Challenge: Language models (LMs) perform complex mathematical reasoning in Physics where physical context requires that any symbolic manipulation satisfies complex semantics.
Approach: They systematically remove crucial context from prompts to force instances where model inference may be algebraically coherent, yet unphysical.
Outcome: The proposed models perform poorly in this domain, and their reasoning is not physics-informed.

Similar Papers

Fine-Grained Detection of Context-Grounded Hallucinations Using LLMs (2026.findings-acl)

Copied to clipboard

Challenge: Existing representations of hallucinations limit the types of errors that can be expressed, so we propose a new representation based on free-form textual descriptions, capturing the full range of possible errors.
Approach: They propose a benchmark for localizing hallucinations using LLMs with a human annotation of over 1,000 examples and a protocol to verify its quality in a humans evaluation.
Outcome: The proposed representation captures the full range of possible errors, and the best model achieves an F1 score of 0.67.
Out-of-Context Reasoning in Large Language Models (2025.findings-emnlp)

Copied to clipboard

Challenge: a lightweight technique trains only new token embeddings on axioms and evaluates them on unseen tasks.
Approach: They propose a lightweight technique that trains only new token embeddings on axioms . they train only new embeddables and evaluate them on unseen tasks .
Outcome: The proposed technique trains only new token embeddings on axioms and evaluates them on unseen tasks.
NEWTON: Are Large Language Models Capable of Physical Reasoning? (2023.findings-emnlp)

Copied to clipboard

Challenge: Large Language Models have been shown to encapsulate syntactic, semantic, word sense, and common-sense knowledge, but limited exploration of their physical reasoning abilities has been conducted.
Approach: They propose a repository and benchmark to evaluate LLMs' physical reasoning skills . they use a pipeline to generate a variant of the benchmark customized to the objects and attributes relevant for their application.
Outcome: The proposed benchmark examines the reasoning capabilities of language models across reasoning tasks.
Tiered Reasoning for Intuitive Physics: Toward Verifiable Commonsense Language Understanding (2021.findings-emnlp)

Copied to clipboard

Challenge: Large-scale, pre-trained language models (LMs) have achieved human-level performance on a breadth of language understanding tasks.
Approach: They propose a commonsense reasoning dataset with dense annotations that allows multi-tiered evaluation of machines’ reasoning process.
Outcome: The proposed model can achieve high end performance but struggle to support predictions with valid supporting evidence.
How Well Do Large Language Models Truly Ground? (2024.naacl-long)

Copied to clipboard

Challenge: Existing research defines “grounding” as having the correct answer, which does not ensure the reliability of the entire response.
Approach: They propose a stricter definition of grounding: fully utilizes the necessary knowledge from the provided context and stays within the limits of that knowledge.
Outcome: The proposed model can be ground on external contexts and maintain its correct answer.
Code over Words: Overcoming Semantic Inertia via Code-Grounded Reasoning (2026.findings-acl)

Copied to clipboard

Challenge: Large Language Models struggle with semantic inertia, a problem that is often attributed to natural language encoding, which entangles descriptive semantics and logical rules, leading to persistent hallucinations of familiar physics despite explicit contradictory rules.
Approach: They propose a framework that decouples logical dynamics from visual priors via amortized theory induction and counterfactual contrastive alignment.
Outcome: The proposed framework outperforms expensive inference-time search methods in both efficiency and accuracy.
STEM-POM: Evaluating Language Models Math-Symbol Reasoning in Document Parsing (2025.findings-acl)

Copied to clipboard

Challenge: Advances in large language models have spurred research into enhancing their reasoning capabilities, particularly in math-rich STEM documents.
Approach: They propose a benchmark dataset to evaluate LLMs’ reasoning abilities on math symbols within contextual scientific text.
Outcome: The proposed dataset demonstrates that state-of-the-art LLMs achieve an average accuracy of 20-60% under in-context learning and 50-60% with fine-tuning, highlighting a substantial gap in their ability to classify mathematical symbols.
Can Large Language Models Understand Context? (2024.findings-eacl)

Copied to clipboard

Challenge: Existing evaluation methodologies for Large Language Models (LLMs) have been inadequate to evaluate their ability to understand contextual features.
Approach: They propose a benchmark to assess large language models' ability to understand context by adapting existing datasets to suit their evaluation.
Outcome: The proposed model performs better under the in-context learning pretraining scenario than state-of-the-art models.
PhysNLU: A Language Resource for Evaluating Natural Language Understanding and Explanation Coherence in Physics (2022.lrec-1)

Copied to clipboard

Challenge: physicists use mathematics to reason and explain, separates their field from other disciplines, including mathematics.
Approach: They present a dataset to evaluate the performance of language models in physics . they find that language models are challenged by coherence related tasks in physicists .
Outcome: The proposed models are able to perform well on coherence-related tasks even when trained on natural language objectives.
Born Pragmatic, Trained to Hallucinate? Quantifying the Origins of Contextual Bias in LLMs via the PaCE Benchmark (2026.findings-acl)

Copied to clipboard

Challenge: Large language models excel at capturing communicative intent, but they have a side effect: pragmatic hallucination.
Approach: They propose a benchmark to quantify the impact of pragmatic hallucination on large language models . they propose RLHF and SFT to induce a strong tendency for pragmatic over-attribution .
Outcome: The proposed model outperforms existing models in predicting pragmatic hallucinations . the evaluations show that current alignment paradigms lack precise control over pragmatic boundaries .

What is GenGO?

GenGO is an NLP powered publication search system. It currenctly indexes 30k+ papers from ACL Anthology, and implements multi-aspect summarization, semantic search, and more!

Information

About
Limitations