Pith. sign in

REVIEW 4 major objections 5 minor 29 references

NeSy-RAG: Neuro-Symbolic RAG for Explainable Question Answering

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read This paper claims that a modular neuro-symbolic RAG system, which synthesizes source-attributed Prolog modules from retrieved text chunks and uses symbolic knowledge-gap detection, can outperform a pure-LLM RAG baseline on the ShARC…

desk verdict New mechanism (0-arity predicate abstraction, knowledge-gap detection) is real, but the headline accuracy gap rests on an asymmetric simulated-user protocol and a test-set-selected k; worth a serious referee, but the empirical claim is not yet clean. read the letter →

arxiv 2608.06292 v1 pith:BNQNP2HF submitted 2026-08-06 cs.CL cs.SC

Jonas Gann , Michael Gertz This is my paper
classification cs.CLcs.SC
keywords retrieval-augmentedgenerationneuro-symbolicreasoningPrologsynthesisexplainablequestionansweringknowledge-gapdetectionShARCbenchmark0-aritypredicatespredicateretrieval
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 proposes NeSy-RAG, a retrieval-augmented question-answering system that replaces opaque LLM reasoning with executable logic. For every retrieved text chunk, an LLM writes a small Prolog module whose rules and facts are traceable to that exact source. A key abstraction turns the module's rules into 0-arity Boolean predicates, so questions can be matched to relevant logic via code embeddings and composed into a Prolog query without loading the whole knowledge base. During execution, dynamic predicates whose values depend on the user trigger follow-up questions, and if the answer depends on a fact the user cannot provide, the system returns 'more' instead of guessing. On the ShARC benchmark, without any domain-specific training, NeSy-RAG reaches 61.1% accuracy, beating a same-model RAG baseline at 42.8%, with 61% of genuinely under-determined cases correctly flagged versus 19% for the baseline.

What carries the argument

The system's load-bearing pieces are: (1) per-chunk Prolog module synthesis, in which an LLM converts a single text snippet into rules and facts that are uniquely attributable to that snippet; (2) 0-arity predicate abstraction, which flattens each module into self-contained Boolean claims such as `could_claim_married_couples_allowance`, so that query construction becomes a matter of selecting and composing a few predicates instead of searching a large knowledge base; (3) joint natural-language–code embedding retrieval (using a dedicated code-embedding model) that ranks 0-arity rules by similarity to the user question; (4) dynamic facts—Boolean predicates whose truth depends on the user—that are evaluated at inference time through a foreign function interface, triggering follow-up questions and enabling knowledge-gap detection; and (5) a module-aware execution trace that maps every rule and fact back to its source chunk. The combination of hierarchical retrieval (chunks first, then predicates) and deterministic Prolog execution is what carries the accuracy and explainability gains.

What would settle it

A random-sample audit of synthesized Prolog modules on ShARC snippets, comparing every rule and dynamic fact against the source conditions, would determine whether semantic synthesis errors, rather than retrieval or query construction, drive the gap to the upper baseline. If that audit finds frequent semantic mismatches that correlate with incorrect answers, the core synthesis assumption fails; alternatively, restricting dynamic predicate generation to only the conditions explicitly stated in the snippet and observing whether the 31% 'yes-to-more' error rate drops would test the paper's diagnosis of superfluous dynamic predicates.

Watch

Extended reading notes

Core claim

The central claim is that a modular neuro-symbolic RAG pipeline can outperform pure LLM retrieval-generation on Boolean question answering while producing fully attributable reasoning. The paper argues that by synthesizing one Prolog module per retrieved chunk, exposing each module's functionality through 0-arity rules, and retrieving those rules with a joint natural-language–code embedding, the system can construct and execute a deterministic Prolog query whose every step references the originating source text. On the ShARC test set, with no domain-specific training, this yields 61.1% accuracy against 42.8% for a same-model RAG baseline; the gain concentrates in correctly returning 'more' when user information is insufficient (61% vs 19%), which the paper attributes to its symbolic knowledge-gap detection. The authors further claim that the modular, attribution-aware design keeps execution time lower than the RAG baseline (7.4s vs 11.4s) by reusing modules and keeping query-generation context small.

Load-bearing premise

The whole pipeline rests on the assumption that an LLM given a single text chunk can produce a semantically faithful Prolog knowledge base—rules and dynamic facts that correctly encode the snippet's conditions—since the paper only measures syntactic validity (2.2% failures), not whether the synthesized logic matches the source's meaning.

Editorial extensions

If this is right

  • If the reported numbers hold, replacing the final generation step of a RAG pipeline with synthesized, attributable Prolog reasoning yields higher accuracy and lower latency than pure LLM inference on interactive Boolean QA, at least on data shaped like ShARC.
  • The 61% versus 19% 'more' classification rate implies that a symbolic trace can detect missing user context that LLMs tend to commit through; this would matter in high-stakes domains where an honest 'more' is safer than a false answer.
  • Because modules are synthesized once per chunk and reused, the per-question cost drops as chunks are reused, pointing toward a scalable design where a growing corpus adds modules without requiring re-processing.
  • The paper's error analysis suggests that over-generating dynamic predicates is the main accuracy bottleneck; tightening that generation is a direct, in-scope next step that would close part of the gap to the upper baseline.

Reading between the lines

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

  • If the synthesis assumption holds—that an LLM faithfully encodes each snippet's conditions—the approach could be extended to multi-hop reasoning by composing queries across modules; the paper leaves this as future work but its modular design makes it a natural next test.
  • The knowledge-gap mechanism might transfer to other symbolic backends, such as answer-set programming or first-order logic solvers, where the same dynamic-fact pattern could drive interaction; that is an inference, not a claim in the paper.
  • A semantic audit of synthesized Prolog modules on a sample of ShARC snippets would separate two explanations for the remaining errors: faulty logic synthesis versus retrieval noise; the paper reports only syntactic failures (2.2%) and does not measure semantic correctness.
  • Because the baseline gap is largest on 'more' instances, a version that forces LLM baselines to explicitly ask for missing facts (rather than guessing) might narrow the gap; the paper's setup suggests this is a testable hypothesis.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper proposes NeSy-RAG, a modular retrieval-augmented generation framework that synthesizes Prolog modules from retrieved text chunks, uses 0-arity predicates for constrained query construction, and includes a symbolic knowledge-gap detection mechanism that triggers follow-up questions. On the ShARC benchmark, without domain-specific training, it reports 61.1% accuracy, outperforming a same-model LLM RAG baseline at 42.8%, and claims improved detection of 'more' instances (61% vs 19%). The paper also claims lower execution time and transparent, attributable reasoning traces compared to the baseline.

Significance. If the empirical claims hold, the 0-arity predicate abstraction and symbolic gap detection would be a meaningful step toward explainable and interactive question answering, with a clean modular design that can be plugged into existing RAG systems. The paper's strengths are a concrete prototype, an external benchmark, transparent prompts in Appendix B, and an honest discussion of the dominant error pattern (over-generation of dynamic predicates). However, the experimental protocol has a serious asymmetry in the simulated-user oracle, and the absence of statistical validation and semantic fidelity analysis means the central accuracy and knowledge-gap claims are not yet established.

major comments (4)
  1. [Section 4 and Appendix B] The simulated-user prompt for NeSy-RAG ('Predicate Answering Prompt') supplies the source snippet, the initial question, and the full list of dynamic predicates, whereas the simulated-user prompt for the LLM RAG baseline ('Follow-up Question Answering Prompt') supplies only the generated follow-up question and the user context. This asymmetry gives NeSy-RAG's oracle access to the snippet, enabling it to resolve predicates that may be ambiguous or under-specified in the synthesized Prolog; the baseline's oracle cannot use the snippet to answer. The 18.3-point accuracy gap may therefore reflect a more informative oracle rather than superior reasoning. Please give both oracles the same information (e.g., context only) or provide an ablation showing the result holds without this oracle advantage.
  2. [Section 4.3, Table 2] The hyperparameter k (number of top 0-arity predicates) is selected on the test set: the paper reports accuracy for k=3, 10, and 20 and highlights k=10 as the best. This constitutes test-set tuning. In addition, no error bars, confidence intervals, or significance tests are reported for any comparison. Because all components use stochastic LLM calls, the 61.1% vs. 42.8% gap could be within run-to-run variation. Please select k on a development split, report variance over repeated runs (or bootstrap), and test the main effect statistically.
  3. [Section 4.3, 'more' analysis] The claim that correct classification of 61% of more instances provides 'strong empirical evidence' for the knowledge-gap detection mechanism is confounded by the same over-generation of dynamic predicates that causes 31% of yes instances to be misclassified as more. A high recall on more could be a byproduct of a conservative bias rather than correct detection of genuinely missing information. Please provide a finer-grained evaluation of more predictions, such as annotating whether the missing fact actually changes the query outcome, or report precision on more and an ablation that removes spurious dynamic predicates.
  4. [Sections 3.1 and 4.3] The paper reports only a 2.2% syntactic failure rate for Prolog synthesis, not semantic fidelity. Since the entire pipeline's correctness depends on the LLM encoding legal conditions correctly as rules and dynamic predicates, the absence of a semantic evaluation (e.g., comparing synthesized module answers to gold answers on a sample with full context, or manually inspecting a sample of modules) leaves the grounding of the accuracy and explainability claims unvalidated. Please add a semantic correctness measure for the synthesized modules.
minor comments (5)
  1. [Section 4.2] The 'LLM Upper Baseline' is not an upper bound for the interactive setting because it receives the complete user context upfront; consider renaming it to 'full-context baseline' to avoid confusion.
  2. [Figure 2] The normalized confusion matrices obscure class counts; consider reporting raw counts or a confusion matrix table as well, since ShARC classes are imbalanced.
  3. [Section 4.1] The statement that 'scaling the number of snippets is not expected to affect accuracy' is an assumption; please justify it or provide a small experiment varying the number of distractors.
  4. [Section 4, first paragraph] Please clarify that the ShARC 'more' label indicates insufficient information, and that the modified protocol intentionally withholds context to simulate an interactive setting.
  5. [Header, page 2] The running header reads 'Gann Gertz' rather than the author names; this appears to be a formatting artifact that should be corrected.

Circularity Check

1 steps flagged · score 2.0 of 10

No load-bearing self-citation or fitted-prediction circularity; the central claim rests on the external ShARC test split. The only quasi-circular element is an asymmetric simulated-user protocol that gives the NeSy-RAG oracle the source snippet and the full predicate list, so part of the reported accuracy and 61% 'more' recall reduces to an information leak in the evaluation harness.

  1. other [Section 4 simulated-user setup; Appendix B, 'NeSy-RAG Evaluation - Predicate Answering Prompt (Simulating User)' versus 'LLM RAG Baseline Evaluation - Follow-up Question Answering Prompt']
    "Your task is to answer a yes/no dynamic predicate about a user based on the given user context. Snippet: {{snippet}} Initial user question: {{question}} User context: {{user context}} All available dynamic predicates: {{dynamic predicates}} Answer the following predicate as true/false, or return None if the information is not stated (unknown)."

    The NeSy-RAG simulated user receives the source snippet and every generated dynamic predicate, while the baseline's simulated user receives only the follow-up question and user context. Dynamic-predicate values are the inputs to the deterministic Prolog execution, and the oracle can resolve them by reading the same snippet from which the modules were synthesized, effectively performing for free the semantic grounding that Section 3.1 attributes to the synthesis step. The paper reports only a 2.2% syntactic failure rate and never measures semantic fidelity, so the headline 61.1% vs 42.8% gap and the 61% 'more' recall are partly artifacts of an oracle with the source text rather than an independent verification of the symbolic derivation.

full rationale

No significant circularity is found in the main derivation chain. NeSy-RAG is evaluated against the external ShARC test split (8,276 entries) with gold labels that are independent of the system's components; no parameter is fitted to a target quantity and then renamed as a prediction, and no claim reduces to a fitted constant by construction. The reference list contains no works by the present authors, so the self-citation patterns (load-bearing self-citation, uniqueness imported from authors, ansatz smuggled via citation) do not apply; external resources such as the jina-code-embeddings-1.5b model and OpenAI/Ollama backends are cited as independent implementations. The framework-level claims of attribution and traceability are design properties built into the module system, not empirical predictions, and the ShARC results are therefore measured rather than forced. Two evaluation-protocol concerns do not rise to circularity but warrant a score of 2: (1) the choice of k=10 on the test split is mild test-set tuning, and the paper itself reports only a small accuracy drop for k=3 and k=20; and (2) the simulated-user prompt for NeSy-RAG includes the source snippet and all dynamic predicates, while the LLM RAG baseline's simulated user sees only the question and context, creating an information asymmetry that can inflate the measured 'more' detection and overall accuracy without testing semantic fidelity of the synthesized Prolog. The paper explicitly admits a related limitation in Section 5: the dominant error pattern is 31% of yes instances classified as more due to superfluous dynamic predicates, which is internal evidence that the system is not tuned to the test distribution. Weighing these factors, the central comparison retains independent empirical content, so the appropriate circularity score is 2, not higher.

Assumptions & free parameters 2 free parameters · 5 assumptions · 0 invented entities

The central claim rests on the LLM's Prolog synthesis fidelity and on the embedding-based retrieval of 0-arity predicates; these are domain assumptions rather than standard math. The only fitted quantity is k, chosen on the test set. The paper introduces software abstractions (0-arity predicates, dynamic facts, interaction methods) but no physical or formal entities.

free parameters (2)
  • k (number of top 0-arity predicates retrieved) = 10
    Selected as best among 3, 10, 20 on the ShARC test set (Table 2); no validation split is reported.
  • Number of snippets passed to the system = 3
    Fixed by the evaluation protocol (target snippet plus two distractors); not ablated, and the authors argue per-snippet filtering makes it insensitive.
assumptions (5)
  • domain assumption LLM-synthesized Prolog modules from a single text chunk are semantically faithful to the source text and sufficient to answer questions derived from it.
    The whole pipeline relies on synthesis quality (Section 3.1, Appendix B). Only syntax errors are reported (2.2%); semantic faithfulness is not measured.
  • domain assumption 0-arity predicates expose all question-relevant Boolean claims and are ranked correctly by the joint natural language-code embedding model.
    Query construction uses only the top-k 0-arity rules (Section 3.3.1); if a needed predicate is not retrieved or ranked poorly, the query cannot be built.
  • domain assumption The simulated user LLM answers follow-up questions truthfully and completely from the withheld context.
    The evaluation of knowledge-gap detection depends on this simulation (Appendix B, Predicate Answering Prompt); inaccurate simulation would confound the more results.
  • domain assumption Each ShARC question can be answered from a single relevant text snippet.
    The evaluation restricts to a single target snippet and explicitly puts multi-hop retrieval out of scope (Section 4).
  • domain assumption Dynamic facts with a single Boolean argument are sufficient for the target QA tasks.
    The knowledge-gap mechanism is implemented only for Boolean dynamic facts (Section 3.2).

how reviews work

0 comments
Cite this review

Pith. "Pith review of NeSy-RAG: Neuro-Symbolic RAG for Explainable Question Answering." pith.science (2026). https://pith.science/paper/BNQNP2HF

@misc{pith2026260806292,
  author       = {Pith},
  title        = {Pith review of: NeSy-RAG: Neuro-Symbolic RAG for Explainable Question Answering},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/BNQNP2HF}},
  note         = {Machine review of arXiv:2608.06292}
}
read the original abstract

Retrieval-augmented generation (RAG) improves question answering by grounding large language models (LLMs) in external knowledge such as text corpora. However, its reasoning process remains largely opaque: intermediate reasoning steps are difficult to verify and cannot be reliably attributed to specific evidence. Moreover, missing user-specific context is rarely detected systematically, often leading to incomplete or incorrect output. We propose NeSy-RAG, a modular neuro-symbolic RAG framework that synthesizes attributable Prolog modules from retrieved text chunks. For each chunk, the system generates semantically meaningful predicates that encode Boolean claims, which may depend on user facts. Using joint natural language-code embeddings, predicates are retrieved and composed into Prolog queries. To address incomplete user context, we introduce a symbolic knowledge-gap detection mechanism that identifies missing user facts whose truth values affect the query outcome and automatically triggers follow-up interactions. Executing the resulting Prolog queries yields deterministic answers together with transparent execution traces that link each reasoning step to its originating source. On the ShARC benchmark, without domain-specific training, NeSy-RAG achieves 61.1% accuracy, outperforming a same-model RAG baseline that achieves 42.8% accuracy.

Figures

Figures reproduced from arXiv: 2608.06292 by the authors.

Figure 1
Figure 1. Overview of the proposed NeSy-RAG framework [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Normalized Confusion Matrices of NeSy-RAG and LLM Baselines [PITH_FULL_IMAGE:figures/full_fig_p009_2.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

29 extracted references · 25 canonical work pages

  1. [1]

    Reliable Reasoning Beyond Natural Language

    Nasim Borazjanizadeh and Steven Piantadosi. Reliable Reasoning Beyond Natural Language . ArXiv, abs/2407.11373, 2024

    arXiv 2024

  2. [2]

    Denison, John Schulman, Arushi Somani, Peter Hase, Misha Wagner, Fabien Roger, Vladimir Mikulik, Sam Bowman, Jan Leike, Jared Kaplan, and Ethan Perez

    Yanda Chen, Joe Benton, Ansh Radhakrishnan, Jonathan Uesato, Carson E. Denison, John Schulman, Arushi Somani, Peter Hase, Misha Wagner, Fabien Roger, Vladimir Mikulik, Sam Bowman, Jan Leike, Jared Kaplan, and Ethan Perez. Reasoning Models Don't Always Say What They Think . ArXiv, abs/2505.05410, 2025

    arXiv 2025

  3. [3]

    LoRP: LLM-based Logical Reasoning via Prolog

    Zhengkun Di, Chaoli Zhang, Hongtao Lv, Lizhen Cui, and Lei Liu. LoRP: LLM-based Logical Reasoning via Prolog . Knowl. Based Syst., 2025

    work page 2025

  4. [4]

    Efficient Code Embeddings from Code Generation Models

    Daria Kryvosheieva, Saba Sturua, Michael G \"u nther, Scott Martens, and Han Xiao. Efficient Code Embeddings from Code Generation Models . ArXiv, abs/2508.21290, 2025

    arXiv 2025

  5. [5]

    MediQ: Question-Asking LLMs and a Benchmark for Reliable Interactive Clinical Reasoning

    Shuyue S Li, Vidhisha Balachandran, Shangbin Feng, Jonathan S Ilgen, Emma Pierson, Pang W Koh, and Yulia Tsvetkov. MediQ: Question-Asking LLMs and a Benchmark for Reliable Interactive Clinical Reasoning . Advances in Neural Information Processing Systems, 37: 0 28858--28888, 2024

    work page 2024

  6. [6]

    Explicit Alignment and Many-to-many Entailment Based Reasoning for Conversational Machine Reading

    Yangyang Luo, Shiyu Tian, Caixia Yuan, and Xiaojie Wang. Explicit Alignment and Many-to-many Entailment Based Reasoning for Conversational Machine Reading . In Conference on Empirical Methods in Natural Language Processing, 2023

    work page 2023

  7. [7]

    Logic-LM: Empowering Large Language Models with Symbolic Solvers for Faithful Logical Reasoning

    Liangming Pan, Alon Albalak, Xinyi Wang, and William Wang. Logic-LM: Empowering Large Language Models with Symbolic Solvers for Faithful Logical Reasoning . In Findings of the Association for Computational Linguistics: EMNLP 2023, pages 3806--3824, 2023

    work page 2023

  8. [8]

    Interpretation of Natural Language Rules in Conversational Machine Reading

    Marzieh Saeidi, Max Bartolo, Patrick Lewis, Sameer Singh, Tim Rockt \"a schel, Mike Sheldon, Guillaume Bouchard, and Sebastian Riedel. Interpretation of Natural Language Rules in Conversational Machine Reading . In Conference on Empirical Methods in Natural Language Processing, 2018

    work page 2018

Show all 29 references
  1. [9]

    Retrieval Augmentation Reduces Hallucination in Conversation

    Kurt Shuster, Spencer Poff, Moya Chen, Douwe Kiela, and Jason Weston. Retrieval Augmentation Reduces Hallucination in Conversation . In Conference on Empirical Methods in Natural Language Processing, 2021

    2021

  2. [10]

    Prolog-Driven Rule-Based Diagnostics with Large Language Models for Precise Clinical Decision Support

    Xiaoyu Tan, Bin Li, Weidi Xu, Chao Qu, Wei Chu, Yinghui Xu, Yuan Qi, and Xihe Qiu. Prolog-Driven Rule-Based Diagnostics with Large Language Models for Precise Clinical Decision Support . In International Conference on Medical Image Computing and Computer-Assisted Intervention, 2025

    2025

  3. [11]

    Priyesh Vakharia, Abigail Kufeldt, Max Meyers, Ian Lane, and Leilani H. Gilpin. ProSLM: A Prolog Synergized Language Model for explainable Domain Specific Knowledge Based Question Answering . In International Workshop on Neural-Symbolic Learning and Reasoning, 2024

    2024

  4. [12]

    Neuro-Symbolic Integration Brings Causal and Reliable Reasoning Proofs

    Sen Yang, Xin Li, Leyang Cui, Li Bing, and Wai Lam. Neuro-Symbolic Integration Brings Causal and Reliable Reasoning Proofs . Findings of the Association for Computational Linguistics, 2025

    2025

  5. [13]

    Prolog-RAG: A Symbolic Reasoning Approach to Retrieval-Augmented Generation

    Bailing Zhang, Jiajie Li, Kang Peng, Shuchang Zheng, and Kai Meng. Prolog-RAG: A Symbolic Reasoning Approach to Retrieval-Augmented Generation . International Conference on Computer Engineering and Application, 2025

    2025

  6. [14]

    Yujia Zhou, Yan Liu, Xiaoxi Li, Jiajie Jin, Hongjin Qian, Zheng Liu, Chaozhuo Li, Zhicheng Dou, Tsung-Yi Ho, and Philip S. Yu. Trustworthiness in Retrieval-Augmented Generation Systems: A Survey . ArXiv, abs/2409.10102, 2024

    2024 arXiv

  7. [15]

    Yangyang Luo and Shiyu Tian and Caixia Yuan and Xiaojie Wang , booktitle=

  8. [16]

    Conference on Empirical Methods in Natural Language Processing , year=

    Marzieh Saeidi and Max Bartolo and Patrick Lewis and Sameer Singh and Tim Rockt. Conference on Empirical Methods in Natural Language Processing , year=

  9. [17]

    ArXiv , year=

    Daria Kryvosheieva and Saba Sturua and Michael G. ArXiv , year=

  10. [18]

    2020 , volume=

    Zhangyin Feng and Daya Guo and Duyu Tang and Nan Duan and Xiaocheng Feng and Ming Gong and Linjun Shou and Bing Qin and Ting Liu and Daxin Jiang and Ming Zhou , journal=. 2020 , volume=

    2020

  11. [19]

    Kurt Shuster and Spencer Poff and Moya Chen and Douwe Kiela and Jason Weston , booktitle=

  12. [20]

    Yanda Chen and Joe Benton and Ansh Radhakrishnan and Jonathan Uesato and Carson E. Denison and John Schulman and Arushi Somani and Peter Hase and Misha Wagner and Fabien Roger and Vladimir Mikulik and Sam Bowman and Jan Leike and Jared Kaplan and Ethan Perez , journal=. 2025 , volume=

    2025

  13. [21]

    Yu , journal=

    Yujia Zhou and Yan Liu and Xiaoxi Li and Jiajie Jin and Hongjin Qian and Zheng Liu and Chaozhuo Li and Zhicheng Dou and Tsung-Yi Ho and Philip S. Yu , journal=. 2024 , volume=

    2024

  14. [22]

    Xiaoyu Tan and Bin Li and Weidi Xu and Chao Qu and Wei Chu and Yinghui Xu and Yuan Qi and Xihe Qiu , booktitle=

  15. [23]

    Gilpin , booktitle=

    Priyesh Vakharia and Abigail Kufeldt and Max Meyers and Ian Lane and Leilani H. Gilpin , booktitle=

  16. [24]

    2024 , volume=

    Nasim Borazjanizadeh and Steven Piantadosi , journal=. 2024 , volume=

    2024

  17. [25]

    Zhengkun Di and Chaoli Zhang and Hongtao Lv and Lizhen Cui and Lei Liu , journal=

  18. [26]

    Sen Yang and Xin Li and Leyang Cui and Li Bing and Wai Lam , journal=

  19. [27]

    Bailing Zhang and Jiajie Li and Kang Peng and Shuchang Zheng and Kai Meng , journal=

  20. [28]

    Pan, Liangming and Albalak, Alon and Wang, Xinyi and Wang, William , booktitle=

  21. [29]

    Li, Shuyue S and Balachandran, Vidhisha and Feng, Shangbin and Ilgen, Jonathan S and Pierson, Emma and Koh, Pang W and Tsvetkov, Yulia , journal=

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.