Tag
该论文将评估器引导的测试时推理建模为一个计算感知的三动作控制器（trust、fallback、expand），通过学习并校准动作特定的置信度估计来优化推理计算分配，在 Qwen3-8B 上实现约 30% 和 16% 的计算节省。
OpenAI据报正加紧确保其下一代旗舰模型的安全发布，此前已因安全顾虑取消了一个较弱的模型版本，并将10%的算力 diverted 到安全工作；其新型循环架构使AI推理更难监控，美国政府的测试可能导致发布延迟最多30天。
This workshop paper audits how a cumulative token cap falling inside a mathematical derivation affects test-time controllers, comparing strict (stop at cap) versus advisory (finish current attempt) stopping across 19,200 archived traces. It finds that budget curves must jointly report cap, realized cost, eligible candidates, and selector information, since candidate coverage and nominal budgets alone can mislead.
The paper introduces DEEP, a benchmark for evaluating compute-escalation allocators in perception systems against a decision-oracle baseline, finding that 34–54% of escalations that change downstream loss actually make it worse on KITTI, nuScenes, and nuPlan.
This paper investigates test-time scaling for scientific equation discovery, formulating it as an iterative search process and finding that search width is the dominant allocation parameter for improving performance and efficiency under compute budgets.
Approximately 10% or more of compute at frontier AI labs is dedicated to monitoring training runs to prevent AI agents from going rogue during reinforcement learning rollouts, with a suggestion that this should be increased.
Gambit improves reasoning model efficiency by using thought-level beam search to dynamically allocate compute to promising reasoning traces under fixed hardware budgets, yielding significant accuracy and throughput gains.
This paper studies the compute allocation problem in RL post-training for foundation models, proposing a FLOP-accounting framework for GRPO post-training. It finds conditional allocation frontiers depending on model size, budget, and reward system, and introduces RACE as a diagnostic protocol.
This paper proposes consequence-aware test-time compute allocation, routing higher-stakes software engineering tasks to larger compute budgets based on predicted failure cost rather than just difficulty. Evaluated on SWE-bench Lite and Multi-SWE-bench mini, the approach reduces cost-weighted loss by 22–33% compared to difficulty-aware routing.
This paper studies compute allocation in LLM-guided evolutionary search, identifies empirical regularities, and proposes BaSE, a multi-armed bandit algorithm that improves mean fitness and reliability across multiple models and tasks.
Sam Altman announces that a program offering compute capacity will be available until the current allocation sells out, with plans to resume later while reserving capacity for ChatGPT and Codex.
The article critiques AI labs for prioritizing commercial interests like advertising over solving critical global challenges such as climate change and energy. It argues that society should demand AI resources be directed toward beneficial societal goals rather than profit-driven applications.