VLRewardBench: A Challenging Benchmark for Vision-Language Generative Reward Models

Vision-language generative reward models (VL-GenRMs) play a crucial role in aligning and evaluating multimodal AI systems, yet their own evaluation remains under-explored. Current assessment methods primarily rely on AI-annotated preference labels from traditional VL tasks, which can introduce biases and often fail to effectively challenge state-of-the-art models. To address these limitations, we introduce VL-RewardBench, a comprehensive benchmark spanning general multimodal queries, visual hallucination detection, and complex reasoning tasks. Through our AI-assisted annotation pipeline combining sample selection with human verification, we curate 1,250 high-quality examples specifically designed to probe model limitations. Comprehensive evaluation across 16 leading large vision-language models, demonstrates VL-RewardBench's effectiveness as a challenging testbed, where even GPT-4o achieves only 65.4% accuracy, and state-of-the-art open-source models such as Qwen2-VL-72B, struggle to surpass random-guessing. Importantly, performance on VL-RewardBench strongly correlates (Pearson's r > 0.9) with MMMU-Pro accuracy using Best-of-N sampling with VL-GenRMs. Analysis experiments uncover three critical insights for improving VL-GenRMs: (i) models predominantly fail at basic visual perception tasks rather than reasoning tasks; (ii) inference-time scaling benefits vary dramatically by model capacity; and (iii) training VL-GenRMs to learn to judge substantially boosts judgment capability (+14.7% accuracy for a 7B VL-GenRM). We believe VL-RewardBench along with the experimental insights will become a valuable resource for advancing VL-GenRMs.

Stronger Models are NOT Stronger Teachers for Instruction Tuning

Instruction tuning has been widely adopted to ensure large language models (LLMs) follow user instructions effectively. The resulting instruction-following capabilities of LLMs heavily rely on the instruction datasets used for tuning. Recently, synthetic instruction datasets have emerged as an economically viable solution to provide LLMs diverse and high-quality instructions. However, existing approaches typically assume that larger or stronger models are stronger teachers for instruction tuning, and hence simply adopt these models as response generators to the synthetic instructions. In this paper, we challenge this commonly-adopted assumption. Our extensive experiments across five base models and twenty response generators reveal that larger and stronger models are not necessarily stronger teachers of smaller models. We refer to this phenomenon as the Larger Models' Paradox. We observe that existing metrics cannot precisely predict the effectiveness of response generators since they ignore the compatibility between teachers and base models being fine-tuned. We thus develop a novel metric, named as Compatibility-Adjusted Reward (CAR) to measure the effectiveness of response generators. Our experiments across five base models demonstrate that CAR outperforms almost all baselines.

On Memorization of Large Language Models in Logical Reasoning

Large language models (LLMs) achieve good performance on challenging reasoning benchmarks, yet could also make basic reasoning mistakes. This contrasting behavior is puzzling when it comes to understanding the mechanisms behind LLMs' reasoning capabilities. One hypothesis is that the increasingly high and nearly saturated performance on common reasoning benchmarks could be due to the memorization of similar problems. In this paper, we systematically investigate this hypothesis with a quantitative measurement of memorization in reasoning tasks, using a dynamically generated logical reasoning benchmark based on Knights and Knaves (K&K) puzzles. We found that LLMs could interpolate the training puzzles (achieving near-perfect accuracy) after fine-tuning, yet fail when those puzzles are slightly perturbed, suggesting that the models heavily rely on memorization to solve those training puzzles. On the other hand, we show that while fine-tuning leads to heavy memorization, it also consistently improves generalization performance. In-depth analyses with perturbation tests, cross difficulty-level transferability, probing model internals, and fine-tuning with wrong answers suggest that the LLMs learn to reason on K&K puzzles despite training data memorization. This phenomenon indicates that LLMs exhibit a complex interplay between memorization and genuine reasoning abilities. Finally, our analysis with per-sample memorization score sheds light on how LLMs switch between reasoning and memorization in solving logical puzzles. Our code and data are available at https://memkklogic.github.io.

Latent Action Pretraining from Videos

We introduce Latent Action Pretraining for general Action models (LAPA), an unsupervised method for pretraining Vision-Language-Action (VLA) models without ground-truth robot action labels. Existing Vision-Language-Action models require action labels typically collected by human teleoperators during pretraining, which significantly limits possible data sources and scale. In this work, we propose a method to learn from internet-scale videos that do not have robot action labels. We first train an action quantization model leveraging VQ-VAE-based objective to learn discrete latent actions between image frames, then pretrain a latent VLA model to predict these latent actions from observations and task descriptions, and finally finetune the VLA on small-scale robot manipulation data to map from latent to robot actions. Experimental results demonstrate that our method significantly outperforms existing techniques that train robot manipulation policies from large-scale videos. Furthermore, it outperforms the state-of-the-art VLA model trained with robotic action labels on real-world manipulation tasks that require language conditioning, generalization to unseen objects, and semantic generalization to unseen instructions. Training only on human manipulation videos also shows positive transfer, opening up the potential for leveraging web-scale data for robotics foundation model.

CulturalBench: a Robust, Diverse and Challenging Benchmark on Measuring the (Lack of) Cultural Knowledge of LLMs

To make large language models (LLMs) more helpful across diverse cultures, it is essential to have effective cultural knowledge benchmarks to measure and track our progress. Effective benchmarks need to be robust, diverse, and challenging. We introduce CulturalBench: a set of 1,227 human-written and human-verified questions for effectively assessing LLMs' cultural knowledge, covering 45 global regions including the underrepresented ones like Bangladesh, Zimbabwe, and Peru. Questions - each verified by five independent annotators - span 17 diverse topics ranging from food preferences to greeting etiquettes. We evaluate models on two setups: CulturalBench-Easy and CulturalBench-Hard which share the same questions but asked differently. We find that LLMs are sensitive to such difference in setups (e.g., GPT-4o with 27.3% difference). Compared to human performance (92.6% accuracy), CulturalBench-Hard is more challenging for frontier LLMs with the best performing model (GPT-4o) at only 61.5% and the worst (Llama3-8b) at 21.4%. Moreover, we find that LLMs often struggle with tricky questions that have multiple correct answers (e.g., What utensils do the Chinese usually use?), revealing a tendency to converge to a single answer. Our results also indicate that OpenAI GPT-4o substantially outperform other proprietary and open source models in questions related to all but one region (Oceania). Nonetheless, all models consistently underperform on questions related to South America and the Middle East.

Visual Perception in Text Strings

Understanding visual semantics embedded in consecutive characters is a crucial capability for both large language models (LLMs) and multi-modal large language models (MLLMs). This type of artifact possesses the unique characteristic that identical information can be readily formulated in both texts and images, making them a significant proxy for analyzing modern LLMs' and MLLMs' capabilities in modality-agnostic vision understanding. In this work, we select ASCII art as a representative artifact, where the lines and brightness used to depict each concept are rendered by characters, and we frame the problem as an ASCII art recognition task. We benchmark model performance on this task by constructing an evaluation dataset with an elaborate categorization tree and also collect a training set to elicit the models' visual perception ability. Through a comprehensive analysis of dozens of models, results reveal that although humans can achieve nearly 100% accuracy, the state-of-the-art LLMs and MLLMs lag far behind. Models are capable of recognizing concepts depicted in the ASCII arts given only text inputs indicated by over 60% accuracy for some concepts, but most of them achieves merely around 30% accuracy when averaged across all categories. When provided with images as inputs, GPT-4o gets 82.68%, outperforming the strongest open-source MLLM by 21.95%. Although models favor different kinds of ASCII art depending on the modality provided, none of the MLLMs successfully benefit when both modalities are supplied simultaneously. Moreover, supervised fine-tuning helps improve models' accuracy especially when provided with the image modality, but also highlights the need for better training techniques to enhance the information fusion among modalities.

HAICOSYSTEM: An Ecosystem for Sandboxing Safety Risks in Human-AI Interactions

AI agents are increasingly autonomous in their interactions with human users and tools, leading to increased interactional safety risks. We present HAICOSYSTEM, a framework examining AI agent safety within diverse and complex social interactions. HAICOSYSTEM features a modular sandbox environment that simulates multi-turn interactions between human users and AI agents, where the AI agents are equipped with a variety of tools (e.g., patient management platforms) to navigate diverse scenarios (e.g., a user attempting to access other patients' profiles). To examine the safety of AI agents in these interactions, we develop a comprehensive multi-dimensional evaluation framework that uses metrics covering operational, content-related, societal, and legal risks. Through running 1840 simulations based on 92 scenarios across seven domains (e.g., healthcare, finance, education), we demonstrate that HAICOSYSTEM can emulate realistic user-AI interactions and complex tool use by AI agents. Our experiments show that state-of-the-art LLMs, both proprietary and open-sourced, exhibit safety risks in over 50\% cases, with models generally showing higher risks when interacting with simulated malicious users. Our findings highlight the ongoing challenge of building agents that can safely navigate complex interactions, particularly when faced with malicious users. To foster the AI agent safety ecosystem, we release a code platform that allows practitioners to create custom scenarios, simulate interactions, and evaluate the safety and performance of their agents.

SimulBench: Evaluating Language Models with Creative Simulation Tasks

We introduce SimulBench, a benchmark designed to evaluate large language models (LLMs) across a diverse collection of creative simulation scenarios, such as acting as a Linux terminal or playing text games with users. While these simulation tasks serve as effective measures of an LLM's general intelligence, they are seldom incorporated into existing benchmarks. A major challenge is to develop an evaluation framework for testing different LLMs fairly while preserving the multi-round interactive nature of simulation tasks between users and AI. To tackle this issue, we suggest using a fixed LLM as a user agent to engage with an LLM to collect dialogues first under different tasks. Then, challenging dialogue scripts are extracted for evaluating different target LLMs. To facilitate automatic assessment on \DataName{}, GPT-4 is employed as the evaluator, tasked with reviewing the quality of the final response generated by the target LLMs given multi-turn dialogue scripts. Our comprehensive experiments indicate that these simulation tasks continue to pose a significant challenge with their unique natures and show the gap between proprietary models and the most advanced open LLMs. For example, GPT-4-turbo outperforms LLaMA-3-70b-Chat on 18.55\% more cases.

OfficeBench: Benchmarking Language Agents across Multiple Applications for Office Automation

Office automation significantly enhances human productivity by automatically finishing routine tasks in the workflow. Beyond the basic information extraction studied in much of the prior document AI literature, the office automation research should be extended to more realistic office tasks which require to integrate various information sources in the office system and produce outputs through a series of decision-making processes. We introduce OfficeBench, one of the first office automation benchmarks for evaluating current LLM agents' capability to address office tasks in realistic office workflows. OfficeBench requires LLM agents to perform feasible long-horizon planning, proficiently switch between applications in a timely manner, and accurately ground their actions within a large combined action space, based on the contextual demands of the workflow. Applying our customized evaluation methods on each task, we find that GPT-4 Omni achieves the highest pass rate of 47.00%, demonstrating a decent performance in handling office tasks. However, this is still far below the human performance and accuracy standards required by real-world office workflows. We further observe that most issues are related to operation redundancy and hallucinations, as well as limitations in switching between multiple applications, which may provide valuable insights for developing effective agent frameworks for office automation.

The Good, The Bad, and The Greedy: Evaluation of LLMs Should Not Ignore Non-Determinism

Current evaluations of large language models (LLMs) often overlook non-determinism, typically focusing on a single output per example. This limits our understanding of LLM performance variability in real-world applications. Our study addresses this issue by exploring key questions about the performance differences between greedy decoding and sampling, identifying benchmarks' consistency regarding non-determinism, and examining unique model behaviors. Through extensive experiments, we observe that greedy decoding generally outperforms sampling methods for most evaluated tasks. We also observe consistent performance across different LLM sizes and alignment methods, noting that alignment can reduce sampling variance. Moreover, our best-of-N sampling approach demonstrates that smaller LLMs can match or surpass larger models such as GPT-4-Turbo, highlighting the untapped potential of smaller LLMs. This research shows the importance of considering non-determinism in LLM evaluations and provides insights for future LLM development and evaluation.