While large language models (LLMs) have showcased impressive capabilities, they struggle with addressing legal queries due to the intricate complexities and specialized expertise required in the legal field. In this paper, we introduce InternLM-Law, a specialized LLM tailored for addressing diverse legal queries related to Chinese laws, spanning from responding to standard legal questions (e.g., legal exercises in textbooks) to analyzing complex real-world legal situations. We meticulously construct a dataset in the Chinese legal domain, encompassing over 1 million queries, and implement a data filtering and processing pipeline to ensure its diversity and quality. Our training approach involves a novel two-stage process: initially fine-tuning LLMs on both legal-specific and general-purpose content to equip the models with broad knowledge, followed by exclusive fine-tuning on high-quality legal data to enhance structured output generation. InternLM-Law achieves the highest average performance on LawBench, outperforming state-of-the-art models, including GPT-4, on 13 out of 20 subtasks. We make InternLM-Law and our dataset publicly available to facilitate future research in applying LLMs within the legal domain.

Large language models (LLMs) have recently achieved notable success in code‑generation benchmarks such as HumanEval and LiveCodeBench. However, a detailed examination reveals that these evaluation suites often comprise only a limited number of homogeneous test cases, resulting in subtle faults going undetected. This not only artificially inflates measured performance but also compromises accurate reward estimation in reinforcement learning frameworks utilizing verifiable rewards (RLVR). To address these critical shortcomings, we systematically investigate the test-case generation (TCG) task by proposing multi-dimensional metrics designed to rigorously quantify test-suite thoroughness. Furthermore, we introduce a human-LLM collaborative method (SAGA), leveraging human programming expertise with LLM reasoning capability, aimed at significantly enhancing both the coverage and the quality of generated test cases. In addition, we develop a TCGBench to facilitate the study of the TCG task. Experiments show that SAGA achieves a detection rate of 90.62\% and a verifier accuracy of 32.58\% on TCGBench. The Verifier Accuracy (Verifier Acc) of the code generation evaluation benchmark synthesized by SAGA is 10.78\% higher than that of LiveCodeBench-v6. These results demonstrate the effectiveness of our proposed method. We hope this work contributes to building a scalable foundation for reliable LLM code evaluation, further advancing RLVR in code generation, and paving the way for automated adversarial test synthesis and adaptive benchmark integration.

We introduce InternVL 3.5, a new family of open-source multimodal models that significantly advances versatility, reasoning capability, and inference efficiency along the InternVL series. A key innovation is the Cascade Reinforcement Learning (Cascade RL) framework, which enhances reasoning through a two-stage process: offline RL for stable convergence and online RL for refined alignment. This coarse-to-fine training strategy leads to substantial improvements on downstream reasoning tasks, e.g., MMMU and MathVista. To optimize efficiency, we propose a Visual Resolution Router (ViR) that dynamically adjusts the resolution of visual tokens without compromising performance. Coupled with ViR, our Decoupled Vision-Language Deployment (DvD) strategy separates the vision encoder and language model across different GPUs, effectively balancing computational load. These contributions collectively enable InternVL3.5 to achieve up to a +16.0% gain in overall reasoning performance and a 4.05× inference speedup compared to its predecessor, i.e., InternVL3. In addition, InternVL3.5 supports novel capabilities such as GUI interaction and embodied agency. Notably, our largest model, i.e., InternVL3.5-241B-A28B, attains state-of-the-art results among open-source MLLMs across general multimodal, reasoning, text, and agentic tasks – narrowing the performance gap with leading commercial models like GPT-5. All models and code are publicly released.

In recent years, a plethora of open-source foundation models have emerged, achieving remarkable progress in some widely attended fields, with performance being quite close to that of closed-source models. However, in high-value but more challenging scientific professional fields, either the fields still rely on expert models, or the progress of general foundation models lags significantly compared to those in popular areas, far from sufficient for transforming scientific research and leaving substantial gap between open-source models and closed-source models in these scientific domains. To mitigate this gap and explore a step further toward Artificial General Intelligence (AGI), we introduce Intern-S1, a specialized generalist equipped with general understanding and reasoning capabilities with expertise to analyze multiple science modal data. Intern-S1 is a multimodal Mixture-of-Experts (MoE) model with 28 billion activated parameters and 241 billion total parameters, continually pre-trained on 5T tokens, including over 2.5T tokens from scientific domains. In the post-training stage, Intern-S1 undergoes offline and then online reinforcement learning (RL) in InternBootCamp, where we propose Mixture-of-Rewards (MoR) to synergize the RL training on more than 1000 tasks simultaneously. Through integrated innovations in algorithms, data, and training systems, Intern-S1 achieved top-tier performance in online RL training. On comprehensive evaluation benchmarks, Intern-S1 demonstrates competitive performance on general reasoning tasks among open-source models and significantly outperforms open-source models in scientific domains, surpassing closed-source state-of-the-art models in professional tasks, such as molecular synthesis planning, reaction condition prediction, predicting thermodynamic stabilities for crystals. Our models are available at https://huggingface.co/internlm/Intern-S1.

Recent advancements in open-source multi-modal large language models (MLLMs) have primarily focused on enhancing foundational capabilities, leaving a significant gap in human preference alignment. This paper introduces OmniAlign-V, a comprehensive dataset of 200K high-quality training samples featuring diverse images, complex questions, and varied response formats to improve MLLMs’ alignment with human preferences. We also present MM-AlignBench, a human-annotated benchmark specifically designed to evaluate MLLMs’ alignment with human values. Experimental results show that finetuning MLLMs with OmniAlign-V, using Supervised Fine-Tuning (SFT) or Direct Preference Optimization (DPO), significantly enhances human preference alignment while maintaining or enhancing performance on standard VQA benchmarks, preserving their fundamental capabilities.

Large language models (LLMs) have achieved remarkable progress in code generation, yet their true programming competence remains underexplored. We introduce the Code Triangle framework, which systematically evaluates LLMs across three fundamental dimensions: editorial analysis, code implementation, and test case generation. Through extensive experiments on competitive programming benchmarks, we reveal that while LLMs can form a self-consistent system across these dimensions, their solutions often lack the diversity and robustness of human programmers. We identify a significant distribution shift between model cognition and human expertise, with model errors tending to cluster due to training data biases and limited reasoning transfer. Our study demonstrates that incorporating human-generated editorials, solutions, and diverse test cases, as well as leveraging model mixtures, can substantially enhance both the performance and robustness of LLMs. Furthermore, we reveal both the consistency and inconsistency in the cognition of LLMs that may facilitate self-reflection and self-improvement, providing a potential direction for developing more powerful coding models.

Efficient and accurate evaluation is crucial for the continuous improvement of large language models (LLMs). Among various assessment methods, subjective evaluation has garnered significant attention due to its superior alignment with real-world usage scenarios and human preferences. However, human-based evaluations are costly and lack reproducibility, making precise automated evaluators (judgers) vital in this process. In this report, we introduce CompassJudger-1, the first open-source all-in-one judge LLM. CompassJudger-1 is a general-purpose LLM that demonstrates remarkable versatility. It is capable of: 1. Performing unitary scoring and two-model comparisons as a reward model; 2. Conducting evaluations according to specified formats; 3. Generating critiques; 4. Executing diverse tasks like a general LLM. To assess the evaluation capabilities of different judge models under a unified setting, we have also established JudgerBench, a new benchmark that encompasses various subjective evaluation tasks and covers a wide range of topics. CompassJudger-1 offers a comprehensive solution for various evaluation tasks while maintaining the flexibility to adapt to diverse requirements. Both CompassJudger and JudgerBench are released and available to the research community athttps://github.com/open-compass/CompassJudger. We believe that by open-sourcing these tools, we can foster collaboration and accelerate progress in LLM evaluation methodologies.

The evolution of Large Language Models (LLMs) like ChatGPT and GPT-4 has sparked discussions on the advent of Artificial General Intelligence (AGI). However, replicating such advancements in open-source models has been challenging. This paper introduces InternLM2, an open-source LLM that outperforms its predecessors in comprehensive evaluations across 6 dimensions and 30 benchmarks, long-context modeling, and open-ended subjective evaluations through innovative pre-training and optimization techniques. The pre-training process of InternLM2 is meticulously detailed, highlighting the preparation of diverse data types including text, code, and long-context data. InternLM2 efficiently captures long-term dependencies, initially trained on 4k tokens before advancing to 32k tokens in pre-training and fine-tuning stages, exhibiting remarkable performance on the 200k “Needle-in-a-Haystack" test. InternLM2 is further aligned using Supervised Fine-Tuning (SFT) and a novel Conditional Online Reinforcement Learning from Human Feedback (COOL RLHF) strategy that addresses conflicting human preferences and reward hacking. By releasing InternLM2 models in different training stages and model sizes, we provide the community with insights into the model's evolution.

Automated detection of cervical abnormal cells from Thin-prep cytologic test (TCT) images is crucial for efficient cervical abnormal screening using computer-aided diagnosis systems. However, the construction of the detection model is hindered by the preparation of the training images, which usually suffers from issues of class imbalance and incomplete annotations. Additionally, existing methods often overlook the visual feature correlations among cells, which are crucial in cervical lesion cell detection as pathologists commonly rely on surrounding cells for identification. In this paper, we propose a distillation framework that utilizes a patch-level pre-training network to guide the training of an image-level detection network, which can be applied to various detectors without changing their architectures during inference. The main contribution is three-fold: (1) We propose the Balanced Pre-training Model (BPM) as the patch-level cervical cell classification model, which employs an image synthesis model to construct a class-balanced patch dataset for pre-training. (2) We design the Score Correction Loss (SCL) to enable the detection network to distill knowledge from the BPM model, thereby mitigating the impact of incomplete annotations. (3) We design the Patch Correlation Consistency (PCC) strategy to exploit the correlation information of extracted cells, consistent with the behavior of cytopathologists. Experiments on public and private datasets demonstrate the superior performance of the proposed distillation method, as well as its adaptability to various detection architectures.

Automatic examination of thin-prep cytologic test (TCT) slides can assist pathologists in finding cervical abnormality for accurate and efficient cancer screening. Current solutions mostly need to localize suspicious cells and classify abnormality based on local patches, concerning the fact that whole slide images of TCT are extremely large. It thus requires many annotations of normal and abnormal cervical cells, to supervise the training of the patch-level classifier for promising performance. In this paper, we propose CellGAN to synthesize cytopathological images of various cervical cell types for augmenting patch-level cell classification. Built upon a lightweight backbone, CellGAN is equipped with a non-linear class mapping network to effectively incorporate cell type information into image generation. We also propose the Skip-layer Global Context module to model the complex spatial relationship of the cells, and attain high fidelity of the synthesized images through adversarial learning. Our experiments demonstrate that CellGAN can produce visually plausible TCT cytopathological images for different cell types. We also validate the effectiveness of using CellGAN to greatly augment patch-level cell classification performance. Our code and model checkpoint are available at https://github.com/ZhenrongShen/CellGAN .

Plan-and-Write is a common hierarchical approach in long-form narrative text generation, which first creates a plan to guide the narrative writing. Following this approach, several studies rely on simply prompting large language models for planning, which often yields suboptimal results. In this paper, we propose a new framework called Evaluation-guided Iterative Plan Extraction for long-form narrative text generation (EIPE-text), which extracts plans from the corpus of narratives and utilizes the extracted plans to construct a better planner. EIPE-text has three stages: plan extraction, learning, and inference. In the plan extraction stage, it iteratively extracts and improves plans from the narrative corpus and constructs a plan corpus. We propose a question answer (QA) based evaluation mechanism to automatically evaluate the plans and generate detailed plan refinement instructions to guide the iterative improvement. In the learning stage, we build a better planner by fine-tuning with the plan corpus or in-context learning with examples in the plan corpus. Finally, we leverage a hierarchical approach to generate long-form narratives. We evaluate the effectiveness of EIPE-text in the domains of novels and storytelling. Both GPT-4-based evaluations and human evaluations demonstrate that our method can generate more coherent and relevant long-form narratives. Our code will be released in the future.

Deep learning has been instrumental in improving the accuracy of cervical cancer screening using whole-slide images (WSIs) in recent years. Due to the complexity of the computer-aided screening task, the pipeline typically involves detecting “abnormal” cervical cells, and runs classification at the patch and sample levels, respectively. While the patch-level classification for normal or abnormal cells cannot be perfect, the errors may accumulate across individual patches and make the subsequent sample-level classification even more difficult. To address these issues, we propose a Patch-to-Sample (P2S) reasoning method to screen the cervical abnormality in this paper. We first improve the patch-level classifier by the hard patch mining strategy, such that the classifier is not only more accurate but also more powerful to represent suspicious cells in local patches. Then, we propose score embedding and token pooling to a transformer network, which aggregates multiple patches and derives the diagnosis result at the sample level. Experiments show that our P2S method can more effectively utilize the key patches in individual samples, and thus outperforms existing methods.

Vulvovaginal candidiasis (VVC) is the most prevalent human candidal infection, estimated to afflict approximately 75 % of all women at least once in their lifetime. It will lead to several symptoms including pruritus, vaginal soreness, and so on. Automatic whole slide image (WSI) classification is highly demanded, for the huge burden of disease control and prevention. However, the WSI-based computer-aided VCC screening method is still vacant due to the scarce labeled data and unique properties of candida. Candida in WSI is challenging to be captured by conventional classification models due to its distinctive elongated shape, the small proportion of their spatial distribution, and the style gap from WSIs. To make the model focus on the candida easier, we propose an attention-guided method, which can obtain a robust diagnosis classification model. Specifically, we first use a pre-trained detection model as prior instruction to initialize the classification model. Then we design a Skip Self-Attention module to refine the attention onto the fined-grained features of candida. Finally, we use a contrastive learning method to alleviate the overfitting caused by the style gap of WSIs and suppress the attention to false positive regions. Our experimental results demonstrate that our framework achieves state-of-the-art performance. Code and example data are available at https://github.com/caijd2000/MICCAI2023-VVC-Screening .

Cervical cancer is a significant health burden worldwide, and computer-aided diagnosis (CAD) pipelines have the potential to improve diagnosis efficiency and treatment outcomes. However, traditional CAD pipelines have limitations due to the requirement of a detection model trained on a large annotated dataset, which can be expensive and time-consuming. They also have a clear performance limit and low data utilization efficiency. To address these issues, we introduce a two-stage detection-free pipeline, incorporating pooling transformer and MoCo pretraining strategies, that optimizes data utilization for whole slide images (WSIs) while relying solely on sample-level diagnosis labels for training. The experimental results demonstrate the effectiveness of our approach, with performance scaling up as the amount of data increases. Overall, our novel pipeline has the potential to fully utilize massive data in WSI classification and can significantly improve cancer diagnosis and treatment. By reducing the reliance on expensive data labeling and detection models, our approach could enable more widespread and cost-effective implementation of CAD pipelines in clinical settings. Our code and model is available at https://github.com/thebestannie/Detection-free-MICCAI2023 .

Automated detection of cervical abnormal cells from Thin-prep cytologic test (TCT) images is essential for efficient cervical abnormal screening by computer-aided diagnosis system. However, the detection performance is influenced by noise samples in the training dataset, mainly due to the subjective differences among cytologists in annotating the training samples. Besides, existing detection methods often neglect visual feature correlation information between cells, which can also be utilized to aid the detection model. In this paper, we propose a cervical abnormal cell detection method optimized by a novel distillation strategy based on local-scale consistency refinement. Firstly, we use a vanilla RetinaNet to detect top-K suspicious cells and extract region-of-interest (ROI) features. Then, a pre-trained Patch Correction Network (PCN) is leveraged to obtain local-scale features and conduct further refinement for these suspicious cell patches. We design a classification ranking loss to utilize refined scores for reducing the effects of the noisy label. Furthermore, the proposed ROI-correlation consistency loss is computed between extracted ROI features and local-scale features to exploit correlation information and optimize RetinaNet. Our experiments demonstrate that our distillation method can greatly optimize the performance of cervical abnormal cell detection without changing the detector's network structure in the inference. The code is publicly available at https://github.com/feimanman/Cervical-Abnormal-Cell-Detection .

Cervical cancer is one of the primary factors that endanger women's health, and Thin-prep cytologic test (TCT) has been widely applied for early screening. Automatic whole slide image (WSI) classification is highly demanded, as it can significantly reduce the workload of pathologists. Current methods are mainly based on suspicious lesion patch extraction and classification, which ignore the intrinsic relationships between suspicious patches and neglect the other patches apart from the suspicious patches, and therefore limit their robustness and generalizability. Here we propose a novel method to solve the problem, which is based on graph attention network (GAT) and supervised contrastive learning. First, for each WSI, we extract and rank a large number of representative patches based on suspicious cell detection. Then, we select the top-K and bottom-K suspicious patches to construct two graphs seperately. Next, we introduce GAT to aggregate the features from each node, and use supervised contrastive learning to obtain valuable representations of graphs. Specifically, we design a novel contrastive loss so that the latent distances between two graphs are enlarged for positive WSIs and reduced for negative WSIs. Experimental results show that the proposed GAT method outperforms conventional methods, and also demonstrate the effectiveness of supervised contrastive learning.

Deep learning has proven to be an effective approach to read whole-slide cytologic images for computer-assisted cervical cancer screening. To construct such a pipeline, it is often necessary to train a classifier, which decides if a patch (for example, sized 224 × 224 ) provides a positive cell reading or not. Following the clinical guidance, pathologists must label many such patches of both negative (N) and positive (P) cells. Then, a deep network can be trained for the binary N-P classification problem. In this paper, we take advantage of the complexity and illegibility of the cells that are negative according to current clinical guidance, to further improve the classification performance. We claim that the negative patches can be divided into two types: easy-negative (EN) and hard-negative (HN). The initial N-P binary classification can then be converted to an EN-HN-P triple-class problem. We also align the EN-HN and HN-P decision planes in parallel in the latent feature space where all input patches are encoded. The dual planes perform parallel classification then, following a well-planned curriculum learning scheme. Our results show that the proposed method can greatly enhance the performance of classifying positive patches and negative patches by using the better learned latent space and the related encoding of each patch.