Sparse autoencoders (SAEs) are designed to extract interpretable features from language models by enforcing a sparsity constraint. Ideally, training an SAE would yield latents that are both sparse and semantically meaningful. However, many SAE latents activate frequently (i.e., are \emph{dense}), raising concerns that they may be undesirable artifacts of the training procedure. In this work, we systematically investigate the geometry, function, and origin of dense latents and show that they are not only persistent but often reflect meaningful model representations. We first demonstrate that dense latents tend to form antipodal pairs that reconstruct specific directions in the residual stream, and that ablating their subspace suppresses the emergence of new dense features in retrained SAEs -- suggesting that high density features are an intrinsic property of the residual space. We then introduce a taxonomy of dense latents, identifying classes tied to position tracking, context binding, entropy regulation, letter-specific output signals, part-of-speech, and principal component reconstruction. Finally, we analyze how these features evolve across layers, revealing a shift from structural features in early layers, to semantic features in mid layers, and finally to output-oriented signals in the last layers of the model. Our findings indicate that dense latents serve functional roles in language model computation and should not be dismissed as training noise.

While large language models (LLMs) are increasingly used for generating parallel scientific codes, most efforts emphasize functional correctness, often overlooking performance, especially energy efficiency. We propose LASSI-EE, an automated LLM-based refactoring framework that generates energy-efficient parallel codes through a multi-stage, iterative approach integrating runtime power profiling, energy-aware prompting, self-correcting feedback loops, and an LLM-as-a-Judge agent for automated screening of code solutions. We introduce energy-reduction@k, a novel metric that quantifies expected energy reduction when generating k code candidates and selecting the most energy-efficient, enabling systematic evaluation of multi-attempt generation strategies. Evaluating 20 HeCBench applications and two miniApps on NVIDIA A100 and AMD MI100 GPUs, a single run (k=1) with LASSI-EE delivers refactored parallel codes with an average 29% expected energy reduction at an 81% pass rate, representing a 2.8x improvement over vanilla LLM prompting. Multiple runs (k=3) achieve an average 48% expected energy reduction at a 97% pass rate. These results are consistent across devices, demonstrating LASSI-EE's effectiveness across diverse hardware architectures.

Artificial intelligence (AI) systems hold great promise for advancing various scientific disciplines, and are increasingly used in real-world applications. Despite their remarkable progress, further capabilities are expected in order to achieve more general types of intelligence. A critical distinction in this context is between factual knowledge, which can be evaluated against true or false answers (e.g., "what is the capital of England?"), and probabilistic knowledge, reflecting probabilistic properties of the real world (e.g., "what is the sex of a computer science graduate in the US?"). In this paper, our goal is to build a benchmark for understanding the capabilities of LLMs in terms of knowledge of probability distributions describing the real world. Given that LLMs are trained on vast amounts of text, it may be plausible that they internalize aspects of these distributions. Indeed, LLMs are touted as powerful universal approximators of real-world distributions. At the same time, classical results in statistics, known as curse of dimensionality, highlight fundamental challenges in learning distributions in high dimensions, challenging the notion of universal distributional learning. In this work, we develop the first benchmark to directly test this hypothesis, evaluating whether LLMs have access to empirical distributions describing real-world populations across domains such as economics, health, education, and social behavior. Our results demonstrate that LLMs perform poorly overall, and do not seem to internalize real-world statistics naturally. When interpreted in the context of Pearl's Causal Hierarchy (PCH), our benchmark demonstrates that language models do not contain knowledge on observational distributions (Layer 1 of PCH), and thus the Causal Hierarchy Theorem implies that interventional (Layer 2) and counterfactual (Layer 3) knowledge of these models is also limited.

When you purchase through links in our articles, we may earn a small commission. Microsoft announced these same features about a month ago, though they're still in testing. Google has finally begun adding support for its own internal services to Google Gemini, just a month after Microsoft began offering the same capabilities to Windows testers. Google said Wednesday that Gemini Deep Research can now connect to Gmail, Google Drive, and Chat, along with Docs, Slides, Sheets, and PDF files stored within those services. "This powerful new capability is now available for all Gemini users," Google said in a blog post on Wednesday.

SoftBank Group's Japanese mobile unit and OpenAI are set to launch AI services for local companies next year. SoftBank Group's Japanese mobile unit and OpenAI will launch AI services for local companies next year, seeking to realize real revenue in the face of growing concerns over sky-high valuations. SoftBank Corp. and Open AI are still fine-tuning the products the two companies are co-developing for Japanese enterprises, said Junichi Miyakawa, president of the country's third-largest mobile carrier. Miyakawa said he has seen a test version of the services, which once launched would "completely change" the speed in which business is done. One feature is voice recognition that would allow users to rely less on manual typing, he said.

This year, we've seen a real-time experiment playing out across the technology industry, one in which AI's software engineering capabilities have been put to the test against human technologists. And although 2025 may have started with AI looking strong, the transition from vibe coding to what's being termed context engineering shows that while the work of human developers is evolving, they nevertheless remain absolutely critical. This is captured in the latest volume of the " Thoughtworks Technology Radar," a report on the technologies used by our teams on projects with clients. In it, we see the emergence of techniques and tooling designed to help teams better tackle the problem of managing context when working with LLMs and AI agents. Taken together, there's a clear signal of the direction of travel in software engineering and even AI more broadly. After years of the industry assuming progress in AI is all about scale and speed, we're starting to see that what matters is the ability to handle context effectively.

ChatGPT nickname and Trump tariffs nominated for Japan's buzzword of 2025 Women hold toys depicting Myaku-Myaku, the official character for the 2025 Osaka Expo, during a media day ahead of the event's public opening day in Osaka in April. A comment made by Prime Minister Sanae Takaichi immediately after she was elected president of the Liberal Democratic Party last month -- that she would "work, work, work, work and work," which triggered both praise for her determination and criticism amid efforts to improve work-life balance -- also made the list, along with (female prime minister). "There weren't many buzzwords in the first half of the year, but after (U.S.) President Donald Trump returned to office, many phrases went viral regarding tariffs," publisher Jiyukokuminsha, which hosts the award, said in a statement. In a time of both misinformation and too much information, quality journalism is more crucial than ever. By subscribing, you can help us get the story right.

In this work, we introduce Progressive Growing of Patch Size, an automatic curriculum learning approach for 3D medical image segmentation. Our approach progressively increases the patch size during model training, resulting in an improved class balance for smaller patch sizes and accelerated convergence of the training process. We evaluate our curriculum approach in two settings: a resource-efficient mode and a performance mode, both regarding Dice score performance and computational costs across 15 diverse and popular 3D medical image segmentation tasks. The resource-efficient mode matches the Dice score performance of the conventional constant patch size sampling baseline with a notable reduction in training time to only 44%. The performance mode improves upon constant patch size segmentation results, achieving a statistically significant relative mean performance gain of 1.28% in Dice Score. Remarkably, across all 15 tasks, our proposed performance mode manages to surpass the constant patch size baseline in Dice Score performance, while simultaneously reducing training time to only 89%. The benefits are particularly pronounced for highly imbalanced tasks such as lesion segmentation tasks. Rigorous experiments demonstrate that our performance mode not only improves mean segmentation performance but also reduces performance variance, yielding more trustworthy model comparison. Furthermore, our findings reveal that the proposed curriculum sampling is not tied to a specific architecture but represents a broadly applicable strategy that consistently boosts performance across diverse segmentation models, including UNet, UNETR, and SwinUNETR. In summary, we show that this simple yet elegant transformation on input data substantially improves both Dice Score performance and training runtime, while being compatible across diverse segmentation backbones.

Retrieval-Augmented Generation (RAG) has emerged as the dominant paradigm for grounding large language model outputs in verifiable evidence. However, as modern AI agents transition from static knowledge bases to continuous multimodal streams encompassing text, images, video, and audio, two critical challenges arise: maintaining index freshness without prohibitive re-indexing costs, and preserving cross-modal semantic consistency across heterogeneous embedding spaces. We present LUMA-RAG, a lifelong multimodal agent architecture featuring three key innovations: (i) a streaming, multi-tier memory system that dynamically spills embeddings from a hot HNSW tier to a compressed IVFPQ tier under strict memory budgets; (ii) a streaming CLAP->CLIP alignment bridge that maintains cross-modal consistency through incremental orthogonal Procrustes updates; and (iii) stability-aware retrieval telemetry providing Safe@k guarantees by jointly bounding alignment drift and quantization error. Experiments demonstrate robust text-to-image retrieval (Recall@10 = 0.94), graceful performance degradation under product quantization offloading, and provably stable audio-to-image rankings (Safe@1 = 1.0), establishing LUMA-RAG as a practical framework for production multimodal RAG systems.

Large Language Models (LLMs) trained with reinforcement learning and verifiable rewards have achieved strong results on complex reasoning tasks. Recent work extends this paradigm to a multi-agent setting, where a meta-thinking agent proposes plans and monitors progress while a reasoning agent executes subtasks through sequential conversational turns. Despite promising performance, we identify a critical limitation: lazy agent behavior, in which one agent dominates while the other contributes little, undermining collaboration and collapsing the setup to an ineffective single agent. In this paper, we first provide a theoretical analysis showing why lazy behavior naturally arises in multi-agent reasoning. We then introduce a stable and efficient method for measuring causal influence, helping mitigate this issue. Finally, as collaboration intensifies, the reasoning agent risks getting lost in multi-turn interactions and trapped by previous noisy responses. To counter this, we propose a verifiable reward mechanism that encourages deliberation by allowing the reasoning agent to discard noisy outputs, consolidate instructions, and restart its reasoning process when necessary. Extensive experiments demonstrate that our framework alleviates lazy agent behavior and unlocks the full potential of multi-agent framework for complex reasoning tasks. Techniques such as chain-of-thought prompting (Wei et al., 2022; Kojima et al., 2022) and structured methods like Tree-of-Thoughts and Graph-of-Thoughts (Y ao et al., 2023; Besta et al., 2024) expand the space for deliberation. More recently, multi-agent frameworks enable LLMs with specialized roles to collaborate via planning, delegation, and debate, echoing human team dynamics (Li et al., 2023; Wu et al., 2024a; Chen et al., 2023; Du et al., 2023; Y uan & Xie). To support multi-agent and multi-turn reinforcement learning, multi-turn Group Relative Preference Optimization (GRPO) (Wan et al., 2025; Shi et al., 2025; Wei et al., 2025) and its variants (Guo et al., 2025b; Zhang et al., 2025c; Ning et al., 2025; Xue et al., 2025) compute advantages and importance ratios at the turn level, enabling finer-grained optimization and more precise credit assignment. Building on this foundation, ReMA (Wan et al., 2025) introduces a multi-agent LLM reasoning framework with two specialized roles: a meta-thinking agent, which decomposes tasks, sets intermediate goals, and adapts based on feedback, and a reasoning agent, which performs step-by-step 1 The agents alternate sequentially, but since only a final outcome reward is available, ReMA computes a group advantage following GRPO (Shao et al., 2024) and uniformly assigns this trajectory-level signal to every turn in the rollout.