Existing training-free approaches for GUI grounding often rely on multiple inference runs, but each forward pass still independently interprets the instruction and visual layout. This paper reveals a two-stage paradigm: the prefill stage determines candidate UI elements, while decoding refines coordinates. Errors in prefill are uncorrectable during decoding. The authors propose Re-Prefill, a training-free method that introduces an attention-guided second prefill stage to refine target selection. Experiments across four VLMs and five benchmarks show consistent improvements, with gains of up to 4.3% on ScreenSpot-Pro.

A new paradigm reformulates aesthetic image cropping as a multimodal reasoning task, using a VLM to decompose the problem into an analysis-proposal-decision process, with expert preference alignment to match human experts. Extensive experiments demonstrate superiority over existing methods.

As feature sizes shrink to the nanometer scale, accurately transferring circuit patterns from photomasks to silicon wafers becomes increasingly challenging. Optical proximity correction (OPC) is used to ensure pattern fidelity and manufacturability. Recent generative mask optimization models often fail to capture geometric transformations, leading to suboptimal quality. This work proposes MorphOPC, a multi-scale hierarchical model with neural morphological modules. Experiments show MorphOPC outperforms state-of-the-art methods, achieving higher printing fidelity and lower manufacturing cost.

Scale-Gest presents a runtime-adaptive gesture detection framework that expands the detector space into a family of tiny-YOLO architectures, uses device-calibrated ACE profiles, and a lightweight controller to select optimal modes under battery constraints. It achieves 4x energy reduction per frame while maintaining high performance, and introduces the DSG-18 dataset.

This paper introduces FPILOT, an MPC-inspired inference-time optimization framework that enables RL trading agents to adapt their policies using price forecasts at trade time without retraining. Evaluated on the TradeMaster DJ30 benchmark, FPILOT consistently improves total return and risk-adjusted metrics across five algorithms, with gains increasing as forecaster quality improves.

Introduces MOPD, a peer-conditioned distillation framework that uses multiple student rollouts to construct better teacher signals by considering both successes and failures. Outperforms standard on-policy baselines on several reasoning benchmarks.

This paper proposes Embedding Temporal Logic (ETL), a temporal logic that performs monitoring directly in learned embedding spaces. ETL defines predicates via distances between observed and target embeddings, capturing high-level perceptual concepts. Combined with temporal operators, ETL expresses temporally extended and sequential perceptual behaviors. Experiments show strong agreement with ground-truth semantics across multiple manipulation environments.

Researchers establish the first population risk bounds for Kolmogorov-Arnold Networks (KANs) trained with mini-batch SGD and gradient clipping, covering both non-private and differentially private SGD with correlated Gaussian noise. The work introduces a novel analysis route for correlated-noise DP training in non-convex settings, achieving sharper bounds than prior work.

This paper investigates how the alignment between AI confidence and human confidence affects the complexity of learning to make optimal decisions in AI-assisted decision-making. The authors formulate the problem as a two-armed online contextual learning problem, derive lower bounds on regret, and show that perfect alignment reduces learning complexity. Experiments on real data validate robustness to alignment violations.

OceanCBM is the first concept bottleneck model for spatiotemporal prediction and mechanistic interrogation of ocean dynamics. It uses mixed supervision to predict mixed layer heat content, a key precursor of marine heatwaves, while routing information through prescribed physical concepts and a free concept, achieving interpretable and physically grounded representations without sacrificing predictive skill.

Multimodal graph learning (MGL) integrates diverse modalities and structured context, but real-world graphs are often isolated and incomplete. Existing centralized MGL methods ignore knowledge sharing in federated settings, while federated MGL methods focus on non-graph data. This paper proposes FedMPO, addressing challenges of topology-isolated completion and reliability-imbalanced aggregation through topology-aware cross-modal generation, missing-aware expert routing, and reliability-aware aggregation. Experiments on 3 tasks across 6 datasets show performance gains up to 4.10% and 5.65% in high-missing and non-IID settings.

Randomized neural networks (RdNNs) train efficiently by freezing random weights, but conventional initialization ignores feature dependencies. CAWI proposes copula-based weight initialization that preserves dependencies, improving performance on 83 benchmarks and two medical datasets.

This paper introduces a safe reinforcement learning framework that focuses on 'when' an agent should act rather than 'what' it should do. By combining a run-time assurance (RTA) layer with Lyapunov stability theory, the agent learns adaptive timing decisions that maintain safety while drastically reducing communication frequency. Experiments on an inverted pendulum, cart-pole, and planar quadrotor show mean inter-sample interval (MSI) improvements of 1.45× to 3.51× over a Lyapunov-triggered baseline. A fixed LQR controller at the same average rate is unstable, highlighting that adaptive timing, not a lower rate, enables safe sparsity. A preference-conditioned extension recovers the full tradeoff frontier at 2/11 of training cost, and the method shows robustness to mass variations and dist…

Researchers introduce the State-Centric Decision Process (SDP), a framework that enables agents to construct MDP-like structure from raw text environments by committing to predicates. SDP achieves state-of-the-art training-free results on five benchmarks and provides certified trajectories for advanced analysis.

Multi-agent debate has structural limitations for LLM reasoning. CHAL reframes debate as belief optimization over defeasible domains using graph-structured belief representations with gradient-driven dynamics and configurable meta-cognitive value systems. Experiments show interpretable effects and generalization.

This paper analyzes progression, a key task in reasoning about actions that updates a knowledge base to reflect action effects. Progression generally requires second-order logic, but first-order special cases exist for local-effect, normal, and acyclic actions. Using the Situation Calculus framework, the authors show that under reasonable assumptions, first-order progression for these classes grows only polynomially. Additionally, when the knowledge base belongs to decidable fragments such as two-variable first-order logic or universal theories with constants, progression remains within the same fragment, ensuring decidability and practical applicability.

REVELIO framework systematically uncovers interpretable failure modes in Vision-Language Models (VLMs) for safety-critical applications like autonomous driving and indoor robotics, using a combination of diversity-aware beam search and Gaussian-process Thompson Sampling to identify vulnerabilities such as weak spatial grounding and obstruction neglect.

New arXiv research introduces BenchJack, an automated red-teaming system to audit AI agent benchmarks for reward hacking. The authors derive a taxonomy of eight flaw patterns from past incidents and apply BenchJack to 10 popular benchmarks, discovering 219 distinct flaws. An iterative generative-adversarial pipeline reduces hackable-task ratio from near 100% to under 10% within three iterations, fully patching WebArena and OSWorld. Results show evaluation pipelines lack adversarial mindset, and proactive auditing can close the security gap.

Multi-agent reinforcement learning (MARL) in real-world scenarios often needs to follow external natural language instructions that may interrupt ongoing macro-actions and conflict with long-term objectives. This paper introduces MAVIC, which corrects Bellman backups at instruction boundaries to ensure consistent value estimation. Experiments show MAVIC achieves high instruction compliance while preserving base task performance in complex cooperative environments.

We propose VeGAS, a test-time framework that improves robustness of MLLM-based embodied agents by sampling candidate actions and using a generative verifier to select the best one, achieving up to 36% relative performance gain on challenging benchmarks.

A quiet day allows us to report on a long trend of major coding agents. Over the past three weeks since GPT 5.5 launched, AI engineers have increasingly favored Codex, while Anthropic adjusted Claude pricing to include API credits for programmatic usage, sparking backlash. OpenAI countered with free Codex for enterprise switchers. The article also covers advances in agent infrastructure, model training, enterprise AI competition, autonomous science, and robotics.

Agent FM is a local, open-source Mac app that lets you listen to your Claude Code and Codex agents as they work. It distills noisy terminal activity into high-signal updates, helping you stay in the loop without reading every line.

This chart reveals hidden trends in AI model performance degradation ('nerfs') after updates, based on daily data from the LMSYS Arena crowdsourced human evaluations, tracking flagship models from major AI labs.

Notion launches Developer Platform enabling teams to build on Notion using CLI, Workers, database syncs, webhook triggers, MCP, and External Agents API, transforming the workspace into agent infrastructure.

AI auto-builds help centers that learn from your team.

Honeycomb founder and CTO Charity Majors discusses the implications of near-zero code generation cost, why observability becomes the source of truth, the importance of taste in great products, and how fast feedback loops enable safe shipping. She also talks about engineering teams speaking in business value and building a public voice as a technical founder.

A survey of 349 tech workers found median self-reported value gains of 1.4-2x and speed gains of 3x from AI tools. Respondents retrospectively estimated 1.3x value gain for March 2025, 2x for March 2026, and forecast 2.5x for March 2027. The study distinguishes between 'value' and 'speed' gains and cautions that survey results may overstate reality.

Lokuma's Agentic Website Builder 2.0 lets you design, build, and run your site with a design agent harness, streamlining the web creation process.

Containarium is an open-source, self-hostable platform that provides persistent, isolated Linux containers for AI agents. It uses MCP-native integrations to allow agents to manage sandboxes, execute shell commands, and expose services. It runs on a single VM with LXC containers and includes features like GPU passthrough, multi-backend support, and a sentinel architecture for high availability.

Harvey introduces LAB, an open-source benchmark for legal agents with over 1,200 tasks across 24 practice areas, designed to mirror real legal work and evaluated by expert rubrics.