Tag
PORA 提出了一种配对结果引导的推理分配方法，让 LLM 智能体学习在何种状态与剩余预算下使用长推理更划算，在 BFCL memory 等任务上大幅提升性能并显著节省 token 与模型请求。
This paper proposes a trajectory-aware decoding control method for diffusion vision-language models to address reasoning-budget mismatch by routing examples based on their decoding state, improving robustness across benchmarks.
Benchmark results for MindControl, a sampler-level guided reasoning budget for llama.cpp, show multi-stage signaling reduces token consumption without accuracy loss, and sometimes improves it on easier tasks.