REVIEW 4 major objections 4 minor 25 references
Do LLMs Understand Ambiguity in Text? A Case Study in Open-world Question Answering
T0 review · 4 major / 4 minor · reviewed 2026-08-12 · deepseek-v4-flash
Pith's one-line read Adding model-generated context or rewriting an ambiguous question with "what" improves how off-the-shelf LLMs answer ambiguous open-domain questions; small-scale fine-tuning and lower temperature do not.
desk verdict The disambiguation gains vanish once you look at the paper's own baseline variability; still a useful prompt-engineering datapoint, but the evaluation needs rigor. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
The mechanism is a two-pass prompting pipeline. In pass one, the LLM rewrites the ambiguous question either to begin with "what" (a rephrasing prompt) or to include a generated block of extra context (a contextual-enrichment prompt); in pass two, the modified question is fed back to the LLM for a concise answer. The paper's evaluation machinery is cosine similarity between text-embedding-3-large embeddings of the model's answer and the dataset's ground-truth answer, supplemented by similarity measures between the rewritten question and the human-provided disambiguated question (question coherence) and between naive and disambiguated answers (naive answer overlap). The ground-truth disambiguated questions serve as an upper-bound reference for how much improvement is possible.
What would settle it
Score the same 1,000-question outputs with exact string match (or human annotation) instead of embedding cosine similarity; if the context-enriched and "what" conditions do not beat the naive condition on that stricter measure, the paper's central claim is refuted. A faster check is to compute bootstrap confidence intervals for the reported GT Answer Overlap means—overlapping intervals would show the gains are indistinguishable from noise.
Extended reading notes
Core claim
The paper's central claim is that explicit, training-free disambiguation at the prompt level improves LLM accuracy on ambiguous open-domain questions. In their experiments, adding context to the ambiguous question raised the mean cosine similarity between the model's answer and the ground-truth answer from 0.759 to 0.789 for GPT-4o and from 0.692 to 0.710 for GPT-4o-mini; rephrasing with "what" produced smaller gains. The same experiments show that a 50-example fine-tune of GPT-4o-mini actually lowered the score (0.643 vs. 0.626 for the base model under naive prompting), and that temperature 0.2 versus 1.0 made no meaningful difference. The authors conclude that even though LLMs often add wrong context when left to their own devices, the contexts they add on questions humans could disambiguate align with ground truth, so simple prompting strategies are a promising direction for reducing ambiguity-driven errors.
Load-bearing premise
The results are only as valid as the assumption that cosine similarity between the model's answer embedding and the ground-truth answer embedding measures answer correctness; the paper does not validate that metric against exact matches or human judgment, and the reported improvements are small mean differences without significance testing.
Editorial extensions
If this is right
- A simple preprocessing step that rewrites user questions or adds context could improve open-domain QA accuracy without retraining, which matters for deployed systems where users rarely disambiguate.
- Context enrichment is a high-variance strategy: it helps when the model happens to add relevant context and is held back when it adds irrelevant context, so systems using it need a way to detect when the added context is off-target.
- Small-scale fine-tuning on ambiguous QA pairs is not an obviously better use of resources than prompt-level disambiguation, at least at the 50-example scale tested.
- Lowering the temperature is not an effective ambiguity remedy; the model's misinterpretation persists under more deterministic decoding.
Reading between the lines
- My inference: the reported differences are small on a 0–1 cosine scale and the paper reports no confidence intervals or significance tests, so the practical size of the effect is uncertain even if the direction is real.
- My inference: because the evaluation metric is embedding similarity, a stricter exact-match test might rank the methods differently; the paper's "accuracy" language goes beyond what cosine similarity alone establishes.
- My inference: the fine-tuning negative result applies only to a 50-example, single-configuration tune; larger or more targeted fine-tuning could still learn disambiguation, so the paper's training-free conclusion is scale-limited.
- My inference: the finding that context enrichment works better exactly when the human disambiguator also found the correct reading suggests a testable extension—filtering generated context for entity-bearing phrases before answering could improve the hit rate.
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The paper studies whether GPT-4o and GPT-4o-mini can handle ambiguous open-domain questions, comparing naive prompting with two prompt-level disambiguation methods (rephrasing via 'what' and adding contextual enrichment) on 1,000 randomly sampled AmbigQA questions. It reports gains in 'GT Answer Overlap' (cosine similarity between the model answer embedding and the ground-truth answer embedding) for both methods, a small-scale fine-tuning experiment on 50 examples, and a temperature ablation. The central claim is that simple, training-free, prompt-based disambiguation improves LLM performance on ambiguous question answering.
Significance. If the central claim were statistically supported, the paper would offer a low-cost, practical intervention for ambiguous QA and a useful comparison of disambiguation strategies. The paper has clear strengths: it uses a standard public dataset (AmbigQA), reports explicit prompt templates, evaluates two commercially relevant models, and includes an honest limitations section. However, the empirical support is currently insufficient: the reported effect sizes are small, no uncertainty quantification is provided, the evaluation metric is not validated as a measure of answer correctness, and the prompt-selection procedure risks circularity. The central finding is plausible but not established by the reported numbers.
major comments (4)
- [Section V, RQ2 (with Table II)] The paper's own data show that the naive GPT-4o-mini baseline is not stable at the scale of the claimed improvements. Table II reports a naive GT Answer Overlap of 0.692, while the RQ2 evaluation of the same naive condition on a fresh 1,000-question sample reports 0.643; the 0.049 gap is larger than the largest reported disambiguation gain (0.030 for GPT-4o and 0.018 for GPT-4o-mini). No confidence intervals, standard errors, or paired significance tests are reported anywhere. As presented, the headline claim that disambiguation improves accuracy cannot be distinguished from sampling variability. Please provide paired bootstrap confidence intervals or significance tests, report the number of independent API runs, and state how the random samples were drawn.
- [Section IV-D] The evaluation metric is an uncalibrated cosine similarity between text-embedding-3-large embeddings of the model answer and the human ground-truth answer. The paper does not show that a 0.01-0.03 cosine shift corresponds to a meaningful correctness improvement; no threshold, calibration against exact match, or human judgment is provided. Because all conclusions rest on this metric, please validate it on a labeled sample (e.g., by comparing cosine similarity against human-rated correctness or exact-match scores) and, if possible, report results under an additional metric such as exact match or token-level F1.
- [Section IV-C] The disambiguation prompts were chosen after 'experimenting with a variety of prompts' with no description of a held-out validation split. If the same 1,000-question sample was used both to select the prompts and to report the results, the reported improvements are at least partially an artifact of selection on the evaluation set. Please describe the prompt-search procedure, use a separate validation set for prompt selection, and report results on a held-out test sample.
- [Section V, RQ2] The fine-tuning conclusion that 'fine-tuning ... does not provide any improvement' is based on a comparison of 0.643 (naive 4o-mini) versus 0.626 (fine-tuned 4o-mini) on a fresh sample of 1,000 questions. This 0.017 difference is comparable in magnitude to the disambiguation gains the paper highlights, and no error bars or significance tests are reported for it. In addition, the comparison may be confounded by sampling a different set of questions for the fine-tuned evaluation. Please treat this negative result with the same statistical rigor as the main comparisons.
minor comments (4)
- [Abstract and Section IV-C] The methods are described as 'token-level disambiguation,' but both prompts operate on whole questions or on appended context passages; please correct this terminology to avoid overstating the granularity of the interventions.
- [Section V, 'Problem with naive contextual enrichment'] The text says 'why the average is not going up when an LLM is prompted to insert context into a question,' but Tables I and II report that context disambiguation gives the highest GT Answer Overlap (0.789 and 0.710). This subsection appears to refer to the distributions in Figures 2 and 3 rather than the means in the tables; please rewrite to avoid the apparent contradiction.
- [Section V, RQ2] The fine-tuning data format example is not valid JSON: the 'messages' value is not shown as an array and the assistant content string has mismatched quotes. If this is meant to be illustrative, please correct it or mark it as pseudocode.
- [Tables I and II] The 'Upper-bound (via GT disamb. questions)' column is not defined in the text; please explain how this upper bound is computed (for example, whether it averages over all human disambiguations for each question) and why it constitutes an upper bound.
Circularity Check
The reported disambiguation gains are partly fitted: the two prompts were selected after experimenting on the same 1,000-question sample later used for evaluation, so the central 'improvement' is the selection criterion rather than an independent prediction.
-
fitted input called prediction
[Section IV.C (Disambiguation Methods); results in Section V RQ1 and Tables I-II]
"We experimented with a variety of prompts, and finally selected two appropriate prompts for disambiguating a given question using the two GPT models. ... For our random sample of 1,000 questions in these experiments, the average length of the rephrased question was 13.01 words, which is about 1.45 times longer than the original question."
The paper reports no held-out prompt-selection split; the only sample described is the 1,000-question evaluation sample on which Tables I and II are computed. The two disambiguation prompts were retained after 'experimenting with a variety of prompts,' so the GT Answer Overlap values used as evidence (0.759 to 0.789 for GPT-4o; 0.692 to 0.710 for GPT-4o-mini) are indistinguishable from the criterion used to choose the prompts. A prompt that failed to raise overlap would have been discarded, making the reported improvement statistically forced by the selection procedure rather than an out-of-sample prediction.
full rationale
The central claim that training-free disambiguation improves ambiguous-question answering rests on prompt-selection over the same evaluation sample, with no described validation split. That is a fitted-input-called-prediction circularity and is the main reason for the score. The rest of the paper is not circular: the fine-tuning experiment (RQ2) uses a separately drawn 1,000-question sample and reports an external comparison (0.643 vs. 0.626), the temperature ablation (RQ3) is a genuine manipulation, and the upper-bound comparison with human disambiguated questions provides an external anchor. The absence of significance testing and the unvalidated cosine-similarity metric are statistical robustness concerns, not circularity, and are therefore noted only here. Because the headline improvement reduces, at least in part, to prompt selection on the evaluation data, partial circularity is present.
Assumptions & free parameters
free parameters (2)
- Rephrase via 'what' prompt template =
Rewrite this question replacing all questions with a what, but retain the meaning by specifying what entity or what…
- Contextual enrichment prompt template =
Add extra information to the following question.
assumptions (3)
- domain assumption Cosine similarity between text-embedding-3-large embeddings is a valid proxy for answer correctness
- domain assumption The random sample of 1,000 questions from AmbigQA train is representative of ambiguous open-domain QA
- domain assumption AmbigQA ground truth answers are correct and complete
Cite this review
Pith. "Pith review of Do LLMs Understand Ambiguity in Text? A Case Study in Open-world Question Answering." pith.science (2026). https://pith.science/paper/4UONWVGD
@misc{pith2026241112395,
author = {Pith},
title = {Pith review of: Do LLMs Understand Ambiguity in Text? A Case Study in Open-world Question Answering},
year = {2026},
howpublished = {\url{https://pith.science/paper/4UONWVGD}},
note = {Machine review of arXiv:2411.12395}
}
read the original abstract
Ambiguity in natural language poses significant challenges to Large Language Models (LLMs) used for open-domain question answering. LLMs often struggle with the inherent uncertainties of human communication, leading to misinterpretations, miscommunications, hallucinations, and biased responses. This significantly weakens their ability to be used for tasks like fact-checking, question answering, feature extraction, and sentiment analysis. Using open-domain question answering as a test case, we compare off-the-shelf and few-shot LLM performance, focusing on measuring the impact of explicit disambiguation strategies. We demonstrate how simple, training-free, token-level disambiguation methods may be effectively used to improve LLM performance for ambiguous question answering tasks. We empirically show our findings and discuss best practices and broader impacts regarding ambiguity in LLMs.
Figures
Reference graph
Works this paper leans on
-
[1]
56% of college students have used AI on assignments or exams: Bestcolleges,
J. Nam, “56% of college students have used AI on assignments or exams: Bestcolleges,” Nov 2023. [Online]. Available: https://www. bestcolleges.com/research/most-college-students-have-used-ai-survey/
work page 2023
- [2]
-
[3]
Siren’s song in the ai ocean: A survey on hallucination in large language models,
Y . Zhang, Y . Li, L. Cui, D. Cai, L. Liu, T. Fu, X. Huang, E. Zhao, Y . Zhang, Y . Chen, L. Wang, A. T. Luu, W. Bi, F. Shi, and S. Shi, “Siren’s song in the ai ocean: A survey on hallucination in large language models,” 2023. [Online]. Available: https://arxiv.org/abs/2309.01219
arXiv 2023
-
[4]
Attention is all you need,
A. Vaswani, “Attention is all you need,” Advances in Neural Information Processing Systems, 2017
2017
-
[5]
Sparks of artificial general intelligence: Early experiments with gpt-4,
S. Bubeck, V . Chandrasekaran, R. Eldan, J. Gehrke, E. Horvitz, E. Ka- mar, P. Lee, Y . T. Lee, Y . Li, S. Lundberg et al. , “Sparks of artificial general intelligence: Early experiments with gpt-4,” arXiv preprint arXiv:2303.12712, 2023
arXiv 2023
-
[6]
Language models are unsupervised multitask learners,
A. Radford, J. Wu, R. Child, D. Luan, D. Amodei, I. Sutskever et al. , “Language models are unsupervised multitask learners,” OpenAI blog , vol. 1, no. 8, p. 9, 2019
2019
-
[7]
Language models are few-shot learners,
T. B. Brown, “Language models are few-shot learners,” arXiv preprint arXiv:2005.14165, 2020
arXiv 2005
-
[8]
J. Achiam, S. Adler, S. Agarwal, L. Ahmad, I. Akkaya, F. L. Aleman, D. Almeida, J. Altenschmidt, S. Altman, S. Anadkat et al. , “Gpt-4 technical report,” arXiv preprint arXiv:2303.08774 , 2023
arXiv 2023
Show all 25 references
-
[9]
Llama: Open and efficient foundation language models,
H. Touvron, T. Lavril, G. Izacard, X. Martinet, M.-A. Lachaux, T. Lacroix, B. Rozi `ere, N. Goyal, E. Hambro, F. Azhar et al. , “Llama: Open and efficient foundation language models,” arXiv preprint arXiv:2302.13971, 2023
2023 arXiv
-
[10]
Llama 2: Open foundation and fine-tuned chat models,
H. Touvron, L. Martin, K. Stone, P. Albert, A. Almahairi, Y . Babaei, N. Bashlykov, S. Batra, P. Bhargava, S. Bhosale et al. , “Llama 2: Open foundation and fine-tuned chat models,” arXiv preprint arXiv:2307.09288, 2023
2023 arXiv
-
[11]
Instruction tuning for large language models: A survey,
S. Zhang, L. Dong, X. Li, S. Zhang, X. Sun, S. Wang, J. Li, R. Hu, T. Zhang, F. Wu et al. , “Instruction tuning for large language models: A survey,” arXiv preprint arXiv:2308.10792 , 2023
2023
-
[12]
Fighting fire with fire: can chatgpt detect ai-generated text?
A. Bhattacharjee and H. Liu, “Fighting fire with fire: can chatgpt detect ai-generated text?” ACM SIGKDD Explorations Newsletter , vol. 25, no. 2, pp. 14–21, 2024
2024
-
[13]
Mathbench: Evaluating the theory and application proficiency of llms with a hierarchical mathematics benchmark,
H. Liu, Z. Zheng, Y . Qiao, H. Duan, Z. Fei, F. Zhou, W. Zhang, S. Zhang, D. Lin, and K. Chen, “Mathbench: Evaluating the theory and application proficiency of llms with a hierarchical mathematics benchmark,” arXiv preprint arXiv:2405.12209, 2024
2024 arXiv
-
[14]
The butterfly effect of altering prompts: How small changes and jailbreaks affect large language model perfor- mance,
A. Salinas and F. Morstatter, “The butterfly effect of altering prompts: How small changes and jailbreaks affect large language model perfor- mance,” arXiv preprint arXiv:2401.03729 , 2024
2024 arXiv
-
[15]
Quantifying language models’ sensitivity to spurious features in prompt design or: How i learned to start worrying about prompt formatting,
M. Sclar, Y . Choi, Y . Tsvetkov, and A. Suhr, “Quantifying language models’ sensitivity to spurious features in prompt design or: How i learned to start worrying about prompt formatting,” in The Twelfth International Conference on Learning Representations , 2024. [Online]. Av...
2024
-
[16]
Task ambiguity in humans and language models,
A. Tamkin, K. Handa, A. Shrestha, and N. Goodman, “Task ambiguity in humans and language models,” arXiv preprint arXiv:2212.10711 , 2022
2022 arXiv
-
[17]
E. Davis. (na) Notes on ambiguity. [Online]. Available: https: //cs.nyu.edu/∼davise/ai/ambiguity.html
-
[18]
The winograd schema challenge,
H. Levesque, E. Davis, and L. Morgenstern, “The winograd schema challenge,” in Thirteenth international conference on the principles of knowledge representation and reasoning , 2012
2012
-
[19]
Gpt-4o system card,
OpenAI, “Gpt-4o system card,” 2024. [Online]. Available: https: //openai.com/index/gpt-4o-system-card/
2024
-
[20]
Natural questions: a benchmark for question answering research,
T. Kwiatkowski, J. Palomaki, O. Redfield, M. Collins, A. Parikh, C. Alberti, D. Epstein, I. Polosukhin, M. Kelcey, J. Devlin, K. Lee, K. N. Toutanova, L. Jones, M.-W. Chang, A. Dai, J. Uszkoreit, Q. Le, and S. Petrov, “Natural questions: a benchmark for question answering rese...
2019
-
[21]
AmbigQA: Answering ambiguous open-domain questions,
S. Min, J. Michael, H. Hajishirzi, and L. Zettlemoyer, “AmbigQA: Answering ambiguous open-domain questions,” in EMNLP, 2020
2020
-
[22]
Scope ambiguities in large language models,
G. Kamath, S. Schuster, S. Vajjala, and S. Reddy, “Scope ambiguities in large language models,” Transactions of the Association for Compu- tational Linguistics , vol. 12, pp. 738–754, 2024
2024
-
[23]
Mistral 7b (2023),
A. Jiang, A. Sablayrolles, A. Mensch, C. Bamford, D. Chaplot, D. de las Casas, F. Bressand, G. Lengyel, G. Lample, L. Saulnier et al. , “Mistral 7b (2023),” arXiv preprint arXiv:2310.06825 , 2023
2023 arXiv
-
[24]
An empirical study of catastrophic forgetting in large language models during continual fine-tuning,
Y . Luo, Z. Yang, F. Meng, Y . Li, J. Zhou, and Y . Zhang, “An empirical study of catastrophic forgetting in large language models during continual fine-tuning,” 2024. [Online]. Available: https://arxiv.org/abs/2308.08747
2024 arXiv
-
[25]
Preserving principal subspaces to reduce catastrophic forgetting in fine-tuning,
J. K. Franke, M. Hefenbrock, and F. Hutter, “Preserving principal subspaces to reduce catastrophic forgetting in fine-tuning,” in ICLR 2024 Workshop on Mathematical and Empirical Understanding of F oundation Models, 2024. [Online]. Available: https://openreview.net/ forum?id=X...
2024
Reviewed August 12, 2026 · model on record in the stance chip above.
Discussion (0). Continue with ORCID to comment.