Search and filter through extracted claims from AI researchers.
Showing 2001-2020 of 4537 claims
Change2Task can derive multiple tasks grounded in developer evidence from maintained environments by converting merged pull requests into verified tasks
The fundamental goal of agentic visual reasoning should be improving success rates on complex tasks rather than merely equipping models with sophisticated yet inefficient reasoning paradigms
Existing agentic visual reasoning systems fail to optimize for Mode Adaptiveness and Tool Effect, leading to inefficient tool use
Safety fine-tuning suppresses LLMs' tendencies to attribute minds not only to themselves but also to non-human animals and natural objects, while reducing spiritual belief
Current safety alignment efforts inadvertently alter models' representations of mindedness in other entities alongside human beliefs and values
Restoring internal consciousness representations through ablation or steering recovers broad mind attribution and produces more human-like responses without impairing Theory of Mind capabilities
Clinical risk models routinely achieve strong aggregate performance while producing materially different error rates across patient subgroups
Significance in disparity detection tracks each axis's gap against its own minimum detectable effect, with standardization improving correlation from 0.56 to 0.78
Vision-language models as judges of computer-using agent trajectories have not been systematically evaluated for reliability
The field is increasingly turning to VLMs as judges of CUA trajectories for evaluation, data curation, and reinforcement learning at scale
The quadratic cost of full attention makes it prohibitive for visual generation requiring high-resolution images, long videos, and multimodal context
Chimera combines KDA for long-context state tracking with O(N) complexity, MLA for global interaction, and sparse MoE to expand capacity while controlling compute
HeteroP enables principled scaling of heterogeneous architectures by transferring hyperparameters according to each tensor's functional fan-in and model depth
System prompts are rarely disclosed to the public or regulators, creating a serious trust and accountability gap in AI system deployment
System prompt design varies substantially across products and developers, with some averaging over 60 protective instructions while others average fewer
Existing chemistry literature-search systems primarily return ranked document lists, which requires scientists and AI agents to manually locate relevant information, verify provenance, and assemble cross-paper answers
AskChem's claim-centered infrastructure changes the unit of retrieval from the paper to the provenance-carrying claim, converting each paper into atomic, typed claims grounded by source DOI and verbatim quotes
AskChem currently indexes 2.4M claims from 147 papers
A fixed onboard camera alone is adequate for humanoid dodgeball evasion, with the deployed policy coming within a few points of a privileged state oracle
Usable barrier structure in control-barrier functions depends on perceptual observability, with Joint-CBF performing best with accurate ball states but degrading under fixed-camera observations
Pipeline data may be stale or degraded.
Last synthesis: 2026-09-20. 9,017 pending.