Multimodal foundation models, such as GPT-4o, have recently made remarkable progress, but it is not clear where exactly these models stand in terms of understanding vision. In this paper, we benchmark the performance of popular multimodal foundation models (GPT-4o, o4-mini, Gemini 1.5 Pro and Gemini 2.0 Flash, Claude 3.5 Sonnet, Qwen2-VL, Llama 3.2) on standard computer vision tasks (semantic segmentation, object detection, image classification, depth and surface normal prediction) using established datasets (e.g., COCO, ImageNet and its variants, etc). The main challenges to performing this are: 1) most models are trained to output text and cannot natively express versatile domains, such as segments or 3D geometry, and 2) many leading models are proprietary and accessible only at an API level, i.e., there is no weight access to adapt them. We address these challenges by translating standard vision tasks into equivalent text-promptable and API-compatible tasks via prompt chaining to create a standardized benchmarking framework. We observe that 1) the models are not close to the state-of-the-art specialist models at any task. However, 2) they are respectable generalists; this is remarkable as they are presumably trained on primarily image-text-based tasks. 3) They perform semantic tasks notably better than geometric ones. 4) While the prompt-chaining techniques affect performance, better models exhibit less sensitivity to prompt variations. 5) GPT-4o performs the best among non-reasoning models, securing the top position in 4 out of 6 tasks, 6) reasoning models, e.g. o3, show improvements in geometric tasks, and 7) a preliminary analysis of models with native image generation, like the latest GPT-4o, shows they exhibit quirks like hallucinations and spatial misalignments.

Attention mechanisms underpin the success of large language models (LLMs), yet their substantial computational and memory overhead poses challenges for optimizing efficiency and performance. A critical bottleneck arises as KV cache and attention computations scale rapidly with text length, challenging deployment on hardware with limited computational and memory resources. We observe that attention mechanisms exhibit substantial redundancy, since the KV cache can be significantly compressed and attention maps across heads display high similarity, revealing that much of the computation and storage is unnecessary. Leveraging these insights, we propose \textbf{G}rouped-Head Laten\textbf{T} \textbf{A}ttention (GTA), a novel attention mechanism that reduces memory usage and computational complexity while maintaining performance. GTA comprises two components: (1) a shared attention map mechanism that reuses attention scores across multiple heads, decreasing the key cache size; and (2) a nonlinear value decoder with learned projections that compresses the value cache into a latent space, further cutting memory needs. GTA cuts attention computation FLOPs by up to \emph{62.5\%} versus Grouped-Query Attention and shrink the KV cache by up to \emph{70\%}, all while avoiding the extra overhead of Multi-Head Latent Attention to improve LLM deployment efficiency. Consequently, GTA models achieve a \emph{2x} increase in end-to-end inference speed, with prefill benefiting from reduced computational cost and decoding benefiting from the smaller cache footprint.

In this paper, we evaluate the capacity of current language technologies to understand Basque and Spanish language varieties. We use Natural Language Inference (NLI) as a pivot task and introduce a novel, manually-curated parallel dataset in Basque and Spanish, along with their respective variants. Our empirical analysis of crosslingual and in-context learning experiments using encoder-only and decoder-based Large Language Models (LLMs) shows a performance drop when handling linguistic variation, especially in Basque. Error analysis suggests that this decline is not due to lexical overlap, but rather to the linguistic variation itself. Further ablation experiments indicate that encoder-only models particularly struggle with Western Basque, which aligns with linguistic theory that identifies peripheral dialects (e.g., Western) as more distant from the standard. All data and code are publicly available.

Serving large language models (LLMs) is important for cloud providers, and caching intermediate results (KV\$) after processing each request substantially improves serving throughput and latency. However, there is limited understanding of how LLM serving benefits from KV\$ caching, where system design decisions like cache eviction policies are highly workload-dependent. In this paper, we present the first systematic characterization of the KV\$ workload patterns from one of the leading LLM service providers. We draw observations that were not covered by previous studies focusing on synthetic workloads, including: KV\$ reuses are skewed across requests, where reuses between single-turn requests are equally important as multi-turn requests; the reuse time and probability are diverse considering all requests, but for a specific request category, the pattern tends to be predictable; and the overall cache size required for an ideal cache hit ratio is moderate. Based on the characterization, we further propose a workload-aware cache eviction policy that improves the serving performance under real-world traces, especially with limited cache capacity.

Prompt engineering represents a critical bottleneck to harness the full potential of Large Language Models (LLMs) for solving complex tasks, as it requires specialized expertise, significant trial-and-error, and manual intervention. This challenge is particularly pronounced for tasks involving subjective quality assessment, where defining explicit optimization objectives becomes fundamentally problematic. Existing automated prompt optimization methods falter in these scenarios, as they typically require well-defined task-specific numerical fitness functions or rely on generic templates that cannot capture the nuanced requirements of complex use cases. We introduce DEEVO (DEbate-driven EVOlutionary prompt optimization), a novel framework that guides prompt evolution through a debate-driven evaluation with an Elo-based selection. Contrary to prior work, DEEVOs approach enables exploration of the discrete prompt space while preserving semantic coherence through intelligent crossover and strategic mutation operations that incorporate debate-based feedback, combining elements from both successful and unsuccessful prompts based on identified strengths rather than arbitrary splicing. Using Elo ratings as a fitness proxy, DEEVO simultaneously drives improvement and preserves valuable diversity in the prompt population. Experimental results demonstrate that DEEVO significantly outperforms both manual prompt engineering and alternative state-of-the-art optimization approaches on open-ended tasks and close-ended tasks despite using no ground truth feedback. By connecting LLMs reasoning capabilities with adaptive optimization, DEEVO represents a significant advancement in prompt optimization research by eliminating the need of predetermined metrics to continuously improve AI systems.

Speech is a noninvasive digital phenotype that can offer valuable insights into mental health conditions, but it is often treated as a single modality. In contrast, we propose the treatment of patient speech data as a trimodal multimedia data source for depression detection. This study explores the potential of large language model-based architectures for speech-based depression prediction in a multimodal regime that integrates speech-derived text, acoustic landmarks, and vocal biomarkers. Adolescent depression presents a significant challenge and is often comorbid with multiple disorders, such as suicidal ideation and sleep disturbances. This presents an additional opportunity to integrate multi-task learning (MTL) into our study by simultaneously predicting depression, suicidal ideation, and sleep disturbances using the multimodal formulation. We also propose a longitudinal analysis strategy that models temporal changes across multiple clinical interactions, allowing for a comprehensive understanding of the conditions' progression. Our proposed approach, featuring trimodal, longitudinal MTL is evaluated on the Depression Early Warning dataset. It achieves a balanced accuracy of 70.8%, which is higher than each of the unimodal, single-task, and non-longitudinal methods.

Recent advancements in adversarial jailbreak attacks have exposed critical vulnerabilities in Large Language Models (LLMs), enabling the circumvention of alignment safeguards through increasingly sophisticated prompt manipulations. Based on our experiments, we found that the effectiveness of jailbreak strategies is influenced by the comprehension ability of the attacked LLM. Building on this insight, we propose a capability-aware Multi-Encryption Framework (MEF) for evaluating vulnerabilities in black-box LLMs. Specifically, MEF first categorizes the comprehension ability level of the LLM, then applies different strategies accordingly: For models with limited comprehension ability, MEF adopts the Fu+En1 strategy, which integrates layered semantic mutations with an encryption technique, more effectively contributing to evasion of the LLM's defenses at the input and inference stages. For models with strong comprehension ability, MEF uses a more complex Fu+En1+En2 strategy, in which additional dual-ended encryption techniques are applied to the LLM's responses, further contributing to evasion of the LLM's defenses at the output stage. Experimental results demonstrate the effectiveness of our approach, achieving attack success rates of 98.9% on GPT-4o (29 May 2025 release) and 99.8% on GPT-4.1 (8 July 2025 release). Our work contributes to a deeper understanding of the vulnerabilities in current LLM alignment mechanisms.

Recent advancements in large language models (LLMs) have demonstrated impressive chain-of-thought reasoning capabilities, with reinforcement learning (RL) playing a crucial role in this progress. While "aha moment" patterns--where models exhibit self-correction through reflection--are often attributed to emergent properties from RL, we first demonstrate that these patterns exist in multimodal LLMs (MLLMs) prior to RL training but may not necessarily correlate with improved reasoning performance. Building on these insights, we present a comprehensive study on enhancing multimodal reasoning through a two-stage approach: (1) supervised fine-tuning (SFT) as a cold start with structured chain-of-thought reasoning patterns, followed by (2) reinforcement learning via GRPO to further refine these capabilities. Our extensive experiments show that this combined approach consistently outperforms both SFT-only and RL-only methods across challenging multimodal reasoning benchmarks. The resulting models achieve state-of-the-art performance among open-source MLLMs at both 3B and 7B scales, with our 7B model showing substantial improvements over base models (e.g., 66.3 %$\rightarrow$73.4 % on MathVista, 62.9 %$\rightarrow$70.4 % on We-Math) and our 3B model achieving performance competitive with several 7B models. Overall, this work provides practical guidance for building advanced multimodal reasoning models. Our code is available at https://github.com/waltonfuture/RL-with-Cold-Start.

Operating home appliances, among the most common tools in every household, is a critical capability for assistive home robots. This paper presents ApBot, a robot system that operates novel household appliances by "reading" their user manuals. ApBot faces multiple challenges: (i) infer goal-conditioned partial policies from their unstructured, textual descriptions in a user manual document, (ii) ground the policies to the appliance in the physical world, and (iii) execute the policies reliably over potentially many steps, despite compounding errors. To tackle these challenges, ApBot constructs a structured, symbolic model of an appliance from its manual, with the help of a large vision-language model (VLM). It grounds the symbolic actions visually to control panel elements. Finally, ApBot closes the loop by updating the model based on visual feedback. Our experiments show that across a wide range of simulated and real-world appliances, ApBot achieves consistent and statistically significant improvements in task success rate, compared with state-of-the-art large VLMs used directly as control policies. These results suggest that a structured internal representations plays an important role in robust robot operation of home appliances, especially, complex ones.

Training data cleaning is a new application for generative model-based speech restoration (SR). This paper introduces Miipher-2, an SR model designed for million-hour scale data, for training data cleaning for large-scale generative models like large language models. Key challenges addressed include generalization to unseen languages, operation without explicit conditioning (e.g., text, speaker ID), and computational efficiency. Miipher-2 utilizes a frozen, pre-trained Universal Speech Model (USM), supporting over 300 languages, as a robust, conditioning-free feature extractor. To optimize efficiency and minimize memory, Miipher-2 incorporates parallel adapters for predicting clean USM features from noisy inputs and employs the WaveFit neural vocoder for waveform synthesis. These components were trained on 3,000 hours of multi-lingual, studio-quality recordings with augmented degradations, while USM parameters remained fixed. Experimental results demonstrate Miipher-2's superior or comparable performance to conventional SR models in word-error-rate, speaker similarity, and both objective and subjective sound quality scores across all tested languages. Miipher-2 operates efficiently on consumer-grade accelerators, achieving a real-time factor of 0.0078, enabling the processing of a million-hour speech dataset in approximately three days using only 100 such accelerators.