Tag
该论文将生成模型的推理 effort 视为'信任品'，通过反事实实验量化不同 effort 级别对成本与准确率的影响，发现从低 effort 起步、失败后再升级的策略在多数模型和基准上以更低成本达到或超过固定最高 effort 的效果。
作者认为顶级AI实验室近期的短期重点应是让模型具备真实音频能力并在“taste（品味/质量感）”上实现类似Opus 5.5的跃升，从而让人们能自娱自乐，把算力更多用于实际生产力而非处理复杂社交障碍。
Toby Ord 的文章探讨 AI 模型通过更多 token 扩展的效果是否与人类通过更多时间扩展的效果相当，分析了 METR 的对比数据，并讨论了这一问题对 AI 自我改进与智能爆炸前景的影响。
This study evaluates whether test-time reasoning in large language models (like DeepSeek, GPT, and Gemini) improves net portfolio returns in trading, finding that additional reasoning does not reliably enhance economic outcomes across various conditions.
This paper introduces critique refinement and DISH to enhance the realism of alignment evaluations by mitigating evaluation awareness, demonstrating that combining both techniques yields significant improvements over using either alone.
Introduces 'Machine Studying' as a new formulation of continual learning where AI systems autonomously develop expertise from a corpus, and presents StudyBench for evaluation.
Introduces 'Machine Studying' as a problem where AI agents must autonomously develop expertise from a corpus, beyond RAG or long-context, and presents the StudyBench benchmark for evaluation.
This paper systematically studies how inference-time compute (token budgets, context compaction, repeated submissions) affects frontier LLM performance on challenging benchmarks, demonstrating that scores are protocol-dependent and advocating for evaluations that report capability as a function of inference compute.
TMAS introduces a multi-agent framework that enhances large language model reasoning by scaling test-time compute through structured collaboration and hierarchical memory systems. The approach uses specialized agents, cross-trajectory information flow, and hybrid reward reinforcement learning to improve iterative scaling and stability on challenging reasoning benchmarks.