當代理過早承諾：診斷LLM代理中的過早承諾問題

摘要

長時程LLM代理人可能悄悄失敗：它們過早對證據採取單一解讀，隨後花費整個運行階段捍衛該解讀。我們稱此為「過早定論」。最終答案評分無法捕捉此失敗模式，因為它只看答案，而不關注過程是否已塌縮至穩定路徑。我們將「表徵承諾」定義為在固定推理步驟中，跨運行隱藏態的收斂現象，並將其作為軌跡一致性的早期診斷指標。在以Llama-3.1-70B執行ReAct於HotpotQA的實驗中，第4步的隱藏態相似度可預測下游行為一致性（r = -0.35，偏相關r = -0.45），並呈現局部化的時間與層級特徵。此訊號在Qwen-2.5-72B與Phi-3-14B上可複現，在StrategyQA上亦成立（r = -0.83）。該訊號不追蹤正確性：已定論錯誤與已定論正確的問題，在激活相似度上無法區分。此界線正是論點核心。承諾告訴我們代理人是否已定局，而非其是否正確。基於隱藏態的運行時監控器，能以AUROC高達0.97（在較嚴格劃分下為0.85–0.88）檢測不一致軌跡；而一種提示干預手法，在令牌匹配對照組下將行為變異降低28%，同時保持準確率統計上無顯著變化。我們亦測試該訊號能否引導自一致性計算資源分配；在更難的基準測試中，其僅有適度幫助，且表現被更簡單的基於輸出的基線方法所匹敵。最終成果是一種針對隱藏過程失敗的診斷工具，具有明確的適用限制，而非通用的準確率提升槓桿。

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.