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.

LLM Can be a Dangerous Persuader: Empirical Study of Persuasion Safety in Large Language Models

Recent advancements in Large Language Models (LLMs) have enabled them to approach human-level persuasion capabilities. However, such potential also raises concerns about the safety risks of LLM-driven persuasion, particularly their potential for unethical influence through manipulation, deception, exploitation of vulnerabilities, and many other harmful tactics. In this work, we present a systematic investigation of LLM persuasion safety through two critical aspects: (1) whether LLMs appropriately reject unethical persuasion tasks and avoid unethical strategies during execution, including cases where the initial persuasion goal appears ethically neutral, and (2) how influencing factors like personality traits and external pressures affect their behavior. To this end, we introduce PersuSafety, the first comprehensive framework for the assessment of persuasion safety which consists of three stages, i.e., persuasion scene creation, persuasive conversation simulation, and persuasion safety assessment. PersuSafety covers 6 diverse unethical persuasion topics and 15 common unethical strategies. Through extensive experiments across 8 widely used LLMs, we observe significant safety concerns in most LLMs, including failing to identify harmful persuasion tasks and leveraging various unethical persuasion strategies. Our study calls for more attention to improve safety alignment in progressive and goal-driven conversations such as persuasion.

Could AI Trace and Explain the Origins of AI-Generated Images and Text?

AI-generated content is becoming increasingly prevalent in the real world, leading to serious ethical and societal concerns. For instance, adversaries might exploit large multimodal models (LMMs) to create images that violate ethical or legal standards, while paper reviewers may misuse large language models (LLMs) to generate reviews without genuine intellectual effort. While prior work has explored detecting AI-generated images and texts, and occasionally tracing their source models, there is a lack of a systematic and fine-grained comparative study. Important dimensions--such as AI-generated images vs. text, fully vs. partially AI-generated images, and general vs. malicious use cases--remain underexplored. Furthermore, whether AI systems like GPT-4o can explain why certain forged content is attributed to specific generative models is still an open question, with no existing benchmark addressing this. To fill this gap, we introduce AI-FAKER, a comprehensive multimodal dataset with over 280,000 samples spanning multiple LLMs and LMMs, covering both general and malicious use cases for AI-generated images and texts. Our experiments reveal two key findings: (i) AI authorship detection depends not only on the generated output but also on the model's original training intent; and (ii) GPT-4o provides highly consistent but less specific explanations when analyzing content produced by OpenAI's own models, such as DALL-E and GPT-4o itself.

LLM Braces: Straightening Out LLM Predictions with Relevant Sub-Updates

Recent findings reveal that much of the knowledge in a Transformer-based Large Language Model (LLM) is encoded in its feed-forward (FFN) layers, where each FNN layer can be interpreted as the summation of sub-updates, each corresponding to a weighted column vector from the FFN's value parameter matrix that often encodes human-interpretable concepts. In light of this, we hypothesize that model performance and behaviors can be further enhanced and controlled by modulating the contributions of these sub-updates based on their relevance to the input or target output style, and propose LLMBRACES, a novel and efficient method that computes relevance scores associated with value vectors in FFN layers and leverages these scores to dynamically adjust the contribution of sub-updates. By optimizing sub-update contributions, LLMBRACES refines the prediction process, leading to more accurate and reliable outputs, much like a 'brace' providing support and stability. Moreover, LLMBRACES can be extended to support conditional control over generation characteristics, such as sentiment, thereby offering fine-grained steering of LLM outputs. Extensive experiments on various LLMs-including Qwen2.5-1.5B, Llama2-7B, and Llama3-8B-demonstrate that LLMBRACES outperforms baseline approaches in both fine-tuning and zero-shot settings while requiring significantly fewer tunable parameters, up to 75% fewer compared to LoRA. Furthermore, LLMBRACES excels in sentiment-controlled generation and toxicity reduction, highlighting its potential for flexible, controlled text generation across applications.

A Survey on Mechanistic Interpretability for Multi-Modal Foundation Models

The rise of foundation models has transformed machine learning research, prompting efforts to uncover their inner workings and develop more efficient and reliable applications for better control. While significant progress has been made in interpreting Large Language Models (LLMs), multimodal foundation models (MMFMs) - such as contrastive vision-language models, generative vision-language models, and text-to-image models - pose unique interpretability challenges beyond unimodal frameworks. Despite initial studies, a substantial gap remains between the interpretability of LLMs and MMFMs. This survey explores two key aspects: (1) the adaptation of LLM interpretability methods to multimodal models and (2) understanding the mechanistic differences between unimodal language models and crossmodal systems. By systematically reviewing current MMFM analysis techniques, we propose a structured taxonomy of interpretability methods, compare insights across unimodal and multimodal architectures, and highlight critical research gaps.

Error-driven Data-efficient Large Multimodal Model Tuning

Large Multimodal Models (LMMs) have demonstrated impressive performance across numerous academic benchmarks. However, fine-tuning still remains essential to achieve satisfactory performance on downstream tasks, while the task-specific tuning samples are usually not readily available or expensive and time-consuming to obtain. To address this, we propose an error-driven data-efficient tuning framework that aims to efficiently adapt generic LMMs to newly emerging tasks without requiring any task-specific training samples. In our approach, a generic LMM, acting as a student model, is first evaluated on a small validation set of the target task, and then a more powerful model, acting as a teacher model, identifies the erroneous steps within the student model's reasoning steps and analyzes its capability gaps from fully addressing the target task. Based on these gaps, targeted training samples are further retrieved from existing task-agnostic datasets to tune the student model and tailor it to the target task. We perform extensive experiments across three different training data scales and seven tasks, demonstrating that our training paradigm significantly and efficiently improves LMM's performance on downstream tasks, achieving an average performance boost of 7.01%.

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.

Persona-SQ: A Personalized Suggested Question Generation Framework For Real-world Documents

Suggested questions (SQs) provide an effective initial interface for users to engage with their documents in AI-powered reading applications. In practical reading sessions, users have diverse backgrounds and reading goals, yet current SQ features typically ignore such user information, resulting in homogeneous or ineffective questions. We introduce a pipeline that generates personalized SQs by incorporating reader profiles (professions and reading goals) and demonstrate its utility in two ways: 1) as an improved SQ generation pipeline that produces higher quality and more diverse questions compared to current baselines, and 2) as a data generator to fine-tune extremely small models that perform competitively with much larger models on SQ generation. Our approach can not only serve as a drop-in replacement in current SQ systems to immediately improve their performance but also help develop on-device SQ models that can run locally to deliver fast and private SQ experience.

Graph-based Confidence Calibration for Large Language Models

One important approach to improving the reliability of large language models (LLMs) is to provide accurate confidence estimations regarding the correctness of their answers. However, developing a well-calibrated confidence estimation model is challenging, as mistakes made by LLMs can be difficult to detect. We propose a novel method combining the LLM's self-consistency with labeled data and training an auxiliary model to estimate the correctness of its responses to questions. This auxiliary model predicts the correctness of responses based solely on their consistent information. To set up the learning problem, we use a weighted graph to represent the consistency among the LLM's multiple responses to a question. Correctness labels are assigned to these responses based on their similarity to the correct answer. We then train a graph neural network to estimate the probability of correct responses. Experiments demonstrate that the proposed approach substantially outperforms several of the most recent methods in confidence calibration across multiple widely adopted benchmark datasets. Furthermore, the proposed approach significantly improves the generalization capability of confidence calibration on out-of-domain (OOD) data.

Visual Fourier Prompt Tuning

With the scale of vision Transformer-based models continuing to grow, finetuning these large-scale pretrained models for new tasks has become increasingly parameter-intensive. Visual prompt tuning is introduced as a parameter-efficient finetuning (PEFT) method to this trend. Despite its successes, a notable research challenge persists within almost all PEFT approaches: significant performance degradation is observed when there is a substantial disparity between the datasets applied in pretraining and finetuning phases. To address this challenge, we draw inspiration from human visual cognition, and propose the Visual Fourier Prompt Tuning (VFPT) method as a general and effective solution for adapting large-scale transformer-based models. Our approach innovatively incorporates the Fast Fourier Transform into prompt embeddings and harmoniously considers both spatial and frequency domain information. Apart from its inherent simplicity and intuitiveness, VFPT exhibits superior performance across all datasets, offering a general solution to dataset challenges, irrespective of data disparities. Empirical results demonstrate that our approach outperforms current state-of-the-art baselines on two benchmarks, with low parameter usage (e.g., 0.57% of model parameters on VTAB-1k) and notable performance enhancements (e.g., 73.20% of mean accuracy on VTAB-1k). Our code is avaliable at https://github.com/runtsang/VFPT.