AI agents with advanced reasoning and tool use capabilities have demonstrated impressive performance in web browsing for deep search. While existing benchmarks such as BrowseComp evaluate these browsing abilities, they primarily focus on textual information, overlooking the prevalence of multimodal content. To bridge this gap, we introduce MM-BrowseComp, a novel benchmark comprising 224 challenging, hand-crafted questions specifically designed to assess agents' multimodal retrieval and reasoning capabilities. These questions often incorporate images in prompts, and crucial information encountered during the search and reasoning process may also be embedded within images or videos on webpages. Consequently, methods relying solely on text prove insufficient for our benchmark. Additionally, we provide a verified checklist for each question, enabling fine-grained analysis of multimodal dependencies and reasoning paths. Our comprehensive evaluation of state-of-the-art models on MM-BrowseComp reveals that even top models like OpenAI o3 with tools achieve only 29.02\% accuracy, highlighting the suboptimal multimodal capabilities and lack of native multimodal reasoning in current models.

Data contamination refers to the leakage of evaluation data into model training data, resulting in overfitting to supposedly held-out test sets and compromising test validity. We identify an analogous issue, search-time contamination (STC), in evaluating search-based LLM agents which use tools to gather information from online sources when answering user queries. STC occurs when the retrieval step surfaces a source containing the test question (or a near-duplicate) alongside its answer, enabling agents to copy rather than genuinely infer or reason, undermining benchmark integrity. We find that HuggingFace, an online platform hosting evaluation datasets, appears among retrieved sources in search based agent logs. Consequently, agents often explicitly acknowledge discovering question answer pairs from HuggingFace within their reasoning chains. On three commonly used capability benchmarks: Humanity's Last Exam (HLE), SimpleQA, and GPQA, we demonstrate that for approximately 3% of questions, search-based agents directly find the datasets with ground truth labels on HuggingFace. When millions of evaluation queries target the same benchmark, even small, repeated leaks can accelerate the benchmark's obsolescence, shortening its intended lifecycle. After HuggingFace is blocked, we observe a drop in accuracy on the contaminated subset of approximately 15%. We further show through ablation experiments that publicly accessible evaluation datasets on HuggingFace may not be the sole source of STC. To this end, we conclude by proposing best practices for benchmark design and result reporting to address this novel form of leakage and ensure trustworthy evaluation of search-based LLM agents. To facilitate the auditing of evaluation results, we also publicly release the complete logs from our experiments.

Large Language Models (LLMs) face fundamental limitations in context management despite recent advances extending context windows to millions of tokens. We propose Cognitive Workspace, a novel paradigm that transcends traditional Retrieval-Augmented Generation (RAG) by emulating human cognitive mechanisms of external memory use. Drawing from cognitive science foundations including Baddeley's working memory model, Clark's extended mind thesis, and Hutchins' distributed cognition framework, we demonstrate that current passive retrieval systems fail to capture the dynamic, task-driven nature of human memory management. Our analysis of 2024-2025 developments reveals that while techniques like Infini-attention and StreamingLLM achieve impressive context lengths, they lack the metacognitive awareness and active planning capabilities essential for true cognitive extension. Cognitive Workspace addresses these limitations through three core innovations: (1) active memory management with deliberate information curation, (2) hierarchical cognitive buffers enabling persistent working states, and (3) task-driven context optimization that dynamically adapts to cognitive demands. Empirical validation demonstrates Cognitive Workspace achieves an average 58.6% memory reuse rate (ranging from 54-60% across different tasks) compared to 0% for traditional RAG, with 17-18% net efficiency gain despite 3.3x higher operation counts. Statistical analysis confirms these advantages with p < 0.001 and Cohen's d > 23 across multiple task types, establishing the first quantitative evidence for active memory superiority in LLM systems. We present a comprehensive theoretical framework synthesizing insights from 50+ recent papers, positioning Cognitive Workspace as a fundamental shift from information retrieval to genuine cognitive augmentation.

--In particular, large-scale deep learning and artificial intelligence model training uses a lot of computational power and energy, so it poses serious sustainability issues. The fast rise in model complexity has resulted in exponential increases in energy consumption, increasing the demand for techniques maximizing computational efficiency and lowering environmental impact. This work explores environmentally driven performance optimization methods especially intended for advanced GPU architectures from NVIDIA, AMD, and other emerging GPU architectures. Our main focus is on investigating hardware-software co-design techniques meant to significantly increase memory-level and kernel-level operations, so improving performance-per-watt measures. Our thorough research encompasses evaluations of specialized tensor and matrix cores, advanced memory optimization methods, and creative integration approaches that taken together result in notable energy efficiency increases. We also discuss important software-level optimizations that augment hardware capability including mixed-precision arithmetic, advanced energy-aware scheduling algorithms, and compiler-driven kernel enhancements. Moreover, we methodically point out important research gaps and suggest future directions necessary to create really sustainable artificial intelligence systems. This paper emphasizes how major increases in training efficiency can be obtained by co-design of hardware and software, so lowering the environmental impact of artificial intelligence without compromising performance. T o back up our analysis, we use real-world case studies from top companies like Meta, Google, Amazon, and others that show how these sustainable AI training methods are used in the real world. With this thorough analysis, we show that a comprehensive co-design approach can significantly increase training efficiency and lower the carbon footprint of AI without compromising performance. Training modern artificial intelligence models calls for enormous computational resources and energy. Training a single large language model (LLM) such GPT -3 was projected to consume almost 1,300 MWh of electricity--equivalent to the annual power consumption of roughly 130 U.S. homes [1].

--AI hardware design is advancing rapidly, driven by the promise of design automation to make chip development faster, more efficient, and more accessible to a wide range of users. However, their potential is hindered by data privacy concerns and the lack of domain-specific training. T o address this, we introduce FedChip, a Fed erated fine-tuning approach that enables multiple Chip design parties to collaboratively enhance a shared LLM dedicated for automated hardware design generation while protecting proprietary data. FedChip enables parties to train the model on proprietary local data and improve the shared LLM's performance. T o exemplify FedChip's deployment, we create and release APTPU-Gen, a dataset of 30k design variations spanning various performance metric values such as power, performance, and area (PPA). T o encourage the LLM to generate designs that achieve a balance across multiple quality metrics, we propose a new design evaluation metric, Chip@k, which statistically evaluates the quality of generated designs against predefined acceptance criteria. Experimental results show that FedChip improves design quality by more than 77% over high-end LLMs while maintaining data privacy. The global Artificial Intelligence (AI) hardware market, fueled by innovations in deep learning accelerators, autonomous systems, and data-intensive applications, is expected to be worth $84.9 billion by 2031 [1].

Large Language Models (LLMs) have demonstrated great potential in automating the generation of Verilog hardware description language code for hardware design. This automation is critical to reducing human effort in the complex and error-prone process of hardware design. However, existing approaches predominantly rely on human intervention and fine-tuning using curated datasets, limiting their scalability in automated design workflows. Although recent iterative search techniques have emerged, they often fail to explore diverse design solutions and may underperform simpler approaches such as repeated prompting. To address these limitations, we introduce EvoVerilog, a novel framework that combines the reasoning capabilities of LLMs with evolutionary algorithms to automatically generate and refine Verilog code. EvoVerilog utilizes a multiobjective, population-based search strategy to explore a wide range of design possibilities without requiring human intervention. Extensive experiments demonstrate that EvoVerilog achieves state-of-the-art performance, with pass@10 scores of 89.1 and 80.2 on the VerilogEval-Machine and VerilogEval-Human benchmarks, respectively. Furthermore, the framework showcases its ability to explore diverse designs by simultaneously generating a variety of functional Verilog code while optimizing resource utilization.

The rapid expansion of medical literature presents growing challenges for structuring and integrating domain knowledge at scale. Knowledge Graphs (KGs) offer a promising solution by enabling efficient retrieval, automated reasoning, and knowledge discovery. However, current KG construction methods often rely on supervised pipelines with limited generalizability or naively aggregate outputs from Large Language Models (LLMs), treating biomedical corpora as static and ignoring the temporal dynamics and contextual uncertainty of evolving knowledge. To address these limitations, we introduce MedKGent, a LLM agent framework for constructing temporally evolving medical KGs. Leveraging over 10 million PubMed abstracts published between 1975 and 2023, we simulate the emergence of biomedical knowledge via a fine-grained daily time series. MedKGent incrementally builds the KG in a day-by-day manner using two specialized agents powered by the Qwen2.5-32B-Instruct model. The Extractor Agent identifies knowledge triples and assigns confidence scores via sampling-based estimation, which are used to filter low-confidence extractions and inform downstream processing. The Constructor Agent incrementally integrates the retained triples into a temporally evolving graph, guided by confidence scores and timestamps to reinforce recurring knowledge and resolve conflicts. The resulting KG contains 156,275 entities and 2,971,384 relational triples. Quality assessments by two SOTA LLMs and three domain experts demonstrate an accuracy approaching 90%, with strong inter-rater agreement. To evaluate downstream utility, we conduct RAG across seven medical question answering benchmarks using five leading LLMs, consistently observing significant improvements over non-augmented baselines. Case studies further demonstrate the KG's value in literature-based drug repurposing via confidence-aware causal inference.

Legal Artificial Intelligence (LegalAI) has achieved notable advances in automating judicial decision-making with the support of Large Language Models (LLMs). However, existing legal LLMs still struggle to generate reliable and interpretable reasoning processes. They often default to fast-thinking behavior by producing direct answers without explicit multi-step reasoning, limiting their effectiveness in complex legal scenarios that demand rigorous justification. To address this challenge, we propose Legal$Δ$, a reinforcement learning framework designed to enhance legal reasoning through chain-of-thought guided information gain. During training, Legal$Δ$ employs a dual-mode input setup-comprising direct answer and reasoning-augmented modes-and maximizes the information gain between them. This encourages the model to acquire meaningful reasoning patterns rather than generating superficial or redundant explanations. Legal$Δ$ follows a two-stage approach: (1) distilling latent reasoning capabilities from a powerful Large Reasoning Model (LRM), DeepSeek-R1, and (2) refining reasoning quality via differential comparisons, combined with a multidimensional reward mechanism that assesses both structural coherence and legal-domain specificity. Experimental results on multiple legal reasoning tasks demonstrate that Legal$Δ$ outperforms strong baselines in both accuracy and interpretability. It consistently produces more robust and trustworthy legal judgments without relying on labeled preference data. All code and data will be released at https://github.com/NEUIR/LegalDelta.

Bioacoustics, the study of sounds produced by living organisms, plays a vital role in conservation, biodiversity monitoring, and behavioral studies. Many tasks in this field, such as species, individual, and behavior classification and detection, are well-suited to machine learning. However, they often suffer from limited annotated data, highlighting the need for a general-purpose bioacoustic encoder capable of extracting useful representations for diverse downstream tasks. Such encoders have been proposed before, but are often limited in scope due to a focus on a narrow range of species (typically birds), and a reliance on a single model architecture or training paradigm. Moreover, they are usually evaluated on a small set of tasks and datasets. In this work, we present a large-scale empirical study that covers aspects of bioacoustics that are relevant to research but have previously been scarcely considered: training data diversity and scale, model architectures and training recipes, and the breadth of evaluation tasks and datasets. We obtain encoders that are state-of-the-art on the existing and proposed benchmarks. We also identify what matters for training these encoders, such that this work can be extended when more data are available or better architectures are proposed. Specifically, across 26 datasets with tasks including species classification, detection, individual ID, and vocal repertoire discovery, we find self-supervised pre-training followed by supervised post-training on a mixed bioacoustics + general-audio corpus yields the strongest in- and out-of-distribution performance. We show the importance of data diversity in both stages. To support ongoing research and application, we will release the model checkpoints.

Large language models (LLMs) remain acutely vulnerable to prompt injection and related jailbreak attacks; heuristic guardrails (rules, filters, LLM judges) are routinely bypassed. We present Contextual Integrity Verification (CIV), an inference-time security architecture that attaches cryptographically signed provenance labels to every token and enforces a source-trust lattice inside the transformer via a pre-softmax hard attention mask (with optional FFN/residual gating). CIV provides deterministic, per-token non-interference guarantees on frozen models: lower-trust tokens cannot influence higher-trust representations. On benchmarks derived from recent taxonomies of prompt-injection vectors (Elite-Attack + SoK-246), CIV attains 0% attack success rate under the stated threat model while preserving 93.1% token-level similarity and showing no degradation in model perplexity on benign tasks; we note a latency overhead attributable to a non-optimized data path. Because CIV is a lightweight patch -- no fine-tuning required -- we demonstrate drop-in protection for Llama-3-8B and Mistral-7B. We release a reference implementation, an automated certification harness, and the Elite-Attack corpus to support reproducible research.