斯坦福等团队出了个不用微调的验证器，用 token logits 读连续分数，SWE-Bench 干到 78.2%，还能给强化学习当奖励信号，很实用。
Stanford、NVIDIA、UC Berkeley 提出一种无需微调的验证器，直接从评分 token logits 读取连续校准分数。通过三个旋钮——分数粒度、重复评估、标准分解——提升准确率。在 Terminal-Bench V2 达 86.5%，SWE-Bench Verified 78.2%，RoboRewardBench 87.4%，MedAgentBench 73.3%。该连续分数还可作为 SAC 和 GRPO 的密集奖励，并已集成到 Claude Code 扩展中用于任务进度信号。论文见 arxiv.org/abs/2607.05391。
NEW AI paper worth bookmarking. This is something I called early, and this paper confirms it: verif...
NEW AI paper worth bookmarking. This is something I called early, and this paper confirms it: verification has emerged as a new important scaling axis. Here is the simple explainer and what this paper shows. We have seen lots of progress in scaling pre-training, post-training, and test-time compute. For post-training and test-time compute, we are still in its early phases. But one of the most important new directions is using LLMs as verifiers. Verifiers are fundamental to scaling AI. This work from Stanford, NVIDIA, and UC Berkeley builds a training-free verifier that reads a continuous, calibrated score straight off the scoring-token logits instead of trusting a discrete grade. Three knobs move accuracy without any fine-tuning. Score granularity for cleaner separation, repeated evaluation for lower variance, and criteria decomposition for lower complexity. The numbers land across very different domains. 86.5% on Terminal-Bench V2, 78.2% on SWE-Bench Verified, 87.4% on RoboRewardBench, and 73.3% on MedAgentBench. The same continuous score doubles as dense reward for SAC and GRPO and as a task-progress signal shipped in a Claude Code extension. Paper: arxiv.org/abs/2607.05391 Learn to build effective AI agents in our academy: academy.dair.ai 💬 8 🔄 8 ❤️ 41 👀 3906 📊 22 ⚡