Papers11月9日 13:13

arXiv 论文提出并行智能体小组方法提升测试时计算扩展效率

Test-Time Compute Scales Better with Parallel Agent Teams

arXiv cs.AI@Junlin Wang et al. (5)

Paper

Why we picked it

研究发现推理时别让 100 个智能体单干，也别挤成一个大组，分成 10 个小组并行跑效果最好，一半任务还能赢过前沿模型。

一篇 arXiv 论文提出 parallel agent teams 方法：在推理时并行采样多个小规模协作智能体小组，保留最优小组的结果。实验使用 27B 模型、100 个智能体的预算，在 12 个长程任务（GPU kernel、形式化证明、漏洞利用开发）上，10 个一组的并行小组在 12 个任务中的 10 个上胜过并行独立智能体，并在所有 4 个可比任务上胜过单个 100 人团队。在相同美元成本下，该方法在一半任务上超过一个前沿模型。论文解释其优势来自兼顾高均值与方差收益。

Original · arXiv cs.AI

Test-Time Compute Scales Better with Parallel Agent Teams

Test-time scaling improves the performance of a fixed model by spending more computation at inference. There are two common ways to spend it on agents when a verifier can score every attempt. Best-of-N sampling runs many independent agents and keeps the best result but saturates quickly. One large team of collaborating agents can score higher but stops improving as it grows. Can we get the best of both? Our method, parallel agent teams, does so by sampling many small teams in parallel and keeping the best team's result. We test it on 12 long-horizon tasks from GPU kernels and formal proofs to exploit development. With a 27B model and a budget of 100 agents, parallel teams of ten beat parallel agents on 10 of 12 tasks and beat one team of 100 on all four tasks where we ran such large teams. At equal dollar cost the same parallel teams beat a frontier model on half the tasks. A simple statistical model shows that individual agents lose because their many samples have a low mean and that one large team loses because a single sample gets no gain from variance. Sampling many small teams in parallel balances a high mean with the gain from variance. Parallel agent teams thus offer a more effective way to scale test-time compute.

#parallel agent teams #test-time scaling #智能体 #推理模型 #paper

Keep reading

Recent stories with the same tag or section

Papers Builders13:10

Chinese contentCPU-ARC：将 LLM 推理轨迹编译为可复用缓存，调用成本降至 1/16-1/291

一篇挺有意思的论文：让 LLM 只推理一次，之后把解题过程存成 CPU 能直接跑的缓存，便宜 16 到 291 倍还不用 GPU。做推理成本优化的可以看看。

#CPU-ARC #推理缓存 #LLM推理成本 #开源模型

arXiv cs.AI@Peter Belcak et al. (5)Source

Papers Builders13:10

Chinese contentPC-Dr.DPO：让多数投票伪标签的测试时强化学习更可靠

用多数投票当伪标签做测试时强化学习容易翻车，这篇论文给出了误差分解和修复方法，做无标注数据适配的可以看看。

#TTRL #DPO #伪标签 #推理模型

arXiv cs.AI@Haoyu Wang et al. (5)Source

Papers Builders13:10

Chinese contentGRAC：基于图的检索增强方法用于手术场景描述

医生写手术报告太费时间，这篇论文用图结构加检索生成场景描述，三元组准确率比 SGT 高了快 20 个点，做医疗 AI 的可以看看思路。

#GRAC #RAG #手术场景描述 #EndoVis-18

arXiv cs.AI@Rochak Dhakal et al. (6)Source

Papers Builders13:09

Chinese content无参考场景下用多模型一致性评估 LLM 代码摘要可信度

写 COBOL 这种没人愿意人工标注的老代码文档时，这篇给了个思路：让几个模型各写摘要互相投票，不用人也能估准不准。

#LLM-as-judge #代码摘要 #COBOL #置信度估计

arXiv cs.AI@Marco Colombo et al. (8)Source

Papers BuildersSelected13:09

BitDecoding-FP4 Cuts KV Cache to FP4 With 2.4-3.0x Throughput Gains

The co-design removes the trade-off that previously forced a choice between quantization accuracy and decode speed: system-level changes like transposed softmax eliminate lookup overheads that plagued earlier FP4 KV methods. Teams running long-context inference on NVIDIA Blackwell hardware can apply it directly, since the implementation is open-sourced and works across dense and MoE models with GQA and MLA, and it even holds up in RL training rollouts.

#BitDecoding-FP4 #NVFP4 #Qwen3 #KV cache

arXiv cs.AI@Dayou Du et al. (6)Source

Papers Builders71°13:09

Chinese contentLLM 智能体在 Lean 中自动发现经过形式化验证的量子程序

researchers 让 LLM 智能体在 Lean 里写量子程序，还能数学证明验证对不对，量子算术直接提升 290 倍，值得看看这套自动化发现流程怎么搭的。

#Lean #LLM智能体 #量子计算 #形式化验证

arXiv cs.AI@Sergey Shirobokov et al. (25)Source