Search and filter through extracted claims from AI researchers.
Showing 1981-2000 of 4537 claims
Diffusion Transformers with intent-driven fusion guided by pathology-aware diagnostic intents can improve medical image fusion
Frontis-MA1 35B model improved performance on MLE-Bench Lite from 39.39% to 60.61% Medal Average under resource-constrained conditions (12-hour budget, single RTX 4090 with 12GB VRAM)
Coupling learning and evolution in a single loop through meta-evolution agents can enable recursive self-improvement in machine learning engineering tasks
The speed of the fusion module that combines before-and-after satellite image views largely sets the cost of every change-detection search query in Earth observation archives
Inference-time scaling in local computer-use agents often yields diminishing returns while changing failure modes rather than consistently improving performance
The effectiveness of inference-time scaling for resource-constrained local models remains poorly understood compared to frontier models
Advanced reasoning methods (planning, self-critique, reflection, debate) that generate more text than simple chain-of-thought don't necessarily improve accuracy due to their approach - the improvement may come simply from generating more tokens
Simple repeated sampling with majority voting often outperforms complex reasoning methods when token budgets are equalized across 36 experimental comparisons
Current multimodal RAG systems struggle with complex multi-hop reasoning because they focus on instance-level matching and fail to capture relationships across modalities and documents
Graph-enhanced multimodal methods face a fundamental dilemma: fine-grained visual features cause graph expansion and noise, while coarse-grained representations lose critical local evidence
High-fidelity 3D generation's reliance on scaling model capacity and data incurs prohibitive computational costs and overlooks rich priors in discriminative 3D foundation models
Leveraging discriminative 3D foundation models' understanding can significantly reduce the training cost of 3D generation
On-policy self-distillation (OPSD) remains brittle in practice and requires substantial engineering effort to work reliably
Vanilla OPSD's brittleness stems from it being the β=1 case of a broader policy-optimization family, and introducing β as a controllable parameter yields better regularization
13.6% of SWE-bench Verified instances exhibit PR-Issue misalignment across five patterns, undermining benchmark reliability
SWE-bench-like benchmarks suffer from systematic misalignment due to the complexity of PR-Issue pairing in large repositories
Existing post-training quantization methods for Vision Transformers use uniform bit-widths and overlook heterogeneous sensitivity to quantization, leading to inefficient precision allocation
Multimodal on-policy distillation's next-token corrections are source-mixed, combining visual signals with linguistic priors and teacher-specific effects, making it difficult to isolate visual evidence
The key challenge in multimodal distillation is estimating which corrections are supported by visual evidence rather than where or how strongly to distill
Scaling coding agents requires a continuing supply of executable data for training, benchmarking, and continuous evaluation
Pipeline data may be stale or degraded.
Last synthesis: 2026-09-20. 9,017 pending.