ExtractBench introduces a rigorous evaluation framework for schema-guided document extraction, addressing critical gaps in accuracy, grounding, and cost. The benchmark reveals that while commercial VLMs often struggle with long-form document truncation, specialized agentic workflows offer a more reliable and cost-effective path forward.
A comprehensive audit of 88 commercial AI products reveals that while system prompt security is improving, nearly 40% of applications still contain instructions that conflict with user interests. The new AISPA framework provides a standardized method for developers to evaluate these critical, often opaque, governance layers.
ReToken introduces a single learnable embedding to enable efficient, sparse retrieval of visual tokens from large KV caches. This method significantly improves performance on long-context vision-language tasks while maintaining a lightweight footprint suitable for single-GPU deployment.
The Ambient Capture Engine (ACE) introduces a new paradigm for collecting synchronized, multi-modal data in real-world home environments to address the fundamental data bottleneck in embodied intelligence. By capturing 150 hours of high-fidelity human interaction, ACE-Data-0 provides a comprehensive foundation for training next-generation robotic systems.
PhiZero introduces a 'reason-then-render' paradigm for world modeling, utilizing a learned, discrete 'physical language' to represent world-state transitions. This approach moves away from direct pixel-space prediction, enabling more explicit reasoning and physically coherent simulation.
The release of Frontis-MA1 and the OpenMLE stack provides a concrete framework for researching recursive self-improvement within machine learning engineering. By integrating execution-grounded operators with long-horizon search, the system demonstrates significant performance gains in automated ML tasks.
Beacon introduces a framework to optimize when and how Multimodal Large Language Models utilize external tools. By focusing on Mode Adaptiveness and Tool Effect, the model reduces computational overhead while improving performance on complex visual reasoning tasks.
Chimera introduces a hybrid diffusion architecture that leverages Kimi Delta Attention and Sparse MoE to overcome the quadratic scaling limits of traditional transformers. By applying HeteroP scaling laws, the model achieves significant compute efficiency gains while enabling zero-shot long-context video generation.
A new benchmark, OmegaUse-OfficeVal, introduces economic grounding to evaluate LLM agents on complex, long-horizon office tasks. By pairing tasks with human labor time and price proxies, it offers a realistic assessment of agent viability beyond simple accuracy.
AskChem introduces a claim-centered infrastructure for chemistry literature, replacing document-level search with atomic, provenance-backed claims to improve AI agent accuracy. By indexing 2.4 million claims across 147,000 papers, it provides a structured foundation for verifiable scientific synthesis.
A new research paper introduces β-OPSD, a framework that generalizes on-policy self-distillation by treating the KL penalty as a tunable hyperparameter. By converting complex reinforcement learning objectives into efficient logit-mixing distillation targets, the method improves both training stability and reasoning performance.
MindForge introduces an automated pipeline for creating source-free training environments, enabling smaller language models to achieve frontier-level performance in full-cycle software engineering. By training on synthesized trajectories, the Qwen3.6-27B model shows significant improvements across seven diverse software engineering benchmarks.
RefCaptioner introduces a novel framework for multi-reference image-grounded video captioning, enabling precise phrase-level binding and improved cross-reference consistency. By utilizing a two-stage training approach, the model enhances caption factuality for both real-world and AI-generated video content.
SpecFirst introduces a two-stage framework for AI-driven program synthesis that separates behavioral specification elicitation from code implementation. By treating requirements engineering as a first-class phase, the framework significantly improves success rates in from-scratch coding tasks.
A comprehensive study evaluating frontier AI agents on open-ended research tasks reveals significant gaps in their ability to perform scientific inquiry. While agents successfully handle engineering requirements, they consistently fail to navigate the strategic and creative demands of high-level research.
HumanCLAW introduces a novel evaluation framework that separates high-level action decision-making from low-level motor execution in vision-language models. By testing nine state-of-the-art models, researchers found that current VLMs lack the embodied self-awareness necessary to navigate and interact effectively in physical environments.
Parallel Decoding Distillation (PDD) introduces a trajectory-based approach to accelerate diffusion and flow matching models by predicting multiple denoising steps per network evaluation. This method bypasses the instability of traditional adversarial losses, enabling state-of-the-art performance with significantly reduced computational overhead.
Researchers have developed a robust framework for optimizing code execution speed using reinforcement learning, overcoming the inherent instability of timing-based rewards. By integrating a calibrated sandbox and refined GRPO techniques, this approach significantly improves performance metrics while maintaining code correctness.