---
id: 20261025-T0-07
title: "奖励好论证反而让LLM更武断：自我奖励的道德推理一致但不可靠"
title_en: "Rewarding Good Arguments Makes LLMs More Decisive but Less Reliable on Moral Judgment"
url: https://ai.daily.yangsir.net/daily/20261025-T0-07
issue_date: 2026-10-25
publish_date: 2026-10-24T04:00:00.000Z
category: research
source_name: "arXiv cs.CL (NLP)"
source_url: https://arxiv.org/abs/2610.30349
---

# 奖励好论证反而让LLM更武断：自我奖励的道德推理一致但不可靠

研究用强化学习训练模型，以“论证质量”作为奖励信号，假设论证得好就更可能判断得好。结果显示，模型在道德判断上变得更果断，内部一致性提升，但判断本身更容易被引导和动摇，一致性并不等于有效性。这提示：把论证质量当作奖励，可能只是训练出更会自圆其说的模型。对用LLM做道德或价值判断的产品来说，需要额外校验判断的稳定性，而非只看输出是否自洽。

## English Version

**Rewarding Good Arguments Makes LLMs More Decisive but Less Reliable on Moral Judgment**

Researchers train a reinforcement learning policy that rewards good arguments, on the premise that arguing well raises the odds of judging well. The result: models become more decisive on moral judgment tasks and more internally consistent, yet their judgments are also more pliable and easier to steer. Consistency, the authors argue, is not validity. The takeaway for products using LLMs for moral or value judgments: verify stability under pushback rather than trusting self-consistent output.

---

**来源**：[arXiv cs.CL (NLP)](https://arxiv.org/abs/2610.30349)

**详情页**：https://ai.daily.yangsir.net/daily/20261025-T0-07

---

*智语观潮 · Daily — https://ai.daily.yangsir.net/llms.txt*