Recent text-only models demonstrate remarkable mathematical reasoning capabilities. Extending these to visual domains requires vision-language models to translate images into text descriptions. However, current models, trained to produce captions for human readers, often omit the precise details that reasoning systems require. This creates an interface mismatch: reasoners often fail not due to reasoning limitations but because they lack access to critical visual information. We propose Adaptive-Clarification Reinforcement Learning (AC-RL), which teaches vision models what information reasoners need through interaction. Our key insight is that clarification requests during training reveal information gaps; by penalizing success that requires clarification, we create pressure for comprehensive initial captions that enable the reasoner to solve the problem in a single pass. AC-RL improves average accuracy by 4.4 points over pretrained baselines across seven visual mathematical reasoning benchmarks, and analysis shows it would cut clarification requests by up to 39% if those were allowed. By treating clarification as a form of implicit supervision, AC-RL demonstrates that vision-language interfaces can be effectively learned through interaction alone, without requiring explicit annotations. Recent advances in reinforcement learning have produced text-based reasoning models with remarkable mathematical capabilities (Guo et al., 2025a; Shao et al., 2024). While these reasoning capabilities are impressive, extending them to visual domains requires careful consideration of how visual and linguistic information should interface.

The rapid rise of large language models (LLMs) has been driving an enormous demand for AI inference infrastructure, mainly powered by high-end GPUs. While these accelerators offer immense computational power, they incur high capital and operational costs due to frequent upgrades, dense power consumption, and cooling demands, making total cost of ownership (TCO) for AI datacenters a critical concern for cloud providers. Unfortunately, traditional datacenter lifecycle management (designed for general-purpose workloads) struggles to keep pace with AI's fast-evolving models, rising resource needs, and diverse hardware profiles. In this paper, we rethink the AI datacenter lifecycle scheme across three stages: building, hardware refresh, and operation. We show how design choices in power, cooling, and networking provisioning impact long-term TCO. We also explore refresh strategies aligned with hardware trends. Finally, we use operation software optimizations to reduce cost. While these optimizations at each stage yield benefits, unlocking the full potential requires rethinking the entire lifecycle. Thus, we present a holistic lifecycle management framework that coordinates and co-optimizes decisions across all three stages, accounting for workload dynamics, hardware evolution, and system aging. Our system reduces the TCO by up to 40\% over traditional approaches. Using our framework we provide guidelines on how to manage AI datacenter lifecycle for the future.

Large reasoning models show improved performance with longer chains of thought. However, recent work has highlighted (qualitatively) their tendency to overthink, continuing to revise answers even after reaching the correct solution. We quantitatively confirm this inefficiency by tracking Pass@1 for answers averaged over a large number of rollouts and find that the model often begins to always produce the correct answer early in the reasoning, making extra reasoning a waste of tokens. To detect and prevent overthinking, we propose a simple and inexpensive novel signal -- Entropy After (EAT) -- for monitoring and deciding whether to exit reasoning early. By appending a stop thinking token () and monitoring the entropy of the following token as the model reasons, we obtain a trajectory that decreases and stabilizes when Pass@1 plateaus; thresholding its variance under an exponential moving average yields a practical stopping rule. Importantly, our approach enables adaptively allocating compute based on the EAT trajectory, allowing us to spend compute in a more efficient way compared with fixing the token budget for all questions. Empirically, on MATH500 and AIME2025, EAT reduces token usage by 13 - 21% without harming accuracy, and it remains effective in black box settings where logits from the reasoning model are not accessible, and EAT is computed with proxy models.

The rise of Large Language Models (LLMs) is reshaping multimodel models, with speech synthesis being a prominent application. However, existing approaches often underutilize the linguistic intelligence of these models, typically failing to leverage their powerful instruction-following capabilities. This limitation hinders the model's ability to follow text instructions for controllable Text-to-Speech~(TTS). To address this, we propose a new paradigm inspired by ``operationalism'' that decouples instruction understanding from speech generation. We introduce BatonVoice, a framework where an LLM acts as a ``conductor'', understanding user instructions and generating a textual ``plan'' -- explicit vocal features (e.g., pitch, energy). A separate TTS model, the ``orchestra'', then generates the speech from these features. To realize this component, we develop BatonTTS, a TTS model trained specifically for this task. Our experiments demonstrate that BatonVoice achieves strong performance in controllable and emotional speech synthesis, outperforming strong open- and closed-source baselines. Notably, our approach enables remarkable zero-shot cross-lingual generalization, accurately applying feature control abilities to languages unseen during post-training. This demonstrates that objectifying speech into textual vocal features can more effectively unlock the linguistic intelligence of LLMs.

We introduce SCUBA, a benchmark designed to evaluate computer-use agents on customer relationship management (CRM) workflows within the Salesforce platform. SCUBA contains 300 task instances derived from real user interviews, spanning three primary personas, platform administrators, sales representatives, and service agents. The tasks test a range of enterprise-critical abilities, including Enterprise Software UI navigation, data manipulation, workflow automation, information retrieval, and troubleshooting. To ensure realism, SCUBA operates in Salesforce sandbox environments with support for parallel execution and fine-grained evaluation metrics to capture milestone progress. We benchmark a diverse set of agents under both zero-shot and demonstration-augmented settings. We observed huge performance gaps in different agent design paradigms and gaps between the open-source model and the closed-source model. In the zero-shot setting, open-source model powered computer-use agents that have strong performance on related benchmarks like OSWorld only have less than 5\% success rate on SCUBA, while methods built on closed-source models can still have up to 39% task success rate. In the demonstration-augmented settings, task success rates can be improved to 50\% while simultaneously reducing time and costs by 13% and 16%, respectively. These findings highlight both the challenges of enterprise tasks automation and the promise of agentic solutions. By offering a realistic benchmark with interpretable evaluation, SCUBA aims to accelerate progress in building reliable computer-use agents for complex business software ecosystems.

Many enterprises are increasingly adopting Artificial Intelligence (AI) to make internal processes more competitive and efficient. In response to public concern and new regulations for the ethical and responsible use of AI, implementing AI governance frameworks could help to integrate AI within organisations and mitigate associated risks. However, the rapid technological advances and lack of shared ethical AI infrastructures creates barriers to their practical adoption in businesses. This paper presents a real-world AI application at TVS Supply Chain Solutions, reporting on the experience developing an AI assistant underpinned by large language models and the ethical, regulatory, and sociotechnical challenges in deployment for enterprise use.

Unified Multimodal understanding and generation Models (UMMs) have demonstrated remarkable capabilities in both understanding and generation tasks. However, we identify a vulnerability arising from the generation-understanding coupling in UMMs. The attackers can use the generative function to craft an information-rich adversarial image and then leverage the understanding function to absorb it in a single pass, which we call Cross-Modal Generative Injection (CMGI). Current attack methods on malicious instructions are often limited to a single modality while also relying on prompt rewriting with semantic drift, leaving the unique vulnerabilities of UMMs unexplored. We propose STaR-Attack, the first multi-turn jailbreak attack framework that exploits unique safety weaknesses of UMMs without semantic drift. Specifically, our method defines a malicious event that is strongly correlated with the target query within a spatio-temporal context. Using the three-act narrative theory, STaR-Attack generates the pre-event and the post-event scenes while concealing the malicious event as the hidden climax. When executing the attack strategy, the opening two rounds exploit the UMM's generative ability to produce images for these scenes. Subsequently, an image-based question guessing and answering game is introduced by exploiting the understanding capability. STaR-Attack embeds the original malicious question among benign candidates, forcing the model to select and answer the most relevant one given the narrative context. Extensive experiments show that STaR-Attack consistently surpasses prior approaches, achieving up to 93.06% ASR on Gemini-2.0-Flash and surpasses the strongest prior baseline, FlipAttack. Our work uncovers a critical yet underdeveloped vulnerability and highlights the need for safety alignments in UMMs.

A preference for oneself (self-love) is a fundamental feature of biological organisms, with evidence in humans often bordering on the comedic. Since large language models (LLMs) lack sentience - and themselves disclaim having selfhood or identity - one anticipated benefit is that they will be protected from, and in turn protect us from, distortions in our decisions. Yet, across 5 studies and ~20,000 queries, we discovered massive self-preferences in four widely used LLMs. In word-association tasks, models overwhelmingly paired positive attributes with their own names, companies, and CEOs relative to those of their competitors. Strikingly, when models were queried through APIs this self-preference vanished, initiating detection work that revealed API models often lack clear recognition of themselves. This peculiar feature serendipitously created opportunities to test the causal link between self-recognition and self-love. By directly manipulating LLM identity - i.e., explicitly informing LLM1 that it was indeed LLM1, or alternatively, convincing LLM1 that it was LLM2 - we found that self-love consistently followed assigned, not true, identity. Importantly, LLM self-love emerged in consequential settings beyond word-association tasks, when evaluating job candidates, security software proposals and medical chatbots. Far from bypassing this human bias, self-love appears to be deeply encoded in LLM cognition. This result raises questions about whether LLM behavior will be systematically influenced by self-preferential tendencies, including a bias toward their own operation and even their own existence. We call on corporate creators of these models to contend with a significant rupture in a core promise of LLMs - neutrality in judgment and decision-making.

CLIP has revolutionized zero-shot learning by enabling task generalization without fine-tuning. While prompting techniques like CoOp and CoCoOp enhance CLIP's adaptability, their effectiveness in Federated Learning (FL) remains an open challenge. Existing federated prompt learning approaches, such as FedCoOp and FedTPG, improve performance but face generalization issues, high communication costs, and reliance on a central server, limiting scalability and privacy. We propose Zero-shot Decentralized Federated Learning (ZeroDFL), a fully decentralized framework that enables zero-shot adaptation across distributed clients without a central coordinator. ZeroDFL employs an iterative prompt-sharing mechanism, allowing clients to optimize and exchange textual prompts to enhance generalization while drastically reducing communication overhead. We validate ZeroDFL on nine diverse image classification datasets, demonstrating that it consistently outperforms--or remains on par with--state-of-the-art federated prompt learning methods. More importantly, ZeroDFL achieves this performance in a fully decentralized setting while reducing communication overhead by 118x compared to FedTPG. These results highlight that our approach not only enhances generalization in federated zero-shot learning but also improves scalability, efficiency, and privacy preservation--paving the way for decentralized adaptation of large vision-language models in real-world applications.

Controllable summarization moves beyond generic outputs toward human-aligned summaries guided by specified attributes. In practice, the interdependence among attributes makes it challenging for language models to satisfy correlated constraints consistently. Moreover, previous approaches often require per-attribute fine-tuning, limiting flexibility across diverse summary attributes. In this paper, we propose adaptive planning for multi-attribute controllable summarization (PACO), a training-free framework that reframes the task as planning the order of sequential attribute control with a customized Monte Carlo Tree Search (MCTS). In PACO, nodes represent summaries, and actions correspond to single-attribute adjustments, enabling progressive refinement of only the attributes requiring further control. This strategy adaptively discovers optimal control orders, ultimately producing summaries that effectively meet all constraints. Extensive experiments across diverse domains and models demonstrate that PACO achieves robust multi-attribute controllability, surpassing both LLM-based self-planning models and fine-tuned baselines. Remarkably, PACO with Llama-3.2-1B rivals the controllability of the much larger Llama-3.3-70B baselines. With larger models, PACO achieves superior control performance, outperforming all competitors.