A new preprint demonstrates that temperature scaling, a widely used model calibration technique, systematically misrepresents model reliability when ground-truth labels are soft or distributional, such as those from crowd-sourced human annotations. Evaluating nine model configurations on the CIFAR-10H and ChaosNLI datasets, the study finds that temperature scaling calibrated on hard labels consistently underperforms an oracle calibrated on soft labels, with calibration gaps notably larger in language tasks (mean 0.079) than in vision tasks (mean 0.003). The results hold across model scales and with an alternative calibration method, multiclass isotonic regression.
Why it matters: The findings highlight that standard calibration protocols relying on majority-vote labels can give a misleading sense of model reliability in real-world scenarios with inherent label ambiguity, posing risks for safety-critical AI deployments.
Researchers introduce EXPLORE, a framework that integrates simulator-guided Monte Carlo Tree Search with transformer-based language model decoding to improve analog circuit topology generation. On a 6-component benchmark with tight tolerance, EXPLORE achieves a 65% success rate, outperforming one-shot generation (12%) and sampling-and-filter baselines (33%). The framework also reduces mean squared error by over 20% compared to sampling-and-filter under the same search budget.
Why it matters: This work demonstrates a significant advance in automating analog circuit design by enabling language models to generate complex topologies more reliably through structured search.
Researchers have introduced HEDGEHOG, a six-stage filtration benchmark designed to rigorously evaluate generative molecular models for drug discovery. When 23 different generators were tested on 230,000 molecules, only 0.65% of the generated compounds passed all stages, which include medicinal chemistry, synthesis feasibility, and 3D docking constraints. This finding highlights that current AI models rarely produce molecules that meet all practical requirements for drug candidates.
Why it matters: HEDGEHOG exposes a significant gap between the theoretical capabilities of AI-driven molecular generators and their practical utility in real-world drug discovery.
Agora is a system that enables efficient pretraining of large language models using heterogeneous, individually owned GPUs connected via the internet. By combining bandwidth-efficient pipeline parallelism with fault-tolerant collective operations, Agora successfully trained an 8.6B-parameter model on 500B tokens using 330 contributor nodes over 40 days. The system achieved 63% of the efficiency of a centralized H100 GPU cluster, demonstrating the feasibility of large-scale, decentralized model training.
Why it matters: This work shows that large-scale AI model training can be decentralized and permissionless, potentially broadening access to frontier AI development beyond traditional data centers.
A new method called DIVE is introduced for compressing language-model embeddings using a residual compression adapter that incorporates a self-limiting hinge loss and geometry distillation. In experiments on five BEIR benchmarks with LLM2Vec backbones, DIVE consistently outperforms six baseline methods, including PCA and autoencoders, at both 128- and 256-dimensional outputs.
Why it matters: DIVE enables more efficient storage and retrieval in large-scale information retrieval systems by compressing embeddings without sacrificing retrieval quality.
Policy & Safety→Official→arXiv Computers and Society
A preprint study analyzing 29 language models across 177 occupations finds that these models incorporate demographic information into simulated hiring decisions, advantaging female and Black candidates while penalizing disabled candidates. The research shows that post-training alignment—intended to make models more helpful and aligned with human preferences—substantially amplifies these demographic effects, with the female and Black advantage increasing by nearly 400% and the disability penalty worsening by over 150%.
Why it matters: The findings highlight that alignment processes, while designed to improve AI behavior, can unintentionally exacerbate certain forms of discrimination, particularly against disabled individuals, in high-stakes contexts like hiring.
Policy & Safety→Official→arXiv Computers and Society
A new preprint analyzes the environmental impacts of sovereign AI infrastructure in the Global South, focusing on water, energy, and carbon emissions. The study finds that a 1,024-GPU cluster using evaporative cooling in the UAE would consume over 30 million liters of water annually, despite the country's extremely high water stress. The authors identify a 'sovereignty-sustainability trilemma' and propose design principles such as mandatory water usage reporting and prioritizing smaller, more efficient language models.
Why it matters: The research underscores the urgent need for policymakers in water- and climate-vulnerable regions to consider environmental sustainability when planning AI infrastructure.
This preprint reports the first classroom deployment of LEA, an adaptive AI tutoring agent, with real students and evaluates its scalability across three different courses. The study finds that synthetic (simulated) evaluation does not fully predict real-world classroom performance: while answer relevancy and context precision remain stable across courses, faithfulness of responses declines as the curriculum diverges from the system's original subject. These results highlight the need for further research into making AI tutoring systems fully course-agnostic.
Why it matters: This work provides early empirical evidence on the challenges of deploying AI tutoring systems in real classrooms and exposes the limitations of relying solely on synthetic evaluation for predicting real-world performance.
Researchers introduce CtrlBench-Rec, a collaborative multi-agent framework designed to systematically assess the controllability of recommender systems, which are often treated as black boxes. The framework formalizes three key tasks—target content discovery, interest profile shaping, and popularity bias mitigation—to measure how well recommender systems can be steered by explicit or implicit user guidance. Experiments across real-world datasets and models show that CtrlBench-Rec quantifies controllability and highlights persistent challenges, such as resistance to promoting long-tail content.
Why it matters: CtrlBench-Rec provides the first standardized toolkit for evaluating and auditing the controllability of recommender systems, addressing a critical gap in algorithmic transparency and user empowerment.
OvisOCR2 is a 0.8B parameter end-to-end document parsing model that converts document page images into Markdown in natural reading order, handling text, formulas, tables, and visual regions. It achieves state-of-the-art scores of 96.58 on OmniDocBench v1.6 and 75.06 on PureDocBench, surpassing previous pipeline-based methods. The model's training involves a data engine combining real and synthetic data, reinforcement learning, and model fusion.
Why it matters: This result shows that compact end-to-end models can outperform complex pipeline methods in document parsing, potentially simplifying deployment and improving accuracy for document understanding tasks.
A new preprint investigates how large language models (LLMs) acquire and adjust to human values during post-training. The study finds that supervised fine-tuning (SFT) largely determines a model's value alignment, while subsequent preference optimization rarely changes these values significantly. Experiments with Llama-3 and Qwen-3 models further show that different preference optimization algorithms can result in different value alignment outcomes, even when using the same data.
Why it matters: Understanding when and how LLMs learn human values can guide better data curation and algorithm choices for improved model alignment.
Researchers investigate cross-rubric generalization in automated essay scoring, where models trained on essays labeled with one set of rubrics are evaluated on essays scored with previously unseen rubrics. By introducing rubric-agnostic intermediate representations called 'traits' and using a fine-tuning framework, they achieve a 5.0% macro F1 improvement over baselines in the most challenging setting. Their best open-source Llama-based model also outperforms GPT-5-mini prompting by 2.1% macro F1.
Why it matters: This work demonstrates a method for automated essay scoring systems to adapt to new or revised scoring rubrics without retraining, addressing a practical challenge in educational assessment.
Researchers have introduced L2-Bench, an open-source benchmark comprising over 1,000 task-response pairs designed to evaluate large language models (LLMs) in second language (L2) education. The benchmark covers 12 competencies and 31 subcompetencies, validated by more than 200 expert practitioners, and uses a rubric-based evaluation methodology. Results show that Claude Opus 4.7 achieves the highest overall performance at 85.5%, though it is marginally outperformed on some specific tasks. The study also finds that model performance declines on more challenging tasks.
Why it matters: L2-Bench offers a rigorous, pedagogy-driven framework for assessing AI models in language education, enabling stakeholders to make more informed decisions about AI adoption in this field.
A study analyzing Kaggle contest submissions from 2019 to mid-2026 finds that AI coding assistants have led to substantial syntactic homogenization—code structure and literal syntax have become more similar—while semantic diversity, reflecting problem-solving approaches, has remained stable or even expanded. The research also documents widespread convergence toward the random seed value 42, consistent with LLMs reinforcing established programming conventions. These findings are based on both surface-level (TF-IDF) and semantic (embedding-based) analyses of code similarity.
Why it matters: This suggests that while AI coding assistants standardize implementation details, they do not currently reduce the diversity of problem-solving strategies, informing debates about software monoculture and innovation.
A new survey of 45 studies on Generative Engine Optimization (GEO) finds that while interventions can affect how already-retrieved content is cited or used by generative engines, no technique demonstrates a stable, long-term, cross-platform causal effect on organic discoverability. The paper introduces a multistage formal model and a reproducible protocol for evaluating GEO methods, highlighting the complexity and variability of the field. The review also notes that commonly cited gains are often conditional and do not generalize to broader visibility or traffic outcomes.
Why it matters: This survey clarifies the limited empirical support for GEO strategies, helping practitioners and researchers focus on interventions with proven effects and avoid overestimating their impact on generative engine visibility.
Researchers identify 'blind spots' in neural retrievers used in retrieval-augmented generation (RAG) systems, where relevant entities are not retrieved due to low embedding similarity. They introduce a Retrieval Probability Score (RPS) to predict these blind spots before indexing, and present ARGUS, a pipeline that augments documents from a knowledge base to improve retrievability. Experiments across multiple retrievers and datasets show consistent improvements in retrieval metrics, with average gains of +3.4 nDCG@5 and +4.5 nDCG@10.
Why it matters: This work tackles a key reliability challenge in RAG systems by enabling the detection and mitigation of retrieval blind spots, which is important for building more robust and trustworthy AI applications.
A new preprint audits nine leading large language models (LLMs) for asset-specific preferences and finds that Bitcoin's ranking among money-like instruments is highly dependent on the scenario presented, with models ranking it higher in crisis and autonomous-agent contexts. The study identifies a dominant internal feature in Gemma 3 that selectively represents Bitcoin; manipulating this feature can causally shift the model's portfolio allocation toward or away from Bitcoin by up to 5.2 percentage points. This demonstrates that internal model representations can be both audited and causally linked to real financial decisions.
Why it matters: This work provides a concrete method for auditing and influencing asset-specific preferences in financial LLMs, laying groundwork for transparency and accountability as such models are deployed in real-world financial decision-making.
Researchers present CwA (Cluster with Auctions), a method that jointly learns balanced database partitions and a neural probing function for approximate nearest neighbor search. Unlike traditional approaches that use the same assignment function for both queries and database vectors, CwA decouples these processes and directly optimizes search performance for the query distribution. The method employs a parallelizable auction algorithm for combinatorial optimization and demonstrates up to 4.7× throughput improvement over state-of-the-art methods when query and database distributions differ.
Why it matters: This work offers a significant advance in vector search efficiency, particularly in scenarios where query and database distributions are not aligned, by optimizing both partitioning and probing functions jointly.
A new preprint decomposes the sources of non-determinism in large language model (LLM) brand measurements, revealing that query language accounts for 26.5% of the variance in brand sentiment responses, while brand identity itself contributes only 1.5%. The authors recommend that, to improve measurement reliability, practitioners should allocate resources to sampling across different languages and models rather than simply repeating prompts.
Why it matters: This work provides a rigorous framework for understanding and mitigating reproducibility issues in LLM-based brand measurement, offering actionable guidance for more reliable evaluation.
A new preprint demonstrates that open-vocabulary object detectors maintain nearly constant confidence scores even as the visible portion of a target object drops to one-eighth, and often report the target present in about 90% of scenes, even when only a same-category distractor is visible. This indicates that detector confidence signals the presence of the object category somewhere in the scene, rather than the visibility of the specific target. The finding holds across three detectors, nine object categories, multiple simulators, and real video, and exposes a fundamental flaw in using detector confidence as a proxy for occlusion or visibility.
Why it matters: This work challenges a common assumption in robotics and computer vision, showing that detector confidence cannot reliably indicate object visibility, which impacts evaluation, active perception, and downstream applications.