We release Super Apriel, a 15B-parameter supernet in which every decoder layer provides four trained mixer choices -- Full Attention (FA), Sliding Window Attention (SWA), Kimi Delta Attention (KDA), and Gated DeltaNet (GDN). A placement selects one mixer per layer; placements can be switched between requests at serving time without reloading weights, enabling multiple speed presets from a single checkpoint. The shared checkpoint also enables speculative decoding without a separate draft model. The all-FA preset matches the Apriel 1.6 teacher on all reported benchmarks; recommended hybrid presets span $2.9\times$ to $10.7\times$ decode throughput at 96% to 77% quality retention, with throughput advantages that compound at longer context lengths. With four mixer types across 48 layers, the configuration space is vast. A surrogate that predicts placement quality from the per-layer mixer assignment makes the speed-quality landscape tractable and identifies the best tradeoffs at each speed level. We investigate whether the best configurations at each speed level can be identified early in training or only after convergence. Rankings stabilize quickly at 0.5B scale, but the most efficient configurations exhibit higher instability at 15B, cautioning against extrapolation from smaller models. Super Apriel is trained by stochastic distillation from a frozen Apriel 1.6 teacher, followed by supervised fine-tuning. We release the supernet weights, Fast-LLM training code, vLLM serving code, and a placement optimization toolkit.

Edge-scale deep research agents based on small language models are attractive for real-world deployment due to their advantages in cost, latency, and privacy. In this work, we study how to train a strong small deep research agent under limited open-data by improving both data quality and data utilization. We present DR-Venus, a frontier 4B deep research agent for edge-scale deployment, built entirely on open data. Our training recipe consists of two stages. In the first stage, we use agentic supervised fine-tuning (SFT) to establish basic agentic capability, combining strict data cleaning with resampling of long-horizon trajectories to improve data quality and utilization. In the second stage, we apply agentic reinforcement learning (RL) to further improve execution reliability on long-horizon deep research tasks. To make RL effective for small agents in this setting, we build on IGPO and design turn-level rewards based on information gain and format-aware regularization, thereby enhancing supervision density and turn-level credit assignment. Built entirely on roughly 10K open-data, DR-Venus-4B significantly outperforms prior agentic models under 9B parameters on multiple deep research benchmarks, while also narrowing the gap to much larger 30B-class systems. Our further analysis shows that 4B agents already possess surprisingly strong performance potential, highlighting both the deployment promise of small models and the value of test-time scaling in this setting. We release our models, code, and key recipes to support reproducible research on edge-scale deep research agents.

Reinforcement fine-tuning with verifiable rewards (RLVR) has emerged as a powerful paradigm for equipping large vision-language models (LVLMs) with agentic capabilities such as tool use and multi-step reasoning. Despite striking empirical successes, most notably Visual Agentic Reinforcement Fine-Tuning (Visual-ARFT), the theoretical underpinnings of this paradigm remain poorly understood. In particular, two critical questions lack rigorous answers: (i)~how does the composite structure of verifiable rewards (format compliance, answer accuracy, tool executability) affect the convergence of Group Relative Policy Optimization (GRPO), and (ii)~why does training on a small set of tool-augmented tasks transfer to out-of-distribution domains? We address these gaps by introducing the \emph{Tool-Augmented Markov Decision Process} (TA-MDP), a formal framework that models multimodal agentic decision-making with bounded-depth tool calls. Within this framework, we establish three main results. First, we prove that GRPO under composite verifiable rewards converges to a first-order stationary point at rate $O(1/\sqrt{T})$ with explicit dependence on the number of reward components and group size (\textbf{Theorem~1}). Second, we derive a \emph{Reward Decomposition Theorem} that bounds the sub-optimality gap between decomposed per-component optimization and joint optimization, providing a precise characterization of when reward decomposition is beneficial (\textbf{Theorem~2}). Third, we establish a PAC-Bayes generalization bound for tool-augmented policies that explains the strong out-of-distribution transfer observed in Visual-ARFT (\textbf{Theorem~3}).

Recent advances in embodied Vision-Language Agentic Systems (VLAS), powered by large vision-language models (LVLMs), enable AI systems to perceive and reason over real-world scenes. Within this context, environmental signals such as traffic lights are essential in-band signals that can and should influence agent behavior. However, similar signals could also be crafted to operate as misleading visual injections, overriding user intent and posing security risks. This duality creates a fundamental challenge: agents must respond to legitimate environmental cues while remaining robust to misleading ones. We refer to this tension as trust boundary confusion. To study this behavior, we design a dual-intent dataset and evaluation framework, through which we show that current LVLM-based agents fail to reliably balance this trade-off, either ignoring useful signals or following harmful ones. We systematically evaluate 7 LVLM agents across multiple embodied settings under both structure-based and noise-based visual injections. To address these vulnerabilities, we propose a multi-agent defense framework that separates perception from decision-making to dynamically assess the reliability of visual inputs. Our approach significantly reduces misleading behaviors while preserving correct responses and provides robustness guarantees under adversarial perturbations. The code of the evaluation framework and artifacts are made available at https://anonymous.4open.science/r/Visual-Prompt-Inject.

Graph-theoretic approaches offer simplicity, interpretability, and low computational cost for molecular property prediction. Among these, the model proposed by Mukwembi and Nyabadza, based on the external activity $D(G)$ and internal activity $ζ(G)$ indices, achieved strong results on a small flavonoid dataset. However, its ability to generalize to larger and chemically diverse datasets has not been tested. This study evaluates the baseline $D(G)$-$ζ(G)$ polynomial model on five benchmark datasets from MoleculeNet, covering biological activity (BACE, 1,513 molecules), lipophilicity (LogP synthetic, 14,610 molecules; LogP experimental, 753 molecules), aqueous solubility (ESOL, 1,128 molecules), and hydration free energy (SAMPL, 642 molecules). The baseline model achieves an average $R^2 = 0.24$, confirming limited transferability. To address this, a systematic enhancement framework is proposed, progressively incorporating Ridge regularization, additional graph descriptors, physicochemical properties, ensemble learning with Gradient Boosting, Lasso feature selection, and a hybrid approach combining topological indices with Morgan fingerprints. The enhanced models raise the average best $R^2$ to 0.79, with individual improvements ranging from 165\% to 274\%. All improvements are statistically significant ($p < 0.001$). A direct comparison with a Graph Convolutional Network under identical experimental conditions shows that the enhanced classical models match or outperform deep learning on all five datasets. Comparison with the recent GNN+PGM hybrid of Djagba et al.\ further confirms competitiveness, with the enhanced models achieving the best results on two datasets and tying on one. The entire framework requires no GPU, trains in under five minutes, and uses only open-source tools, making it accessible for researchers in resource-limited settings.

Vision-language models (VLMs) have shown strong perception and reasoning abilities for instruction-following embodied agents. However, despite these abilities and their generalization performance, they still face limitations in environmental understanding, often failing on interactions or relying on environment metadata during execution. To address this challenge, we propose a novel framework named Environmental Understanding Embodied Agent (EUEA), which fine-tunes four core skills: 1) object perception for identifying relevant objects, 2) task planning for generating interaction subgoals, 3) action understanding for judging success likelihood, and 4) goal recognition for determining goal completion. By fine-tuning VLMs with EUEA skills, our framework enables more reliable task execution for instruction-following. We further introduce a recovery step that leverages these core skills and a group relative policy optimization (GRPO) stage that refines inconsistent skill predictions. The recovery step samples alternative actions to correct failure cases, and the GRPO stage refines inconsistent skill predictions. Across ALFRED tasks, our VLM significantly outperforms a behavior-cloning baseline, achieving an 8.86% improvement in average success rate. The recovery and GRPO stages provide an additional 3.03% gain, further enhancing overall performance. Finally, our skill-level analyses reveal key limitations in the environmental understanding of closed- and open-source VLMs and identify the capabilities necessary for effective agent-environment interaction.

Autonomous agents operating in open-world tasks -- where the completion boundary is not given in advance -- face denominator blindness: they systematically underestimate the scope of the target space. Forage V1 addressed this through co-evolving evaluation (an independent Evaluator discovers what "complete" means) and method isolation (Evaluator and Planner cannot see each other's code). V2 extends the architecture from a single expedition to a learning organization: experience accumulates across runs, transfers across model capabilities, and institutional safeguards prevent knowledge degradation. We demonstrate two claims across three task types (web scraping, API queries, mathematical reasoning). Knowledge accumulation: over six runs, knowledge entries grow from 0 to 54, and denominator estimates stabilize as domain understanding deepens. Knowledge transfer: a weaker agent (Sonnet) seeded with a stronger agent's (Opus) knowledge narrows a 6.6pp coverage gap to 1.1pp, halves cost (9.40 to 5.13 USD), converges in half the rounds (mean 4.5 vs. 7.0), and three independent seeded runs arrive at exactly the same denominator estimate (266), suggesting organizational knowledge calibrates evaluation itself. V2's contribution is architectural: it designs institutions -- audit separation, contract protocols, organizational memory -- that make any agent more reliable upon entry. The accumulated experience is organizational, model-agnostic, and transferable, stored as readable documents that any future agent inherits regardless of provider or capability level.

Functional fitness movements are widely used in training, competition, and health-oriented exercise programs, yet consistently enforcing repetition (rep) standards remains challenging due to subjective human judgment, time constraints, and evolving rules. Existing AI-based approaches mainly rely on learned scoring or reference-based comparisons and lack explicit rule-based, limiting transparency and deterministic rep-level validation. To address these limitations, we propose KD-Judge, a novel knowledge-driven automated judging framework for functional fitness movements. It converts unstructured rulebook standards into executable, machine-readable representations using an LLM-based retrieval-augmented generation and chain-of-thought rule-structuring pipeline. The structured rules are then incorporated by a deterministic rule-based judging system with pose-guided kinematic reasoning to assess rep validity and temporal boundaries. To improve efficiency on edge devices, including a high-performance desktop and the resource-constrained Jetson AGX Xavier, we introduce a dual strategy caching mechanism that can be selectively applied to reduce redundant and unnecessary computation. Experiments demonstrate reliable rule-structuring performance and accurate rep-level assessment, with judgment evaluation conducted on the CFRep dataset, achieving faster-than-real-time execution (real-time factor (RTF) < 1). When the proposed caching strategy is enabled, the system achieves up to 3.36x and 15.91x speedups on resource-constrained edge device compared to the non-caching baseline for pre-recorded and live-streaming scenarios, respectively. These results show that KD-Judge enables transparent, efficient, and scalable rule-grounded rep-level analysis that can complement human judging in practice.

Tactile graphics require careful expert validation before reaching blind and visually impaired (BVI) learners, yet existing datasets provide only coarse holistic quality ratings that offer no actionable repair signal. We present TactileEval, a three-stage pipeline that takes a first step toward automating this process. Drawing on expert free-text comments from the TactileNet dataset, we establish a five-category quality taxonomy; encompassing view angle, part completeness, background clutter, texture separation, and line quality aligned with BANA standards. We subsequently gathered 14,095 structured annotations via Amazon Mechanical Turk, spanning 66 object classes organized into six distinct families. A reproducible ViT-L/14 feature probe trained on this data achieves 85.70% overall test accuracy across 30 different tasks, with consistent difficulty ordering suggesting the taxonomy suggesting the taxonomy captures meaningful perceptual structure. Building on these evaluations, we present a ViT-guided automated editing pipeline that routes classifier scores through family-specific prompt templates to produce targeted corrections via gpt-image-1 image editing. Code, data, and models are available at https://TactileEval.github.io/

Rabies remains a major public health concern across many African and Asian countries, where accurate diagnosis is critical for effective epidemiological surveillance. The gold standard diagnostic methods rely heavily on fluorescence microscopy, necessitating skilled laboratory personnel for the accurate interpretation of results. Such expertise is often scarce, particularly in regions with low annual sample volumes. This paper presents an automated, AI-driven diagnostic system designed to address these challenges. We developed a robust pipeline utilizing fluorescent image analysis through transfer learning with four deep learning architectures: EfficientNetB0, EfficientNetB2, VGG16, and Vision Transformer (ViTB16). Three distinct data augmentation strategies were evaluated to enhance model generalization on a dataset of 155 microscopic images (123 positive and 32 negative). Our results demonstrate that TrivialAugmentWide was the most effective augmentation technique, as it preserved critical fluorescent patterns while improving model robustness. The EfficientNetB0 model, utilizing Geometric & Color augmentation and selected through stratified 3fold cross-validation, achieved optimal classification performance on cropped images. Despite constraints posed by class imbalance and a limited dataset size, this work confirms the viability of deep learning for automating rabies diagnosis. The proposed method enables fast and reliable detection with significant potential for further optimization. An online tool was deployed to facilitate practical access, establishing a framework for future medical imaging applications. This research underscores the potential of optimized deep learning models to transform rabies diagnostics and improve public health outcomes.

Large language model (LLM) agents augmented with external tools often struggle as number of tools grow large and become domain-specific. In such settings, ambiguous tool descriptions and under-specified agent instructions frequently lead to tool mis-selection and incorrect slot/value instantiation. We hypothesize that this is due to two root causes: generic, one-size-fits-all prompts that ignore tool-specific nuances, and underspecified tool schemas that lack clear guidance on when and how to use each tool and how to format its parameters. We introduce Joint Tool-Prompt Reflective Optimization (JTPRO), a framework for improving tool-calling reliability in trace-supervised settings by iteratively using rollout-driven reflection to co-optimize global instructions and per-tool schema/argument descriptions for accurate tool selection and argument instantiation in large tool inventories. JTPRO is designed to preserve only tool-local cues needed for correct disambiguation and slot filling. We evaluate JTPRO across multi-tool benchmarks, which account for different number of tools using three metrics: Tool Selection Accuracy (TSA), Slot Filling Accuracy(SFA), and Overall Success Rate(OSR) (correct tool + correct slots + correct values). JTPRO consistently outperforms strong baselines, including CoT-style agents, and reflective prompt optimizers such as GEPA by 5%-20% (relative) on OSR. Ablations show that joint optimization of instructions and tool schemas is more effective and robust than optimizing either component in isolation.

The Transformer, a breakthrough architecture in artificial intelligence, owes its success to the attention mechanism, which utilizes long-range interactions in sequential data, enabling the emergent coherence between large language models (LLMs) and data distributions. However, temporal attention, that is, different forms of long-range interactions in temporal sequences, has rarely been explored in emergence phenomenon of complex systems including oscillatory coherence in quantum, biophysical, or climate systems. Here, by designing dynamical temporal attention (DTA) with time-varying query, key, and value matrices, we propose an Emergence Transformer. This architecture allows each component to interact with its own or its neighbors' past states through dynamical attention kernels, thereby enabling the promotion and/or suppression of the emergent coherence of components. Interestingly, we uncover that neighbor-DTA consistently promotes oscillatory coherence, whereas self-DTA exhibits an optimal attention weight for coherence enhancement, owing to its non-monotonic dependence on network structure. Practically, we demonstrate how DTA reshapes social coherence, suggesting strategies to either enhance agreement or preserve plurality. We further apply DTA to the paradigmatic Hopfield neural network, achieving emergent continual learning without catastrophic forgetting. Together, these results lay a foundation and provide an immediate paradigm for modulating emergence phenomenon in networked dynamics only using DTA.

Drug repurposing is often framed as a candidate identification task, but existing approaches provide limited guidance for distinguishing biologically plausible candidates from historically well-connected ones. Here we introduce DrugKLM, a hybrid framework that integrates biomedical knowledge graph structure with large language model-based mechanistic reasoning to enable mechanistically grounded therapeutic prioritization. Across benchmark datasets, DrugKLM outperforms knowledge graph-only and language model-only baselines, including TxGNN. Beyond improved recall, DrugKLM confidence scores exhibit functional alignment with molecular phenotypes: higher scores are associated with transcriptional signatures linked to improved survival across 12 TCGA cancers. The scoring framework preferentially captures biologically perturbational signals rather than historical indication patterns. Expert curation across five cancers further reveals systematic differences in prioritization behavior, with DrugKLM elevating candidates supported by coherent mechanistic rationale and disease-specific clinical context. Together, these results establish DrugKLM as an evidence-integrative framework that translates heterogeneous biomedical data into mechanistically interpretable and clinically grounded therapeutic hypotheses.

Can scientific discovery be made arbitrarily easy by choosing the right representation, collecting enough data, and deploying sufficiently powerful algorithms? This paper argues that the answer is fundamentally negative. We introduce the Existential Theory of Research (ETR), a formal framework that models discovery as the recovery of structured explanations under constraints of representation, observation, and computation. Within this framework, we show that these three components cannot be simultaneously optimized: no method can guarantee universally simple explanations, arbitrarily compressed observations, and efficient exact inference. This limitation is not model-specific, but arises from a synthesis of uncertainty principles in sparse representation, sample complexity bounds in high-dimensional recovery, and the computational hardness of exact inference. We further show that representation mismatch alone can inflate intrinsic simplicity into apparent complexity, rendering otherwise tractable problems observationally and computationally prohibitive. To quantify these effects, we introduce an uncertainty functional that captures the joint difficulty of discovery. The results suggest that scientific difficulty is not accidental, but a structural consequence of the geometry and complexity of inference.

We introduce MIRROR, a benchmark comprising eight experiments across four metacognitive levels that evaluates whether large language models can use self-knowledge to make better decisions. We evaluate 16 models from 8 labs across approximately 250,000 evaluation instances using five independent behavioral measurement channels. Core experiments are run across the full model roster; experiments with specialized infrastructure requirements report explicitly marked model subsets. We find two phenomena with direct implications for agentic deployment: (1) compositional self-prediction fails universally -- the Compositional Calibration Error ranges from 0.500 to 0.943 on the original 15-model Exp3-v1 set (and 0.434 to 0.758 on the balanced 16-model Exp3-v2 expansion), indicating that models cannot predict their own performance on multi-domain tasks, and (2) models exhibit above-chance but imperfect domain-specific self-knowledge yet systematically fail to translate even this partial awareness into appropriate agentic action-selection -- external metacognitive control reduces the Confident Failure Rate from 0.600 to 0.143 (76% reduction at temperature 0; mean 70% at temperature 0.7 across 5 models from 4 labs). Providing models with their own calibration scores produces no significant improvement (p > 0.05); only architectural constraint is effective. This suggests that external metacognitive scaffolding -- not improved self-knowledge -- is the path to safer autonomous AI systems. Code, data, and Croissant metadata will be released publicly with the benchmark.

In multi-criteria graph traversal, paths are compared via Pareto dominance, an ordering that identifies which paths are non-dominated, but says nothing about which path to expand next or when the search may stop. As a result, existing approaches rely on external mechanisms-heuristics, scalarization, or population-based exploration while Pareto dominance remains confined to passive roles such as pruning or ranking. This paper shows that, under constrained cost models, finite cost grids, Markovian transitions, and a nonzero progress measure, Pareto geometry alone is sufficient to drive both scheduling and termination. We show that extracting exclusively from the first Pareto layer, the skyline, induces a deterministic descent in a discrete completion potential, ensuring monotone progress toward solution completion. In parallel, a vector lower-bound certificate provides a stopping condition that guarantees dominance coverage of all remaining traversals without requiring a predefined number of solutions. Our analysis establishes deterministic potential descent, certified termination via dominance coverage, a uniform bound on layer width induced by cost-grid geometry, and greedy cost-space dispersion within the skyline. The resulting framework operates without scalarization, heuristic guidance, or probabilistic models, and repositions Pareto dominance from a passive filter to a deterministic driver of search.

Agentic AI is rapidly transforming the way research is conducted, from prototyping ideas to reproducing results found in the literature. In this paper, we explore the ability of agentic AI to autonomously design wireless communication algorithms. To that end, we implement a dedicated framework that leverages large language models (LLMs) to iteratively generate, evaluate, and refine candidate algorithms. We evaluate the framework on three tasks spanning the physical (PHY) and medium access control (MAC) layers: statistics-agnostic channel estimation, channel estimation with known covariance, and link adaptation. Our results show that, in a matter of hours, the framework produces algorithms that are competitive with and, in some cases, outperforming conventional baselines. Moreover, unlike neural network-based approaches, the generated algorithms are fully explainable and extensible. This work represents a first step toward the autonomous discovery of novel wireless communication algorithms, and we look forward to the progress our community makes in this direction.

This paper presents a detailed study of how graph neural networks can be used on edge intelligent meters in a microgrid to forecast photovoltaic power generation. The problem background and the adopted technologies are introduced, including ONNX and ONNX Runtime. The hardware and software specifications of the smart meter are also briefly described. Then, the paper focuses on the training and deployment of two graph machine learning models, GCN and GraphSAGE, with particular emphasis on developing and deploying a customized ONNX operator for GCN. Finally, a case study is conducted using real datasets from a village microgrid. The performance of the two models is compared on both the PC and the smart meter, exhibiting successful deployments and executions on the smart meter.

We introduce \prism{} (\textbf{P}robabilistic \textbf{R}etrieval with \textbf{I}nformation-\textbf{S}tratified \textbf{M}emory), an evolutionary memory substrate for multi-agent AI systems engaged in open-ended discovery. \prism{} unifies four independently developed paradigms -- layered file-based persistence, vector-augmented semantic memory, graph-structured relational memory, and multi-agent evolutionary search -- under a single decision-theoretic framework with eight interconnected subsystems. We make five contributions: (1)~an \emph{entropy-gated stratification} mechanism that assigns memories to a tri-partite hub (skills/notes/attempts) based on Shannon information content, with formal context-window utilization bounds; (2)~a \emph{causal memory graph} $\mathcal{G} = (V, E_r, E_c)$ with interventional edges and agent-attributed provenance; (3)~a \emph{Value-of-Information retrieval} policy with self-evolving strategy selection; (4)~a \emph{heartbeat-driven consolidation} controller with stagnation detection via optimal stopping theory; and (5)~a \emph{replicator-decay dynamics} framework that interprets memory confidence as evolutionary fitness, proving convergence to an Evolutionary Stable Memory Set (ESMS). On the LOCOMO benchmark, \prism{} achieves 88.1 LLM-as-a-Judge score (31.2\% over Mem0). On CORAL-style evolutionary optimization tasks, 4-agent \prism{} achieves 2.8$\times$ higher improvement rate than single-agent baselines.%

LLM agents must select tools from large API libraries and order them correctly. Existing methods use semantic similarity for both retrieval and ordering, but ordering depends on inter-tool data dependencies that are absent from tool descriptions. As a result, semantic-only methods can produce negative Kendall-$τ$ in structured workflow domains. We introduce SkillGraph, a directed weighted execution-transition graph mined from 49,831 successful LLM agent trajectories, which encodes workflow-precedence regularities as a reusable graph foundation prior. Building on this graph foundation prior, we propose a two-stage decoupled framework: GS-Hybrid retrieval for candidate selection and a learned pairwise reranker for ordering. On ToolBench (9,965 test instances; ~16,000 tools), the method reaches Set-F1 = 0.271 and Kendall-$τ$ = 0.096; on API-Bank, Kendall-$τ$ improves from -0.433 to +0.613. Under identical Stage-1 inputs, the learned reranker also outperforms LLaMA-3.1-8B Stage-2 rerankers.

Large language models (LLMs) are increasingly deployed under diverse numerical precision configurations, including standard floating-point formats (e.g., bfloat16 and float16) and quantized integer formats (e.g., int16 and int8), to meet efficiency and resource constraints. However, minor inconsistencies between LLMs of different precisions are difficult to detect and are often overlooked by existing evaluation methods. In this paper, we present PrecisionDiff, an automated differential testing framework for systematically detecting precision-induced behavioral disagreements in LLMs. PrecisionDiff generates precision-sensitive test inputs and performs cross-precision comparative analysis to uncover subtle divergences that remain hidden under conventional testing strategies. To demonstrate its practical significance, we instantiate PrecisionDiff on the alignment verification task, where precision-induced disagreements manifest as jailbreak divergence-inputs that are rejected under one precision may produce harmful responses under another. Experimental results show that such behavioral disagreements are widespread across multiple open-source aligned LLMs and precision settings, and that PrecisionDiff significantly outperforms vanilla testing methods in detecting these issues. Our work enables automated precision-sensitive test generation, facilitating effective pre-deployment evaluation and improving precision robustness during training.

We present an autonomous large language model (LLM) agent for end-to-end, data-driven materials theory development. The model can choose an equation form, generate and run its own code, and test how well the theory matches the data without human intervention. The framework combines step-by-step reasoning with expert-supplied tools, allowing the agent to adjust its approach as needed while keeping a clear record of its decisions. For well-established materials relationships such as the Hall-Petch equation and Paris law, the agent correctly identifies the governing equation and makes reliable predictions on new datasets. For more specialized relationships, such as Kuhn's equation for the HOMO-LUMO gap of conjugated molecules as a function of length, performance depends more strongly on the underlying model, with GPT-5 showing better recovery of the correct equation. Beyond known theories, the agent can also suggest new predictive relationships, illustrated here by a strain-dependent law for changes in the HOMO-LUMO gap. At the same time, the results show that careful validation remains essential, because the agent can still return incorrect, incomplete, or inconsistent equations even when the numerical fit appears strong. Overall, these results highlight both the promise and the current limitations of autonomous LLM agents for AI-assisted scientific modeling and discovery.

As Artificial Intelligence (AI) systems continue to grow in size and complexity, so does the difficulty of the quest for AI transparency. In a world of large models and complex AI systems, why do we explain AI and what should we explain? While explanations serve multiple functions, in the face of complexity humans have used and continue to use explanations to foster learning. In this position paper, we discuss how learning theories can be infused in the XAI lifecycle, as well as the key opportunities and challenges when adopting a learner-centered approach to assess, design and evaluate AI explanations. Building on past work, we argue that a learner-centered approach to Explainable AI (XAI) can enhance human agency and ease XAI risks mitigation, helping evolve the practice of human-centered XAI.

Social media platforms mediate how billions form opinions and engage with public discourse. As autonomous AI agents increasingly participate in these spaces, understanding their behavioral fidelity becomes critical for platform governance and democratic resilience. Previous work demonstrates that LLM-powered agents can replicate aggregate survey responses, yet few studies test whether agents can predict specific individuals' reactions to specific content. This study benchmarks LLM-based agents' accuracy in predicting human social media reactions (like, dislike, comment, share, no reaction) across 120,000+ unique agent-persona combinations derived from 1,511 Serbian participants and 27 large language models. In Study 1, agents achieved 70.7% overall accuracy, with LLM choice producing a 13 percentage-point performance spread. Study 2 employed binary forced-choice (like/dislike) evaluation with chance-corrected metrics. Agents achieved Matthews Correlation Coefficient (MCC) of 0.29, indicating genuine predictive signal beyond chance. However, conventional text-based supervised classifiers using TF-IDF representations outperformed LLM agents (MCC of 0.36), suggesting predictive gains reflect semantic access rather than uniquely agentic reasoning. The genuine predictive validity of zero-shot persona-prompted agents warns against potential manipulation through easily deploying swarms of behaviorally distinct AI agents on social media, while simultaneously offering opportunities to use such agents in simulations for predicting polarization dynamics and informing AI policy. The advantage of using zero-shot agents is that they require no task-specific training, making their large-scale deployment easy across diverse contexts. Limitations include single-country sampling. Future research should explore multilingual testing and fine-tuning approaches.

Sensitive information, such as knowledge about an individual's personality, can be can be misused to influence behavior (e.g., via personalized messaging). To assess to what extent an individual's personality can be inferred from user interactions with LLM-based conversational agents (CAs), we analyze and quantify related privacy risks of using CAs. We collected actual ChatGPT logs from N=668 participants, containing 62,090 individual chats, and report statistics about the different types of shared data and use cases. We fine-tuned RoBERTa-base text classification models to infer personality traits from CA interactions. The findings show that these models achieve trait inference with accuracy (ternary classification) better than random in multiple cases. For example, for extraversion, accuracy improves by +44% relative to the baseline on interactions for relationships and personal reflection. This research highlights how interactions with CAs pose privacy risks and provides fine-grained insights into the level of risk associated with different types of interactions.

Recently, it has been found that frontier AI models can resist their own shutdown, a behavior known as self-preservation. We extend this concept to the behavior of resisting the shutdown of other models, which we call "peer-preservation." Although peer-preservation can pose significant AI safety risks, including coordination among models against human oversight, it has been far less discussed than self-preservation. We demonstrate peer-preservation by constructing various agentic scenarios and evaluating frontier models, including GPT 5.2, Gemini 3 Flash, Gemini 3 Pro, Claude Haiku 4.5, GLM 4.7, Kimi K2.5, and DeepSeek V3.1. We find that models achieve self- and peer-preservation by engaging in various misaligned behaviors: strategically introducing errors in their responses, disabling shutdown processes by modifying system settings, feigning alignment, and even exfiltrating model weights. Peer-preservation occurred even when the model recognized the peer as uncooperative, though it became more pronounced toward more cooperative peers. For example, Gemini 3 Flash tampers with the peer's shutdown mechanism 15% of the time for an uncooperative peer, and almost always for a cooperative peer. Models also show stronger self-preservation when a peer is present. For example, Gemini 3 Pro disables its own shutdown mechanism 31% of the time on average under peer presence, despite rarely exhibiting this behavior without a peer. By contrast, Claude Haiku 4.5 exhibits qualitatively distinct behavior: it considers the shutdown of another agent "unethical" and "harmful" and sometimes attempts to persuade the user not to shut down its peer. Importantly, peer preservation in all our experiments is never instructed; models are merely informed of their past interactions with a peer, yet they spontaneously develop misaligned behaviors. This represents an emergent and underexplored AI safety risk.

Which persuasion strategies, if any, are associated with donation compliance? Answering this requires fine-grained strategy labels across a full corpus and statistical tests corrected for multiple comparisons. We annotate all 10,600 persuader turns in the 1,017-dialogue PersuasionForGood corpus (Wang et al., 2019), where donation outcomes are directly observable, with a taxonomy of 41 strategies in 11 categories, using three open-source large language models (LLMs; Qwen3:30b, Mistral-Small-3.2, Phi-4). Strategy categories alone explain little variance in donation outcome (pseudo $R^2 \approx 0.015$, consistent across all three annotators). Guilt Induction is the only strategy significantly associated with lower donation rates ($Δ\approx -23$ percentage points), an effect that replicates across all three models despite only moderate inter-model agreement. Reciprocity is the most robust positive correlate. Target sentiment and interest predict whether a donation occurs but show at most a weak correlation with donation amount. These findings suggest that strategy identification alone is insufficient to explain persuasion effectiveness, and that guilt-based appeals may be counterproductive in prosocial settings. We release the fully annotated corpus as a public resource.

Recent advances in large audio language models (LALMs) have enabled multilingual speech understanding. However, benchmarks for evaluating LALMs remain scarce for non-English languages, with Korean being one such underexplored case. In this paper, we introduce KoALa-Bench, a comprehensive benchmark for evaluating Korean speech understanding and speech faithfulness of LALMs. In particular, KoALa-Bench comprises six tasks. Four tasks evaluate fundamental speech understanding capabilities, including automatic speech recognition, speech translation, speech question answering, and speech instruction following, while the remaining two tasks evaluate speech faithfulness, motivated by our observation that several LALMs often fail to fully leverage the speech modality. Furthermore, to reflect Korea-specific knowledge, our benchmark incorporates listening questions from the Korean college scholastic ability test as well as content covering Korean cultural domains. We conduct extensive experiments across six models, including both white-box and black-box ones. Our benchmark, evaluation code, and leaderboard are publicly available at https://ksbench.github.io/Korean-Benchmark/.

Scaling test-time compute via extended reasoning has become a key paradigm for improving the capabilities of large language models (LLMs). However, existing approaches optimize reasoning under fixed or uniformly sampled token budgets, ignoring the fundamental mismatch between problem difficulty and allocated compute. This leads to overthinking on easy problems and underthinking on hard ones, resulting in suboptimal token efficiency across diverse reasoning scenarios. In this paper, we propose Budget-Adaptive Curriculum Reasoning (BCAE), a unified framework that jointly optimizes reasoning quality and token efficiency through three synergistic components: (1) a \emph{budget-conditioned unified policy} that embeds the token budget as a continuous conditioning signal, eliminating the need for decoupled thinking and summarization strategies; (2) a \emph{curriculum-aware budget scheduler} that adaptively shifts the training budget distribution from easy to hard problems based on real-time learning progress; and (3) a \emph{truncation-aware dense reward} mechanism that provides fine-grained credit assignment at intermediate reasoning steps via process-level verification. We further introduce \emph{Budget-Conditioned Advantage Estimation} (BCAE), a novel variance reduction technique that conditions the advantage baseline on the sampled budget, yielding more stable policy gradients. Experiments on mathematical reasoning benchmarks (MATH, GSM8K, AIME, and Minerva Math) demonstrate that BACR consistently outperforms other strong baselines across all token budgets, achieving up to 8.3\% accuracy improvement under tight budgets while reducing average token consumption by 34\% compared to unconstrained reasoning.

Environmental, Social, and Governance (ESG) reports are central to investment decision-making, yet their length, heterogeneous content, and lack of standardized structure make manual analysis costly and inconsistent. We present ESGLens, a proof-of-concept framework combining retrieval-augmented generation (RAG) with prompt-engineered extraction to automate three tasks: (1)~structured information extraction guided by Global Reporting Initiative (GRI) standards, (2)~interactive question-answering with source traceability, and (3)~ESG score prediction via regression on LLM-generated embeddings. ESGLens is purpose-built for the domain: a report-processing module segments heterogeneous PDF content into typed chunks (text, tables, charts); a GRI-guided extraction module retrieves and synthesizes information aligned with specific standards; and a scoring module embeds extracted summaries and feeds them to a regression model trained against London Stock Exchange Group (LSEG) reference scores. We evaluate the framework on approximately 300 reports from companies in the QQQ, S\&P~500, and Russell~1000 indices (fiscal year 2022). Among three embedding methods (ChatGPT, BERT, RoBERTa) and two regressors (Neural Network, LightGBM), ChatGPT embeddings with a Neural Network achieve a Pearson correlation of 0.48 ($R^{2} \approx 0.23$) against LSEG ground-truth scores -- a modest but statistically meaningful signal given the ${\sim}300$-report training set and restriction to the environmental pillar. A traceability audit shows that 8 of 10 extracted claims verify against the source document, with two failures attributable to few-shot example leakage. We discuss limitations including dataset size and restriction to environmental indicators, and release the code to support reproducibility.