TL-GRPO: Turn-Level RL for Reasoning-Guided Iterative Optimization

Large language models have demonstrated strong reasoning capabilities in complex tasks through tool integration, which is typically framed as a Markov Decision Process and optimized with trajectory-level RL algorithms such as GRPO. However, a common class of reasoning tasks, iterative optimization, presents distinct challenges: the agent interacts with the same underlying environment state across turns, and the value of a trajectory is determined by the best turn-level reward rather than cumulative returns. Existing GRPO-based methods cannot perform fine-grained, turn-level optimization in such settings, while black-box optimization methods discard prior knowledge and reasoning capabilities. To address this gap, we propose Turn-Level GRPO (TL-GRPO), a lightweight RL algorithm that performs turn-level group sampling for fine-grained optimization. We evaluate TL-GRPO on analog circuit sizing (ACS), a challenging scientific optimization task requiring multiple simulations and domain expertise. Results show that TL-GRPO outperforms standard GRPO and Bayesian optimization methods across various specifications. Furthermore, our 30B model trained with TL-GRPO achieves state-of-the-art performance on ACS tasks under same simulation budget, demonstrating both strong generalization and practical utility.

Mixing Expert Knowledge: Bring Human Thoughts Back To the Game of Go

Large language models (LLMs) have demonstrated exceptional performance in reasoning tasks such as mathematics and coding, matching or surpassing human capabilities. However, these impressive reasoning abilities face significant challenges in specialized domains. Taking Go as an example, although AlphaGo has established the high performance ceiling of AI systems in Go, mainstream LLMs still struggle to reach even beginner-level proficiency, let alone perform natural language reasoning. This performance gap between general-purpose LLMs and domain experts is significantly limiting the application of LLMs on a wider range of domain-specific tasks. In this work, we aim to bridge the divide between LLMs' general reasoning capabilities and expert knowledge in domain-specific tasks. We perform mixed fine-tuning with structured Go expertise and general long Chain-of-Thought (CoT) reasoning data as a cold start, followed by reinforcement learning to integrate expert knowledge in Go with general reasoning capabilities. Through this methodology, we present \textbf{LoGos}, a powerful LLM that not only maintains outstanding general reasoning abilities, but also conducts Go gameplay in natural language, demonstrating effective strategic reasoning and accurate next-move prediction. LoGos achieves performance comparable to human professional players, substantially surpassing all existing LLMs. Through this work, we aim to contribute insights on applying general LLM reasoning capabilities to specialized domains. We will release the first large-scale Go dataset for LLM training, the first LLM Go evaluation benchmark, and the first general LLM that reaches human professional-level performance in Go at: https://github.com/Entarochuan/LoGos.

InternBootcamp Technical Report: Boosting LLM Reasoning with Verifiable Task Scaling

Large language models (LLMs) have revolutionized artificial intelligence by enabling complex reasoning capabilities. While recent advancements in reinforcement learning (RL) have primarily focused on domain-specific reasoning tasks (e.g., mathematics or code generation), real-world reasoning scenarios often require models to handle diverse and complex environments that narrow-domain benchmarks cannot fully capture. To address this gap, we present InternBootcamp, an open-source framework comprising 1000+ domain-diverse task environments specifically designed for LLM reasoning research. Our codebase offers two key functionalities: (1) automated generation of unlimited training/testing cases with configurable difficulty levels, and (2) integrated verification modules for objective response evaluation. These features make InternBootcamp fundamental infrastructure for RL-based model optimization, synthetic data generation, and model evaluation. Although manually developing such a framework with enormous task coverage is extremely cumbersome, we accelerate the development procedure through an automated agent workflow supplemented by manual validation protocols, which enables the task scope to expand rapidly. % With these bootcamps, we further establish Bootcamp-EVAL, an automatically generated benchmark for comprehensive performance assessment. Evaluation reveals that frontier models still underperform in many reasoning tasks, while training with InternBootcamp provides an effective way to significantly improve performance, leading to our 32B model that achieves state-of-the-art results on Bootcamp-EVAL and excels on other established benchmarks. In particular, we validate that consistent performance gains come from including more training tasks, namely \textbf{task scaling}, over two orders of magnitude, offering a promising route towards capable reasoning generalist.

Image-to-Text Logic Jailbreak: Your Imagination can Help You Do Anything

Large Visual Language Models (VLMs) such as GPT-4 have achieved remarkable success in generating comprehensive and nuanced responses, surpassing the capabilities of large language models. However, with the integration of visual inputs, new security concerns emerge, as malicious attackers can exploit multiple modalities to achieve their objectives. This has led to increasing attention on the vulnerabilities of VLMs to jailbreak. Most existing research focuses on generating adversarial images or nonsensical image collections to compromise these models. However, the challenge of leveraging meaningful images to produce targeted textual content using the VLMs' logical comprehension of images remains unexplored. In this paper, we explore the problem of logical jailbreak from meaningful images to text. To investigate this issue, we introduce a novel dataset designed to evaluate flowchart image jailbreak. Furthermore, we develop a framework for text-to-text jailbreak using VLMs. Finally, we conduct an extensive evaluation of the framework on GPT-4o and GPT-4-vision-preview, with jailbreak rates of 92.8% and 70.0%, respectively. Our research reveals significant vulnerabilities in current VLMs concerning image-to-text jailbreak. These findings underscore the need for a deeper examination of the security flaws in VLMs before their practical deployment.

Is the System Message Really Important to Jailbreaks in Large Language Models?

The rapid evolution of Large Language Models (LLMs) has rendered them indispensable in modern society. While security measures are typically in place to align LLMs with human values prior to release, recent studies have unveiled a concerning phenomenon named "jailbreak." This term refers to the unexpected and potentially harmful responses generated by LLMs when prompted with malicious questions. Existing research focuses on generating jailbreak prompts but our study aim to answer a different question: Is the system message really important to jailbreak in LLMs? To address this question, we conducted experiments in a stable GPT version gpt-3.5-turbo-0613 to generated jailbreak prompts with varying system messages: short, long, and none. We discover that different system messages have distinct resistances to jailbreak by experiments. Additionally, we explore the transferability of jailbreak across LLMs. This finding underscores the significant impact system messages can have on mitigating LLMs jailbreak. To generate system messages that are more resistant to jailbreak prompts, we propose System Messages Evolutionary Algorithms (SMEA). Through SMEA, we can get robust system messages population that demonstrate up to 98.9% resistance against jailbreak prompts. Our research not only bolsters LLMs security but also raises the bar for jailbreak, fostering advancements in this field of study.