 330000, China (e-mail: ccb02kingdom@gmail.com).

Error correction is an important capability when applying large language models (LLMs) to facilitate user typing on mobile devices. In this paper, we use LLMs to synthesize a high-quality dataset of error correction pairs to evaluate and improve LLMs for mobile applications. We first prompt LLMs with error correction domain knowledge to build a scalable and reliable addition to the existing data synthesis pipeline. We then adapt the synthetic data distribution to match the mobile application domain by reweighting the samples. The reweighting model is learnt by predicting (a handful of) live A/B test metrics when deploying LLMs in production, given the LLM performance on offline evaluation data and scores from a small privacy-preserving on-device language model. Finally, we present best practices for mixing our synthetic data with other data sources to improve model performance on error correction in both offline evaluation and production live A/B testing.

Numerous recent prompt optimization approaches like chain-of-thought, have been demonstrated to significantly improve the quality of content generated by large language models (LLMs). In-context learning (ICL), a recent paradigm where a few representative examples guide content generation has also led to strong improvements in generation quality of LLM generated content. This idea has been applied to great effect in synthetic tabular data generation, where LLMs, through effective use of ICL and prompt optimization, can generate data that approximate samples from complex, heterogeneous distributions based on representative examples. However, ensuring high-fidelity synthetic data often requires a very large number of ICL examples which may be unavailable or costly to obtain. At the same time, as LLMs get larger and larger, their in-built prior knowledge becomes vast and can potentially substitute for specific data examples. In this paper, we introduce Knowledge-Guided Prompting (KGP) as a new knob in prompt optimization and explore the ability of KGP-based prompt optimization to offset the cost of ICL. Specifically, we explore the question `how many examples can a prompt substitute for?' and explore knowledge-guided prompting (KGP) where domain knowledge, either inferred or available, is explicitly injected into the prompt, reducing dependence on ICL examples. Our experiments systematically explore the trade-off between ICL and KGP, revealing an empirical scaling law that quantifies how quality of generated synthetic data varies with increasing domain knowledge and decreasing example count. Our results demonstrate that knowledge-guided prompting can be a scalable alternative, or addition, to in-context examples, unlocking new approaches to synthetic data generation.

Various industries have produced a large number of documents such as industrial plans, technical guidelines, and regulations that are structurally complex and content-wise fragmented. This poses significant challenges for experts and decision-makers in terms of retrieval and understanding. Although existing LLM-based Retrieval-Augmented Generation methods can provide context-related suggestions, they lack quantitative weighting and traceable reasoning paths, making it difficult to offer multi-level and transparent decision support. To address this issue, this paper proposes the RAD method, which integrates Multi-Criteria Decision Making with the semantic understanding capabilities of LLMs. The method automatically extracts key criteria from industry documents, builds a weighted hierarchical decision model, and generates structured reports under model guidance. The RAD framework introduces explicit weight assignment and reasoning chains in decision generation to ensure accuracy, completeness, and traceability. Experiments show that in various decision-making tasks, the decision reports generated by RAD significantly outperform existing methods in terms of detail, rationality, and structure, demonstrating its application value and potential in complex decision support scenarios.

Modern large language model (LLM) services increasingly rely on complex, often abstract operations, such as multi-step reasoning and multi-agent collaboration, to generate high-quality outputs. While users are billed based on token consumption and API usage, these internal steps are typically not visible. We refer to such systems as Commercial Opaque LLM Services (COLS). This position paper highlights emerging accountability challenges in COLS: users are billed for operations they cannot observe, verify, or contest. We formalize two key risks: \textit{quantity inflation}, where token and call counts may be artificially inflated, and \textit{quality downgrade}, where providers might quietly substitute lower-cost models or tools. Addressing these risks requires a diverse set of auditing strategies, including commitment-based, predictive, behavioral, and signature-based methods. We further explore the potential of complementary mechanisms such as watermarking and trusted execution environments to enhance verifiability without compromising provider confidentiality. We also propose a modular three-layer auditing framework for COLS and users that enables trustworthy verification across execution, secure logging, and user-facing auditability without exposing proprietary internals. Our aim is to encourage further research and policy development toward transparency, auditability, and accountability in commercial LLM services.

Recent advances in large reasoning models (LRMs) show strong performance in structured domains such as mathematics and programming; however, they often lack pedagogical coherence and realistic teaching behaviors. To bridge this gap, we introduce Pedagogy-R1, a framework that adapts LRMs for classroom use through three innovations: (1) a distillation-based pipeline that filters and refines model outputs for instruction-tuning, (2) the Well-balanced Educational Benchmark (WBEB), which evaluates performance across subject knowledge, pedagogical knowledge, tracing, essay scoring, and teacher decision-making, and (3) a Chain-of-Pedagogy (CoP) prompting strategy for generating and eliciting teacher-style reasoning. Our mixed-method evaluation combines quantitative metrics with qualitative analysis, providing the first systematic assessment of LRMs' pedagogical strengths and limitations.