In this report, we present the third technical report on the development of slow-thinking models as part of the STILL project. As the technical pathway becomes clearer, scaling RL training has become a central technique for implementing such reasoning models. We systematically experiment with and document the effects of various factors influencing RL training, conducting experiments on both base models and fine-tuned models. Specifically, we demonstrate that our RL training approach consistently improves the Qwen2.5-32B base models, enhancing both response length and test accuracy. Furthermore, we show that even when a model like DeepSeek-R1-Distill-Qwen-1.5B has already achieved a high performance level, it can be further refined through RL training, reaching an accuracy of 39.33% on AIME 2024. Beyond RL training, we also explore the use of tool manipulation, finding that it significantly boosts the reasoning performance of large reasoning models. This approach achieves a remarkable accuracy of 86.67% with greedy search on AIME 2024, underscoring its effectiveness in enhancing model capabilities. We release our resources at the STILL project website: https://github.com/RUCAIBox/Slow_Thinking_with_LLMs.

Multi-modal Large Language Models (MLLMs) integrate visual and linguistic reasoning to address complex tasks such as image captioning and visual question answering. While MLLMs demonstrate remarkable versatility, MLLMs appears limited performance on special applications. But tuning MLLMs for downstream tasks encounters two key challenges: Task-Expert Specialization, where distribution shifts between pre-training and target datasets constrain target performance, and Open-World Stabilization, where catastrophic forgetting erases the model general knowledge. In this work, we systematically review recent advancements in MLLM tuning methodologies, classifying them into three paradigms: (I) Selective Tuning, (II) Additive Tuning, and (III) Reparameterization Tuning. Furthermore, we benchmark these tuning strategies across popular MLLM architectures and diverse downstream tasks to establish standardized evaluation analysis and systematic tuning principles. Finally, we highlight several open challenges in this domain and propose future research directions. To facilitate ongoing progress in this rapidly evolving field, we provide a public repository that continuously tracks developments: https://github.com/WenkeHuang/Awesome-MLLM-Tuning.

Large Language Models (LLMs) excel in reasoning but remain constrained by their Chain-of-Thought (CoT) approach, which struggles with complex tasks requiring more nuanced topological reasoning. We introduce SOLAR, Scalable Optimization of Large-scale Architecture for Reasoning, a framework that dynamically optimizes various reasoning topologies to enhance accuracy and efficiency. Our Topological Annotation Generation (TAG) system automates topological dataset creation and segmentation, improving post-training and evaluation. Additionally, we propose Topological-Scaling, a reward-driven framework that aligns training and inference scaling, equipping LLMs with adaptive, task-aware reasoning. SOLAR achieves substantial gains on MATH and GSM8K: +5% accuracy with Topological Tuning, +9% with Topological Reward, and +10.02% with Hybrid Scaling. It also reduces response length by over 5% for complex problems, lowering inference latency. To foster the reward system, we train a multi-task Topological Reward Model (M-TRM), which autonomously selects the best reasoning topology and answer in a single pass, eliminating the need for training and inference on multiple single-task TRMs (S-TRMs), thus reducing both training cost and inference latency. In addition, in terms of performance, M-TRM surpasses all S-TRMs, improving accuracy by +10% and rank correlation by +9%. To the best of our knowledge, SOLAR sets a new benchmark for scalable, high-precision LLM reasoning while introducing an automated annotation process and a dynamic reasoning topology competition mechanism.

Multimodal summarization integrating information from diverse data modalities presents a promising solution to aid the understanding of information within various processes. However, the application and advantages of multimodal summarization have not received much attention in model-based engineering (MBE), where it has become a cornerstone in the design and development of complex systems, leveraging formal models to improve understanding, validation and automation throughout the engineering lifecycle. UML and EMF diagrams in model-based engineering contain a large amount of multimodal information and intricate relational data. Hence, our study explores the application of multimodal large language models within the domain of model-based engineering to evaluate their capacity for understanding and identifying relationships, features, and functionalities embedded in UML and EMF diagrams. We aim to demonstrate the transformative potential benefits and limitations of multimodal summarization in improving productivity and accuracy in MBE practices. The proposed approach is evaluated within the context of automotive software development, while many promising state-of-art models were taken into account.

Large Language Models (LLMs) are revolutionizing bioinformatics, enabling advanced analysis of DNA, RNA, proteins, and single-cell data. This survey provides a systematic review of recent advancements, focusing on genomic sequence modeling, RNA structure prediction, protein function inference, and single-cell transcriptomics. Meanwhile, we also discuss several key challenges, including data scarcity, computational complexity, and cross-omics integration, and explore future directions such as multimodal learning, hybrid AI models, and clinical applications. By offering a comprehensive perspective, this paper underscores the transformative potential of LLMs in driving innovations in bioinformatics and precision medicine.

Large Language Model (LLM) based judges form the underpinnings of key safety evaluation processes such as offline benchmarking, automated red-teaming, and online guardrailing. This widespread requirement raises the crucial question: can we trust the evaluations of these evaluators? In this paper, we highlight two critical challenges that are typically overlooked: (i) evaluations in the wild where factors like prompt sensitivity and distribution shifts can affect performance and (ii) adversarial attacks that target the judge. We highlight the importance of these through a study of commonly used safety judges, showing that small changes such as the style of the model output can lead to jumps of up to 0.24 in the false negative rate on the same dataset, whereas adversarial attacks on the model generation can fool some judges into misclassifying 100% of harmful generations as safe ones. These findings reveal gaps in commonly used meta-evaluation benchmarks and weaknesses in the robustness of current LLM judges, indicating that low attack success under certain judges could create a false sense of security. Well-known jailbreak attacks on widely used Large Language Models (LLMs) such as ChatGPT have raised concerns about the robustness of these systems to safety violations. As a result, organizations deploying them typically rely on a two-pronged approach to safety: 1) offline benchmarking and red-teaming (Mazeika et al., 2024; Perez et al., 2022; Ganguli et al., 2022), and 2) online guardrails designed to minimize the risk from attacks (Mu et al., 2024; Manczak et al., 2024; Neill et al., 2024).

Recent advancements in slow-thinking reasoning models have shown exceptional performance in complex reasoning tasks. However, these models often exhibit overthinking-generating redundant reasoning steps for simple problems, leading to excessive computational resource usage. While current mitigation strategies uniformly reduce reasoning tokens, they risk degrading performance on challenging tasks that require extended reasoning. This paper introduces Difficulty-Adaptive Slow-Thinking (DAST), a novel framework that enables models to autonomously adjust the length of Chain-of-Thought(CoT) based on problem difficulty. We first propose a Token Length Budget (TLB) metric to quantify difficulty, then leveraging length-aware reward shaping and length preference optimization to implement DAST. DAST penalizes overlong responses for simple tasks while incentivizing sufficient reasoning for complex problems. Experiments on diverse datasets and model scales demonstrate that DAST effectively mitigates overthinking (reducing token usage by over 30\% on average) while preserving reasoning accuracy on complex problems.

The need for interpretability in deep learning has driven interest in counterfactual explanations, which identify minimal changes to an instance that change a model's prediction. Current counterfactual (CF) generation methods require task-specific fine-tuning and produce low-quality text. Large Language Models (LLMs), though effective for high-quality text generation, struggle with label-flipping counterfactuals (i.e., counterfactuals that change the prediction) without fine-tuning. We introduce two simple classifier-guided approaches to support counterfactual generation by LLMs, eliminating the need for fine-tuning while preserving the strengths of LLMs. Despite their simplicity, our methods outperform state-of-the-art counterfactual generation methods and are effective across different LLMs, highlighting the benefits of guiding counterfactual generation by LLMs with classifier information. We further show that data augmentation by our generated CFs can improve a classifier's robustness. Our analysis reveals a critical issue in counterfactual generation by LLMs: LLMs rely on parametric knowledge rather than faithfully following the classifier.

Vision-language models (VLMs) have achieved remarkable advancements, capitalizing on the impressive capabilities of large language models (LLMs) across diverse tasks. Despite this, a critical challenge known as hallucination occurs when models overconfidently describe objects or attributes absent from the image, a problem exacerbated by the tendency of VLMs to rely on linguistic priors. This limitation reduces model reliability in high-stakes applications. In this work, we have observed the characteristic of logits' continuity consistency enhancement and introduced a straightforward and efficient method, Cross-Temporal Prediction Connection (TPC), designed to enhance the semantic consistency of logits by connecting them temporally across timesteps. TPC amplifies information flow and improves coherence, effectively reducing hallucination. Extensive experiments show that TPC surpasses existing representatives, delivering superior performance in both accuracy and efficiency while maintaining robustness in open-ended text generation tasks.

Framing is an essential device in news reporting, allowing the writer to influence public perceptions of current affairs. While there are existing automatic news framing detection datasets in various languages, none of them focus on news framing in the Chinese language which has complex character meanings and unique linguistic features. This study introduces the first Chinese News Framing dataset, to be used as either a stand-alone dataset or a supplementary resource to the SemEval-2023 task 3 dataset. We detail its creation and we run baseline experiments to highlight the need for such a dataset and create benchmarks for future research, providing results obtained through fine-tuning XLM-RoBERTa-Base and using GPT-4o in the zero-shot setting. We find that GPT-4o performs significantly worse than fine-tuned XLM-RoBERTa across all languages. For the Chinese language, we obtain an F1-micro (the performance metric for SemEval task 3, subtask 2) score of 0.719 using only samples from our Chinese News Framing dataset and a score of 0.753 when we augment the SemEval dataset with Chinese news framing samples. With positive news frame detection results, this dataset is a valuable resource for detecting news frames in the Chinese language and is a valuable supplement to the SemEval-2023 task 3 dataset.