Skip to content
NewA calmer, faster ZeroHour

Study tests reasoning lock-in on Qwen3 across aspect and pragmatics

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

arXiv cs.AI / cs.LG / cs.CLby An-Sheng Jhang ()1 min read1 · 1 read
AI summary · x-ai/grok-4.7Summaries by Vani AI

Fifteen of 16 preregistered slots stayed confirmatory on 16,640 formal units at temperature 0.

Four sealed experiments on Qwen3-8B-Q8_0 test whether prompt and readout choices change judgments of aspectual and pragmatic ambiguity. Direct-CoT and multi-candidate protocols diverge sharply on telic completion, atelic NLI, and Moorean coherence. An OpenAI ablation and a five-model extension show discrete NLI-CoT and scalar Direct-CoT judgments dissociate across models. Internal premature lock-in remains a hypothesis; code and stimuli are public.

Unconfirmed: Confirmatory effects concern observable protocols; internal lock-in is not established.

Key facts
  • Direct-CoT versus MCU differs by 25.00 pp on telic completion.
  • NLI-CoT falls 91.25 pp on uninterrupted atelic items.
  • MCU-CF raises Moorean coherence 26.64 pp over Direct-CoT.
  • gpt-4o-mini produced 1,000 of 1,000 valid ablation units.
Explain this AI news with Vani AIWhat was announced, what is new and why it matters, in a new chat

From the original article

1 min read at arxiv.org

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.

arXiv cs.AI / cs.LG / cs.CL

By An-Sheng Jhang · arxiv.org

Read the full article at arXiv cs.AI / cs.LG / cs.CL

A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.