LLM benchmarks show improved reasoning and efficiency

Statements (20)

1. ▲ Bullish

   The technology of minimal-norm univariate two-layer ReLU networks with skip connections achieves exact global optimality for binary classification.

   arxiv.orgSep 23, 2026 17:37 UTC
2. ▲ Bullish

   CompKV framework improves sparse attention speedup for LLM inference by up to 6.85x over full attention.

   arxiv.orgSep 22, 2026 12:11 UTC
3. ▲ Bullish

   AI model training efficiency improves with hybrid algorithms combining sharded data parallelism and federated learning.

   arxiv.orgSep 17, 2026 13:21 UTC
4. ▲ Bullish

   CoRA-NAS framework achieves mean Spearman correlations of 0.786 on TransNAS-Bench-101, indicating improved accuracy prediction for neural architecture search.

   arxiv.orgSep 10, 2026 17:49 UTC
5. ▲ Bullish

   CoRA-NAS framework achieves mean Spearman correlations of 0.715 on NAS-Bench-101, indicating improved accuracy prediction for neural architecture search.

   arxiv.orgSep 10, 2026 17:49 UTC
6. ▲ Bullish

   CoRA-NAS framework achieves mean Spearman correlations of 0.946 on NAS-Bench-201, indicating improved accuracy prediction for neural architecture search.

   arxiv.orgSep 10, 2026 17:49 UTC
7. ▲ Bullish

   CoRA-NAS framework achieves mean Spearman correlations of 0.894 on NATS-SSS, indicating improved accuracy prediction for neural architecture search.

   arxiv.orgSep 10, 2026 17:49 UTC
8. ▲ Bullish

   CoRA-NAS framework achieves mean Spearman correlations of 0.715 on NAS-Bench-201/CIFAR-100, indicating improved accuracy prediction for neural architecture search.

   arxiv.orgSep 10, 2026 17:49 UTC
9. ▲ Bullish

   DSE-VTG achieves state-of-the-art performance on training-free video temporal grounding benchmarks.

   arxiv.orgSep 8, 2026 15:03 UTC
10. ▲ Bullish

    Sliding Window Attention with sinks requires no post-training and is extremely fast and low memory.

    arxiv.orgAug 28, 2026 15:31 UTC
11. ▲ Bullish

    Large Language Models using Sliding Window Attention with sinks achieve 2 to 10 times higher performance than Linear Attention models on long-context reasoning tasks.

    arxiv.orgAug 28, 2026 15:31 UTC
12. ▲ Bullish

    Switching to Sliding Window Attention instead of post-training Linear Attention models reduces inference memory cost.

    arxiv.orgAug 28, 2026 15:31 UTC
13. ▲ Bullish

    Reinforcement learning on tasks with verified statistical reward substantially improves reliability of LLM agents in hypothesis testing.

    arxiv.orgAug 7, 2026 17:22 UTC
14. ▲ Bullish

    Fisher-R1-14B substantially improves over its backbone and outperforms strong proprietary and open-source baselines including GPT-5.4 and DeepSeekV4-Pro.

    arxiv.orgAug 7, 2026 17:22 UTC
15. ▲ Bullish

    Skill Entropy RL improves Qwen3-4B-Instruct on Skill^2-Bench from 34.4% to 68.4%.

    arxiv.orgAug 5, 2026 17:57 UTC
16. ▲ Bullish

    Compass achieves state-of-the-art performance on multimodal crack segmentation with 90% depth modality missing, using only 2.58M parameters.

    arxiv.orgAug 4, 2026 12:27 UTC
17. ● Neutral

    Behavioral Grammar uses a compact 0.88M-parameter causal Transformer for host runtime behavior analysis.

    arxiv.orgAug 1, 2026 16:28 UTC
18. ▲ Bullish

    Microsoft Corporation may benefit from improved mixed-precision post-training quantization performance for Vision Transformers.

    arxiv.orgJul 30, 2026 17:43 UTC
19. ▲ Bullish

    ViT-Small/16 and ResNet-50 models achieve the best validation accuracy when using a spectral inner step with a Muon outer step.

    arxiv.orgJul 28, 2026 17:18 UTC
20. ▲ Bullish

    SpecAHD reduces held-out objective cost by up to 57.7% on four routing problems using LLM backbones.

    arxiv.orgJul 26, 2026 14:22 UTC