NVIDIA提出PivotOPD:教多轮AI智能体从关键错误中恢复
NVIDIA PivotOPD Teaches Multi-Turn AI Agents to Recover From Pivotal Mistakes
Agent容错是落地核心痛点,这篇论文给出了可复现的蒸馏框架与具体参数权衡,做Agent工程的同学值得深入看。
NVIDIA researchers, with Princeton University and the University of Maryland, have introduced PivotOPD, an on-policy distillation method for multi-turn LLM agents. PivotOPD on-policy distillation trains an agent to avoid its most damaging early mistake, and to recover when it happens anyway. Against 13 baselines, it posts the best average on ALFWorld, WebShop and Search-based QA for Qwen3-1.7B and Qwen3-8B students. The takeaway: recovery is learnable, and standard OPD rarely teaches it.
NVIDIA 研究人员与普林斯顿大学及马里兰大学合作,推出了 PivotOPD,这是一种用于多轮大语言模型(LLM)智能体的在线策略蒸馏方法。PivotOPD 在线策略蒸馏训练智能体避免其最具破坏性的早期错误,并在错误发生时进行恢复。在 13 个基线方法的对比中,Qwen3-1.7B 和 Qwen3-8B 学生模型在 ALFWorld、WebShop 和基于搜索的问答任务上取得了最佳平均成绩。核心启示:恢复能力是可学习的,而标准的 OPD 很少教授这一点。
TL;DR
TL;DR(太长不看版)
- Size: A training method, not a model. Tested on Qwen3-1.7B and Qwen3-8B students, plus a Nemotron-3.5-SFT student on SWE-Bench Verified.
- Runs on: Trained on NVIDIA H100 nodes. Adds 0 inference cost, so the trained agent runs wherever its base model runs.
- Performance: First on all 8 per-benchmark averages against 13 baselines, across 3 seeds.
- Best: Recovers from 72.7% of replayed pivotal mistakes, vs 20.3% for standard OPD.
- Worst: 55.9% on ALFWorld “Look” tasks with the 1.7B student, vs 83.9% for SOD.
- Bottom line: Best: teaches recovery that outcome-only RL cannot reach. Worst: depends on replayable environments and a teacher whose pivots match the oracle in 77.8% of failed rollouts.
- 规模:这是一种训练方法,而非模型。在 Qwen3-1.7B 和 Qwen3-8B 学生模型上进行了测试,此外还在 SWE-Bench Verified 上测试了 Nemotron-3.5-SFT 学生模型。
- 运行环境:在 NVIDIA H100 节点上进行训练。不增加任何推理成本,因此训练后的智能体可在其基础模型运行的任何地方部署。
- 性能:在 3 次随机种子实验中,针对 13 个基线方法,在所有 8 项基准测试的平均得分中均排名第一。
- 最佳表现:能够恢复 72.7% 的重放关键错误,而标准 OPD 仅为 20.3%。
- 最差表现:在使用 1.7B 学生模型的 ALFWorld “Look” 任务中得分为 55.9%,而 SOD 为 83.9%。
- 结论:最佳方面:教授了仅靠结果导向强化学习无法达到的恢复能力。最差方面:依赖于可重放的环境,以及教师模型的关键点(pivots)在 77.8% 的失败轨迹中与 oracle 匹配。
What is a pivotal mistake in a multi-turn agent?
什么是多轮智能体中的关键错误?
A pivotal mistake is an action that lengthens the shortest remaining path to finishing a task, or makes it unsolvable. ALFWorld’s symbolic oracle measures this at every turn.
关键错误是指延长了完成任务的最短剩余路径,或导致任务变得不可解的动作。ALFWorld 的符号化 oracle 会在每一步衡量这一点。
Across Qwen3-8B, Qwen3-30B-A3B and Qwen3-235B-A22B, 59% of failed rollouts (155 of 262) contained one. The first pivotal turn arrived early, at a median of turn 8 to 12 out of 30. The agents then wasted 18 to 21 more turns without recovering.
在 Qwen3-8B、Qwen3-30B-A3B 和 Qwen3-235B-A22B 模型中,59% 的失败轨迹(262 次中的 155 次)包含至少一个关键错误。第一个关键步骤出现得很早,在第 30 步的中位数为第 8 到 12 步。随后,智能体在没有恢复的情况下浪费了 18 到 21 个更多步骤。
In replays of Qwen3-8B failures, correcting the pivotal turn raised success from 8% to 59%. Leaving the mistake in place and forcing the right action for the next 2 turns still reached 58%.
在重放 Qwen3-8B 失败案例时,纠正关键步骤将成功率从 8% 提升至 59%。即使保留该错误并强制后续 2 步采取正确动作,成功率仍能达到 58%。
Why does standard on-policy distillation miss it?
为什么标准的在线策略蒸馏会忽略它?
Standard OPD lowered the held-out failure rate from 79% to 56%. Failures after a pivotal turn only fell from 51% to 49%. The correct action stayed below 1% probability at every pivotal turn, so 8 rollouts rarely sample it.
标准 OPD 将留出集(held-out)的错误率从 79% 降低至 56%。但在关键步骤之后的失败率仅从 51% 降至 49%。在每个关键步骤中,正确动作的概率始终低于 1%,因此 8 次轨迹采样很难捕捉到它。
Outcome-based RL shares the blind spot: if every rollout fails, the group-relative advantage is 0.
基于结果的强化学习也存在同样的盲区:如果所有轨迹都失败,组内相对优势值为 0。
How does PivotOPD work?
PivotOPD 是如何工作的?
PivotOPD adds 3 components to group-based RL, combined in a single PPO update.
PivotOPD 向基于组的强化学习添加了 3 个组件,这些组件组合在一个单一的 PPO 更新中。
- Pivot detection: A larger teacher model reads each rollout and its outcome in hindsight. It picks candidate turns and names a gold action at each. A turn counts as pivotal when the student’s action differs from the gold action. On ALFWorld, detected pivots land within 1 turn of the oracle’s pivot in 77.8% of failed rollouts on average.
- Preventive distillation (reverse KL): A frozen copy of the student, hinted with the gold action, re-scores the student’s own response. This pushes the student away from the committed mistake.
- Recovery distillation (forward KL): After each pivot, the teacher names a recovery action for up to K turns. The hinted self-teacher writes recovery responses, and the unhinted student trains on them. Forward KL is mass-covering, so it lifts actions the student almost never samples.
- 枢轴检测:一个更大的教师模型事后阅读每次 rollout 及其结果。它选择候选转折点,并在每个转折点上指定一个黄金动作。当学生的动作与黄金动作不同时,该转折点被视为关键转折点。在 ALFWorld 上,平均而言,在 77.8% 的失败 rollout 中,检测到的枢轴位于oracle枢轴1步之内。
- 预防性蒸馏(反向 KL):学生模型的冻结副本,辅以黄金动作提示,对学生自己的响应重新评分。这将学生推离已犯下的错误。
- 恢复性蒸馏(前向 KL):在每个枢轴之后,教师为最多 K 个步骤指定恢复动作。带有提示的自我教师生成恢复响应,无提示的学生在这些响应上进行训练。前向 KL 具有质量覆盖特性,因此它能提升学生几乎从不采样的动作。
The teacher only names actions. Token-level targets come from the student’s own hinted distribution.
教师仅指定动作。词元级目标来自学生自身的带提示分布。
How does PivotOPD perform on agent benchmarks?
PivotOPD 在智能体基准测试中的表现如何?
With the 1.7B student, PivotOPD averages 73.7% on ALFWorld, 5.5 points above SDAR. It averages 44.5% on Search-based QA, 5.9 points above RLSD. On WebShop it beats RLSD by 1.2 in score but by 14.1 in success rate (76.6%).
使用 1.7B 学生模型,PivotOPD 在 ALFWorld 上平均得分为 73.7%,比 SDAR 高出 5.5 分。在基于搜索的 QA 任务上平均得分为 44.5%,比 RLSD 高出 5.9 分。在 WebShop 上,它在得分上比 RLSD 高出 1.2,但在成功率上高出 14.1(达到 76.6%)。
With the 8B student, it reaches 93.0% on ALFWorld, 47.4% on Search-based QA and 81.9% WebShop success. Margins are smaller, at least 1.8 points.
使用 8B 学生模型,它在 ALFWorld 上达到 93.0%,在基于搜索的 QA 任务上达到 47.4%,WebShop 成功率为 81.9%。优势幅度较小,至少为 1.8 分。
With Qwen3-8B as its own teacher, PivotOPD still wins all 3 benchmarks by at least 1.5 points, 3.9 on average.
以 Qwen3-8B 作为其自身教师时,PivotOPD 仍在所有 3 个基准测试中以至少 1.5 分的优势获胜,平均高出 3.9 分。
On SWE-Bench Verified, a Nemotron-3.5-SFT student taught by Nemotron-3-Super went from 62.8% to 66.0%. Standard OPD reached 63.0%, and the teacher scores 73.0%.
在 SWE-Bench Verified 上,由 Nemotron-3-Super 指导的 Nemotron-3.5-SFT 学生模型从 62.8% 提升至 66.0%。标准 OPD 达到 63.0%,而教师模型的得分为 73.0%。
Recovery is the standout. Across 72 replayed pivotal mistakes, PivotOPD recovered 72.7% of the time, vs 8.3% for the base model, 20.3% for standard OPD and 45.8% for preventive-only. It averaged 9.7 turns to recover, against an optimal 6.2.
恢复能力最为突出。在 72 次重放的枢轴错误中,PivotOPD 有 72.7% 的概率实现恢复,而基础模型为 8.3%,标准 OPD 为 20.3%,仅含预防性的方法为 45.8%。恢复所需的平均步数为 9.7 步,而最优值为 6.2 步。
更进一步:量化金融体系
看懂新闻只是起点——沿量化金融路径,把它变成能交付的工程能力