When Bob's mom leaves for a work trip, he thinks it's going to be a quiet day at home. However, Bob discovers something terrifying: Krad, the King of Crogzeins, is coming to conquer Earth. His soldiers are already here, hunting for the seven scattered pieces of the legendary Sword of Ghix, the only weapon that can kill Krad. If Bob assembles the sword first, he might save everything. This is an unfinished game with WASD/arrow keys and Z/X/C/Space controls.

Syncpen is a markdown writing workspace that lets AI agents (Claude Code, Cursor, claude.ai, Cowork) read, draft, and suggest edits directly in your documents. Changes are signed and tracked; you approve before they take effect. Features real-time collaboration, native markdown, one-click CMS publishing, and built-in research tools.

After US lifted export controls, Anthropic will restore Fable 5 on July 1, but with usage limits and per-credit billing after July 7. The company details the jailbreak incident, improved safety classifiers, and continued government collaboration.

This paper presents a method for synthesizing control inputs for discrete-time linear systems subject to Time Window Temporal Logic (TWTL) specifications. By encoding robust satisfaction as mixed-integer linear constraints, the authors formulate synthesis as a Mixed-Integer Linear Program (MILP) that maximizes robustness. They propose both open-loop and closed-loop (MPC) formulations, with the MPC employing a task-adaptive horizon to reduce computational cost.

This paper presents TAPE, a tether-aware path planning method for autonomous exploration of unknown 3D cavities using a tethered aerial robot. It employs a two-level hierarchical architecture with global frontier-based planning solving a TSP for distance minimization and local planning minimizing path cost and tether length via an adjustable decision function. Simulations and field tests show the method ensures tether length stays within limits in 100% of cases with only a 4.1% increase in distance traveled.

Researchers introduce Derail, an adversarial framework that exploits the scoring head in generative end-to-end autonomous driving planners, causing trajectory selection to flip from safe to unsafe with collision rates up to 50% and score drops of 39-80%, outperforming generic attacks.

This study compares continuous and discrete quadrotor dynamics on SE(3) for wind velocity estimation using EKF and UKF. Results show that discrete SE(3) dynamics with UKF achieves higher accuracy with low-cost sensors, highlighting potential for high-accuracy tasks.

This research explores leveraging large-scale dexterous grasp datasets to support articulated tool use in robotics. The authors construct a 355k-trajectory pretraining dataset, adopt a hierarchical imitation learning framework, and achieve significant improvements in task success rates in both simulation and real-world experiments.

This paper proposes a Vision-Language Procedural Reasoning (VL-PR) framework for autonomous guidewire navigation in robotic-assisted endovascular interventions. It integrates a multimodal large language model to interpret real-time visual observations for high-level navigation context inference and dynamically adjusts reward component importance. Experiments on a physical robotic platform demonstrate enhanced task reliability and efficiency over static-reward methods.

ViTL framework uses LLMs to compile natural language commands into Linear Temporal Logic formulas, which are converted into Deterministic Finite Automata to coordinate multi-channel value maps. It introduces a directional score for navigation, enabling zero-shot completion of multi-target, temporally constrained navigation tasks. Experiments on HM3D demonstrate its effectiveness.

This paper argues that despite rapid performance gains of Vision-Language-Action (VLA) models on robot manipulation benchmarks, current evaluation metrics cannot distinguish semantic from physical generalization, thus failing to verify physical reasoning abilities. The authors propose evaluation designs with controlled variation to separately measure these two capabilities.

This paper introduces a locker-based truck-drone routing problem (LTDRP-PDNF) that integrates pickups, deliveries, battery constraints, and no-fly zones, aiming to minimize total operational cost. The authors propose a two-stage deep reinforcement learning heuristic that outperforms baselines with short computation times.

This paper presents a benchmark for Arabic--Russian scientific translation, including a hybrid parallel corpus of about 27,000 sentence pairs compiled from scientific abstracts and general-domain texts. Three multilingual language models were fine-tuned using LoRA with various ranks. The Qwen2.5-7B model with QLoRA (rank 8) achieved the best results: BLEU 23.15, chrF 43.89, BERTScore 0.906, COMET 0.758, outperforming zero-shot baseline by +4.36 BLEU and +0.051 COMET. Few-shot prompting did not improve performance, indicating the necessity of domain-specific fine-tuning. The models, corpus, and evaluation code are released publicly, aiming to lower language barriers for scientific knowledge exchange between Arabic and Russian speakers and contributing to UN SDGs 9 and 17.

This study evaluates the robustness of Bangla event detection systems in real-world noisy conditions by introducing a benchmark of 9,979 annotated sentences across 40 event subtypes, including clean text, ASR transcripts, and orthographically corrupted text. Results show a trade-off: encoder models excel on clean text but degrade under noise, while decoder-only LLMs are more robust. Instruction tuning with annotation guidelines improves baseline performance on noise but not consistently. Model scaling boosts decoder robustness, and mixed training narrows the robustness gap, especially benefiting encoders.

This paper proposes a framework for therapeutic response generation driven by multi-dimensional, human-aligned evaluation. Stage I introduces TheraJudge, an open-source therapeutic evaluator trained via preference-based optimization on human-annotated data, producing reliable judgments across 7 psychological dimensions. Stage II introduces TheraAgent, which operationalizes the evaluations through a coordinated refinement process with Critic, Coach, and Therapist roles. TheraJudge achieves strong agreement with clinician ratings (ICC=0.87-0.95), and TheraAgent yields a +0.43 improvement in therapeutic quality, with low-quality responses improving by +2.45 points and a 94% recovery rate.

This paper describes a submission to SemEval-2026 Task 9 on detecting multilingual, multicultural, and multievent online polarization. The approach uses transformer-based models (RoBERTa-base for English, AfroXLMR-base for Swahili) with class-weighted loss functions to handle severe label imbalance and per-label threshold tuning for multi-label classification. On the test set, F1 macro scores are 0.7901 (English) and 0.7910 (Swahili) for Subtask 1, 0.4615 and 0.4808 for Subtask 2, and 0.4791 and 0.5830 for Subtask 3, showing competitive performance. Error analysis reveals struggles with dehumanization detection and lack of empathy.

This work proposes Outcome Reward Models (ORMs) as learned semantic scoring functions for test-time verification in Text-to-SQL. The GradeSQL framework automates ORM training via candidate generation and execution labeling, eliminating manual annotation. ORMs outperform execution-based Best-of-N and Majority Voting by up to 4.33% on BIRD and 2.10% on Spider, demonstrating scalability and stronger improvements on complex queries.

This study evaluates transformer language models on 'impossible' languages, finding that while grammatical sensitivity degrades gradually, generative capacity suffers pronounced failures, suggesting generative deficiency as a key factor in non-attestation.

A new study introduces ACE, an evaluation framework that controls for accuracy differences when comparing LLM calibration, revealing that many previously observed calibration advantages vanish and rankings can reverse.

This paper proposes a framework using generative AI agents as behavioral engines for black-box auditing of personalization algorithms. In a case study on X after the 2024 U.S. election with 1,120 agents, they find that the algorithmic feed amplifies toxic, polarizing, political, and right-leaning content compared to the chronological feed, with amplification varying by user ideology. Counterfactual analysis shows demographic signals affect content delivery in persona-dependent ways.

The Indi-RomCoM benchmark covers seven instruction-following tasks, four Indic languages, and three code-mixing intensity levels to systematically evaluate LLMs on Romanized Code-Mixed instructions. Results show LLMs consistently underperform, with performance degrading as code-mixing density increases; reasoning tasks degrade less than detection tasks.

A single LLM rewrite using false-positive and false-negative cases can match manually tuned skill descriptions, reducing engineering effort by 32x while maintaining routing accuracy.

Probabilistic downscaling, crucial for atmospheric science and climate modeling, often uses a mean-residual framework. However, in real-world applications, this approach frequently yields biased and under-dispersive ensembles. The root cause is identified as residual target misspecification: the residual distribution during training differs systematically from that required at test time due to downscaling bias. To address this, ReMatch (Residual Distribution Matching) aligns the training residual distribution to the test-time regime via optimal transport in a low-dimensional PCA space. Experiments on synthetic benchmarks and a real-world HRRR-ERA5 wind field downscaling task show ReMatch substantially reduces under-dispersion, improves calibration, and outperforms strong baselines.

This paper introduces Depth-wise Gradient Augmentation, a paradigm that transforms optimizer updates along depth to leverage inter-layer structure. A simple instantiation, Gradient Smoothing with a window operator, consistently improves optimization and generalization across diverse tasks (LM pretraining, RL post-training, diffusion, ViT) without architectural changes. It promotes structured representation evolution, interpreted as depth-wise preconditioning.

This paper develops a first-principles reduced-order model of GRPO training dynamics, subsuming the empirical single-exponential saturation law as its overdamped limit and adding an inertial term to capture the slow-start phase. It yields predictions tied to independently measurable quantities such as group-size invariance, a sharp stability threshold in the refresh interval, and an overdamped-to-oscillatory transition. The closed-form trajectory fits training reward with R² ≥ 0.91 across three models and two group sizes, and the predicted group-size invariance holds on both the reward curve and out-of-distribution transfer to eight math benchmarks. Additionally, the model furnishes diagnostics that separate failure modes conflated by the reward curve alone, such as reward hacking, advantage degeneracy, policy concentration, and dynamical instability.

This paper introduces process sidecars, a two-coefficient edit family for revoking learned state in language models after safety training, proving second-order accuracy over naive methods and demonstrating improvements across three models.

ReactionAtlas introduces a machine learning framework that autonomously builds chemical reaction networks from a small set of seed molecules without hand-crafted rules. Using a generative model and a DFT-trained machine learned force field, it discovers approximately 47,000 reactions among 12,000 compounds, achieving near-DFT accuracy for transition states. This enables new insights into prebiotic chemistry, particularly the formose cycle.

Hierarchical Global Attention (HGA) is a drop-in replacement for dense causal attention in pretrained long-context transformers, enabling 64K token context on a single RTX 5090 without retraining or calibration, with minimal quality loss.

This paper reveals that deterministic few-step generation fails on continuous text latents due to geometric constraints: smooth deterministic maps cannot resolve discrete branch choices before sharp categorical readouts. Diagnostics DABI and CCI quantify the sharpness gap between text and image decoders. Two escape mechanisms are identified: categorical commitment (autoregressive) and stochastic re-injection.

Large language models (LLMs) excel at many tasks but often lack structural consistency in their solutions. A new approach called MetaFlow treats workflow generation as a meta-learning problem, training LLMs to compose solution strategies using a two-stage process: supervised fine-tuning on synthetic workflows, followed by reinforcement learning with verifiable rewards. MetaFlow achieves strong performance on in-domain tasks and demonstrates remarkable zero-shot generalization to out-of-domain tasks and novel operator sets.