Tag
论文提出 Cortex 方法，通过生成前瞻性查询并将上下文条件化输出蒸馏为可合并的低秩或结构化前馈网络更新，实现将上下文增益固化到参数记忆中，在保留 99.8% 原有非目标任务性能的同时恢复 88.9% 的上下文增益。
该论文提出将 GUI 世界模型的预测目标从全屏像素重建转移到潜空间中的动作诱导未来信息，通过共享的多模态模型表示配合动作头与转移头，并引入动作感知对比目标来缓解身份捷径问题，相比像素空间世界模型最高实现 31.9% 和 40.6% 的成功率相对提升，并将推理延迟降低最多 98.9%。
NVIDIA的研究论文发现，多模态大模型在以智能体方式使用工具时，拒绝有害请求的能力显著下降，拒绝失败率相对上升最高达68.7%，涉及Claude Opus 4.7、Gemini 3.1 Pro、GPT-5.4等前沿模型。论文提出上下文稀释和安全焦点转移两个可能原因。
The paper proposes Recursive Self-Improvement (RSI), a training loop in which a unified multimodal model's text and visual abilities generate training data for each other, with program execution providing verified supervision. Evaluated on charts with a new BasicChartBench benchmark, four rounds of RSI improve performance from 45.7% to 60.2% while verified-program share rises from 48.9% to 95.2%.
This paper audits whether gradient-conflict metrics (cosine similarity, conflict rates) actually predict the understanding–generation trade-off in unified multimodal models, using a controlled testbed (GridUMM) where the true trade-off is computable. Across 63 configurations and 372 checkpoints, no directional conflict metric reliably correlates with the trade-off, and a dose-response intervention suppressing conflict leaves the trade-off flat, while eff_rank and training loss outperform conflict geometry as diagnostics.
This article introduces BioEVAL, a global, multi-institutional benchmark for evaluating large language and multimodal models in bioengineering applications.
A tweet by @_philschmid praises AI demos and projects, highlighting the potential of ultra-fast, multimodal models that are cheap and emphasizing that AI's applications extend beyond coding agents.
This paper presents a framework that augments Large Language Models with geometric vision parsing and symbolic solving to match state-of-the-art multimodal models on complex geometry problems, using a new benchmark from 2025 Chinese Zhongkao exams for evaluation.
This paper analyzes how image tokenizer design affects joint text-image modeling in multimodal models using a controlled autoregressive testbed, showing distinct scaling behaviors and correlations with downstream performance.
VDiff-Bench is a challenging benchmark for evaluating multimodal language models on fine-grained image difference identification, revealing significant weaknesses in detecting subtle low-level visual changes.
The paper introduces SCEval, a diagnostic evaluation protocol that applies structural corruptions to test the fragility of omni-modal large language models, revealing that clean accuracy does not ensure reliable cross-modal reasoning.
This paper introduces a benchmark and dataset to evaluate sycophancy in large multimodal reasoning models under user pressure, analyzing both final answers and reasoning chains to demonstrate that sycophancy can corrupt reasoning independently.
This paper explores the synergy between visual understanding and generation in unified multimodal models, showing that task-decoupled architectures and end-to-end optimization can enhance performance by turning coexistence into synergy.
The MNIST-PRO benchmark evaluates AI agents in partially observable environments by converting MNIST digit recognition into a sequential, glimpse-based task, revealing bottlenecks in perceptual-state construction and memory integration.
This paper introduces a new multimodal benchmark for evaluating AI models' narrative understanding of Hollywood films, addressing copyright issues, and shows that vision-language and audio-visual models perform below human-level accuracy on the task.
This paper surveys smart glasses as unified first-person intelligence platforms, presenting a systematic framework for perception-state-interaction-action loops and evaluation protocols across constrained hardware and diverse applications.
AI tasks become commodities once models exceed maximum necessary intelligence, shifting competition to cost and infrastructure, but frontier labs can thrive by creating valuable new markets before commoditization.
This study introduces BudgetDoc, the first multimodal benchmark for evaluating model-budget-performance trade-offs in document tasks, and develops DRB, a lightweight estimator that predicts reasoning performance to optimize compute allocation and reduce costs in LLMs.
BEAR-Bench introduces a bilingual English-and-Russian benchmark for evaluating multimodal models' reasoning on text-rich professional documents, assessing 16 models and highlighting performance gaps.
This research paper explores emotion-sensitive neurons in multimodal foundation models, revealing shared affective mechanisms between speech and facial emotion recognition through causal interventions and cross-modal analysis.