LLMEval-3: A Large-Scale Longitudinal Study on Robust and Fair Evaluation of Large Language Models

Existing evaluation of Large Language Models (LLMs) on static benchmarks is vulnerable to data contamination and leaderboard overfitting, critical issues that obscure true model capabilities. To address this, we introduce LLMEval-3, a framework for dynamic evaluation of LLMs. LLMEval-3 is built on a proprietary bank of 220k graduate-level questions, from which it dynamically samples unseen test sets for each evaluation run. Its automated pipeline ensures integrity via contamination-resistant data curation, a novel anti-cheating architecture, and a calibrated LLM-as-a-judge process achieving 90% agreement with human experts, complemented by a relative ranking system for fair comparison. An 20-month longitudinal study of nearly 50 leading models reveals a performance ceiling on knowledge memorization and exposes data contamination vulnerabilities undetectable by static benchmarks. The framework demonstrates exceptional robustness in ranking stability and consistency, providing strong empirical validation for the dynamic evaluation paradigm. LLMEval-3 offers a robust and credible methodology for assessing the true capabilities of LLMs beyond leaderboard scores, promoting the development of more trustworthy evaluation standards.

LLMEval-Med: A Real-world Clinical Benchmark for Medical LLMs with Physician Validation

Evaluating large language models (LLMs) in medicine is crucial because medical applications require high accuracy with little room for error. Current medical benchmarks have three main types: medical exam-based, comprehensive medical, and specialized assessments. However, these benchmarks have limitations in question design (mostly multiple-choice), data sources (often not derived from real clinical scenarios), and evaluation methods (poor assessment of complex reasoning). To address these issues, we present LLMEval-Med, a new benchmark covering five core medical areas, including 2,996 questions created from real-world electronic health records and expert-designed clinical scenarios. We also design an automated evaluation pipeline, incorporating expert-developed checklists into our LLM-as-Judge framework. Furthermore, our methodology validates machine scoring through human-machine agreement analysis, dynamically refining checklists and prompts based on expert feedback to ensure reliability. We evaluate 13 LLMs across three categories (specialized medical models, open-source models, and closed-source models) on LLMEval-Med, providing valuable insights for the safe and effective deployment of LLMs in medical domains. The dataset is released in https://github.com/llmeval/LLMEval-Med.

Progressive Mastery: Customized Curriculum Learning with Guided Prompting for Mathematical Reasoning

Large Language Models (LLMs) have achieved remarkable performance across various reasoning tasks, yet post-training is constrained by inefficient sample utilization and inflexible difficulty samples processing. To address these limitations, we propose Customized Curriculum Learning (CCL), a novel framework with two key innovations. First, we introduce model-adaptive difficulty definition that customizes curriculum datasets based on each model's individual capabilities rather than using predefined difficulty metrics. Second, we develop "Guided Prompting," which dynamically reduces sample difficulty through strategic hints, enabling effective utilization of challenging samples that would otherwise degrade performance. Comprehensive experiments on supervised fine-tuning and reinforcement learning demonstrate that CCL significantly outperforms uniform training approaches across five mathematical reasoning benchmarks, confirming its effectiveness across both paradigms in enhancing sample utilization and model performance.

Compression Hacking: A Supplementary Perspective on Informatics Metric of Language Models from Geometric Distortion

Recently, the concept of ``compression as intelligence'' has provided a novel informatics metric perspective for language models (LMs), emphasizing that highly structured representations signify the intelligence level of LMs. However, from a geometric standpoint, the word representation space of highly compressed LMs tends to degenerate into a highly anisotropic state, which hinders the LM's ability to comprehend instructions and directly impacts its performance. We found this compression-anisotropy synchronicity is essentially the ``Compression Hacking'' in LM representations, where noise-dominated directions tend to create the illusion of high compression rates by sacrificing spatial uniformity. Based on this, we propose three refined compression metrics by incorporating geometric distortion analysis and integrate them into a self-evaluation pipeline. The refined metrics exhibit strong alignment with the LM's comprehensive capabilities, achieving Spearman correlation coefficients above 0.9, significantly outperforming both the original compression and other internal structure-based metrics. This confirms that compression hacking substantially enhances the informatics interpretation of LMs by incorporating geometric distortion of representations.

Two Minds Better Than One: Collaborative Reward Modeling for LLM Alignment

Reward models (RMs) play a pivotal role in aligning large language models (LLMs) with human values. However, noisy preferences in human feedback can lead to reward misgeneralization - a phenomenon where reward models learn spurious correlations or overfit to noisy preferences, which poses important challenges to the generalization of RMs. This paper systematically analyzes the characteristics of preference pairs and aims to identify how noisy preferences differ from human-aligned preferences in reward modeling. Our analysis reveals that noisy preferences are difficult for RMs to fit, as they cause sharp training fluctuations and irregular gradient updates. These distinctive dynamics suggest the feasibility of identifying and excluding such noisy preferences. Empirical studies demonstrate that policy LLM optimized with a reward model trained on the full preference dataset, which includes substantial noise, performs worse than the one trained on a subset of exclusively high quality preferences. To address this challenge, we propose an online Collaborative Reward Modeling (CRM) framework to achieve robust preference learning through peer review and curriculum learning. In particular, CRM maintains two RMs that collaboratively filter potential noisy preferences by peer-reviewing each other's data selections. Curriculum learning synchronizes the capabilities of two models, mitigating excessive disparities to promote the utility of peer review. Extensive experiments demonstrate that CRM significantly enhances RM generalization, with up to 9.94 points improvement on RewardBench under an extreme 40\% noise. Moreover, CRM can seamlessly extend to implicit-reward alignment methods, offering a robust and versatile alignment strategy.

Improving RL Exploration for LLM Reasoning through Retrospective Replay

Reinforcement learning (RL) has increasingly become a pivotal technique in the post-training of large language models (LLMs). The effective exploration of the output space is essential for the success of RL. We observe that for complex problems, during the early stages of training, the model exhibits strong exploratory capabilities and can identify promising solution ideas. However, its limited capability at this stage prevents it from successfully solving these problems. The early suppression of these potentially valuable solution ideas by the policy gradient hinders the model's ability to revisit and re-explore these ideas later. Consequently, although the LLM's capabilities improve in the later stages of training, it still struggles to effectively address these complex problems. To address this exploration issue, we propose a novel algorithm named Retrospective Replay-based Reinforcement Learning (RRL), which introduces a dynamic replay mechanism throughout the training process. RRL enables the model to revisit promising states identified in the early stages, thereby improving its efficiency and effectiveness in exploration. To evaluate the effectiveness of RRL, we conduct extensive experiments on complex reasoning tasks, including mathematical reasoning and code generation, and general dialogue tasks. The results indicate that RRL maintains high exploration efficiency throughout the training period, significantly enhancing the effectiveness of RL in optimizing LLMs for complicated reasoning tasks. Moreover, it also improves the performance of RLHF, making the model both safer and more helpful.

PFDial: A Structured Dialogue Instruction Fine-tuning Method Based on UML Flowcharts

Process-driven dialogue systems, which operate under strict predefined process constraints, are essential in customer service and equipment maintenance scenarios. Although Large Language Models (LLMs) have shown remarkable progress in dialogue and reasoning, they still struggle to solve these strictly constrained dialogue tasks. To address this challenge, we construct Process Flow Dialogue (PFDial) dataset, which contains 12,705 high-quality Chinese dialogue instructions derived from 440 flowcharts containing 5,055 process nodes. Based on PlantUML specification, each UML flowchart is converted into atomic dialogue units i.e., structured five-tuples. Experimental results demonstrate that a 7B model trained with merely 800 samples, and a 0.5B model trained on total data both can surpass 90% accuracy. Additionally, the 8B model can surpass GPT-4o up to 43.88% with an average of 11.00%. We further evaluate models' performance on challenging backward transitions in process flows and conduct an in-depth analysis of various dataset formats to reveal their impact on model performance in handling decision and sequential branches. The data is released in https://github.com/KongLongGeFDU/PFDial.

Measuring Data Diversity for Instruction Tuning: A Systematic Analysis and A Reliable Metric

Data diversity is crucial for the instruction tuning of large language models. Existing studies have explored various diversity-aware data selection methods to construct high-quality datasets and enhance model performance. However, the fundamental problem of precisely defining and measuring data diversity remains underexplored, limiting clear guidance for data engineering. To address this, we systematically analyze 11 existing diversity measurement methods by evaluating their correlation with model performance through extensive fine-tuning experiments. Our results indicate that a reliable diversity measure should properly account for both inter-sample differences and the information distribution in the sample space. Building on this, we propose NovelSum, a new diversity metric based on sample-level "novelty." Experiments on both simulated and real-world data show that NovelSum accurately captures diversity variations and achieves a 0.97 correlation with instruction-tuned model performance, highlighting its value in guiding data engineering practices. With NovelSum as an optimization objective, we further develop a greedy, diversity-oriented data selection strategy that outperforms existing approaches, validating both the effectiveness and practical significance of our metric.

Predicting Large Language Model Capabilities on Closed-Book QA Tasks Using Only Information Available Prior to Training

The GPT-4 technical report from OpenAI suggests that model performance on specific tasks can be predicted prior to training, though methodologies remain unspecified. This approach is crucial for optimizing resource allocation and ensuring data alignment with target tasks. To achieve this vision, we focus on predicting performance on Closed-book Question Answering (CBQA) tasks, which are closely tied to pre-training data and knowledge retention. We address three major challenges: 1) mastering the entire pre-training process, especially data construction; 2) evaluating a model's knowledge retention; and 3) predicting task-specific knowledge retention using only information available prior to training. To tackle these challenges, we pre-train three large language models (i.e., 1.6B, 7B, and 13B) using 560k dollars and 520k GPU hours. We analyze the pre-training data with knowledge triples and assess knowledge retention using established methods. Additionally, we introduce the SMI metric, an information-theoretic measure that quantifies the relationship between pre-training data, model size, and task-specific knowledge retention. Our experiments reveal a strong linear correlation ($\text{R}^2 > 0.84$) between the SMI metric and the model's accuracy on CBQA tasks across models of varying sizes (i.e., 1.1B, 1.6B, 7B, and 13B). The dataset, model, and code are available at https://github.com/yuhui1038/SMI.

DocFusion: A Unified Framework for Document Parsing Tasks

Document parsing is essential for analyzing complex document structures and extracting fine-grained information, supporting numerous downstream applications. However, existing methods often require integrating multiple independent models to handle various parsing tasks, leading to high complexity and maintenance overhead. To address this, we propose DocFusion, a lightweight generative model with only 0.28B parameters. It unifies task representations and achieves collaborative training through an improved objective function. Experiments reveal and leverage the mutually beneficial interaction among recognition tasks, and integrating recognition data significantly enhances detection performance. The final results demonstrate that DocFusion achieves state-of-the-art (SOTA) performance across four key tasks.