发现大模型推理过早固化判断，导致后续错误难以纠正。
When Agents Commit Too Soon: Diagnosing Premature Commitment in LLM Agents
- 用隐藏状态收敛度量早期决策固化，捕捉推理过程是否已锁定路径。
- 在多个模型和数据集上，该信号可提前预测行为一致性（相关系数达-0.83）。
- 能识别错误路径但无法判断对错，适合用于运行时监控与纠错干预。
长时序大模型代理常因过早固化判断而失败：它们在早期选定一种证据解读后，后续推理全程都在维护这一选择。我们称此为‘过早承诺’。最终答案评分无法察觉此失败模式，因其只关注结果而非过程是否已坍塌。本文定义‘表征承诺’为固定推理步骤下跨运行的隐藏状态收敛，并以此作为轨迹一致性的早期诊断指标。在 Llama-3.1-70B 上运行 ReAct 于 HotpotQA 时，第4步隐藏状态相似性可预测下游行为一致性（皮尔逊相关系数 r = -0.35，偏相关 r = -0.45），且具有明确的时间与层间特征。该信号在 Qwen-2.5-72B 与 Phi-3-14B 上复现，并在 StrategyQA 上达到 r = -0.83。该信号不反映正确性：错误承诺与正确承诺在激活相似性上无法区分。这正是核心主张——承诺仅指示是否已固化，而非是否正确。运行时监测器基于隐藏状态实现最高 AUROC 0.97（严格划分下为 0.85–0.88）。提示干预使行为方差降低 28%，而准确率无显著变化。测试其是否可引导自洽计算资源时，仅在更难任务中表现平庸，且被更简单的输出基线超越。结论是，该方法提供对隐藏过程失效的诊断工具，有明确边界，非通用提效手段。
原文摘要 · Abstract (English)
Long-horizon LLM agents can fail quietly: they settle on one reading of the evidence early, then spend the rest of the run defending it. We call this premature commitment. Final-answer scoring misses the failure mode because it sees only the answer, not whether the process has already collapsed to a stable path. We define representational commitment as cross-run hidden-state convergence at a fixed reasoning step, and use it as an early diagnostic of trajectory consistency. On Llama-3.1-70B running ReAct on HotpotQA, step-4 hidden-state similarity predicts downstream behavioral consistency (r = -0.35, partial r = -0.45), with a localized temporal and layer-wise signature. The signal replicates across Qwen-2.5-72B and Phi-3-14B, and on StrategyQA (r = -0.83). It does not track correctness: committed-wrong and committed-correct questions are not separable in activation similarity. That boundary is central to the claim. Commitment tells us whether an agent has settled, not whether it is right. A runtime monitor detects inconsistent trajectories from hidden states at AUROC up to 0.97 (0.85--0.88 under a stricter split), and a prompting intervention cuts behavioral variance by 28% against a token-matched control while leaving accuracy statistically unchanged. We also test whether the signal can route self-consistency compute; on a harder benchmark it helps only modestly and is matched by a simpler output-based baseline. The result is a diagnostic for a hidden process failure, with clear limits rather than a general accuracy lever.
Thank you to arXiv for use of its open access interoperability. PaperDance 不是 arXiv 官方产品；中文卡片由大模型生成，请以原文为准。