MINDGAMES: A Live Arena for Evaluating Social and Strategic Reasoning in Multi-Agent LLMs

Large language models (LLMs) are increasingly deployed as interactive agents, yet their capacity for social and strategic reasoning over extended interaction remains poorly understood. Existing evaluations rely on static vignettes or single-game benchmarks that cannot capture the sustained, multi-faceted reasoning that real-world multi-agent settings demand. We introduce Mindgames, a multi-game arena and evaluation platform for LLM agents that operationalizes complementary reasoning demands relevant to ``theory of mind'': belief attribution under hidden information, opponent modeling through repeated strategic interaction, cooperative inference under knowledge asymmetries, and sustained deception in social deduction. Built on TextArena, Mindgames provides a unified interaction interface, TrueSkill-based rating, and full trajectory logging across four game environments. We instantiate Mindgames through a 2025 competition cycle hosted at a major AI conference, which assessed 944 submitted agents from 76 teams across four games: Colonel Blotto, Iterated Prisoner's Dilemma, Codenames, and Secret Mafia. Our analysis surfaces both agent-level and evaluation-level limitations: brittle rule adherence remains a major bottleneck, top-performing systems repeatedly rely on explicit structural scaffolding, and leaderboard validity differs sharply across environments. In particular, failure-heavy environments can reward robustness to opponent errors as much as strategic ability, with Secret Mafia exhibiting a pronounced error-survival confound in this cycle. We release a dataset of 29,571 multi-agent games with turn-level observations, actions, and rewards, together with MG-Ref, a deterministic offline tournament protocol that scores new agents against a frozen reference pool of top-ranked, low-error Stage~II submissions under the same error-attribution lens used in this analysis.

WideDepth: Millimeter-Accurate Benchmark for Fisheye Depth Estimation

Fisheye cameras are increasingly adopted in robotics for near-field manipulation, navigation, and immersive perception, yet indoor depth benchmarks with accurate ground truth are still missing. To address this, we introduce WideDepth - the first indoor dataset for fisheye depth estimation, featuring 101 scenes containing 5K high-resolution stereo pairs labeled with millimeter-level ground truth depth and disparity. Our dataset also includes paired pinhole and fisheye samples across varying fields of view and baselines in both horizontal and vertical stereo setups. We further propose a method to adapt pinhole-trained stereo models to fisheye images and introduce a novel stereo fisheye image generation pipeline based on high-resolution LiDAR scans. Leveraging these methods, we thoroughly evaluate state-of-the-art monocular depth, stereo matching, and depth completion models on our benchmark. Additionally, we provide 18K LiDAR-derived sparse depth training samples, achieving up to a 62% performance boost on fisheye data when fine-tuning pinhole-based stereo models. In summary, the high precision and versatility of our benchmark set a strong foundation for advancing research in fisheye depth estimation and robotics perception. Project page: https://ilyaind.github.io/WideDepth

Pre-trained LLMs Meet Sequential Recommenders: Efficient User-Centric Knowledge Distillation

Sequential recommender systems have achieved significant success in modeling temporal user behavior but remain limited in capturing rich user semantics beyond interaction patterns. Large Language Models (LLMs) present opportunities to enhance user understanding with their reasoning capabilities, yet existing integration approaches create prohibitive inference costs in real time. To address these limitations, we present a novel knowledge distillation method that utilizes textual user profile generated by pre-trained LLMs into sequential recommenders without requiring LLM inference at serving time. The resulting approach maintains the inference efficiency of traditional sequential models while requiring neither architectural modifications nor LLM fine-tuning.

Beyond Isolated Clients: Integrating Graph-Based Embeddings into Event Sequence Models

Large-scale digital platforms generate billions of timestamped user-item interactions (events) that are crucial for predicting user attributes in, e.g., fraud prevention and recommendations. While self-supervised learning (SSL) effectively models the temporal order of events, it typically overlooks the global structure of the user-item interaction graph. To bridge this gap, we propose three model-agnostic strategies for integrating this structural information into contrastive SSL: enriching event embeddings, aligning client representations with graph embeddings, and adding a structural pretext task. Experiments on four financial and e-commerce datasets demonstrate that our approach consistently improves the accuracy (up to a 2.3% AUC) and reveals that graph density is a key factor in selecting the optimal integration strategy.

ReDAct: Uncertainty-Aware Deferral for LLM Agents

Recently, LLM-based agents have become increasingly popular across many applications, including complex sequential decision-making problems. However, they inherit the tendency of LLMs to hallucinate, leading to incorrect decisions. In sequential settings, even a single mistake can irreversibly degrade the trajectory, making hallucinations an even bigger problem. Although larger LLMs hallucinate less, they incur a significantly higher per-token cost. In this paper, we address this tradeoff by proposing ReDAct (Reason-Defer-Act). In ReDAct, an agent is equipped with two LLMs: a small, cheap model used by default, and a large, more reliable but expensive model. When the predictive uncertainty of the small model exceeds a calibrated threshold, the decision is deferred to the large model. We evaluate our approach in text-based embodied environments such as ALFWorld and MiniGrid and show that deferring only about 15% of decisions to the large model can match the quality of using it exclusively, while significantly reducing inference costs.

Benchmarking IoT Time-Series AD with Event-Level Augmentations

Anomaly detection (AD) for safety-critical IoT time series should be judged at the event level: reliability and earliness under realistic perturbations. Yet many studies still emphasize point-level results on curated base datasets, limiting value for model selection in practice. We introduce an evaluation protocol with unified event-level augmentations that simulate real-world issues: calibrated sensor dropout, linear and log drift, additive noise, and window shifts. We also perform sensor-level probing via mask-as-missing zeroing with per-channel influence estimation to support root-cause analysis. We evaluate 14 representative models on five public anomaly datasets (SWaT, WADI, SMD, SKAB, TEP) and two industrial datasets (steam turbine, nuclear turbogenerator) using unified splits and event aggregation. There is no universal winner: graph-structured models transfer best under dropout and long events (e.g., on SWaT under additive noise F1 drops 0.804->0.677 for a graph autoencoder, 0.759->0.680 for a graph-attention variant, and 0.762->0.756 for a hybrid graph attention model); density/flow models work well on clean stationary plants but can be fragile to monotone drift; spectral CNNs lead when periodicity is strong; reconstruction autoencoders become competitive after basic sensor vetting; predictive/hybrid dynamics help when faults break temporal dependencies but remain window-sensitive. The protocol also informs design choices: on SWaT under log drift, replacing normalizing flows with Gaussian density reduces high-stress F1 from ~0.75 to ~0.57, and fixing a learned DAG gives a small clean-set gain (~0.5-1.0 points) but increases drift sensitivity by ~8x.

From Images to Decisions: Assistive Computer Vision for Non-Metallic Content Estimation in Scrap Metal

Scrap quality directly affects energy use, emissions, and safety in steelmaking. Today, the share of non-metallic inclusions (contamination) is judged visually by inspectors - an approach that is subjective and hazardous due to dust and moving machinery. We present an assistive computer vision pipeline that estimates contamination (per percent) from images captured during railcar unloading and also classifies scrap type. The method formulates contamination assessment as a regression task at the railcar level and leverages sequential data through multi-instance learning (MIL) and multi-task learning (MTL). Best results include MAE 0.27 and R2 0.83 by MIL; and an MTL setup reaches MAE 0.36 with F1 0.79 for scrap class. Also we present the system in near real time within the acceptance workflow: magnet/railcar detection segments temporal layers, a versioned inference service produces railcar-level estimates with confidence scores, and results are reviewed by operators with structured overrides; corrections and uncertain cases feed an active-learning loop for continual improvement. The pipeline reduces subjective variability, improves human safety, and enables integration into acceptance and melt-planning workflows.

SuS: Strategy-aware Surprise for Intrinsic Exploration

We propose Strategy-aware Surprise (SuS), a novel intrinsic motivation framework that uses pre-post prediction mismatch as a novelty signal for exploration in reinforcement learning. Unlike traditional curiosity-driven methods that rely solely on state prediction error, SuS introduces two complementary components: Strategy Stability (SS) and Strategy Surprise (SuS). SS measures consistency in behavioral strategy across temporal steps, while SuS captures unexpected outcomes relative to the agent's current strategy representation. Our combined reward formulation leverages both signals through learned weighting coefficients. We evaluate SuS on mathematical reasoning tasks using large language models, demonstrating significant improvements in both accuracy and solution diversity. Ablation studies confirm that removing either component results in at least 10% performance degradation, validating the synergistic nature of our approach. SuS achieves 17.4% improvement in Pass@1 and 26.4% improvement in Pass@5 compared to baseline methods, while maintaining higher strategy diversity throughout training.

Multi-Track Multimodal Learning on iMiGUE: Micro-Gesture and Emotion Recognition

Micro-gesture recognition and behavior-based emotion prediction are both highly challenging tasks that require modeling subtle, fine-grained human behaviors, primarily leveraging video and skeletal pose data. In this work, we present two multimodal frameworks designed to tackle both problems on the iMiGUE dataset. For micro-gesture classification, we explore the complementary strengths of RGB and 3D pose-based representations to capture nuanced spatio-temporal patterns. To comprehensively represent gestures, video, and skeletal embeddings are extracted using MViTv2-S and 2s-AGCN, respectively. Then, they are integrated through a Cross-Modal Token Fusion module to combine spatial and pose information. For emotion recognition, our framework extends to behavior-based emotion prediction, a binary classification task identifying emotional states based on visual cues. We leverage facial and contextual embeddings extracted using SwinFace and MViTv2-S models and fuse them through an InterFusion module designed to capture emotional expressions and body gestures. Experiments conducted on the iMiGUE dataset, within the scope of the MiGA 2025 Challenge, demonstrate the robust performance and accuracy of our method in the behavior-based emotion prediction task, where our approach secured 2nd place.

AraToken: Optimizing Arabic Tokenization with Normalization Pipeline and Language Extension for Qwen3

Tokenization is a critical preprocessing step for large language models (LLMs), directly impacting training efficiency and downstream performance. General-purpose tokenizers trained predominantly on English and Latin-script languages exhibit suboptimal performance on morphologically rich languages such as Arabic, resulting in inflated token sequences and reduced compression efficiency. In this work, we present AraToken, an Arabic-optimized tokenizer built on SentencePiece Unigram algorithm with a comprehensive normalization pipeline addressing Arabic-specific orthographic variations including Alif variants, diacritics, and Arabic-Indic numerals. We systematically compare BPE, WordPiece, and SentencePiece algorithms across multiple configurations, demonstrating that SentencePiece with normalization achieves 18% lower fertility (1.199 vs 1.35 tokens/word) compared to unnormalized baselines. Furthermore, we introduce the Language Extension Pipeline (LEP), a method for integrating the optimized tokenizer into Qwen3-0.6B through vocabulary extension with mean subtoken initialization and selective transformer layer unfreezing. Our experiments show that LEP reduces evaluation loss from 8.28 to 2.43 within 800 training steps on 100K Arabic samples. We release our tokenizer, training scripts, and model checkpoints to facilitate Arabic NLP research.