Pith. sign in

REVIEW 3 major objections 3 minor 30 references

Why Are You Wrong? Counterfactual Explanations for Language Grounding with 3D Objects

T0 review · 3 major / 3 minor · reviewed 2026-08-15 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read When a model misreads a 3D object description, a one-word change can flip its answer and show why it erred.

desk verdict A genuinely new XAI application with solid valid-counterfactual success rates, but the semantic-similarity claim is contradicted by the paper's own LLM evaluation and needs revision. read the letter →

arxiv 2505.06030 v1 pith:R2HWL5P6 submitted 2025-05-09 cs.AI cs.CVcs.LG

Tobias Preintner , Weixuan Yuan , Qi Huang , Adrian König , Thomas Bäck , Elena Raponi , Niki van Stein This is my paper · ORCID
classification cs.AIcs.CVcs.LG
keywords counterfactualexplanationsobjectreferentidentificationlanguagegrounding3DobjectsgeneticalgorithmexplainableAIShapeTalkdatasetsemanticsimilarity
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper sets out to answer, for 3D object referent identification, the question 'Why did the model pick the wrong object?' Its answer: generate a counterfactual utterance—a minimally altered version of the original text description—that makes the model choose the correct object. The paper claims that a sampling-and-genetic-algorithm search over content-word substitutions can find such counterfactuals while keeping the utterance semantically close to the original. Across three different shape-backbone models and the ShapeTalk dataset, the method achieves high success rates in flipping mispredictions, and the LLM-based context-aware samplers produce counterfactuals that are both more grammatically sound and more semantically similar to the original. If the method works as claimed, it gives practitioners a concrete, human-readable explanation of model failure and a recipe for rewording instructions to robots or design tools.

What carries the argument

The load-bearing mechanism is the counterfactual utterance generator: a population of candidate utterances is initialized by sampling a single mutable content word (NVAA) and replacing it using one of four strategies, then a genetic algorithm with tournament selection, single-point crossover, and word-level mutation optimizes a fitness score that rewards flipping the model's class decision (with a penalty for invalid counterfactuals) plus the cosine similarity between Universal Sentence Encoder embeddings of original and candidate. The key design choice is restricting mutations to content words so sentence structure is preserved, and using a context-aware LLM sampler that proposes replacement words aware of the full utterance.

What would settle it

A concrete check: take the 1,000 misclassified ShapeTalk samples used in the paper, and for each test whether any single-word substitution (from the full vocabulary or from LLM suggestions) can flip the model's prediction while keeping the counterfactual's embedding-based cosine similarity above a chosen threshold. If a substantial fraction of samples have no such word and the genetic algorithm's success rate is carried only by easier samples, the central claim would be limited. Alternatively, run the method on a harder dataset with longer, multi-clause descriptions; if the success rate drops sharply, the single-word assumption is the bottleneck.

Watch

Extended reading notes

Core claim

The paper's central claim is that misclassifications in object referent identification can be explained by finding a nearby utterance, differing by one content word (noun, verb, adjective, adverb), that flips the model's prediction from distractor to target. The authors show that their combined sampler-plus-genetic-algorithm procedure produces such valid counterfactuals for the large majority of misclassified samples across three model architectures, and that the resulting counterfactuals keep high embedding-based semantic similarity to the original. The method also exposes systematic biases—for instance, that models link 'small' to horizontal thinness while 'short' is tied to height—and pinpoints weak description words. The paper further claims that replacing words with LLM-suggested context-appropriate alternatives yields counterfactuals that are more grammatically acceptable and semantically faithful than synonym-based or type-preserving random replacements.

Load-bearing premise

The method assumes that replacing a single content word in the original utterance is enough to traverse the model's decision boundary from the wrong object to the right one while keeping the sentence meaningful; if the correct decision requires several coordinated word changes or a syntactic restructuring, the search cannot find a valid, faithful counterfactual.

Editorial extensions

If this is right

  • Practitioners can reword a failed instruction by a single word and the robot or design tool will pick the right object.
  • The generated counterfactuals expose which words the model relies on and which it ignores, allowing engineers to debug and retrain the identifier.
  • The same method could be applied to other grounding tasks beyond pairs of objects, such as selecting among many objects or interpreting spatial-relation descriptions.
  • Using LLM-suggested replacements during search yields counterfactuals that are more natural to human readers, increasing trust in the explanation.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The success rate depends on the assumed single-word edit budget; on harder, multi-word or syntactically ambiguous misclassifications the method may fail, and a valid counterfactual would require editing several coordinated words.
  • The embedding-similarity objective could be replaced by a human-aligned or task-level semantic metric to avoid under-penalizing contextually unrelated insertions, a limitation the paper itself notes.
  • The counterfactual utterances could be used as data augmentation for training; the observed biases suggest that training on such counterfactuals might make models more robust to rephrasings of object descriptions.
  • Because the method is model-agnostic and only queries class probabilities, it could transfer to other black-box language-vision models without retraining.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 3 minor

Summary. The paper proposes a method for generating counterfactual utterances for object referent identification with 3D objects. Given a misclassified sample consisting of two objects and a text description, the method first samples alternative utterances by replacing a single content word using one of four strategies (unaware, word-type-aware, word-aware, or context-aware via small LLMs), then applies a genetic algorithm whose fitness combines a class-flip objective and a USE-cosine similarity objective. Experiments on 1,000 ShapeTalk misclassified samples with three backbone models report high success rates (94.5-100%), normalized Levenshtein distances around 0.2-0.3, and use GPT-4o-mini to rate grammaticality and semantic similarity. The paper claims that the resulting counterfactuals are valid, semantically similar, and meaningful, and that context-aware LLM sampling is superior.

Significance. The problem is timely and the paper is the first to address counterfactual explanations for 3D object language grounding. The method is fully described in Algorithm 1, uses reproducible components, and the code is provided; the success-rate results are strong and consistently above the random-search baseline. However, the central semantic-similarity claim is not supported by the paper's own independent evaluation, and the headline similarity metric is also the optimized objective. As presented, the contribution is better described as a method that produces valid counterfactuals with mixed semantic fidelity, not as a method whose counterfactuals are 'semantically similar and meaningful' as stated in the abstract.

major comments (3)
  1. [Table I and Fig. 3; Abstract; Section V] The central claim that the generated counterfactuals are 'semantically similar and meaningful' is contradicted by the paper's own GPT-4o-mini evaluation. For the best-performing context-aware configuration (ViT+Qwen), only 3.7% of counterfactuals are rated 'equivalent' and 29.4% 'very similar', while 53.2% are 'dissimilar' and 6.2% 'very dissimilar'; across nearly all model/strategy rows in Table I, 'dissimilar' is the modal category, and 'similar or better' totals are around 39% or less. Since this is the paper's own evaluation in Section IV.B, not an external disagreement about the meaning of 'similar', the abstract's unqualified claim is not supported by the presented evidence.
  2. [Section III.C (Eq. (2), Eq. (3)) and Section IV.B (Table I)] The USE cosine similarity in Eq. (2) is one of the two objectives optimized in Eq. (3), yet the same quantity is reported as the primary 'Cosine Similarity' metric in Table I and plotted in Fig. 2. A high cosine score is therefore partly a property of the optimizer, not independent evidence of semantic preservation. The paper's own limitation statement in Section V explicitly concedes that this metric 'under-penalize[s] the insertion of contextually unrelated words,' which is exactly the gap exposed by the GPT-4o-mini evaluation. The evaluation should either separate the optimization objective from the reported metric or justify why the reported cosine similarity is not circular.
  3. [Section III.C vs. Section IV.B] The paper does not consistently define which counterfactual utterance is analyzed. Section III.C states that 'we will always refer to a single counterfactual per sample, defined as the one with the highest cosine similarity,' while Section IV.B's GPT-4o-mini prompt evaluates 'the counterfactual utterance that achieves the best fitness value during the optimization process.' These selection rules can name different utterances, and Section IV.B defines success rate as the fraction of samples with at least one valid counterfactual in the population. As a result, the high success rates and the semantic-similarity distributions in Table I may characterize different utterances than the one a user would receive as the explanation. Please align the selection rule and report both the existence-based and the reported-utterance-based statistics.
minor comments (3)
  1. [Abstract] The last sentence contains a typo: 'Theses insights' should be 'These insights'.
  2. [Section IV.B] The evaluation prompt lists the similarity scale as 'equivalent, very similar, similar, neutral, dissimilar, very dissimilar, or unrelated,' but Table I reports only 'Very dissimilar' and not 'unrelated.' Please clarify whether 'unrelated' is a separate category and how it was mapped in Table I.
  3. [Fig. 4] The word clouds lack a numeric scale or counts, making the statement that 'the larger a word appears, the more frequently it is replaced or inserted' difficult to verify; adding counts or a size legend would improve reproducibility.

Circularity Check

2 steps flagged · score 6.0 of 10

USE cosine is both the optimized objective and a reported evaluation metric; the structure-preservation claim holds by construction.

  1. fitted input called prediction [Section IV.B (Evaluation Metrics), with Eq. (2) and Eq. (3) in Section III.C, and Table I]
    "Semantic similarity between two sentences can be measured by evaluating their embedding-based similarity. A popular choice is the cosine similarity of the USE [6] embeddings. This similarity score is also used in the optimization objective."

    Eq. (2) defines Sim(u,u') as the USE cosine similarity, and Eq. (3) sets Fitness = Classflip(O,u') + Sim(u,u'). The GA is therefore explicitly selecting counterfactuals that maximize USE cosine. Table I then reports 'Cosine Similarity' as a quantitative evaluation metric for the same counterfactuals. The high cosine values are partly a property of the optimizer, not an independent confirmation of semantic closeness. The paper's own Conclusion admits the metric 'under-penalize the insertion of contextually unrelated words as long as semantical key parts of the utterance are retained,' which is exactly the gap that the separate LLM evaluation was intended to fill.

  2. self definitional [Section III.B (Sampling) and Abstract]
    "we restrict the mutable word types to content words, such as nouns, verbs, adjectives, and adverbs (NV AA), to maintain the core sentence structure ... Our counterfactual examples maintain the structure of the original description"

    The method defines counterfactuals as single-word replacements of NVAA content words, leaving function words and word order untouched. Every utterance produced by sampling or by the GA therefore retains the original sentence skeleton by construction. The Abstract and Conclusion present 'maintain the structure of the original description' as an empirical property, but it is a direct restatement of the mutable-word restriction introduced in Section III.B.

full rationale

The class-flip objective is genuinely external: it is computed directly from the identification model's predicted probabilities, so the reported success rates are not circular. The paper also provides an independent GPT-4o-mini evaluation of semantic similarity, which is a real check beyond the optimized metric. However, the USE cosine-similarity column in Table I is not independent: Eq. (3) optimizes exactly that quantity, and Section IV.B reports it as an evaluation metric, so high values are partly optimizer artifacts. The paper's own limitation statement acknowledges that this metric under-penalizes contextually unrelated insertions. Additionally, the claim that counterfactuals 'maintain the structure of the original description' is true by construction because Section III.B restricts mutations to NVAA content words with single-word replacement. These two reductions make part of the semantic-similarity and structure claims circular, while the validity and LLM-judged aspects remain externally grounded.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

The central method rests on hand-chosen GA hyperparameters and on assumptions that single-word edits preserve meaning, that USE cosine similarity measures semantic closeness, and that an LLM judge can stand in for humans. No new physical or formal entities are introduced.

free parameters (4)
  • population_size_N = 20
    Initial population of counterfactual utterances per sample in the GA; chosen by hand, not tuned (Section IV.A).
  • generations_M = 30
    Number of GA generations per sample; chosen by hand (Section IV.A).
  • mutation_rate_p = 10%
    Mutation probability in GA; chosen to balance exploration and exploitation (Section IV.A).
  • evaluation_runs = 5
    Number of repeated GPT-4o-mini evaluation rounds with majority vote; chosen without stated justification (Section IV.B).
assumptions (4)
  • domain assumption Replacing a single NVAA content word preserves utterance meaning and can flip the model decision
    The method mutates one content word at a time and the GA crossover combines such single-word edits; if multiple coordinated edits are needed, the search space is incomplete (Section III.B).
  • domain assumption USE cosine similarity is an adequate proxy for semantic similarity for optimization
    The fitness function maximizes USE cosine similarity with no other meaning-preserving constraint; this choice affects all generated counterfactuals (Section III.C).
  • domain assumption GPT-4o-mini is an acceptable human surrogate for scoring grammaticality and semantic similarity
    The paper uses an LLM evaluator rather than human raters, citing prior validation; this is a stated modeling choice (Section IV.B).
  • domain assumption The three listener models and the binary target/distractor setup represent object referent identification
    Experiments use only pairwise comparisons with two objects and three backbone model variants (Section IV.A).

how reviews work

0 comments
Cite this review

Pith. "Pith review of Why Are You Wrong? Counterfactual Explanations for Language Grounding with 3D Objects." pith.science (2026). https://pith.science/paper/R2HWL5P6

@misc{pith2026250506030,
  author       = {Pith},
  title        = {Pith review of: Why Are You Wrong? Counterfactual Explanations for Language Grounding with 3D Objects},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/R2HWL5P6}},
  note         = {Machine review of arXiv:2505.06030}
}
read the original abstract

Combining natural language and geometric shapes is an emerging research area with multiple applications in robotics and language-assisted design. A crucial task in this domain is object referent identification, which involves selecting a 3D object given a textual description of the target. Variability in language descriptions and spatial relationships of 3D objects makes this a complex task, increasing the need to better understand the behavior of neural network models in this domain. However, limited research has been conducted in this area. Specifically, when a model makes an incorrect prediction despite being provided with a seemingly correct object description, practitioners are left wondering: "Why is the model wrong?". In this work, we present a method answering this question by generating counterfactual examples. Our method takes a misclassified sample, which includes two objects and a text description, and generates an alternative yet similar formulation that would have resulted in a correct prediction by the model. We have evaluated our approach with data from the ShapeTalk dataset along with three distinct models. Our counterfactual examples maintain the structure of the original description, are semantically similar and meaningful. They reveal weaknesses in the description, model bias and enhance the understanding of the models behavior. Theses insights help practitioners to better interact with systems as well as engineers to improve models.

Figures

Figures reproduced from arXiv: 2505.06030 by the authors.

Figure 1
Figure 1. Visual illustration of our method for explaining misclassified samples in object referent identification—the task of [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Left: Success rate as ratio of samples for which at least one valid counterfactual utterance is generated. Right: Average [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Majority vote results obtained from five repetitive but independent evaluation runs using GPT-4o-mini [20] as a human [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Commonly replaced and inserted words by the Qwen sampler to generate counterfactuals for the ViT model. [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: Qualitative results of our method. Given the original utterance, the model misclassifies the left object as the target. [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

30 extracted references · 27 canonical work pages

  1. [1]

    Learning representations and generative models for 3d point clouds

    Panos Achlioptas et al. “Learning representations and generative models for 3d point clouds”. In: Interna- tional conference on machine learning . PMLR. 2018, pp. 40–49

    work page 2018

  2. [2]

    ShapeGlot: Learning lan- guage for shape differentiation

    Panos Achlioptas et al. “ShapeGlot: Learning lan- guage for shape differentiation”. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. 2019, pp. 8938–8947

    work page 2019

  3. [3]

    ShapeTalk: A language dataset and framework for 3d shape edits and deformations

    Panos Achlioptas et al. “ShapeTalk: A language dataset and framework for 3d shape edits and deformations”. In: Proceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition. 2023, pp. 12685– 12694

    work page 2023

  4. [4]

    Towards llm-guided causal explainability for black-box text classifiers

    Amrita Bhattacharjee et al. “Towards llm-guided causal explainability for black-box text classifiers”. In: AAAI 2024 Workshop on Responsible Language Models, Van- couver, BC, Canada. 2024

    work page 2024

  5. [5]

    Zero-shot LLM-guided Counterfactual Generation: A Case Study on NLP Model Evaluation

    Amrita Bhattacharjee et al. “Zero-shot LLM-guided Counterfactual Generation: A Case Study on NLP Model Evaluation”. In: 2024 IEEE International Con- ference on Big Data (BigData) (2024), pp. 1243–1248

    work page 2024

  6. [6]

    Universal sentence encoder for En- glish

    Daniel Cer et al. “Universal sentence encoder for En- glish”. In: Proceedings of the 2018 conference on em- pirical methods in natural language processing: system demonstrations. 2018, pp. 169–174

    work page 2018

  7. [7]

    Can Large Lan- guage Models Be an Alternative to Human Evalua- tions?

    Cheng-Han Chiang and Hung-yi Lee. “Can Large Lan- guage Models Be an Alternative to Human Evalua- tions?” In: Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) . Association for Computational Lin- guistics, July 2023, pp. 15607–15631

    work page 2023

  8. [8]

    V oxel-informed Language Grounding

    Rodolfo Corona et al. “V oxel-informed Language Grounding”. In: Annual Meeting of the Association for Computational Linguistics. 2022

    work page 2022

Show all 30 references
  1. [9]

    Multi-objective counterfactual explanations

    Susanne Dandl et al. “Multi-objective counterfactual explanations”. In: International conference on parallel problem solving from nature . Springer. 2020, pp. 448– 469

    2020

  2. [10]

    The llama 3 herd of models

    Abhimanyu Dubey et al. “The llama 3 herd of models”. In: arXiv preprint arXiv:2407.21783 (2024)

    2024 arXiv

  3. [11]

    Deep residual learning for image recognition

    Kaiming He et al. “Deep residual learning for image recognition”. In: Proceedings of the IEEE conference on computer vision and pattern recognition. 2016, pp. 770– 778

    2016

  4. [12]

    Text classification algorithms: A survey

    Kamran Kowsari et al. “Text classification algorithms: A survey”. In: Information 10.4 (2019), p. 150

    2019

  5. [13]

    Generalized inverse classifica- tion

    Michael T Lash et al. “Generalized inverse classifica- tion”. In: Proceedings of the 2017 SIAM International Conference on Data Mining. SIAM. 2017, pp. 162–170

    2017

  6. [14]

    Binary codes capable of correcting deletions, insertions, and reversals

    VI Levenshtein. “Binary codes capable of correcting deletions, insertions, and reversals”. In: Proceedings of the Soviet physics doklady (1966)

    1966

  7. [15]

    Prompting Large Language Models for Counterfactual Generation: An Empirical Study

    Yongqi Li et al. “Prompting Large Language Models for Counterfactual Generation: An Empirical Study”. In: International Conference on Language Resources and Evaluation. 2023

    2023

  8. [16]

    G-Eval: NLG Evaluation using Gpt- 4 with Better Human Alignment

    Yang Liu et al. “G-Eval: NLG Evaluation using Gpt- 4 with Better Human Alignment”. In: Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing . Association for Computational Linguistics, Dec. 2023, pp. 2511–2522

    2023

  9. [17]

    Genetic algorithms+ data structures= evolution programs

    Zbigniew Michalewicz. Genetic algorithms+ data structures= evolution programs . Springer Science & Business Media, 2013

    2013

  10. [18]

    Which One? Leveraging Con- text Between Objects and Multiple Views for Language Grounding

    Chancharik Mitra et al. “Which One? Leveraging Con- text Between Objects and Multiple Views for Language Grounding”. In: Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technolo- gies (Volume 1: ...

    2024

  11. [19]

    Explaining machine learning classifiers through diverse counterfactual explanations

    Ramaravind K Mothilal, Amit Sharma, and Chenhao Tan. “Explaining machine learning classifiers through diverse counterfactual explanations”. In: Proceedings of the 2020 conference on fairness, accountability, and transparency. 2020, pp. 607–617

    2020

  12. [20]

    GPT-4o mini

    OpenAI. GPT-4o mini. https://openai.com/index/gpt- 4o-mini-advancing-cost-efficient-intelligence/. 2024

    2024

  13. [21]

    Learning Transferable Visual Mod- els From Natural Language Supervision

    Alec Radford et al. “Learning Transferable Visual Mod- els From Natural Language Supervision”. In: ICML. 2021

    2021

  14. [22]

    GeCo: quality counterfac- tual explanations in real time

    Maximilian Schleich et al. “GeCo: quality counterfac- tual explanations in real time”. In: Proceedings of the VLDB Endowment 14.9 (2021), pp. 1681–1693

    2021

  15. [23]

    CERTIFAI: A common framework to provide explana- tions and analyse the fairness and robustness of black- box models

    Shubham Sharma, Jette Henderson, and Joydeep Ghosh. “CERTIFAI: A common framework to provide explana- tions and analyse the fairness and robustness of black- box models”. In: Proceedings of the AAAI/ACM Con- ference on AI, Ethics, and Society . 2020, pp. 166–172

    2020

  16. [24]

    ShapeWalk: Compositional Shape Editing Through Language- Guided Chains

    Habib Slim and Mohamed Elhoseiny. “ShapeWalk: Compositional Shape Editing Through Language- Guided Chains”. In: Proceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition . 2024, pp. 22574–22583

    2024

  17. [25]

    Multi-Task Domain Adaptation for Language Grounding with 3D Objects

    Penglei Sun et al. “Multi-Task Domain Adaptation for Language Grounding with 3D Objects”. In: Euro- pean Conference on Computer Vision . Springer. 2024, pp. 387–404

    2024

  18. [26]

    The Falcon 3 Family of Open Models

    Falcon-LLM Team. The Falcon 3 Family of Open Models. Dec. 2024. URL: https://huggingface.co/blog/ falcon3

    2024

  19. [27]

    Qwen2.5: A Party of Foundation Models

    Qwen Team. Qwen2.5: A Party of Foundation Models . Sept. 2024. URL: https://qwenlm.github.io/blog/qwen2. 5/

    2024

  20. [28]

    Language grounding with 3d objects

    Jesse Thomason et al. “Language grounding with 3d objects”. In: Conference on Robot Learning . PMLR. 2022, pp. 1691–1701

    2022

  21. [29]

    Counterfactual explanations for machine learning: A review

    Sahil Verma, John Dickerson, and Keegan Hines. “Counterfactual explanations for machine learning: A review”. In: arXiv preprint arXiv:2010.10596 2 (2020), p. 1

    2020 arXiv

  22. [30]

    Counterfactual explanations without opening the black box: Automated decisions and the GDPR

    Sandra Wachter, Brent Mittelstadt, and Chris Russell. “Counterfactual explanations without opening the black box: Automated decisions and the GDPR”. In: Harv. JL & Tech. 31 (2017), p. 841

    2017

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.