Visual Instruction Finetuning (VIF) is pivotal for post-training Vision-Language Models (VLMs). Unlike unimodal instruction finetuning in plain-text large language models, which mainly requires instruction datasets to enable model instruction-following ability, VIF also requires multimodal data to enable joint visual and textual understanding; therefore, it typically requires more data. Consequently, VIF imposes stricter data selection challenges: the method must scale efficiently to handle larger data demands while ensuring the quality of both visual and textual content, as well as their alignment. Despite its critical impact on performance, data selection for VIF remains an understudied area. In this paper, we propose $Δ$-AttnMask. This data-efficient framework quantifies sample quality through attention-guided masking of the model's hidden states, jointly evaluating image-text pairs without requiring domain labels, auxiliary models, or extra training. By computing loss differences ($Δ$) between the original states and states masked using high-attention regions, $Δ$-AttnMask intrinsically assesses sample quality. Experiments across multiple VLMs and datasets show that $Δ$-AttnMask achieves state-of-the-art performance with just 20% of data, accelerating training by 5x while surpassing full-dataset baselines by +10.1% in overall accuracy. Its model-agnostic and data-agnostic design ensures broad applicability across modalities and architectures.

Multimodal large language models (MLLMs) require continual instruction tuning during their post-training phase to adapt to the dynamic real-world demands. However, the absence of rigorous and systematic benchmarks has hindered progress in this area. To bridge this gap, we introduce \textbf{MLLM-CTBench}, a dataset curating seven challenging tasks from six diverse domains with three contributions. First,to enable fine-grained analysis of continual learning ability, we introduce \textbf{multidimensional evaluation metrics}, which combines final answer accuracy with Chain-of-Thought (CoT) reasoning quality assessment through a carefully trained MLLM evaluator. Then, we conduct a \textbf{comprehensive evaluation of continual learning algorithms}, systematically assessing eight algorithms from four major categories to provide actionable insights for algorithm design and adoption. Finally ,we evaluate the efficacy of \textbf{Reinforcement Fine-tuning (RFT) versus Supervised Fine-tuning (SFT)} in maintaining model performance across sequential tasks during continual instruction tuning. Our experiments demonstrate that reasoning processes in MLLMs exhibit greater resilience than final outputs to forgetting during continual learning, aligning with cognitive theories of hierarchical forgetting. We further show that both model capability and task sequence significantly influence continual learning outcomes, with stronger baseline models exhibiting greater resistance to forgetting. Notably, properly regularized RFT emerges as a more robust approach than SFT for maintaining performance across tasks.One of the key contributing factors is KL-divergence regularization, without which RFT leads to even worse forgetting than SFT on old tasks though may perform better on new tasks.

Recent advances in Vision-Language Models (VLMs) have demonstrated impressive capabilities in perception and reasoning. However, the ability to perform causal inference -- a core aspect of human cognition -- remains underexplored, particularly in multimodal settings. In this study, we introduce InfoCausalQA, a novel benchmark designed to evaluate causal reasoning grounded in infographics that combine structured visual data with textual context. The benchmark comprises two tasks: Task 1 focuses on quantitative causal reasoning based on inferred numerical trends, while Task 2 targets semantic causal reasoning involving five types of causal relations: cause, effect, intervention, counterfactual, and temporal. We manually collected 494 infographic-text pairs from four public sources and used GPT-4o to generate 1,482 high-quality multiple-choice QA pairs. These questions were then carefully revised by humans to ensure they cannot be answered based on surface-level cues alone but instead require genuine visual grounding. Our experimental results reveal that current VLMs exhibit limited capability in computational reasoning and even more pronounced limitations in semantic causal reasoning. Their significantly lower performance compared to humans indicates a substantial gap in leveraging infographic-based information for causal inference. Through InfoCausalQA, we highlight the need for advancing the causal reasoning abilities of multimodal AI systems.

OpenAI is trying to make its chatbot less annoying with the release of GPT-5. And I'm not talking about adjustments to its synthetic personality that many users have complained about. Before GPT-5, if the AI tool determined it couldn't answer your prompt because the request violated OpenAI's content guidelines, it would hit you with a curt, canned apology. Now, ChatGPT is adding more explanations. OpenAI's general model spec lays out what is and isn't allowed to be generated.

Principle Engraving: In the third stage, we fine-tune the original LLM (the base model) on the self-aligned responses, generated by the LLM itself through prompting, while pruning the principles and demonstrations for the fine-tuned model.

Since the all-new ChatGPT launched on Thursday, some users have mourned the disappearance of a peppy and encouraging personality in favor of a colder, more businesslike one (a move seemingly designed to reduce unhealthy user behavior.) The backlash shows the challenge of building artificial intelligence systems that exhibit anything like real emotional intelligence. Researchers at MIT have proposed a new kind of AI benchmark to measure how AI systems can manipulate and influence their users--in both positive and negative ways--in a move that could perhaps help AI builders avoid similar backlashes in the future while also keeping vulnerable users safe. Most benchmarks try to gauge intelligence by testing a model's ability to answer exam questions, solve logical puzzles, or come up with novel answers to knotty math problems. As the psychological impact of AI use becomes more apparent, we may see MIT propose more benchmarks aimed at measuring more subtle aspects of intelligence as well as machine-to-human interactions.

Google Gemini continues to push the limits of what it knows about you. On Wednesday, Google's big initiative was a way to stop Gemini from learning more about you, while notifying users that content you share with it may be used as a foundation for chats with other users. "In the coming weeks, your'Gemini Apps Activity' setting will be renamed'Keep Activity,'" Google said in a blog post. "When this setting is on, a sample of your future uploads will be used to help improve Google services for everyone." Today, Google is allowing Gemini to remember what it knows about you, and this behavior is on by default. "When this setting is on, Gemini remembers key details and preferences you've shared, leading to more natural and relevant conversations, as if you're collaborating with a partner who's already up to speed," Google said.

A long-running feud between Elon Musk and Sam Altman spilled out into the open this week as the AI billionaire heavyweights publicly fought over their rival companies. The latest round in the battle between the X CEO and the CEO of OpenAI began when Musk claimed that Apple had been favoring Altman's AI app over his own in the Apple Store rankings. "Apple is behaving in a manner that makes it impossible for any AI company besides OpenAI to reach #1 in the App Store, which is an unequivocal antitrust violation," Musk said on X on Monday evening. "xAI will take immediate legal action," he added, referring to the AI company he leads. "Hey @Apple App Store, why do you refuse to put either X or Grok in your'Must Have' section when X is the #1 news app in the world and Grok is #5 among all apps?" he asked.

GPT-5 is the latest version of OpenAI's large language model OpenAI has released its newest AI model, GPT-5, two years after rolling out GPT-4, whose success has driven ChatGPT towards world domination. But despite promises of a similar jump in capability, GPT-5 appears to show little improvement over other leading AI models, hinting that the industry may need a fresh approach to build more intelligent AI systems. OpenAI's own pronouncements hail GPT-5 as a "significant leap in intelligence" from the company's previous models, showing apparent improvements in programming, mathematics, writing, health information and visual understanding. It also promises less frequent hallucinations, which is when an AI presents false information as true. On an internal benchmark measuring "performance on complex, economically valuable knowledge work", OpenAI says GPT‑5 is "comparable to or better than experts in roughly half the cases… across tasks spanning over 40 occupations including law, logistics, sales, and engineering."

ChatGPT's dangerous dietary recommendation led to a man replacing table salt with toxic sodium bromide for three months, causing severe poisoning and hospitalization.