论文

研究:单链推理过早锁定假设在 Qwen3-8B 上获 16 项检验中 15 项支持

From Imperfective Paradox to Moorean Absurdity: Single-Chain Reasoning Lock-In and Multi-Candidate Verbalization Across Aspectual and Pragmatic Ambiguity

精选理由

这篇论文做得很扎实:4 个预注册实验、1.6 万个测试单元,专门验证推理链会不会在模糊语义上过早下结论。它提出的 MCV 方法一次调用拿 5 个候选答案再加权,对做评测或依赖 LLM 语义判断的开发者有直接可抄的工程价值。

一项预注册形式化研究在 Qwen3-8B-Q8_0 上测试了单链推理过早锁定假设,共完成 4 个密封实验、16,640 个形式化单元(temperature=0.0),16 个假设槽位中 15 个保持确认性结论,其中 12 个在 Holm 家族校正 alpha=0.05 下拒绝 H0,仅 E3_H3 降级为描述性结论。

正文

具体效应量上,实验 1 中 Direct-CoT 与 MCU 在有终止目标(telic)高置信完成任务上相差 25.00 个百分点(Holm p=1.31239e-05),而 NLI-CoT 在未中断的 atelic 项目上崩溃 91.25 个百分点。研究提出的 Multi-Candidate Verbalization(MCV)方法可在单次调用中引出 K=5 个加权估计,将提示与无加权(MCU)和自加权(MCW)读出分离。实验 4 显示 MCU-CF 在隐含 Moorean 句上的感知连贯性比 Direct-CoT 高 26.64 个百分点。

扩展实验包括 1,800 次调用的 OpenAI 消融和跨 5 个外部模型、4,240 个单元的 P16.7 阶段延伸,其中 OpenAI gpt-4o-mini 在 top_logprobs=20 下实现 1,000/1,000 有效单元。作者强调确认性效应仅涉及可观察的提示与读出协议,内部过早锁定仍属假设。全部代码、刺激材料与预注册分析计划已在 GitHub 开源。

原文 · arXiv cs.AI

From Imperfective Paradox to Moorean Absurdity: Single-Chain Reasoning Lock-In and Multi-Candidate Verbalization Across Aspectual and Pragmatic Ambiguity

Natural language entailment and coherence judgments often depend on whether an event merely occurred, reached a telic culmination, satisfied a plan, or was evaluated under multiple pragmatic standards. We test observable consequences of the single-chain premature lock-in hypothesis across four preregistered, sealed formal experiments on Qwen3-8B-Q8_0 (16,640 formal units, temperature=0.0): 15 of 16 hypothesis slots retain confirmatory status (12 rejecting H0 under within-experiment Holm control at alpha_fam=0.05; E3_H3 demoted to descriptive status). In Experiment 1, Direct-CoT vs. MCU differs by 25.00 pp on telic high-confidence completion (Holm p=1.31239e-05), while NLI-CoT collapses by -91.25 pp on uninterrupted atelic items (Holm p=1.81805e-21). Multi-Candidate Verbalization (MCV) elicits K=5 weighted estimates in one call, separating the prompt from unweighted (MCU) and self-weighted (MCW) readouts. Mandarin (E2) and English (E3) expose an Interruption-Anchored Realis Paradox under NLI-CoT. In Experiment 4, MCU-CF raises perceived coherence on implicit Moorean sentences by 26.64 pp over Direct-CoT (Holm p=0.0006), alongside Direct-CF - Direct-CoT = 29.15 pp and MCU-CoT - Direct-CoT = 17.31 pp. A 1,800-call OpenAI ablation and a stage="dev" P16.7 cross-model extension (4,240 planned/collected units across five external models: 4,222 valid, 18 format-missing, 0 transport-missing; <2% missing per slice; OpenAI gpt-4o-mini at top_logprobs=20 achieves 1,000/1,000 valid units) show that discrete NLI-CoT and scalar Direct-CoT judgments dissociate across models. Confirmatory effects concern observable prompt-and-readout protocols; internal premature lock-in remains a hypothesis. All code, stimuli, preregistered analysis plans, and verification tools are available at https://github.com/asj252/llm-judgment-studies.

产业坐标与信源核验

本条情报归属 AITOP「论文」核心栏目,重点聚焦中国 AI 产业与国产大模型生态。在跨区域产业观察坐标中,该动态直接关联「AI 开发者工具链、Coding Agent 与工程化实践」研究主线,反映出当前技术演进与商业落地的最新风向。涉及核心实体与技术议题包括:Qwen3-8B、gpt-4o-mini、CoT、NLI,读者可通过下方标签进入对应专题页查阅历史脉络与关联对比。

信源核验说明:本文原始线索采编自「arXiv cs.AI」(@An-Sheng Jhang),经 AITOP 多维评分管线结构化提炼。