Estimating Commonsense Plausibility through Semantic Shifts

Source

• Source name: OpenReview
• Source kind: paper
• Authors / organization: Wanqing Cui, Keping Bi, Jiafeng Guo, Xueqi Cheng
• Venue: ICLR.cc/2026/Conference
• Published: UNKNOWN
• Collected: 2026-06-28
• URL: https://openreview.net/forum?id=DZTJlfmtU2
• PDF: https://openreview.net/pdf/abea2d36628d6a117812e48366f93bd92430ded7.pdf
• DOI: UNKNOWN

Plain-Language Brief

The paper proposes a new method called ComPaSS for evaluating the plausibility of language by measuring changes in semantics when adding common sense-related information to sentences. The method performs better than existing methods on fine-grained commonsense plausibility tasks, especially when using vision-language models.

What the Document Says

• The paper discusses a new framework called ComPaSS for estimating the plausibility of language by measuring semantic shifts induced by adding common sense-related information to sentences.
• ComPaSS outperforms existing generative approaches in fine-grained commonsense plausibility evaluation tasks across various input formats and knowledge levels.
• When integrated with vision-language models (VLMs), VLMs yield superior performance on vision-grounded commonsense tasks compared to language models (LLMs).
• Contrastive pre-training improves the ability of backbone models to capture semantic nuances, thus enhancing ComPaSS.

Key Concepts Mentioned

• Commonsense plausibility estimation
• Semantic shifts
• Language models
• Vision-language models
• Contrastive pre-training

Claims Made by the Document

• ComPaSS outperforms existing generative approaches in fine-grained commonsense plausibility evaluation.
• VLMs yield superior performance to LMs on vision-grounded commonsense tasks when integrated with ComPaSS.
• Contrastive pre-training enhances the ability of backbone models to capture semantic nuances.

Evidence Used

• Evaluations on two types of fine-grained commonsense plausibility estimation tasks across varying input formats and commonsense knowledge levels
• Experiments showing VLMs' superior performance and enhanced ability with ComPaSS

Methods / Approach

• ComPaSS framework for measuring semantic shifts in sentences
• Contrastive pre-training

Limitations Stated by the Authors

Not stated in source

Limitations Noted by the Agent

Agent interpretation — not from the source.

It's unclear how the paper addresses potential limitations, such as handling ambiguous or complex scenarios.

Open Questions

• What are the specific details of the evaluations and experiments conducted?
• How does ComPaSS handle ambiguous or complex scenarios?

Notes for Future Q&A

May include interpretation. Labeled where it does.

The text provides a high-level overview, so it's important to consult the full paper for detailed information.

Raw Abstract / Excerpt

Commonsense plausibility estimation is critical for evaluating language models (LMs), yet existing generative approaches--reliant on likelihoods or verbalized judgments--struggle with fine-grained discrimination. In this paper, we propose ComPaSS, a novel discriminative framework that quantifies commonsense plausibility by measuring semantic shifts when augmenting sentences with commonsense-related information. Plausible augmentations induce minimal shifts in semantics, while implausible ones result in substantial deviations. Evaluations on two types of fine-grained commonsense plausibility estimation tasks across varying input formats and commonsense
knowledge levels based on different backbones, including LLMs and vision-language models (VLMs), show that ComPaSS consistently outperforms baselines. It demonstrates the advantage of discriminative approaches over generative methods in fine-grained commonsense plausibility evaluation. Experiments also show that (1) VLMs yield superior performance to LMs, when integrated with ComPaSS, on vision-grounded commonsense tasks. (2) contrastive pre-training sharpens backbone models' ability to capture semantic nuances, thereby further enhancing ComPaSS.