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.

Stealthy Multi-Task Adversarial Attacks

Deep Neural Networks exhibit inherent vulnerabilities to adversarial attacks, which can significantly compromise their outputs and reliability. While existing research primarily focuses on attacking single-task scenarios or indiscriminately targeting all tasks in multi-task environments, we investigate selectively targeting one task while preserving performance in others within a multi-task framework. This approach is motivated by varying security priorities among tasks in real-world applications, such as autonomous driving, where misinterpreting critical objects (e.g., signs, traffic lights) poses a greater security risk than minor depth miscalculations. Consequently, attackers may hope to target security-sensitive tasks while avoiding non-critical tasks from being compromised, thus evading being detected before compromising crucial functions. In this paper, we propose a method for the stealthy multi-task attack framework that utilizes multiple algorithms to inject imperceptible noise into the input. This novel method demonstrates remarkable efficacy in compromising the target task while simultaneously maintaining or even enhancing performance across non-targeted tasks - a criterion hitherto unexplored in the field. Additionally, we introduce an automated approach for searching the weighting factors in the loss function, further enhancing attack efficiency. Experimental results validate our framework's ability to successfully attack the target task while preserving the performance of non-targeted tasks. The automated loss function weight searching method demonstrates comparable efficacy to manual tuning, establishing a state-of-the-art multi-task attack framework.

Human Motion Instruction Tuning

This paper presents LLaMo (Large Language and Human Motion Assistant), a multimodal framework for human motion instruction tuning. In contrast to conventional instruction-tuning approaches that convert non-linguistic inputs, such as video or motion sequences, into language tokens, LLaMo retains motion in its native form for instruction tuning. This method preserves motion-specific details that are often diminished in tokenization, thereby improving the model's ability to interpret complex human behaviors. By processing both video and motion data alongside textual inputs, LLaMo enables a flexible, human-centric analysis. Experimental evaluations across high-complexity domains, including human behaviors and professional activities, indicate that LLaMo effectively captures domain-specific knowledge, enhancing comprehension and prediction in motion-intensive scenarios. We hope LLaMo offers a foundation for future multimodal AI systems with broad applications, from sports analytics to behavioral prediction. Our code and models are available on the project website: https://github.com/ILGLJ/LLaMo.

DeepSPV: An Interpretable Deep Learning Pipeline for 3D Spleen Volume Estimation from 2D Ultrasound Images

Splenomegaly, the enlargement of the spleen, is an important clinical indicator for various associated medical conditions, such as sickle cell disease (SCD). Spleen length measured from 2D ultrasound is the most widely used metric for characterising spleen size. However, it is still considered a surrogate measure, and spleen volume remains the gold standard for assessing spleen size. Accurate spleen volume measurement typically requires 3D imaging modalities, such as computed tomography or magnetic resonance imaging, but these are not widely available, especially in the Global South which has a high prevalence of SCD. In this work, we introduce a deep learning pipeline, DeepSPV, for precise spleen volume estimation from single or dual 2D ultrasound images. The pipeline involves a segmentation network and a variational autoencoder for learning low-dimensional representations from the estimated segmentations. We investigate three approaches for spleen volume estimation and our best model achieves 86.62%/92.5% mean relative volume accuracy (MRVA) under single-view/dual-view settings, surpassing the performance of human experts. In addition, the pipeline can provide confidence intervals for the volume estimates as well as offering benefits in terms of interpretability, which further support clinicians in decision-making when identifying splenomegaly. We evaluate the full pipeline using a highly realistic synthetic dataset generated by a diffusion model, achieving an overall MRVA of 83.0% from a single 2D ultrasound image. Our proposed DeepSPV is the first work to use deep learning to estimate 3D spleen volume from 2D ultrasound images and can be seamlessly integrated into the current clinical workflow for spleen assessment.

Scaling LLM Inference with Optimized Sample Compute Allocation

Sampling is a basic operation in many inference-time algorithms of large language models (LLMs). To scale up inference efficiently with a limited compute, it is crucial to find an optimal allocation for sample compute budgets: Which sampling configurations (model, temperature, language, etc.) do we use? How many samples do we generate in each configuration? We formulate these choices as a learning problem and propose OSCA, an algorithm that Optimizes Sample Compute Allocation by finding an optimal mix of different inference configurations. Our experiments show that with our learned mixed allocation, we can achieve accuracy better than the best single configuration with 128x less compute on code generation and 25x less compute on 4 reasoning tasks. OSCA is also shown to be effective in agentic workflows beyond single-turn tasks, achieving a better accuracy on SWE-Bench with 3x less compute than the default configuration. Our code and generations are released at https://github.com/LeiLiLab/OSCA.

Anticipating Future with Large Language Model for Simultaneous Machine Translation

Simultaneous machine translation (SMT) takes streaming input utterances and incrementally produces target text. Existing SMT methods only use the partial utterance that has already arrived at the input and the generated hypothesis. Motivated by human interpreters' technique to forecast future words before hearing them, we propose $\textbf{T}$ranslation by $\textbf{A}$nticipating $\textbf{F}$uture (TAF), a method to improve translation quality while retraining low latency. Its core idea is to use a large language model (LLM) to predict future source words and opportunistically translate without introducing too much risk. We evaluate our TAF and multiple baselines of SMT on four language directions. Experiments show that TAF achieves the best translation quality-latency trade-off and outperforms the baselines by up to 5 BLEU points at the same latency (three words).

AutoMIR: Effective Zero-Shot Medical Information Retrieval without Relevance Labels

Medical information retrieval (MIR) is essential for retrieving relevant medical knowledge from diverse sources, including electronic health records, scientific literature, and medical databases. However, achieving effective zero-shot dense retrieval in the medical domain poses substantial challenges due to the lack of relevance-labeled data. In this paper, we introduce a novel approach called Self-Learning Hypothetical Document Embeddings (SL-HyDE) to tackle this issue. SL-HyDE leverages large language models (LLMs) as generators to generate hypothetical documents based on a given query. These generated documents encapsulate key medical context, guiding a dense retriever in identifying the most relevant documents. The self-learning framework progressively refines both pseudo-document generation and retrieval, utilizing unlabeled medical corpora without requiring any relevance-labeled data. Additionally, we present the Chinese Medical Information Retrieval Benchmark (CMIRB), a comprehensive evaluation framework grounded in real-world medical scenarios, encompassing five tasks and ten datasets. By benchmarking ten models on CMIRB, we establish a rigorous standard for evaluating medical information retrieval systems. Experimental results demonstrate that SL-HyDE significantly surpasses existing methods in retrieval accuracy while showcasing strong generalization and scalability across various LLM and retriever configurations. CMIRB data and evaluation code are publicly available at: https://github.com/CMIRB-benchmark/CMIRB.

Cooperative Strategic Planning Enhances Reasoning Capabilities in Large Language Models

Enhancing the reasoning capabilities of large language models (LLMs) is crucial for enabling them to tackle complex, multi-step problems. Multi-agent frameworks have shown great potential in enhancing LLMs' reasoning capabilities. However, the lack of effective cooperation between LLM agents hinders their performance, especially for multi-step reasoning tasks. This paper proposes a novel cooperative multi-agent reasoning framework (CoPlanner) by separating reasoning steps and assigning distinct duties to different agents. CoPlanner consists of two LLM agents: a planning agent and a reasoning agent. The planning agent provides high-level strategic hints, while the reasoning agent follows these hints and infers answers. By training the planning agent's policy through the interactive reasoning process via Proximal Policy Optimization (PPO), the LLaMA-3-8B-based CoPlanner outperforms the previous best method by 9.94\% on LogiQA and 3.09\% on BBH. Our results demonstrate that the guidance from the planning agent and the effective cooperation between the agents contribute to the superior performance of CoPlanner in tackling multi-step reasoning problems.

Why Does the Effective Context Length of LLMs Fall Short?

Advancements in distributed training and efficient attention mechanisms have significantly expanded the context window sizes of large language models (LLMs). However, recent work reveals that the effective context lengths of open-source LLMs often fall short, typically not exceeding half of their training lengths. In this work, we attribute this limitation to the left-skewed frequency distribution of relative positions formed in LLMs pretraining and post-training stages, which impedes their ability to effectively gather distant information. To address this challenge, we introduce ShifTed Rotray position embeddING (STRING). STRING shifts well-trained positions to overwrite the original ineffective positions during inference, enhancing performance within their existing training lengths. Experimental results show that without additional training, STRING dramatically improves the performance of the latest large-scale models, such as Llama3.1 70B and Qwen2 72B, by over 10 points on popular long-context benchmarks RULER and InfiniteBench, establishing new state-of-the-art results for open-source LLMs. Compared to commercial models, Llama 3.1 70B with \method even achieves better performance than GPT-4-128K and clearly surpasses Claude 2 and Kimi-chat.

CA*: Addressing Evaluation Pitfalls in Computation-Aware Latency for Simultaneous Speech Translation

Simultaneous speech translation (SimulST) systems must balance translation quality with response time, making latency measurement crucial for evaluating their real-world performance. However, there has been a longstanding belief that current metrics yield unrealistically high latency measurements in unsegmented streaming settings. In this paper, we investigate this phenomenon, revealing its root cause in a fundamental misconception underlying existing latency evaluation approaches. We demonstrate that this issue affects not only streaming but also segment-level latency evaluation across different metrics. Furthermore, we propose a modification to correctly measure computation-aware latency for SimulST systems, addressing the limitations present in existing metrics.