Sarcasm is a complex linguistic phenomenon that involves a disparity between literal and intended meanings, making it challenging for sentiment analysis and other emotion-sensitive tasks. While traditional sarcasm detection methods primarily focus on text, recent approaches have incorporated multimodal information. However, the application of Large Visual Language Models (LVLMs) in Multimodal Sarcasm Analysis (MSA) remains underexplored. In this paper, we evaluate LVLMs in MSA tasks, specifically focusing on Multimodal Sarcasm Detection and Multimodal Sarcasm Explanation. Through comprehensive experiments, we identify key limitations, such as insufficient visual understanding and a lack of conceptual knowledge. To address these issues, we propose a training-free framework that integrates in-depth object extraction and external conceptual knowledge to improve the model's ability to interpret and explain sarcasm in multimodal contexts. The experimental results on multiple models show the effectiveness of our proposed framework. The code is available at https://github.com/cp-cp/LVLM-MSA.

Recent advancements in large multimodal models have provided blind or visually impaired (BVI) individuals with new capabilities to interpret and engage with the real world through interactive systems that utilize live video feeds. However, the potential benefits and challenges of such capabilities to support diverse real-world assistive tasks remain unclear. In this paper, we present findings from an exploratory study with eight BVI participants. Participants used ChatGPT's Advanced Voice with Video, a state-of-the-art live video AI released in late 2024, in various real-world scenarios, from locating objects to recognizing visual landmarks, across unfamiliar indoor and outdoor environments. Our findings indicate that current live video AI effectively provides guidance and answers for static visual scenes but falls short in delivering essential live descriptions required in dynamic situations. Despite inaccuracies in spatial and distance information, participants leveraged the provided visual information to supplement their mobility strategies. Although the system was perceived as human-like due to high-quality voice interactions, assumptions about users' visual abilities, hallucinations, generic responses, and a tendency towards sycophancy led to confusion, distrust, and potential risks for BVI users. Based on the results, we discuss implications for assistive video AI agents, including incorporating additional sensing capabilities for real-world use, determining appropriate intervention timing beyond turn-taking interactions, and addressing ecological and safety concerns.

Retrieval-Augmented Generation (RAG) systems using Multimodal Large Language Models (MLLMs) show great promise for complex document understanding, yet their development is critically hampered by inadequate evaluation. Current benchmarks often focus on specific part of document RAG system and use synthetic data with incomplete ground truth and evidence labels, therefore failing to reflect real-world bottlenecks and challenges. To overcome these limitations, we introduce Double-Bench: a new large-scale, multilingual, and multimodal evaluation system that is able to produce fine-grained assessment to each component within document RAG systems. It comprises 3,276 documents (72,880 pages) and 5,168 single- and multi-hop queries across 6 languages and 4 document types with streamlined dynamic update support for potential data contamination issues. Queries are grounded in exhaustively scanned evidence pages and verified by human experts to ensure maximum quality and completeness. Our comprehensive experiments across 9 state-of-the-art embedding models, 4 MLLMs and 4 end-to-end document RAG frameworks demonstrate the gap between text and visual embedding models is narrowing, highlighting the need in building stronger document retrieval models. Our findings also reveal the over-confidence dilemma within current document RAG frameworks that tend to provide answer even without evidence support. We hope our fully open-source Double-Bench provide a rigorous foundation for future research in advanced document RAG systems. We plan to retrieve timely corpus and release new benchmarks on an annual basis.

With the advancement of code generation capabilities in large language models (LLMs), their reliance on input premises has intensified. When users provide inputs containing faulty premises, the probability of code generation hallucinations rises significantly, exposing deficiencies in their self-scrutiny capabilities. This paper proposes Faulty Premises Bench (FPBench), the first code generation evaluation framework targeting faulty premises. By systematically constructing three categories of faulty premises and integrating multi-dimensional evaluation metrics, it conducts in-depth assessments of 15 representative LLMs. The key findings are as follows: (1) Most models exhibit poor reasoning abilities and suboptimal code generation performance under faulty premises, heavily relying on explicit prompts for error detection, with limited self-scrutiny capabilities; (2) Faulty premises trigger a point of diminishing returns in resource investment, leading to blindly increasing length fails to enhance quality; (3) The three types of faulty premises respectively activate distinct defect patterns in models, revealing a triple dissociation in the cognitive mechanisms of code generation models. This study not only highlights the urgent need for LLMs to proactively verify premises in code generation but also, through the proposed FPBench framework and multi-dimensional evaluation system, provides a theoretical foundation and practical pathway for developing reliable, human-centric code generation models.

Massive activations are scalar values in transformer hidden states that achieve values orders of magnitude larger than typical activations and have been shown to be critical for model functionality. While prior work has characterized these phenomena in fully trained models, the temporal dynamics of their emergence during training remain poorly understood. We present the first comprehensive analysis of massive activation development throughout transformer training, using the Pythia model family as our testbed. Through systematic analysis of various model sizes across multiple training checkpoints, we demonstrate that massive activation emergence follows predictable mathematical patterns that can be accurately modeled using an exponentially-modulated logarithmic function with five key parameters. We develop a machine learning framework to predict these mathematical parameters from architectural specifications alone, achieving high accuracy for steady-state behavior and moderate accuracy for emergence timing and magnitude. These findings enable architects to predict and potentially control key aspects of massive activation emergence through design choices, with significant implications for model stability, training cycle length, interpretability, and optimization. Our findings demonstrate that the emergence of massive activations is governed by model design and can be anticipated, and potentially controlled, before training begins.

We introduce Goedel-Prover-V2, a series of open-source language models that set a new state-of-the-art in automated theorem proving. Built on the standard expert iteration and reinforcement learning pipeline, our approach incorporates three key innovations: (1) Scaffolded data synthesis: We generate synthetic tasks of increasing difficulty to train the model to master increasingly complex theorems; (2) Verifier-guided self-correction: We enable the model to iteratively revise its proofs by leveraging feedback from the Lean compiler; (3) Model averaging: We merge model checkpoints to mitigate the decrease in model output diversity in later stages of training. Our small model, Goedel-Prover-V2-8B, reaches 84.6% pass@32 on MiniF2F and outperforms DeepSeek-Prover-V2-671B under the same metric, despite being 80X smaller. Our flagship model, Goedel-Prover-V2-32B, achieves 88.1% on MiniF2F at pass@32 in standard mode and 90.4% in self-correction mode, outperforming prior SOTA by a large margin. Additionally, our flagship model solves 86 problems on PutnamBench at pass@184, securing the first place among open-source models on the leaderboard, surpassing DeepSeek-Prover-V2-671B's record of solving 47 problems by pass@1024 with a significantly smaller model size and compute budget. At the time of its release (July-August 2025), Goedel-Prover-V2 achieves the strongest overall performance among all open-source theorem provers. It also ranks among the top-performing models--including closed-source systems with publicly reported performance--under a constrained test-time compute budget. Our models, code, and data are released at https://github.com/Goedel-LM/Goedel-Prover-V2.

ABSTRACT Large Language Models (LLMs) often suffer from performance degradation when faced with domain shifts, primarily due to catastrophic forgetting. In this work, we propose KILO (Knowledge-Instructed Learning for Continual Adaptation), a novel continual learning framework that integrates dynamic knowledge graphs with instruction tuning. By leveraging retrieved domain-specific knowledge as guidance during training, KILO enhances both adaptability to new domains and retention of previously acquired knowledge. We pretrain our model on WikiText-103 and evaluate sequential adaptation across four diverse target domains: BioASQ, SciQ, TweetEval, and MIND. Our experiments demonstrate that KILO consistently outperforms strong baselines, including continual fine-tuning, ERNIE 2.0, and CPT, in terms of backward transfer, forward transfer, F1 score, retention rate, and training efficiency. These results highlight the effectiveness of combining structured knowledge retrieval and instruction prompting to overcome domain shift challenges in continual learning scenarios.

Internet memes, now a staple of digital communication, play a pivotal role in how users engage within online communities and allow researchers to gain insight into contemporary digital culture. These engaging user-generated content are characterised by their reuse of visual elements also found in other memes. Matching instances of memes via these shared visual elements, called Meme Matching, is the basis of a wealth of meme analysis approaches. However, most existing methods assume that every meme consists of a shared visual background, called a Template, with some overlaid text, thereby limiting meme matching to comparing the background image alone. Current approaches exclude the many memes that are not template-based and limit the effectiveness of automated meme analysis and would not be effective at linking memes to contemporary web-based meme dictionaries. In this work, we introduce a broader formulation of meme matching that extends beyond template matching. We show that conventional similarity measures, including a novel segment-wise computation of the similarity measures, excel at matching template-based memes but fall short when applied to non-template-based meme formats. However, the segment-wise approach was found to consistently outperform the whole-image measures on matching non-template-based memes. Finally, we explore a prompting-based approach using a pretrained Multimodal Large Language Model for meme matching. Our results highlight that accurately matching memes via shared visual elements, not just background templates, remains an open challenge that requires more sophisticated matching techniques.

Effectively adapting powerful pretrained foundation models to diverse tasks remains a key challenge in AI deployment. Current approaches primarily follow two paradigms:discrete optimization of text prompts through prompt engineering, or continuous adaptation via additional trainable parameters. Both exhibit limitations-discrete methods lack refinement precision while parameter-based techniques increase complexity and reduce interpretability. To address these constraints, we propose EmbedGrad, a novel framework that optimizes text prompt embeddings through gradient-based refinement. Our approach uniquely decouples training from deployment:during optimization,labeled examples guide precise embedding adjustments while preserving semantic meaning; during inference, only optimized embeddings integrate with user queries. This enables fine-grained calibration impossible in text space, such as enhancing the reasoning capability of prompts like please reason step by step. Comprehensive evaluations across mathematical reasoning, sentiment analysis, and causal judgment tasks demonstrate EmbedGrad's effectiveness:optimizing this reasoning prompt for Qwen2.5-Math-1.5B increased accuracy from 14.74\% to 58.96\% on mathematical problems. Consistent improvements were observed across model scales (0.5B-14B) and all tasks, with particularly significant gains for smaller models on complex problems like causal judgment. By bridging prompt engineering and parameter efficiency without architectural changes, our work establishes embedding refinement as a powerful new paradigm for task adaptation.

Parameter-efficient fine-tuning (PEFT) is essential for reducing the computational overhead of large language models (LLMs). Low-rank family adapters are commonly used to control the parameter size efficiently while maintaining the generative power of LLMs. However, their limited expressiveness due to the rank constraint often restricts their performance on complex tasks. We propose Mixture of Kronecker Adapters (MoKA), a new generation of Kronecker adapters that addresses this limitation by modeling weight updates as a mixture of Kronecker products. Our proposed adapter leverages a gating mechanism that measures the importance of each Kronecker factor, enabling more expressive adaptation. Moreover, MoKA enables a rank flexibility that provides a better trade-off between parameter efficiency and accuracy. To ensure hardware efficiency, we reformulate Kronecker computations using standard matrix operations, allowing seamless deployment on GPU-optimized hardware. We conduct extensive experiments on instruction-tuning and commonsense reasoning tasks using low-bit quantized versions of LLaMA2-7B and LLaMA3-8B models. MoKA not only outperforms PEFT baselines, but also reduces the number of trainable parameters up to 27x, achieving state-of-the-art trade-offs between performance and parameter efficiency.