Toward Generalizable Evaluation in the LLM Era: A Survey Beyond Benchmarks

Large Language Models (LLMs) are advancing at an amazing speed and have become indispensable across academia, industry, and daily applications. To keep pace with the status quo, this survey probes the core challenges that the rise of LLMs poses for evaluation. We identify and analyze two pivotal transitions: (i) from task-specific to capability-based evaluation, which reorganizes benchmarks around core competencies such as knowledge, reasoning, instruction following, multi-modal understanding, and safety; and (ii) from manual to automated evaluation, encompassing dynamic dataset curation and "LLM-as-a-judge" scoring. Yet, even with these transitions, a crucial obstacle persists: the evaluation generalization issue. Bounded test sets cannot scale alongside models whose abilities grow seemingly without limit. We will dissect this issue, along with the core challenges of the above two transitions, from the perspectives of methods, datasets, evaluators, and metrics. Due to the fast evolving of this field, we will maintain a living GitHub repository (links are in each section) to crowd-source updates and corrections, and warmly invite contributors and collaborators.

MetaScale: Test-Time Scaling with Evolving Meta-Thoughts

One critical challenge for large language models (LLMs) for making complex reasoning is their reliance on matching reasoning patterns from training data, instead of proactively selecting the most appropriate cognitive strategy to solve a given task. Existing approaches impose fixed cognitive structures that enhance performance in specific tasks but lack adaptability across diverse scenarios. To address this limitation, we introduce METASCALE, a test-time scaling framework based on meta-thoughts -- adaptive thinking strategies tailored to each task. METASCALE initializes a pool of candidate meta-thoughts, then iteratively selects and evaluates them using a multi-armed bandit algorithm with upper confidence bound selection, guided by a reward model. To further enhance adaptability, a genetic algorithm evolves high-reward meta-thoughts, refining and extending the strategy pool over time. By dynamically proposing and optimizing meta-thoughts at inference time, METASCALE improves both accuracy and generalization across a wide range of tasks. Experimental results demonstrate that MetaScale consistently outperforms standard inference approaches, achieving an 11% performance gain in win rate on Arena-Hard for GPT-4o, surpassing o1-mini by 0.9% under style control. Notably, METASCALE scales more effectively with increasing sampling budgets and produces more structured, expert-level responses.

Making Every Step Effective: Jailbreaking Large Vision-Language Models Through Hierarchical KV Equalization

In the realm of large vision-language models (LVLMs), adversarial jailbreak attacks serve as a red-teaming approach to identify safety vulnerabilities of these models and their associated defense mechanisms. However, we identify a critical limitation: not every adversarial optimization step leads to a positive outcome, and indiscriminately accepting optimization results at each step may reduce the overall attack success rate. To address this challenge, we introduce HKVE (Hierarchical Key-Value Equalization), an innovative jailbreaking framework that selectively accepts gradient optimization results based on the distribution of attention scores across different layers, ensuring that every optimization step positively contributes to the attack. Extensive experiments demonstrate HKVE's significant effectiveness, achieving attack success rates of 75.08% on MiniGPT4, 85.84% on LLaVA and 81.00% on Qwen-VL, substantially outperforming existing methods by margins of 20.43\%, 21.01\% and 26.43\% respectively. Furthermore, making every step effective not only leads to an increase in attack success rate but also allows for a reduction in the number of iterations, thereby lowering computational costs. Warning: This paper contains potentially harmful example data.

Semantic-Clipping: Efficient Vision-Language Modeling with Semantic-Guidedd Visual Selection

Vision-Language Models (VLMs) leverage aligned visual encoders to transform images into visual tokens, allowing them to be processed similarly to text by the backbone large language model (LLM). This unified input paradigm enables VLMs to excel in vision-language tasks such as visual question answering (VQA). To improve fine-grained visual reasoning, recent advancements in vision-language modeling introduce image cropping techniques that feed all encoded sub-images into the model. However, this approach significantly increases the number of visual tokens, leading to inefficiency and potential distractions for the LLM. To address the generalization challenges of image representation in VLMs, we propose a lightweight, universal framework that seamlessly integrates with existing VLMs to enhance their ability to process finegrained details. Our method leverages textual semantics to identify key visual areas, improving VQA performance without requiring any retraining of the VLM. Additionally, it incorporates textual signals into the visual encoding process, enhancing both efficiency and effectiveness. The proposed method, SEMCLIP, strengthens the visual understanding of a 7B VLM, LLaVA-1.5 by 3.3% on average across 7 benchmarks, and particularly by 5.3% on the challenging detailed understanding benchmark V*.

MapQA: Open-domain Geospatial Question Answering on Map Data

Geospatial question answering (QA) is a fundamental task in navigation and point of interest (POI) searches. While existing geospatial QA datasets exist, they are limited in both scale and diversity, often relying solely on textual descriptions of geo-entities without considering their geometries. A major challenge in scaling geospatial QA datasets for reasoning lies in the complexity of geospatial relationships, which require integrating spatial structures, topological dependencies, and multi-hop reasoning capabilities that most text-based QA datasets lack. To address these limitations, we introduce MapQA, a novel dataset that not only provides question-answer pairs but also includes the geometries of geo-entities referenced in the questions. MapQA is constructed using SQL query templates to extract question-answer pairs from OpenStreetMap (OSM) for two study regions: Southern California and Illinois. It consists of 3,154 QA pairs spanning nine question types that require geospatial reasoning, such as neighborhood inference and geo-entity type identification. Compared to existing datasets, MapQA expands both the number and diversity of geospatial question types. We explore two approaches to tackle this challenge: (1) a retrieval-based language model that ranks candidate geo-entities by embedding similarity, and (2) a large language model (LLM) that generates SQL queries from natural language questions and geo-entity attributes, which are then executed against an OSM database. Our findings indicate that retrieval-based methods effectively capture concepts like closeness and direction but struggle with questions that require explicit computations (e.g., distance calculations). LLMs (e.g., GPT and Gemini) excel at generating SQL queries for one-hop reasoning but face challenges with multi-hop reasoning, highlighting a key bottleneck in advancing geospatial QA systems.

ThinkGuard: Deliberative Slow Thinking Leads to Cautious Guardrails

Ensuring the safety of large language models (LLMs) is critical as they are deployed in real-world applications. Existing guardrails rely on rule-based filtering or single-pass classification, limiting their ability to handle nuanced safety violations. To address this, we propose ThinkGuard, a critique-augmented guardrail model that distills knowledge from high-capacity LLMs by generating structured critiques alongside safety labels. Fine-tuned on critique-augmented data, the captured deliberative thinking ability drastically enhances the guardrail's cautiousness and interpretability. Evaluated on multiple safety benchmarks, ThinkGuard achieves the highest average F1 and AUPRC, outperforming all baselines. Compared to LLaMA Guard 3, ThinkGuard improves accuracy by 16.1% and macro F1 by 27.0%. Moreover, it surpasses label-only fine-tuned models, confirming that structured critiques enhance both classification precision and nuanced safety reasoning while maintaining computational efficiency.

AGrail: A Lifelong Agent Guardrail with Effective and Adaptive Safety Detection

The rapid advancements in Large Language Models (LLMs) have enabled their deployment as autonomous agents for handling complex tasks in dynamic environments. These LLMs demonstrate strong problem-solving capabilities and adaptability to multifaceted scenarios. However, their use as agents also introduces significant risks, including task-specific risks, which are identified by the agent administrator based on the specific task requirements and constraints, and systemic risks, which stem from vulnerabilities in their design or interactions, potentially compromising confidentiality, integrity, or availability (CIA) of information and triggering security risks. Existing defense agencies fail to adaptively and effectively mitigate these risks. In this paper, we propose AGrail, a lifelong agent guardrail to enhance LLM agent safety, which features adaptive safety check generation, effective safety check optimization, and tool compatibility and flexibility. Extensive experiments demonstrate that AGrail not only achieves strong performance against task-specific and system risks but also exhibits transferability across different LLM agents' tasks.

Unraveling Indirect In-Context Learning Using Influence Functions

This work introduces a novel paradigm for generalized In-Context Learning (ICL), termed Indirect In-Context Learning. In Indirect ICL, we explore demonstration selection strategies tailored for two distinct real-world scenarios: Mixture of Tasks and Noisy Demonstrations. We systematically evaluate the effectiveness of Influence Functions (IFs) as a selection tool for these settings, highlighting the potential for IFs to better capture the informativeness of examples within the demonstration pool. For the Mixture of Tasks setting, demonstrations are drawn from 28 diverse tasks, including MMLU, BigBench, StrategyQA, and CommonsenseQA. We demonstrate that combining BertScore-Recall (BSR) with an IF surrogate model can significantly improve performance, leading to average absolute accuracy gains of 0.37\% and 1.45\% for 3-shot and 5-shot setups when compared to traditional ICL metrics. In the Noisy Demonstrations setting, we examine scenarios where demonstrations might be mislabeled. Our experiments show that reweighting traditional ICL selectors (BSR and Cosine Similarity) with IF-based selectors boosts accuracy by an average of 2.90\% for Cosine Similarity and 2.94\% for BSR on noisy GLUE benchmarks. In sum, we propose a robust framework for demonstration selection that generalizes beyond traditional ICL, offering valuable insights into the role of IFs for Indirect ICL.

MetaScientist: A Human-AI Synergistic Framework for Automated Mechanical Metamaterial Design

The discovery of novel mechanical metamaterials, whose properties are dominated by their engineered structures rather than chemical composition, is a knowledge-intensive and resource-demanding process. To accelerate the design of novel metamaterials, we present MetaScientist, a human-in-the-loop system that integrates advanced AI capabilities with expert oversight with two primary phases: (1) hypothesis generation, where the system performs complex reasoning to generate novel and scientifically sound hypotheses, supported with domain-specific foundation models and inductive biases retrieved from existing literature; (2) 3D structure synthesis, where a 3D structure is synthesized with a novel 3D diffusion model based on the textual hypothesis and refined it with a LLM-based refinement model to achieve better structure properties. At each phase, domain experts iteratively validate the system outputs, and provide feedback and supplementary materials to ensure the alignment of the outputs with scientific principles and human preferences. Through extensive evaluation from human scientists, MetaScientist is able to deliver novel and valid mechanical metamaterial designs that have the potential to be highly impactful in the metamaterial field.

SleeperMark: Towards Robust Watermark against Fine-Tuning Text-to-image Diffusion Models

Recent advances in large-scale text-to-image (T2I) diffusion models have enabled a variety of downstream applications, including style customization, subject-driven personalization, and conditional generation. As T2I models require extensive data and computational resources for training, they constitute highly valued intellectual property (IP) for their legitimate owners, yet making them incentive targets for unauthorized fine-tuning by adversaries seeking to leverage these models for customized, usually profitable applications. Existing IP protection methods for diffusion models generally involve embedding watermark patterns and then verifying ownership through generated outputs examination, or inspecting the model's feature space. However, these techniques are inherently ineffective in practical scenarios when the watermarked model undergoes fine-tuning, and the feature space is inaccessible during verification ((i.e., black-box setting). The model is prone to forgetting the previously learned watermark knowledge when it adapts to a new task. To address this challenge, we propose SleeperMark, a novel framework designed to embed resilient watermarks into T2I diffusion models. SleeperMark explicitly guides the model to disentangle the watermark information from the semantic concepts it learns, allowing the model to retain the embedded watermark while continuing to be fine-tuned to new downstream tasks. Our extensive experiments demonstrate the effectiveness of SleeperMark across various types of diffusion models, including latent diffusion models (e.g., Stable Diffusion) and pixel diffusion models (e.g., DeepFloyd-IF), showing robustness against downstream fine-tuning and various attacks at both the image and model levels, with minimal impact on the model's generative capability. The code is available at https://github.com/taco-group/SleeperMark.