Pith. sign in

REVIEW 3 major objections 5 minor 42 references

NeSy-CSA: A Neuro-Symbolic Framework for Open-Ended Critical Scenario Attribution

T0 review · 3 major / 5 minor · reviewed 2026-07-11 · grok-4.5

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

Pith's one-line read NeSy-CSA turns open-ended critical-scenario attribution into structured, traceable neuro-symbolic reasoning that improves intervention effectiveness over LLM baselines.

desk verdict Solid engineering pipeline for open-ended critical-scenario attribution that beats LLM baselines on intervention metrics; the causal claim is softer than the abstract implies but the work is still worth engaging. read the letter →

arxiv 2607.03847 v1 pith:MHGCH62H submitted 2026-07-04 cs.LG

Qitong Chu , Xunjie He , Chen Deng , Huaxin Pei , Yufeng Yue This is my paper · ORCID
classification cs.LG
keywords neuro-symbolicreasoningcriticalscenarioattributionscenario-basedtestingtraceableinterventionevaluationdecision-makingagentsLLM
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

Critical scenarios found by scenario-based testing of decision-making agents are only useful if we can explain why they are critical. That attribution problem is open-ended: causes can involve different state variables, interaction patterns, and failure mechanisms across tasks, so fixed diagnostic templates are too narrow and unconstrained LLM explanations are hard to inspect or trust. This paper claims that the gap can be closed by NeSy-CSA, a neuro-symbolic pipeline that first selects a compact set of data-supported factors, then organizes reasoning as a reusable predecessor-dependent subtask graph, and finally executes formalizable steps with dynamically composed symbolic procedures while routing semantic steps to evidence-constrained neural inference. Because open-ended attribution has no gold answers, the authors evaluate both the structural validity of the generated reasoning graph against expert references and the behavioral effect of intervening on parameters named by the attribution under a bounded budget. Across four environments the method raises two intervention metrics, criticality transition rate and reward improvement rate, by about 18% and 14% over LLM-only, tool-augmented, and chain-of-thought baselines, supporting the claim that discovered critical cases can become reusable knowledge for later testing and safety analysis.

What carries the argument

NeSy-CSA: key-factor refinement (knowledge-guided generation, semantic deduplication, statistical and effect-size filtering), a validated predecessor-dependent subtask graph G*, and neuro-symbolic hybrid execution that routes formalizable/verifiable subtasks to symbolic procedures built from atomic functions and the rest to evidence-constrained neural inference.

What would settle it

On held-out critical scenarios from the same four environments, extract parameters from NeSy-CSA conclusions, apply the paper’s bounded interventions, and check whether criticality-transition and reward-improvement rates remain near the reported levels; if they fall to baseline while process-level graph scores stay high, the result-level claim fails.

Watch

Extended reading notes

Core claim

Open-ended critical scenario attribution can be made both flexible and inspectable by constraining it at three levels—factor selection, predecessor-dependent subtask structure, and hybrid symbolic/neural execution—so that conclusions remain grounded in intermediate evidence and can be checked by controlled interventions that change criticality or reward.

Load-bearing premise

The method treats a successful bounded local edit of parameters named in the attribution text—flipping a critical outcome or raising reward—as evidence that those factors were the true drivers of criticality rather than merely correlated or easy-to-game knobs.

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 / 5 minor

Summary. The paper proposes NeSy-CSA, a neuro-symbolic framework for open-ended critical scenario attribution in decision-making agents. Given a critical sample, historical data, and domain priors, it (i) refines LLM-generated candidate factors via semantic deduplication and data-driven filtering (Welch t-test, BH correction, Hedges' g; Eqs. 7–14), (ii) builds a reusable predecessor-dependent subtask graph with deterministic Cont/Pred/IO validation (Eqs. 16–19), and (iii) executes formalizable subtasks via dynamically composed atomic symbolic procedures while routing the rest to evidence-constrained LLM inference (Eqs. 20–27). Evaluation combines process-level Precision/Recall/F1 against expert reference subtasks (Table III) with result-level CTR and RIR under bounded interventions on Extract(R*) parameters (Eqs. 34–36, Fig. 4). Across ACAS Xu, CoopNavi, BipedalWalker, and CARLA, NeSy-CSA reports average CTR/RIR gains of 18.32% and 13.67% over three LLM baselines (Table II), with ablations (Tables V–VI) and a CoopNavi case study (Fig. 7).

Significance. If the intervention protocol is accepted as a proxy for attribution quality, the work is a useful contribution to scenario-based testing: it turns discovered critical cases into structured, reusable explanations rather than isolated failures, and it does so without requiring a fixed causal graph or complete symbolic knowledge base. Strengths include multi-environment evaluation with three-run means and stds, process-level expert matching, component ablations, token-cost analysis of one-time vs per-sample graph construction, and an explicit hybrid routing design that is more inspectable than pure LLM attribution. The CTR/RIR metrics are falsifiable behavioral tests and a practical step beyond reference-free textual explanations. The main significance is methodological—constraining open-ended LLM attribution with factor filtering, validated subtask structure, and selective symbolic execution—rather than a new causal identification theorem.

major comments (3)
  1. §III.E.b, Eqs. (34)–(36), Fig. 4, and Table II: The central effectiveness claim (CTR +18.32%, RIR +13.67%) treats successful bounded interventions on Extract(R*_i) as evidence of correct causal attribution. This only shows that the named parameters are high-leverage editable knobs under budget ε (calibrated ~0.3 in Fig. 6). Without sham-factor / random-parameter controls, or comparison to classical causal/fault-diagnosis baselines discussed in §II.B, the protocol cannot distinguish true drivers from correlated easy-to-edit variables. Extract itself is an unvalidated text step. Please add at least one negative control (e.g., intervene on non-attributed or randomly selected parameters of matched cardinality) and report whether NeSy-CSA still outperforms baselines under that control; otherwise the causal reading of CTR/RIR should be substantially softened.
  2. §IV.A.4 and Table II: Baselines are restricted to LLM-only, LLM+Tool, and LLM+CoT. The related-work section motivates traditional fault diagnosis and causal/counterfactual methods, yet none appear as experimental baselines. For environments where state variables and intervention spaces are structured (especially ACAS Xu), a simple causal or sensitivity baseline would test whether the neuro-symbolic machinery is necessary for the reported gains. Absence of such baselines weakens the claim that NeSy-CSA specifically advances open-ended attribution beyond existing structured methods.
  3. §III.B–D and Algorithm 1: Factor filtering uses the same critical/non-critical partition of D that later defines Dc for CTR/RIR. While re-simulation is partly independent, there is mild circularity: factors are retained precisely because they differ between groups (Eq. 14), then interventions on those factors are scored by flipping criticality. Please clarify train/eval separation (e.g., filter factors on a held-out subset of D, evaluate CTR/RIR only on unseen critical samples) or quantify sensitivity of Table II to this reuse.
minor comments (5)
  1. Fig. 1 and Abstract: The +18.32% / +13.67% figures are averages over four environments relative to the mean of three LLM baselines; state this aggregation explicitly in the figure caption and abstract to avoid reading them as per-environment or vs. the best baseline.
  2. §IV.A.3: Free parameters α, δ, η_d, η_m, ε are fixed after limited calibration (Fig. 6 only for ACAS Xu). A short sensitivity table for δ and η across environments would strengthen robustness claims.
  3. Eq. (20) and §III.D.a: Formalizable(ST_t) ∧ Verifiable(ST_t) is central to routing but only described qualitatively. A brief operational definition or example of the LLM’s routing decision would improve reproducibility.
  4. Table III: Process-level F1 uses the best of three expert reference sets (Eq. 32). Report also mean F1 over the three experts to show sensitivity to reference choice.
  5. Typographical / presentation: arXiv id and some figure labels use mixed underscore styles (ACAS_Xu vs ACAS Xu); unify. Ensure Algorithm 1 line numbers match the narrative references in §III.C–D.

Circularity Check

1 steps flagged · score 1.0 of 10

No load-bearing circular derivation; mild evaluation-design dependence of factor screening on the same criticality labels, not of the CTR/RIR claims.

  1. other [§III.B.c Data-driven Filtering, Eqs. (7)–(14); Algorithm 1 lines 1–3]
    "F∗ = { ã_m ∈ F̃_cand | p̃_m < α, |g_hedge_m| > δ }, where α is the significance threshold and δ is the minimum effect-size threshold. This criterion ensures that the retained factors are not only statistically distinguishable between non-critical and critical samples, but also practically informative for subsequent attribution reasoning."

    Key factors are selected by testing group differences on the same critical/non-critical labels (y_j) that define the attribution problem. This is mild evaluation/design dependence (supervised screening on the target partition), not a claim that recovers those labels as a 'prediction.' CTR/RIR remain independent re-simulation tests and are not forced by the p-values or effect sizes.

full rationale

NeSy-CSA is a methods paper whose central claims are empirical (CTR/RIR gains vs LLM baselines under a shared intervention protocol), not first-principles predictions. Factor refinement (Welch t-test + BH + Hedges g on D_c vs D_nc) uses the criticality labels that define the problem, which is standard supervised screening rather than a self-definitional loop: the paper does not claim to re-predict those labels, and result-level success is measured by re-simulation after bounded edits (Eqs. 34–36), which is not forced by the screening statistics. Process-level F1 selects the best of three fixed expert references—an optimistic reporting choice, not a reduction of the method to its inputs. ε is calibrated for evaluation locality (Fig. 6) and applied uniformly; relative gains are not by construction. No uniqueness theorem, ansatz, or load-bearing self-citation chain underwrites the headline results. Score 1 only for the mild label-reuse in factor filtering; the derivation chain is otherwise self-contained against external benchmarks.

Assumptions & free parameters 4 free parameters · 5 assumptions · 3 invented entities

The central empirical claim depends on standard statistical screening choices, hand-set similarity/intervention thresholds, LLM planners/routers, and the modeling assumption that criticality is recoverable from intervenable scenario parameters under a local budget. No new physical entities; invented constructs are methodological (metrics, graph schema, hybrid executor).

free parameters (4)
  • significance threshold α
    Fixed at 0.05 for Welch tests after BH correction in factor filtering; standard but still a free choice that gates which factors enter F*.
  • effect-size threshold δ (Hedges’ g)
    Set to 0.5 as a medium-effect cutoff; directly controls which statistically significant factors are retained.
  • semantic similarity thresholds η_d, η_m
    Both set to 0.7 for deduplication and process-level matching; change which factors/subtasks count as the same.
  • intervention budget ε
    Calibrated environment-wise (example ACAS Xu sweep in Fig. 6) and set to 0.3; bounds the result-level CTR/RIR evaluation the main claim uses.
assumptions (5)
  • domain assumption Differences in factor means between critical and non-critical historical samples (Welch + BH + effect size) identify a useful attribution space for open-ended causes.
    §III.B data-driven filtering treats association with criticality labels as the basis for F*; causal sufficiency is assumed for downstream reasoning.
  • domain assumption A single validated predecessor-dependent subtask graph per task, built from F*, Q, and one exemplar, is reusable across critical samples of that task.
    §III.C and Table IV; one-time decomposition is load-bearing for efficiency and claimed stability.
  • ad hoc to paper Subtasks that are Formalizable and Verifiable can be correctly routed to symbolic procedures composed from a lightweight atomic library; remaining subtasks are adequately handled by evidence-constrained LLM inference.
    Routing rule Eq. (20) and SymExec/NeuInfer split in §III.D; correctness of routing is not independently proven.
  • ad hoc to paper Bounded re-simulation interventions on Extract(R*) parameters measure attribution effectiveness (CTR/RIR).
    §III.E.b defines the primary success metrics used in the abstract claim.
  • standard math Standard hypothesis-testing and effect-size machinery (Welch’s t, BH FDR, Hedges’ g) is applicable to the factor value distributions under large-sample CLT arguments.
    §III.B.c explicitly invokes these procedures and CLT robustness.
invented entities (3)
  • NeSy-CSA hybrid executor (SymExec from atomic library A + NeuInfer under predecessor evidence)
    purpose: Execute open-ended attribution without a full predefined symbolic rule base while keeping formalizable steps checkable.
    Core methodological construct of the paper; independent evidence is only the reported experiments, not external theory.
  • CTR and RIR intervention metrics
    purpose: Score open-ended attribution when no gold-standard cause labels exist.
    Defined in Eq. (36); useful operational metrics but paper-specific success criteria.
  • Predecessor-dependent subtask graph G* with deterministic Cont/Pred/IO validation
    purpose: Make attribution traceable and constrain explanation space from XF to XG.
    Formalized in §III.C; schema and validators are introduced for this pipeline.

how reviews work

0 comments
Cite this review

Pith. "Pith review of NeSy-CSA: A Neuro-Symbolic Framework for Open-Ended Critical Scenario Attribution." pith.science (2026). https://pith.science/paper/MHGCH62H

@misc{pith2026260703847,
  author       = {Pith},
  title        = {Pith review of: NeSy-CSA: A Neuro-Symbolic Framework for Open-Ended Critical Scenario Attribution},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/MHGCH62H}},
  note         = {Machine review of arXiv:2607.03847}
}
read the original abstract

Understanding why discovered scenarios become critical in scenario-based testing is essential for effectively leveraging them in decision-making systems. Reasoning about such criticality can be formulated as an attribution problem. However, across different decision-making tasks, the causes of criticality may involve diverse state variables, interaction patterns, and failure mechanisms, making attribution an inherently open-ended problem beyond predefined explanation spaces. Existing attribution methods still struggle to balance open-ended reasoning flexibility with the interpretability and traceability required for critical scenario reasoning. To address this limitation, we propose NeSy-CSA, a neuro-symbolic framework that transforms open-ended critical scenario attribution from unconstrained explanation generation into structured and traceable reasoning. NeSy-CSA narrows the attribution space by selecting relevant factors, makes the reasoning process traceable through a dependency-aware evidence graph, and executes symbolic reasoning procedures derived from atomic operations, coordinated with evidence-constrained neural inference to support flexible open-ended attribution. We further introduce a process-level and result-level assessment module to evaluate the structural validity of the attribution process and the behavioral effectiveness of the attribution results under controlled interventions. Experiments across four decision-making environments show that NeSy-CSA improves two intervention-based measures of attribution effectiveness by 18.32% and 13.67% over LLM-based baselines. These results demonstrate its potential to transform discovered critical scenarios into reusable knowledge for subsequent testing and safety analysis.

Figures

Figures reproduced from arXiv: 2607.03847 by the authors.

Figure 1
Figure 1. NeSy-CSA improves critical scenario attribution over LLM-based [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Overview of the NeSy-CSA framework. The pipeline includes key attribution factor refinement, subtask graph construction, and neuro-symbolic hybrid reasoning that combines symbolic computation for formalizable subtasks with neural inference for the remaining subtasks. The resulting attribution is evaluated from both process-level and result-level perspectives. Since critical scenario attribution is open-ended and oft… view at source ↗
Figure 3
Figure 3. Generation and validation of NeSy-CSA’s predecessor-dependent [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Result-level assessment procedure based on bounded intervention, [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: Illustrative critical scenarios across four environments. ACAS Xu: unsafe separation between the ownship (blue) and the intruder (red). Coop￾Navi: incomplete team coordination in reaching assigned landmarks. Bipedal￾Walker: loss of locomotion stability on diverse terra…
Figure 6
Figure 6. Figure 6: Sensitivity analysis of the intervention budget [PITH_FULL_IMAGE:figures/full_fig_p010_6.png]
Figure 7
Figure 7. Figure 7: Case study of traceable neuro-symbolic attribution reasoning. [PITH_FULL_IMAGE:figures/full_fig_p013_7.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

42 extracted references · 4 linked inside Pith

  1. [1]

    Confidence-aware robust dynamical distance constrained reinforcement learning for social robot navigation,

    K. Zhu, T. Xue, and T. Zhang, “Confidence-aware robust dynamical distance constrained reinforcement learning for social robot navigation,” IEEE Transactions on Automation Science and Engineering, 2025

    2025

  2. [2]

    Dynamic path optimization of humanoid welding robots based on environmental modeling and real-time obstacle detection,

    P. Chi, Z. Wang, Y . Liu, H. Liao, X. Wu, J. Wu, and Q. Zhang, “Dynamic path optimization of humanoid welding robots based on environmental modeling and real-time obstacle detection,”IEEE Transactions on Au- tomation Science and Engineering, 2026

    2026

  3. [3]

    Behaviorally-aware multi-agent rl with dynamic optimization for au- tonomous driving,

    H. Taghavifar, C. Hu, C. Wei, A. Mohammadzadeh, and C. Zhang, “Behaviorally-aware multi-agent rl with dynamic optimization for au- tonomous driving,”IEEE Transactions on Automation Science and Engineering, vol. 22, pp. 10 672–10 683, 2025

    2025

  4. [4]

    A hybrid optimal acceleration evaluation model for automated driving based on importance sampling method,

    H. Jiang, Y . Chen, A. Li, and S. Ma, “A hybrid optimal acceleration evaluation model for automated driving based on importance sampling method,”IEEE Transactions on Automation Science and Engineering, vol. 23, pp. 717–727, 2025

    2025

  5. [5]

    Autonomous driving system testing via diversity-oriented driving scenario exploration,

    X. Ji, L. Xue, Z. He, and X. Luo, “Autonomous driving system testing via diversity-oriented driving scenario exploration,”ACM Transactions on Software Engineering and Methodology, vol. 34, no. 7, pp. 1–28, 2025

    2025

  6. [6]

    On the comparabil- ity and optimal aggressiveness of the adversarial scenario-based safety testing of robots,

    B. Weng, G. A. Castillo, W. Zhang, and A. Hereid, “On the comparabil- ity and optimal aggressiveness of the adversarial scenario-based safety testing of robots,”IEEE Transactions on Robotics, vol. 39, no. 4, pp. 3299–3318, 2023

    2023

  7. [7]

    A survey of fault diagnosis and fault-tolerant techniques—part i: Fault diagnosis with model-based and signal-based approaches,

    Z. Gao, C. Cecati, and S. X. Ding, “A survey of fault diagnosis and fault-tolerant techniques—part i: Fault diagnosis with model-based and signal-based approaches,”IEEE transactions on industrial electronics, vol. 62, no. 6, pp. 3757–3767, 2015

    2015

  8. [8]

    Pearl,Causality: Models, Reasoning and Inference, 2nd ed

    J. Pearl,Causality: Models, Reasoning and Inference, 2nd ed. Cam- bridge University Press, 2009

    2009

Show all 42 references
  1. [9]

    Gpt-4 technical report,

    J. Achiam, S. Adler, S. Agarwal, L. Ahmad, I. Akkaya, F. L. Aleman, D. Almeida, J. Altenschmidt, S. Altman, S. Anadkatet al., “Gpt-4 technical report,”arXiv preprint arXiv:2303.08774, 2023

    2023 arXiv

  2. [10]

    Deepseek-v3 technical report,

    A. Liu, B. Feng, B. Xue, B. Wang, B. Wu, C. Lu, C. Zhao, C. Deng, C. Zhang, C. Ruanet al., “Deepseek-v3 technical report,”arXiv preprint arXiv:2412.19437, 2024

    2024 arXiv

  3. [11]

    A survey on llm symbolic reasoning,

    J. Li, Y . Fu, Y . Yang, J. Liu, H. Zhang, H. Li, Y . Yue, and M. Yang, “A survey on llm symbolic reasoning,”Authorea Preprints, 2025

    2025

  4. [12]

    Fast task planning with neuro-symbolic relaxation,

    Q. Du, B. Li, Y . Du, S. Su, T. Fu, Z. Zhan, Z. Zhao, and C. Wang, “Fast task planning with neuro-symbolic relaxation,”IEEE Robotics and Automation Letters, 2026

    2026

  5. [13]

    Predicate invention for bilevel plan- ning,

    T. Silver, R. Chitnis, N. Kumar, W. McClinton, T. Lozano-P ´erez, L. Kaelbling, and J. B. Tenenbaum, “Predicate invention for bilevel plan- ning,” inProceedings of the AAAI Conference on Artificial Intelligence, vol. 37, no. 10, 2023, pp. 12 120–12 129

    2023

  6. [14]

    Bisim- ulation makes analogies in goal-conditioned reinforcement learning,

    P. Hansen-Estruch, A. Zhang, A. Nair, P. Yin, and S. Levine, “Bisim- ulation makes analogies in goal-conditioned reinforcement learning,” in International Conference on Machine Learning. PMLR, 2022, pp. 8407–8426

    2022

  7. [15]

    Learning neuro-symbolic relational transition models for bilevel planning,

    R. Chitnis, T. Silver, J. B. Tenenbaum, T. Lozano-Perez, and L. P. Kael- bling, “Learning neuro-symbolic relational transition models for bilevel planning,” in2022 IEEE/RSJ international conference on intelligent robots and systems (IROS). IEEE, 2022, pp. 4166–4173

    2022

  8. [16]

    Towards safe autonomous driving policies using a neuro-symbolic deep reinforcement learning approach,

    I. Sharifi, M. Yildirim, and S. Fallah, “Towards safe autonomous driving policies using a neuro-symbolic deep reinforcement learning approach,” arXiv preprint arXiv:2307.01316, 2023

    2023

  9. [17]

    Neurosymbolic reinforcement learning and planning: A survey,

    K. Acharya, W. Raza, C. Dourado, A. Velasquez, and H. H. Song, “Neurosymbolic reinforcement learning and planning: A survey,”IEEE Transactions on Artificial Intelligence, vol. 5, no. 5, pp. 1939–1953, 2023

    1939

  10. [18]

    Mdpfuzz: testing models solving markov decision processes,

    Q. Pang, Y . Yuan, and S. Wang, “Mdpfuzz: testing models solving markov decision processes,” inProceedings of the 31st ACM SIGSOFT International Symposium on Software Testing and Analysis, 2022, pp. 378–390

    2022

  11. [19]

    Fuzzing with sequence di- versity inference for sequential decision-making model testing,

    K. Wang, Y . Wang, J. Wang, and Q. Wang, “Fuzzing with sequence di- versity inference for sequential decision-making model testing,” in2023 IEEE 34th International Symposium on Software Reliability Engineering (ISSRE). IEEE, 2023, pp. 706–717

    2023

  12. [20]

    Dicritest: Testing scenario gener- ation for decision-making agents considering diversity and criticality,

    Q. Chu, Y . Yue, D. Yao, and H. Pei, “Dicritest: Testing scenario gener- ation for decision-making agents considering diversity and criticality,” arXiv preprint arXiv:2508.11514, 2025

    2025 arXiv

  13. [21]

    A search-based testing approach for deep reinforcement learning agents,

    A. Zolfagharian, M. Abdellatif, L. C. Briand, M. Bagherzadehet al., “A search-based testing approach for deep reinforcement learning agents,” IEEE Transactions on Software Engineering, vol. 49, no. 7, pp. 3715– 3735, 2023

    2023

  14. [22]

    Dense reinforcement learning for safety validation of autonomous vehicles,

    S. Feng, H. Sun, X. Yan, H. Zhu, Z. Zou, S. Shen, and H. X. Liu, “Dense reinforcement learning for safety validation of autonomous vehicles,” Nature, vol. 615, no. 7953, pp. 620–627, 2023

    2023

  15. [23]

    Generative model-based testing on decision-making poli- cies,

    Z. Li, X. Wu, D. Zhu, M. Cheng, S. Chen, F. Zhang, X. Xie, L. Ma, and J. Zhao, “Generative model-based testing on decision-making poli- cies,” in2023 38th IEEE/ACM International Conference on Automated Software Engineering (ASE). IEEE, 2023, pp. 243–254

    2023

  16. [24]

    Diffscene: Diffusion- based safety-critical scenario generation for autonomous vehicles,

    C. Xu, A. Petiushko, D. Zhao, and B. Li, “Diffscene: Diffusion- based safety-critical scenario generation for autonomous vehicles,” in Proceedings of the AAAI conference on artificial intelligence, vol. 39, no. 8, 2025, pp. 8797–8805

    2025

  17. [25]

    Chatgpt-based sce- nario engineer: A new framework on scenario generation for trajectory prediction,

    X. Li, E. Liu, T. Shen, J. Huang, and F.-Y . Wang, “Chatgpt-based sce- nario engineer: A new framework on scenario generation for trajectory prediction,”IEEE Transactions on Intelligent Vehicles, vol. 9, no. 3, pp. 4422–4431, 2024

    2024

  18. [26]

    Om- nitester: Multimodal large language model driven scenario testing for autonomous vehicles,

    Q. Lu, X. Wang, Y . Jiang, G. Zhao, M. Ma, and S. Feng, “Om- nitester: Multimodal large language model driven scenario testing for autonomous vehicles,”Automotive Innovation, vol. 8, no. 4, pp. 838– 852, 2025

    2025

  19. [27]

    A theory of diagnosis from first principles,

    R. Reiter, “A theory of diagnosis from first principles,”Artificial intelligence, vol. 32, no. 1, pp. 57–95, 1987

    1987

  20. [28]

    Causes and explanations: A structural-model approach. part i: Causes,

    J. Y . Halpern and J. Pearl, “Causes and explanations: A structural-model approach. part i: Causes,”The British journal for the philosophy of science, 2005

    2005

  21. [29]

    Explainable reinforcement learning through a causal lens,

    P. Madumal, T. Miller, L. Sonenberg, and F. Vetere, “Explainable reinforcement learning through a causal lens,” inProceedings of the AAAI conference on artificial intelligence, vol. 34, no. 03, 2020, pp. 2493–2500

    2020

  22. [30]

    Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning,

    D. Guo, D. Yang, H. Zhang, J. Song, P. Wang, Q. Zhu, R. Xu, R. Zhang, S. Ma, X. Biet al., “Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning,”arXiv preprint arXiv:2501.12948, 2025

    2025 arXiv

  23. [31]

    Chain-of-thought prompting elicits reasoning in large language models,

    J. Wei, X. Wang, D. Schuurmans, M. Bosma, F. Xia, E. Chi, Q. V . Le, D. Zhouet al., “Chain-of-thought prompting elicits reasoning in large language models,”Advances in neural information processing systems, vol. 35, pp. 24 824–24 837, 2022

    2022

  24. [32]

    React: Synergizing reasoning and acting in language models,

    S. Yao, J. Zhao, D. Yu, N. Du, I. Shafran, K. Narasimhan, and Y . Cao, “React: Synergizing reasoning and acting in language models,” in11th International Conference on Learning Representations, ICLR 2023, 2023

    2023

  25. [33]

    Toolformer: Language models can teach themselves to use tools,

    T. Schick, J. Dwivedi-Yu, R. Dess `ı, R. Raileanu, M. Lomeli, E. Hambro, L. Zettlemoyer, N. Cancedda, and T. Scialom, “Toolformer: Language models can teach themselves to use tools,”Advances in neural informa- tion processing systems, vol. 36, pp. 68 539–68 551, 2023

    2023

  26. [34]

    Retrieval- augmented generation for knowledge-intensive nlp tasks,

    P. Lewis, E. Perez, A. Piktus, F. Petroni, V . Karpukhin, N. Goyal, H. K ¨uttler, M. Lewis, W.-t. Yih, T. Rockt ¨aschelet al., “Retrieval- augmented generation for knowledge-intensive nlp tasks,”Advances in neural information processing systems, vol. 33, pp. 9459–9474, 2020

    2020

  27. [35]

    Evaluating verifiability in generative search engines,

    N. F. Liu, T. Zhang, and P. Liang, “Evaluating verifiability in generative search engines,” inFindings of the Association for Computational Linguistics: EMNLP 2023, 2023, pp. 7001–7025

    2023

  28. [36]

    Logic-lm: Empowering large language models with symbolic solvers for faithful logical rea- soning,

    L. Pan, A. Albalak, X. Wang, and W. Wang, “Logic-lm: Empowering large language models with symbolic solvers for faithful logical rea- soning,” inFindings of the Association for Computational Linguistics: EMNLP 2023, 2023, pp. 3806–3824

    2023

  29. [37]

    Neuro- symbolic artificial intelligence: a survey,

    B. P. Bhuyan, A. Ramdane-Cherif, R. Tomar, and T. Singh, “Neuro- symbolic artificial intelligence: a survey,”Neural Computing and Appli- cations, vol. 36, no. 21, pp. 12 809–12 844, 2024

    2024

  30. [38]

    Instructflow: Adaptive symbolic constraint- guided code generation for long-horizon planning,

    H. Chi, Z. Feng, Y . Lyu, C. Zheng, L. Luo, Y . S. Ong, I. Tsang, H. Chen, Y . Chang, and H. Yin, “Instructflow: Adaptive symbolic constraint- guided code generation for long-horizon planning,”Advances in Neural Information Processing Systems, vol. 38, pp. 2602–2632, 2026

    2026

  31. [39]

    Imperative learning: A self-supervised neuro-symbolic learning framework for robot autonomy,

    C. Wang, K. Ji, J. Geng, Z. Ren, T. Fu, F. Yang, Y . Guo, H. He, X. Chen, Z. Zhanet al., “Imperative learning: A self-supervised neuro-symbolic learning framework for robot autonomy,”The International Journal of Robotics Research, p. 02783649251353181, 2024

    2024

  32. [40]

    Neuro- symbolic ai for auditable cognitive information extraction from medical reports,

    G. A. Prenosil, T. K. Weitzel, S. C. Bello, C. Mingels, G. Manzini, L. P. Meier, K.-Y . Shi, A. Rominger, and A. Afshar-Oromieh, “Neuro- symbolic ai for auditable cognitive information extraction from medical reports,”Communications Medicine, vol. 5, no. 1, p. 491, 2025

    2025

  33. [41]

    Nesypr: Neurosymbolic proceduraliza- tion for efficient embodied reasoning,

    W. Choi, J. Kim, and H. Woo, “Nesypr: Neurosymbolic proceduraliza- tion for efficient embodied reasoning,”Advances in Neural Information Processing Systems, vol. 38, pp. 174 484–174 504, 2026

    2026

  34. [42]

    Cohen,Statistical Power Analysis for the Behavioral Sciences, 2nd ed

    J. Cohen,Statistical Power Analysis for the Behavioral Sciences, 2nd ed. Lawrence Erlbaum Associates, 1988

    1988

Pith tools

Reviewed July 11, 2026 · model on record in the stance chip above.