Search and filter through extracted claims from AI researchers.
Showing 2201-2220 of 4537 claims
Video re-shooting can be achieved without explicit 3D priors or paired training data by using text-driven semantic viewpoint specification and self-supervised learning of camera dynamics
Lower and middle Transformer layers build semantic representations while upper layers increasingly specialize them for prediction, creating a natural division of labor
CoMem achieves 97.05 on RULER and 38.27 on LoCoMo versus 34.59 for full-context KV-Direct, demonstrating superior long-context performance while using fixed retrieval budget independent of stored-context length
Writing context chunks only through intermediate layers and recomputing upper layers over retrieved states enables efficient long-context processing with memory independent of context length
Existing disaster response datasets like Incidents1M and CrisisMMD suffer from either complete lack of text or severe text-image semantic misalignment
High-fidelity textual descriptions for vision-only datasets can be successfully generated and validated at scale using a combination of dense and MoE language models plus LLM-as-a-Judge validation
An image-blind LLM-as-a-Judge validation approach can ensure generated captions provide reliable semantic anchoring for Data-Free Knowledge Distillation by intentionally obscuring the original image
The Locksmith Loop agentic method consistently improved test coverage beyond input-search plateaus for COBOL-to-Java migration, reaching near-complete branch coverage across programs ranging from 430 to 4,114 lines
Iterative agentic loops combining witness search with parity-preserving mutations can overcome locked paragraphs that prevent deeper program exploration during legacy code migration testing
Most existing memory-augmented agents treat retrieved experiences as static records to replay verbatim, causing negative transfer by ignoring the gap between abstract stored experience and concrete current states
LLM agents can actively harness and reconstruct past experiences based on present context by critiquing and reconstructing retrieved memories conditioned on current state, similar to how humans adapt memories
No benchmark exists dedicated to testing Greek language models' inflectional competence, despite Greek being a richly inflected language
MORFES benchmark with 500 expert-verified items can effectively test recognition and production of Greek inflected forms
Grammatical competence in morphologically rich languages remains under-measured in the rapidly scaling open-weight LLM ecosystem
CNN and Transformer-based hyperspectral image classification methods still face locality constraints and high computational complexity
Combining multi-scale CNN with Mamba can integrate local feature extraction with long sequence modeling advantages for hyperspectral classification
Prior language-integrated monocular depth estimation methods fail to fully harness language potential due to short text input, coarse feature learning, and limited guidance
Detailed long captions can enhance visual perception capabilities of vision-language models and alleviate visual ambiguities in monocular depth estimation
Depth sensors remain underexplored for affordance segmentation in wearable robots despite their potential to complement RGB data
Hardware-aware neural architecture search with depth integration can overcome the limitations of tiny-like approaches that fail in critical scenarios due to severe hardware constraints
Pipeline data may be stale or degraded.
Last synthesis: 2026-09-20. 9,035 pending.