模型即裁判

Tool use represents a critical capability for AI agents, with recent advances focusing on leveraging reinforcement learning (RL) for test-time scaling to achieve better performance through more deliberate reasoning. However, there are some key challenges in current RL-based scaling approaches: (a) direct RL training often struggles to scale up thinking length sufficiently to solve complex problems, and (b) scaled-up models tend to overthink simpler problems, resulting in substantial token inefficiency. To address these challenges, we propose a novel training paradigm that first employs warm-up supervised fine-tuning to help models distinguish between simple and complex problems, followed by RL that enable models to automatically determine appropriate reasoning trajectories. Furthermore, to tackle the issue of automatic thinking-length scaling, we discover that entropy-based optimization objectives effectively maintain model diversity while successfully unlocking the model's scaling capabilities. Based on this insight, we introduce an entropy-based long-short reasoning fusion RL strategy. Our experiments on three benchmarks demonstrate that model successfully achieves auto-scaling for efficient tool use, achieving significant 9.8\% accuracy improvements while reducing computational overhead by ~81\%.

Fulfilling user needs through Large Language Model multi-turn, multi-step tool-use is rarely a straightforward process. Real user interactions are inherently $\textbf{wild}$, being intricate, messy, and flexible. We identify three key challenges from user behaviour: $\textit{compositional tasks}$ that demand efficient orchestration of tool-call topologies, $\textit{implicit intent}$ spread across dialogue turns that require contextual inference, and $\textit{instruction transition}$, which mixes task queries, clarifications, and casual conversation, forcing LLMs to adjust their policies on the fly. Existing benchmarks overlook these behaviors, making the apparent progress of LLMs on tool-use spurious. To address this, we introduce $\textbf{\textit{WildToolBench}}$, an LLM tool-use benchmark grounded in real-world user behavior patterns. Comprehensive evaluations of 57 LLMs reveal that no model achieves an accuracy of more than 15\%, indicating a substantial gap in the robustness of LLMs' agentic ability. Controlled experiments and in-depth analyses further indicate that the real challenge for LLM tool-use lies not in artificially complex tasks, but in the wild nature of user behavior, emphasizing the need to reconsider the interactions among $\textit{LLMs}$, $\textit{users}$, and $\textit{tools}$.

Evaluating user-facing AI applications remains a central challenge, especially in open-ended domains such as travel planning, clinical note generation, or dialogue. The gold standard is user feedback (e.g., thumbs up/down) or behavioral signals (e.g., retention), but these are often scarce in prototypes and research projects, or too-slow to use for system optimization. We present **AutoMetrics**, a framework for synthesizing evaluation metrics under low-data constraints. AutoMetrics combines retrieval from **MetricBank**, a collection of 48 metrics we curate, with automatically generated LLM-as-a-Judge criteria informed by lightweight human feedback. These metrics are composed via regression to maximize correlation with human signal. AutoMetrics takes you from expensive measures to interpretable automatic metrics. Across 5 diverse tasks, AutoMetrics improves Kendall correlation with human ratings by up to 33.4% over LLM-as-a-Judge while requiring fewer than 100 feedback points. We show that AutoMetrics can be used as a proxy reward to equal effect as a verifiable reward. We release the full AutoMetrics toolkit and MetricBank to accelerate adaptive evaluation of LLM applications.

Reward is critical to the evaluation and training of large language models (LLMs). However, existing rule-based or model-based reward methods struggle to generalize to GUI agents, where access to ground-truth trajectories or application databases is often unavailable, and static trajectory-based LLM-as-a-Judge approaches suffer from limited accuracy. To address these challenges, we propose ProRe, a proactive reward system that leverages a general-purpose reasoner and domain-specific evaluator agents (actors). The reasoner schedules targeted state probing tasks, which the evaluator agents then execute by actively interacting with the environment to collect additional observations. This enables the reasoner to assign more accurate and verifiable rewards to GUI agents. Empirical results on over 3K trajectories demonstrate that ProRe improves reward accuracy and F1 score by up to 5.3\% and 19.4\%, respectively. Furthermore, integrating ProRe with state-of-the-art policy agents yields a success rate improvement of up to 22.4\%.

Large language models (LLMs) are increasingly deployed in contexts where their failures have the potential to carry sociopolitical consequences. However, existing safety benchmarks sparsely test vulnerabilities in domains such as political manipulation, propaganda generation, or surveillance and information control. To address this gap, we propose SocialHarmBench, a dataset of 585 prompts spanning 7 sociopolitical categories and 34 countries with real-world events, designed to evaluate LLM vulnerabilities to sociopolitical harms. Using SocialHarmBench, we provide: (1) adversarial evaluation coverage of high-risk domains including authoritarian surveillance, disinformation campaigns, erosion of democratic processes, and crimes against humanity; (2) adversarial evaluations across open-source models, establishing baseline robustness and measuring attack efficiency in politically charged settings; and (3) insights into domain-specific vulnerability comparisons, temporal-wide investigations to trace vulnerable time periods, and region-specific vulnerabilities. Our findings reveal that existing safeguards fail to transfer effectively to sociopolitical contexts, exposing partisan biases and limitations in preserving human rights and democratic values.

Data curriculums have become central to successful LLM training, yet principles governing optimal data placement remain unclear. We introduce the *training re-evaluation curve (TREC)*, a diagnostic that retrospectively evaluates training batches *using the final model weights*. The TREC characterizes how well a trained model retains training data as a function of *when* the data was encountered during training. Analyzing TRECs for models from 111M to 3.9B parameters, we show that placing high-quality data at low points on the TREC significantly improves performance. Importantly, while a TREC is initially observable only after training, we demonstrate it can be *predicted in advance* from AdamW’s implicit EMA coefficients, enabling proactive curriculum design. By predicting TRECs for published training recipes, we explain prior ablations and reveal suboptimal data placements. We also align high-quality data with TREC minima in order to improve continual pre-training of a 3.9B-parameter LLM trained on 900B tokens.

The widespread deployment of large language models (LLMs) across linguistic communities necessitates reliable multilingual safety alignment. However, recent efforts to extend alignment to other languages often require substantial resources, either through large-scale, high-quality supervision in the target language or through pairwise alignment with high-resource languages, which limits scalability. In this work, we propose a resource-efficient method for improving multilingual safety alignment. We introduce a plug-and-play Multi-Lingual Consistency (MLC) loss that can be integrated into existing monolingual alignment pipelines. By improving collinearity between multilingual representation vectors, our method encourages directional consistency at the multilingual semantic level in a single update. This allows simultaneous alignment across multiple languages using only multilingual prompt variants without requiring additional response-level supervision in low-resource languages. We validate the proposed method across different model architectures and alignment paradigms, and demonstrate its effectiveness in enhancing multilingual safety with limited impact on general model utility. Further evaluation across languages and tasks indicates improved cross-lingual generalization, suggesting the proposed approach as a practical solution for multilingual consistency alignment under limited supervision.

Most LLM fingerprinting methods teach the model to respond to a few fixed queries with predefined atypical responses (keys). This memorization often does not survive common deployment steps such as finetuning or quantization, and such keys can be easily detected and filtered from LLM responses, ultimately breaking the fingerprint. To overcome these limitations we introduce *LLM fingerprinting via semantically conditioned watermarks*, replacing fixed query sets with a broad semantic domain, and replacing brittle atypical keys with a statistical watermarking signal diffused throughout each response. After teaching the model to watermark its responses only to prompts from a predetermined domain e.g., French language, the model owner can use queries from that domain to reliably detect the fingerprint and verify ownership. As we confirm in our thorough experimental evaluation, our fingerprint is both stealthy and robust to all common deployment scenarios.

In language tasks requiring extensive human-model interaction, the inference cost of large language models (LLMs) can be substantial. To reduce expenses while preserving the quality of the responses, an LLM router selects among candidate models to balance between the expected response quality and the inference cost. A central challenge in router training is the accuracy and accessibility of reliable supervision. Gold-standard data, obtained from domain experts or benchmark labels, provide accurate quality evaluations of LLM responses but are costly and difficult to scale. In contrast, preference-based data, collected via crowdsourcing or LLM-as-a-judge systems, are cheaper and more scalable, yet often biased in reflecting the true quality of responses. We cast the problem of LLM router training with combined Gold-standard and preference-based data into a causal inference framework by viewing the response evaluation mechanism as the treatment assignment. This perspective further reveals that the bias in preference-based data corresponds to the well-known causal estimand: the conditional average treatment effect (CATE). Based on this new perspective, we develop an integrative causal router training framework that corrects preference-data bias, addresses imbalances between two data sources, and improves routing robustness and efficiency. Numerical experiments demonstrate that our approach delivers more accurate routing and improves the trade-off between cost and quality.

As the demand for comprehensive evaluations of diverse model capabilities steadily increases, benchmark suites have correspondingly grown significantly in scale. Despite notable advances in redundancy reduction and subset-level performance prediction, a systematic framework that effectively integrates these methods to ensure both prediction accuracy and ranking consistency is still largely elusive. In this paper, we first perform a sample-level analysis of benchmark redundancy and identify several highly similar samples that can be eliminated. Besides, we frame benchmark compression as an optimization problem with the aim of score reconstruction. Building on these, we then propose EssenceBench, a coarse-to-fine framework utilizing an iterative Genetic Algorithm (GA), which takes the advantages of fitness-based subset search and attribution-based sample search. Compared to previous methods, our approach yields superior compression results with lower reconstruction error and markedly higher efficiency. In particular, on the HellaSwag benchmark (10K samples), our method preserves the ranking of all models shifting within 5% using 25$\times$ fewer samples, and achieves 95% ranking preservation shifting within 5% using only 200$\times$ fewer samples. The source code will be made available upon acceptance of the paper.

Recent advances in large language models (LLMs) for mathematical reasoning have largely focused on tasks with easily verifiable final answers while generating and verifying natural language math proofs remains an open challenge. We identify the absence of a reliable, fine-grained evaluator for LLM-generated math proofs as a critical gap. To address this, we propose a systematic methodology for developing and validating evaluators that assign fine-grained scores on a 0-7 scale to model-generated math proofs. To enable this study, we introduce ProofBench, the first expert-annotated dataset of fine-grained proof ratings, spanning 145 problems from six major math competitions (USAMO, IMO, Putnam, etc) and 435 LLM-generated solutions from Gemini-2.5-Pro, o3, and DeepSeek-R1. Using ProofBench as a testbed, we systematically explore the evaluator design space across key axes: the backbone model, input context, instructions and evaluation workflow. Our analysis delivers ProofGrader, an evaluator that combines a strong reasoning backbone LM, rich context from reference solutions and marking schemes, and a simple ensembling method; it achieves a low Mean Absolute Error (MAE) of 0.926 against expert scores, significantly outperforming naive baselines. Finally, we demonstrate its practical utility in a best-of-$n$ selection task: at $n=16$, ProofGrader achieves an average score of 4.14/7, closing 78\% of the gap between a naive binary evaluator (2.48) and the human oracle (4.62), highlighting its potential to advance downstream proof generation.

As large language models (LLMs) are increasingly applied to scientific reasoning, the complexity of answer formats and the diversity of equivalent expressions make answer verification a critical yet challenging task. Existing verification studies in scientific domains suffer from two major limitations: (a) the absence of systematic evaluation standards and insufficient disciplinary coverage, which hinders their comprehensive assessment; and (b) heavy reliance on cumbersome rule design or prompt engineering, which reduces their effectiveness in complex reasoning scenarios or limits their cross-disciplinary generalization. To address these challenges, we propose solutions at both the data and model levels. On the data side, we construct **SCI-VerifyBench**, a cross-disciplinary benchmark covering mathematics, physics, biology, chemistry, and general scientific QA. The benchmark is built from real LLM responses and enhanced with domain-specific equivalence transformations that generate challenging and realistic data. Model-based and expert annotations ensure both quality and diversity, enabling rigorous evaluation of verification ability. On the model side, we emphasize the importance of reasoning for verification and introduce **SCI-Verifier**, a unified reasoning-augmented verifier for scientific domains. Through post-training, SCI-Verifier demonstrates strong logical reasoning and equivalence judgment capabilities while maintaining concise and stable outputs. Together, SCI-VerifyBench and SCI-Verifier provide a principled framework for scientific verification, offering both systematic evaluation and practical pathways to enhance the reliability and applicability of LLMs in scientific domains.

Large Language Models (LLMs) have emerged as a promising approach for binary decompilation. However, the existing LLM-based decompilers still are somewhat limited in effectively presenting a program's source-level structure with its original identifiers. To mitigate this, we introduce SK2Decompile, a novel two-phase approach to decompile from the skeleton (semantic structure) to the skin (identifier) of programs. Specifically, we first apply a Structure Recovery model to translate a program's binary code to an Intermediate Representation (IR) as deriving the program's "skeleton", i.e., preserving control flow and data structures while obfuscating all identifiers with generic placeholders. We also apply reinforcement learning to reward the model for producing program structures that adhere to the syntactic and semantic rules expected by compilers. Second, we apply an Identifier Naming model to produce meaningful identifiers which reflect actual program semantics as deriving the program's "skin". We train the Identifier Naming model with a separate reinforcement learning objective that rewards the semantic similarity between its predictions and the reference code. Such a two-phase decompilation process facilitates advancing the correctness and readability of decompilation independently. Our evaluations indicate that SK2Decompile, significantly outperforms the SOTA baselines, achieving 21.6% average re-executability rate gain over GPT-5-mini on the HumanEval dataset and 29.4% average R2I improvement over Idioms on the GitHub2025 benchmark.

This paper investigates the application of Large Language Models (LLMs) in Automated Heuristic Design (AHD), where their integration into evolutionary frameworks reveals a significant gap in global control and long-term learning. We propose the Hindsight-Foresight Prompt (HiFo-Prompt), a novel framework for LLM-based AHD designed to overcome these limitations. This is achieved through two synergistic strategies: Foresight and Hindsight. Foresight acts as a high-level meta-controller, monitoring population dynamics(e.g., stagnation and diversity collapse) to switch the global search strategy between exploration and exploitation explicitly. Hindsight builds a persistent knowledge base by distilling successful design principles from past generations, making this knowledge reusable. This dual mechanism ensures that the LLM is not just a passive operator but an active reasoner, guided by a global plan (Foresight) while continuously improving from its cumulative experience (Hindsight). Empirical results demonstrate that HiFo-Prompt significantly outperforms a comprehensive suite of state-of-the-art AHD methods, discovering higher-quality heuristics with substantially improved convergence speed and query efficiency. Our code is available at https://github.com/Challenger-XJTU/HiFo-Prompt.

Reliable certification of Large Language Models (LLMs)—verifying that failure rates are below a safety threshold—is critical yet challenging. While "LLM-as-a-Judge" offers scalability, judge imperfections, noise, and bias can invalidate statistical guarantees. We introduce a "Noisy but Valid" hypothesis testing framework to address this. By leveraging a small human-labelled calibration set to estimate the judge's True Positive and False Positive Rates (TPR/FPR), we derive a variance-corrected critical threshold applied to a large judge-labelled dataset. Crucially, our framework theoretically guarantees finite-sample Type-I error control (validity) despite calibration uncertainty. This distinguishes our work from Prediction-Powered Inference (PPI), positioning our method as a diagnostic tool that explicitly models judge behavior rather than a black-box estimator. Our contributions include: (1) Theoretical Guarantees: We derive the exact conditions under which noisy testing yields higher statistical power than direct evaluation; (2) Empirical Validation: Experiments on Jigsaw Comment, Hate Speech and SafeRLHF confirm our theory; (3) The Oracle Gap: We reveal a significant performance gap between practical methods and the theoretical "Oracle" (perfectly known judge parameters), quantifying the cost of estimation. Specifically, we provide the first systematic treatment of the imperfect-judge setting, yielding interpretable diagnostics of judge reliability and clarifying how evaluation power depends on judge quality, dataset size, and certification levels. Together, these results sharpen understanding of statistical evaluation with LLM judges, and highlight trade-offs among competing inferential tools.

Large Language Models (LLMs) as judges and LLM-based data synthesis have emerged as two fundamental LLM-driven data annotation methods in model development. While their combination significantly enhances the efficiency of model training and evaluation, little attention has been given to the potential contamination brought by this new model development paradigm. In this work, we expose preference leakage, a contamination problem in LLM-as-a-judge caused by the relatedness between the synthetic data generators and LLM-based evaluators. To study this issue, we first define three common relatednesses between the data generator LLM and the judge LLM: being the same model, having an inheritance relationship, and belonging to the same model family. Through extensive experiments, we empirically confirm the bias of judges towards their related student models caused by preference leakage across multiple LLM baselines and benchmarks. Further analysis suggests that preference leakage is a pervasive and real-world problem that is harder to detect compared to previously identified biases in LLM-as-a-judge scenarios. All of these findings imply that preference leakage is a widespread and challenging problem in the area of LLM-as-a-judge.

A core challenge in modern AI model development is obtaining high-quality evaluation metrics in a cost-effective way. Such evaluation often involves tradeoffs between expensive, high-quality measurements and a variety of lower-quality proxies. We introduce Multiple-Prediction-Powered Inference (MultiPPI), a general framework for constructing statistically efficient estimates by optimally allocating resources across these diverse data sources. We provide theoretical guarantees about the minimax optimality, finite-sample performance, and asymptotic normality of the MultiPPI estimator. Through experiments across three diverse large language model (LLM) evaluation scenarios, we show that MultiPPI consistently achieves lower estimation error than existing baselines. This advantage stems from its budget-adaptive allocation strategy, which strategically combines subsets of models by learning their complex cost and correlation structures.

Identifying the vulnerabilities of large language models (LLMs) is crucial for improving their safety by addressing inherent weaknesses. Jailbreaks, in which adversaries bypass safeguards with crafted input prompts, play a central role in red-teaming by probing LLMs to elicit unintended or unsafe behaviors. Recent optimization-based jailbreak approaches iteratively refine attack prompts by leveraging LLMs. However, they often rely heavily on either binary attack success rate (ASR) signals, which are sparse, or manually crafted scoring templates, which introduce human bias and uncertainty in the scoring outcomes. To address these limitations, we introduce AMIS (Align to MISalign), a meta-optimization framework that jointly evolves jailbreak prompts and scoring templates through a bi-level structure. In the inner loop, prompts are refined using fine-grained and dense feedback using fixed scoring template. In the outer loop, the template is optimized using an ASR alignment score, gradually evolving to better reflect true attack outcomes across queries. This co-optimization process yields progressively stronger jailbreak prompts and more calibrated scoring signals. Evaluations on AdvBench and JBB-Behaviors demonstrate that AMIS achieves state-of-the-art performance, including 88.0% ASR on Claude-3.5-Haiku and 100.0% ASR on Claude-4-Sonnet, outperforming existing baselines by substantial margins.

The rise of AI has fueled growing concerns about ``hype'' in machine learning papers, yet a reliable way to quantify rhetorical style independently of substantive content has remained elusive. Because strong empirical results can justify stronger claims, it is often unclear whether bold language reflects genuine evidence or merely rhetorical style. We introduce a counterfactual, LLM-based framework to disentangle rhetorical style from substantive content: multiple LLM rhetorical personas generate counterfactual writings from the same substantive content, an LLM judge compares them through pairwise evaluations, and the outcomes are aggregated using a Bradley--Terry model. Applying this method to 8,485 ICLR submissions sampled from 2017 to 2025, we generate more than 250,000 counterfactual writings and provide a large-scale quantification of rhetorical style in ML papers. Visionary framing significantly predicts downstream attention, including citations and media coverage, even after controlling for peer-review evaluations. We also observe a sharp rise in rhetorical strength after 2023, and provide evidence showing that this increase is strongly correlated with the adoption of LLM writing assistance. The reliability of our framework is validated by its robustness to the choice of personas and the high correlation between LLM judgments and human annotations. Our work demonstrates that LLMs can serve as instruments for improving how ML research is evaluated.

The LLM-as-a-judge paradigm is widely used in both evaluating free-text model responses and reward modeling for model alignment and finetuning. Recently, finetuning judges with judge-specific data has emerged as an often preferred choice over directly prompting frontier models as judges, as the former achieves better performance with smaller model sizes while being more robust to common biases. However, the standard evaluation ignores several practical concerns of finetuned judges regarding their real world deployment. In this paper, we identify and formalize three aspects that affect the *shelf life* of these judges: *future-proofing* and *backward-compatibility* – how well judges finetuned on responses by today's generator models perform on responses by future models or past models, as well as *question generalization* – how well judges generalize to unseen questions at test time. We study these three aspects under a unified framework with varying train and test distributions in two reasoning datasets, three SFT- and DPO-based fine-tuning algorithms, and three different backbone models. Experiments suggest that future-proofing is challenging for most models, while backward-compatibility is relatively easy, with DPO-trained models consistently *improving* performance. We further find that continual learning provides a more balanced adaptation to shifts between older and newer response distributions than training solely on stronger or weaker responses. Moreover, all models observe certain degrees of performance degradation when moving from questions seen during training to unseen ones, showing that current judges do not fully generalize to unseen questions. These findings provide insights into practical considerations for developing and deploying judge models in the face of ever-changing generators.