Pith. sign in

REVIEW 5 major objections 6 minor 43 references

Causal Prompting for Implicit Sentiment Analysis with Large Language Models

T0 review · 5 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read The paper claims that ranking LLM sentiment outputs by an estimated causal effect of the query on the answer, rather than by majority voting over reasoning chains, produces more accurate and more robust implicit sentiment predictions.

desk verdict Weighted cluster-ensembling works; the front-door causal story doesn't—worth reviewing, but the causal claims need to go or be proven. read the letter →

arxiv 2507.00389 v1 pith:HXN7QMOL submitted 2025-07-01 cs.CL

Jing Ren , Wenhao Zhou , Bowen Li , Mujie Liu , Nguyen Linh Dan Le , Jiade Cen , Liping Chen , Ziqi Xu
show 2 more authors
Xiwei Xu Xiaodong Li
This is my paper · ORCID
classification cs.CL
keywords implicitsentimentanalysiscausalpromptingfront-dooradjustmentchain-of-thoughtlargelanguagemodelsself-consistencyNWGMapproximationdebiasing
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 tries to establish that, for implicit sentiment analysis, an LLM's candidate answers should be ranked by the estimated causal effect of the query on the sentiment label rather than by how often each answer appears across sampled reasoning chains. It argues that majority voting over chain-of-thought paths inherits the model's internal biases and spurious correlations, whereas a front-door-adjusted estimate separates the query's influence on the reasoning from the reasoning's influence on the answer. The proposed framework, CAPITAL, estimates this effect by clustering sampled reasoning paths and weighting each cluster's answer frequencies by cluster size and a retrieval-based expectation over prompts. If the claim holds, prompted LLMs can be made more trustworthy on implied sentiment without fine-tuning.

What carries the argument

The load-bearing device is the front-door adjustment formula (Theorem 2) adapted to LLM prompting, combined with three approximations. First, repeated sampling with temperature variation and K-means clustering on encoder embeddings estimates $P(T\mid do(X))$ by cluster frequencies $|T_k|/A$. Second, the NWGM approximation estimates $P(Y\mid do(T))$ by retrieving semantically similar incorrect chain-of-thought demonstrations, building a revised prompt, querying the LLM $N$ times, and averaging indicator predictions. Third, Equation 19 combines these estimates to rank candidate answers. The front-door criterion is what allows the framework to claim identifiability of the causal effect even though the confounder $Z$ is unobserved.

What would settle it

Generate a synthetic sentiment dataset whose true causal graph includes an unobserved confounder $Z$ with edges into both the query $X$ and the answer $Y$, then add an edge from $Z$ into the chain-of-thought variable $T$; if CAPITAL's estimate of $P(Y\mid do(X))$ no longer matches the known true effect once $Z\to T$ is present, the front-door assumption fails. A cheaper real-data check: adversarially perturb only the bias cues that should change $Z$ while leaving the true sentiment unchanged, and observe whether the estimated causal ranking of the correct answer drops when the confounder changes.

Watch

Extended reading notes

Core claim

On its own terms, the paper's central claim is that the interventional probability $P(Y\mid do(X))$ for a fixed prompt $X$ and sentiment polarity $Y$ can be approximated by the weighted ensemble $$\widehat P(Y\mid do(X)) = \sum_{k=1}^{K} \frac{|T_k|}{A}\cdot \frac{1}{N}\sum_{n=1}^{N}\mathbb{I}(y_{k,n}=y),$$ where $T_k$ is the $k$-th cluster of generated chain-of-thought paths, $A$ is the total number of sampled paths, and $y_{k,n}$ are repeated polarity predictions conditioned on the cluster's centroid path. The paper asserts that the chain-of-thought variable $T$ satisfies all conditions of the front-door criterion relative to $(X,Y)$, so this quantity identifies the causal effect of the query on the answer despite an unobserved confounder $Z$ representing internal LLM bias. Selecting the polarity with the highest estimated causal effect is then argued to be more trustworthy than majority voting. The reported experiments on two aspect-based sentiment benchmarks across three LLMs show consistent F1 gains, with the largest margins on the implicit-sentiment subset and under adversarial perturbation.

Load-bearing premise

The whole causal argument depends on the assumption that the model's hidden bias affects only the final answer and never the chain of thought it produces; if the bias can shape the reasoning steps too, the front-door adjustment no longer estimates a true causal effect.

Editorial extensions

If this is right

  • If Equation 19 approximates the true causal effect, answer selection in chain-of-thought prompting should be based on causal-effect ranking; the reported 4 to 10 point F1 gains over the strongest baselines follow from that shift.
  • The two-stage decomposition from prompt to reasoning and reasoning to output can be reused for other classification tasks where an LLM's reasoning path is a plausible mediator, such as stance detection or irony detection.
  • Because the method requires no fine-tuning and works with both open-source and closed-source models, it offers a portable way to reduce spurious correlations in prompted sentiment systems.
  • The robustness results imply that causal weighting degrades less than majority voting under out-of-distribution adversarial perturbation, extending the method's usefulness beyond standard benchmarks.

Reading between the lines

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

  • Editorial inference: The empirical gains could partly come from the clustering and demonstration-retrieval machinery rather than from front-door adjustment itself; a control that keeps the same clustering and retrieval but replaces the causal weighting with a non-causal similarity weighting would isolate the causal component.
  • Editorial inference: If the hidden bias $Z$ also influences which chains of thought are generated, Equation 19 is not a true interventional estimate; a synthetic test with a known graph that includes $Z\to T$ would show how far the method remains useful as a debiasing heuristic.
  • Editorial inference: The contrastive encoder alignment mentioned in the abstract and conclusion is not specified in the full text, so the clustering step is not fully reproducible until that objective is described.
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

5 major / 6 minor

Summary. The paper proposes CAPITAL, a prompting framework for implicit sentiment analysis that claims to use front-door adjustment to debias LLM predictions. The method generates multiple chain-of-thought (CoT) paths from a fixed prompt, clusters their embeddings, selects centroid CoTs, retrieves similar incorrect demonstrations from a training set to form revised prompts, and queries the LLM multiple times per centroid. The final prediction is chosen by scoring each sentiment polarity with the cluster-weighted average of indicator frequencies, which the paper presents as an estimate of P(Y | do(X)). Experiments on SemEval-2014 Laptop and Restaurant subsets with LLaMA-2, LLaMA-3, and GPT-3.5 report consistent improvements over ICL, CoT, CoT-SC, CAD, and THOR, with additional robustness and ablation studies.

Significance. If the causal identification were correct, CAPITAL would be a valuable integration of front-door adjustment with LLM prompting, and the reported gains across three backbones would be notable. The paper also ships source code and includes a thoughtful ablation design, which are strengths. However, the central methodological claim is not supported: the estimator in Eq. (19) does not implement front-door adjustment, and the asserted front-door conditions are unverified and plausible violated. The empirical recipe may still be a useful heuristic, but the causal interpretation is the main advertised contribution, and that contribution is not established. The inconsistencies between the main tables further undermine confidence in the reported results.

major comments (5)
  1. [IV-C2, Eqs. (12)-(19)] The estimator in Eq. (19) is not a valid estimate of P(Y | do(X)) even under the paper's own DAG. The derivation requires P(Y | do(T)) = E_X[P(Y | T, X)], and a Monte Carlo estimate of this expectation must average over X drawn from P(X). Instead, Eqs. (15)-(18) construct the prompt X by retrieving training demonstrations that are most similar to the CoT T, so X is a function of T and the samples are drawn from q(X | T), not from P(X). Eq. (18) therefore converges to E_{X ~ q(·|T)}[P(Y | T, X)], which equals the required backdoor adjustment only if q(X | T) = P(X). The paper neither states nor justifies this equality, and it is plainly false: retrieval deliberately makes X and T dependent, which is exactly the dependence the adjustment is supposed to remove. Consequently, the final approximation in Eq. (19) does not follow from the front-door formula, and the reported gains may be due to the retrieval heuristic rather than to causal effect estimation.
  2. [IV-C, Fig. 2c] The claim that 'the variable T satisfies all the conditions of the front-door criterion relative to (X, Y)' is asserted without argument. In particular, the front-door criterion requires that there be no unblocked back-door path from T to Y that is not blocked by X; this requires that the latent confounder Z does not affect T. Since Z is described as 'internal bias' of the LLM, it is highly plausible that Z also influences the generated CoT T, which would open a back-door path T <- Z -> Y and invalidate the front-door adjustment. The DAG in Fig. 2c assumes this away by drawing no edge from Z to T, but no empirical or theoretical support is provided. This is a load-bearing assumption, not a harmless simplification.
  3. [Eq. (4), IV-C] Equation (4) is introduced as a 'variant of the front-door adjustment' without derivation or proof. The standard front-door formula in Eq. (2) has an inner summation over X', and the paper replaces this with P(Y | do(T)) in Eq. (4). While this replacement could be justified under the assumed DAG by applying backdoor adjustment to T -> Y, the paper does not provide that derivation. More importantly, the variable X in Eq. (5) (a fixed prompt with demonstrations selected by similarity to the test sentence) is not the same variable as the retrieved prompt X in Eq. (16), which depends on the CoT T. The two are conflated in Eq. (19), making the notation and the claimed estimand ambiguous. A precise causal model and a derivation of Eq. (4) are needed before the method can be described as front-door adjustment.
  4. [IV-C2, Eq. (18) and Fig. 4] The method is described as using the 'NWGM approximation,' but Eq. (18) is an arithmetic mean of indicator functions, not a normalized weighted geometric mean. NWGM, as used in the cited references [34], [35], involves a geometric mean over softmax probabilities; the paper's computation is a frequency average over binary events. This is a conceptual mismatch between the stated theoretical basis and the implemented estimator. The authors should either provide a derivation of how the indicator-frequency average approximates the NWGM or remove the NWGM terminology and justify the average directly as a Monte Carlo estimate under a clearly specified sampling distribution.
  5. [Tables I, II, and III] There are numerical inconsistencies between the tables that undermine the experimental narrative. Table I reports CAPITAL LLaMA-3 Restaurant ISA F1 of 71.63, but Table III reports the corresponding 'ISA' value as 90.17. Similarly, Table I reports LLaMA-3 Laptop ISA as 73.68, while Table III reports 83.56. The values in Table III match the 'Adv.' column of Table II, not the ISA column of Table I. This suggests the ablation study is reporting adversarial (OOD) results under the 'ISA' label, or the tables are mislabeled. Such a discrepancy is not a minor typo; it affects the reader's ability to assess the robustness and the internal consistency of the reported improvements.
minor comments (6)
  1. [Definition 2] The third bullet of Definition 2 states 'No unblocked back-door paths from M to Y,' but the variable M is not defined; it should presumably be W.
  2. [Section VI] The conclusion begins with an extraneous 's' character: 's This paper presents CAPITAL...'.
  3. [Figure 5] Figure 5 appears to be corrupted in the provided manuscript: its content is rendered as a sequence of '/uni...' escape sequences rather than a legible plot, so the hyper-parameter analysis cannot be visually verified.
  4. [Section V-C] The implementation details refer to 'LLaMA-3 253B' accessed via NVIDIA's hosted API, but the tables and text use 'LLaMA-3'; the model name should be specified consistently, especially since 253B is not a standard LLaMA-3 release name.
  5. [Section I and Table I] The problem statement in Section IV-A says the polarity label Y is 'positive, neutral, or negative,' but the prompt template and the SemEval datasets include a fourth label, CONFLICT; the paper should clarify whether CONFLICT is used in the evaluation or mapped to one of the three polarities.
  6. [Tables I-III] The paper reports F1 scores without standard deviations or significance tests across repeated runs, despite the stochastic nature of LLM sampling; adding error bars or a small number of independent runs would strengthen the empirical claims.

Circularity Check

1 steps flagged · score 6.0 of 10

Eq. 18 substitutes a T-dependent retrieval distribution for the marginal P(X) required by the back-door adjustment, so the 'causal effect' in Eq. 19 is by construction a weighted conditional heuristic rather than a front-door estimate.

  1. fitted input called prediction [Section IV-C2, Eqs. 12-18]
    "P (Y | do(T )) = Σ_X P (Y | T, X)P (X) = E_X [P (Y | T, X)]. ... For each t_centroid_k, the final prompt is constructed as: x_t_centroid_k = [t_wrong↑_l , . . . , t_wrong↑_1 , s_test] ... The probability of the predicted polarity can now be approximated as follows: E_x [P (Y | T, X)] ≈ 1/N Σ_{n=1}^N I(y_{k,n}=y)."

    Eq. 13 defines the target as an expectation over the marginal prompt distribution P(X). However, Eqs. 15-16 construct the prompt X by retrieving the top-L training demonstrations ranked by cosine similarity to the fixed CoT T_k, so X is a function of T. The empirical average in Eq. 18 therefore converges to E_{X|T}[P(Y|T,X)] or to P(Y|T,X*) for the single selected prompt, not to E_X[P(Y|T,X)]. The quantity relabeled P(Y|do(T)) is by construction an average over prompts chosen by T, and Eq. 19 inherits this mismatch. The 'causal prediction' is thus a fitted conditional heuristic presented as a front-door estimate.

full rationale

The paper's empirical comparison against external benchmarks is not circular: the final accuracy numbers are measured on held-out sentiment labels, and the main contribution is testable independently of the causal framing. However, the load-bearing causal derivation is partially circular. The front-door adjustment in Eq. 4 requires P(Y|do(T)) = E_X[P(Y|T,X)], but the implementation replaces the marginal distribution P(X) with a retrieval distribution q(X|T) that is explicitly constructed as a function of T (Eqs. 15-16). Consequently Eq. 18 estimates a conditional frequency, and Eq. 19 is a weighted ensemble of those conditional frequencies; calling it an unbiased causal effect is a relabeling of the heuristic, not a consequence of the front-door theorem. This is a fitted-input-called-prediction reduction. Separate correctness concerns, not circularity: the paper asserts without argument that T satisfies all front-door conditions, and Definition 2 contains an undefined variable M ('No unblocked back-door paths from M to Y'), so the invoked criterion is not even fully stated. Self-citations by the corresponding author (e.g., refs. [24]-[28]) appear only as background causal-inference work and are not load-bearing, so they do not raise the score further.

Assumptions & free parameters 6 free parameters · 4 assumptions · 1 invented entities

The ledger shows that the method's causal operation rests on asserted structural assumptions (front-door validity, NWGM approximation, encoder alignment) and several unstated or data-tuned hyperparameters (K, A, L, temperature).

free parameters (6)
  • K = 8 (Restaurant) and 12 (Laptop) in Section V-F; Implementation details says K=8
    Number of CoT clusters; tuned in hyperparameter study (Figure 5) to maximize F1, so it is data-dependent.
  • A = 20
    Number of sampled CoTs; tuned in hyperparameter study.
  • N = 5
    Second-stage LLM queries per centroid; chosen without a stated criterion.
  • R = 3
    Number of in-context demonstrations in the initial prompt; fixed by hand.
  • L = not reported
    Number of retrieved demonstrations in Eq 16; never assigned in the paper, blocking exact replication.
  • temperature = not reported
    Sampling temperature for generating A CoTs; not stated.
assumptions (4)
  • ad hoc to paper The SCM in Figure 2c holds: X->T->Y, and Z confounds X and Y but does not affect T.
    Section IV-C asserts T satisfies the front-door criterion; no evidence or argument is given that hidden bias fails to influence CoT generation.
  • domain assumption Sampling multiple CoTs from the LLM at varied temperature gives a representative interventional distribution P(T|do(X)).
    Section IV-C1 uses cluster frequencies from A=20 samples to approximate P(T|do(X)).
  • domain assumption Retrieving in-context examples by embedding similarity approximates the expectation over X in the backdoor adjustment (NWGM).
    Section IV-C2 replaces E_X[P(Y|T,X)] with an average over retrieved demonstrations; the quality of this approximation is not analyzed.
  • ad hoc to paper The encoder's embeddings are aligned with the LLM's reasoning space.
    The abstract and conclusion claim a contrastive fine-tuning strategy, but it is not described in the methods, so the alignment is unverified.
invented entities (1)
  • Latent confounder Z
    purpose: Models the internal bias of the LLM that confounds query and answer in the SCM.
    Z is a hypothetical variable used to justify front-door adjustment; no falsifiable handle is provided.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Causal Prompting for Implicit Sentiment Analysis with Large Language Models." pith.science (2026). https://pith.science/paper/HXN7QMOL

@misc{pith2026250700389,
  author       = {Pith},
  title        = {Pith review of: Causal Prompting for Implicit Sentiment Analysis with Large Language Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/HXN7QMOL}},
  note         = {Machine review of arXiv:2507.00389}
}
read the original abstract

Implicit Sentiment Analysis (ISA) aims to infer sentiment that is implied rather than explicitly stated, requiring models to perform deeper reasoning over subtle contextual cues. While recent prompting-based methods using Large Language Models (LLMs) have shown promise in ISA, they often rely on majority voting over chain-of-thought (CoT) reasoning paths without evaluating their causal validity, making them susceptible to internal biases and spurious correlations. To address this challenge, we propose CAPITAL, a causal prompting framework that incorporates front-door adjustment into CoT reasoning. CAPITAL decomposes the overall causal effect into two components: the influence of the input prompt on the reasoning chains, and the impact of those chains on the final output. These components are estimated using encoder-based clustering and the NWGM approximation, with a contrastive learning objective used to better align the encoder's representation with the LLM's reasoning space. Experiments on benchmark ISA datasets with three LLMs demonstrate that CAPITAL consistently outperforms strong prompting baselines in both accuracy and robustness, particularly under adversarial conditions. This work offers a principled approach to integrating causal inference into LLM prompting and highlights its benefits for bias-aware sentiment reasoning. The source code and case study are available at: https://github.com/whZ62/CAPITAL.

Figures

Figures reproduced from arXiv: 2507.00389 by the authors.

Figure 1
Figure 1. Examples comparing explicit and implicit sentiment [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Three structural causal models (SCMs) representing [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. An example illustrating internal bias in LLMs. Voting-based methods such as THOR [2] select the most frequent answer [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: The overall framework of CAPITAL, which includes three stages: CoT generation, distribution inference via clustering, [PITH_FULL_IMAGE:figures/full_fig_p004_4.png]
Figure 5
Figure 5. Figure 5: Hyper-parameter analysis of the number of clusters and [PITH_FULL_IMAGE:figures/full_fig_p009_5.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

43 extracted references · 35 canonical work pages

  1. [34]

    Debiasing nlu models via causal intervention and counterfactual reasoning,

    B. Tian, Y . Cao, Y . Zhang, and C. Xing, “Debiasing nlu models via causal intervention and counterfactual reasoning,” in Proceedings of the AAAI Conference on Artificial Intelligence , vol. 36, no. 10, 2022, pp. 11 376–11 384

    work page 2022

  2. [35]

    Causal intervention and counterfactual reasoning for multi-modal fake news detection,

    Z. Chen, L. Hu, W. Li, Y . Shao, and L. Nie, “Causal intervention and counterfactual reasoning for multi-modal fake news detection,” in Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), 2023, pp. 627–638

    work page 2023

  3. [1]

    Exploring public sentiment during covid-19: A cross country analysis,

    S. Yu, S. He, Z. Cai, I. Lee, M. Naseriparsa, and F. Xia, “Exploring public sentiment during covid-19: A cross country analysis,” IEEE Transactions on Computational Social Systems, vol. 10, no. 3, pp. 1083– 1094, 2022

    work page 2022

  4. [2]

    Reasoning implicit sentiment with chain-of-thought prompting,

    H. Fei, B. Li, Q. Liu, L. Bing, F. Li, and T.-S. Chua, “Reasoning implicit sentiment with chain-of-thought prompting,” in Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics , 2023, pp. 1171–1182

    work page 2023

  5. [3]

    Causal prompting: Debiasing large language model prompting based on front-door adjust- ment,

    C. Zhang, L. Zhang, J. Wu, Y . He, and D. Zhou, “Causal prompting: Debiasing large language model prompting based on front-door adjust- ment,” in Proceedings of the AAAI Conference on Artificial Intelligence , vol. 39, no. 24, 2025, pp. 25 842–25 850. 10

    work page 2025

  6. [4]

    Decot: Debiasing chain-of-thought for knowledge-intensive tasks in large language models via causal intervention,

    J. Wu, T. Yu, X. Chen, H. Wang, R. A. Rossi, S. Kim, A. B. Rao, and J. J. McAuley, “Decot: Debiasing chain-of-thought for knowledge-intensive tasks in large language models via causal intervention,” in Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics, 2024, pp. 14 073–14 087

    work page 2024

  7. [5]

    Implicit aspect extraction in sentiment analysis: Review, taxonomy, oppportunities, and open challenges,

    M. Tubishat, N. Idris, and M. A. Abushariah, “Implicit aspect extraction in sentiment analysis: Review, taxonomy, oppportunities, and open challenges,” Information Processing & Management , vol. 54, no. 4, pp. 545–563, 2018

    work page 2018

  8. [6]

    Convolutional neural network for sentence classification,

    Y . Chen, “Convolutional neural network for sentence classification,” Master’s thesis, University of Waterloo, 2015

    work page 2015

Show all 43 references
  1. [7]

    Document modeling with gated recurrent neural network for sentiment classification,

    D. Tang, B. Qin, and T. Liu, “Document modeling with gated recurrent neural network for sentiment classification,” in Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing , 2015, pp. 1422–1432

    2015

  2. [8]

    Bert: Pre-training of deep bidirectional transformers for language understanding,

    J. Devlin, M.-W. Chang, K. Lee, and K. Toutanova, “Bert: Pre-training of deep bidirectional transformers for language understanding,” in Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologi...

    2019

  3. [9]

    A contrastive cross-channel data augmentation framework for aspect-based sentiment analysis,

    B. Wang, L. Ding, Q. Zhong, X. Li, and D. Tao, “A contrastive cross-channel data augmentation framework for aspect-based sentiment analysis,” in Proceedings of the 29th International Conference on Computational Linguistics, 2022, pp. 6691–6704

    2022

  4. [10]

    Learning implicit sentiment in aspect-based sentiment analysis with supervised contrastive pre-training,

    Z. Li, Y . Zou, C. Zhang, Q. Zhang, and Z. Wei, “Learning implicit sentiment in aspect-based sentiment analysis with supervised contrastive pre-training,” in Proceedings of the Conference on Empirical Methods in Natural Language Processing , 2021, pp. 246–256

    2021

  5. [11]

    Relational graph attention network for aspect-based sentiment analysis,

    K. Wang, W. Shen, Y . Yang, X. Quan, and R. Wang, “Relational graph attention network for aspect-based sentiment analysis,” in Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, 2020, pp. 3229–3238

    2020

  6. [12]

    Sentiment analysis in the era of large language models: A reality check,

    W. Zhang, Y . Deng, B. Liu, S. Pan, and L. Bing, “Sentiment analysis in the era of large language models: A reality check,” in Findings of the Association for Computational Linguistics , 2024, pp. 3881–3906

    2024

  7. [13]

    Aspect- based sentiment analysis with explicit sentiment augmentations,

    J. Ouyang, Z. Yang, S. Liang, B. Wang, Y . Wang, and X. Li, “Aspect- based sentiment analysis with explicit sentiment augmentations,” in Proceedings of the AAAI Conference on Artificial Intelligence , vol. 38, no. 17, 2024, pp. 18 842–18 850

    2024

  8. [14]

    Rvisa: reasoning and verification for implicit sentiment analysis,

    W. Lai, H. Xie, G. Xu, and Q. Li, “Rvisa: reasoning and verification for implicit sentiment analysis,” IEEE Transactions on Affective Computing, 2025

    2025

  9. [15]

    Causal prompting: Debiasing large language model prompting based on front-door adjust- ment,

    C. Zhang, L. Zhang, J. Wu, Y . He, and D. Zhou, “Causal prompting: Debiasing large language model prompting based on front-door adjust- ment,” in Proceedings of the AAAI Conference on Artificial Intelligence , vol. 39, no. 24, 2025, pp. 25 842–25 850

    2025

  10. [16]

    A comprehensive survey of prompt engineering techniques in large language models,

    T. Debnath, M. N. A. Siddiky, M. E. Rahman, P. Das, and A. K. Guha, “A comprehensive survey of prompt engineering techniques in large language models,” TechRxiv, 2025

    2025

  11. [17]

    Language models are few-shot learners,

    T. Brown, B. Mann, N. Ryder, M. Subbiah, J. D. Kaplan, P. Dhariwal, A. Neelakantan, P. Shyam, G. Sastry, A. Askellet al., “Language models are few-shot learners,” Advances in Neural Information Processing Systems, vol. 33, pp. 1877–1901, 2020

    1901

  12. [18]

    Prompt programming for large language models: Beyond the few-shot paradigm,

    L. Reynolds and K. McDonell, “Prompt programming for large language models: Beyond the few-shot paradigm,” in Extended Abstracts of the 2021 CHI Conference on Human Factors in Computing Systems , 2021, pp. 1–7

    2021

  13. [19]

    A survey of automatic prompt engineering: An optimization perspective,

    W. Li, X. Wang, W. Li, and B. Jin, “A survey of automatic prompt engineering: An optimization perspective,” ArXiv Preprint ArXiv:2502.11560, 2025

    2025 arXiv

  14. [20]

    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. Zhou et 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

  15. [21]

    Self-consistency improves chain of thought reasoning in language models,

    X. Wang, J. Wei, D. Schuurmans, Q. V . Le, E. H. Chi, S. Narang, A. Chowdhery, and D. Zhou, “Self-consistency improves chain of thought reasoning in language models,” in The Eleventh International Conference on Learning Representations , 2023

    2023

  16. [22]

    Pearl, M

    J. Pearl, M. Glymour, and N. P. Jewell, Causal inference in statistics: A primer. John Wiley & Sons, 2016

    2016

  17. [23]

    Disen- tangled representation learning for causal inference with instruments,

    D. Cheng, J. Li, L. Liu, Z. Xu, W. Zhang, J. Liu, and T. D. Le, “Disen- tangled representation learning for causal inference with instruments,” IEEE Transactions on Neural Networks and Learning Systems, pp. 1–14, 2024

    2024

  18. [24]

    Causal infer- ence with conditional front-door adjustment and identifiable variational autoencoder,

    Z. Xu, D. Cheng, J. Li, J. Liu, L. Liu, and K. Yu, “Causal infer- ence with conditional front-door adjustment and identifiable variational autoencoder,” in The Twelfth International Conference on Learning Representations, 2024

    2024

  19. [25]

    Conditional instrumental variable regression with representation learning for causal inference,

    D. Cheng, Z. Xu, J. Li, L. Liu, J. Liu, and T. D. Le, “Conditional instrumental variable regression with representation learning for causal inference,” in The Twelfth International Conference on Learning Repre- sentations, 2024

    2024

  20. [26]

    Instrumental variable estimation for causal inference in longitudinal data with time-dependent latent confounders,

    D. Cheng, Z. Xu, J. Li, L. Liu, J. Liu, W. Gao, and T. D. Le, “Instrumental variable estimation for causal inference in longitudinal data with time-dependent latent confounders,” in Thirty-Eighth AAAI Conference on Artificial Intelligence , 2024, pp. 11 480–11 488

    2024

  21. [27]

    Causal inference with conditional instruments using deep generative models,

    D. Cheng, Z. Xu, J. Li, L. Liu, J. Liu, and T. D. Le, “Causal inference with conditional instruments using deep generative models,” in Thirty- Seventh AAAI Conference on Artificial Intelligence , 2023, pp. 7122– 7130

    2023

  22. [28]

    Learning condi- tional instrumental variable representation for causal effect estimation,

    D. Cheng, Z. Xu, J. Li, L. Liu, T. D. Le, and J. Liu, “Learning condi- tional instrumental variable representation for causal effect estimation,” in Machine Learning and Knowledge Discovery in Databases: Research Track - European Conference , vol. 14169, 2023, pp. 525–540

    2023

  23. [29]

    Causal intervention improves implicit sentiment analysis,

    S. Wang, J. Zhou, C. Sun, J. Ye, T. Gui, Q. Zhang, and X.-J. Huang, “Causal intervention improves implicit sentiment analysis,” in Proceed- ings of the 29th International Conference on Computational Linguistics , 2022, pp. 6966–6977

    2022

  24. [30]

    Causalqa: A benchmark for causal question answering,

    A. Bondarenko, M. Wolska, S. Heindorf, L. Bl ¨ubaum, A.-C. N. Ngomo, B. Stein, P. Braslavski, M. Hagen, and M. Potthast, “Causalqa: A benchmark for causal question answering,” in Proceedings of the 29th International Conference on Computational Linguistics, 2022, pp. 3296– 3308

    2022

  25. [31]

    Pearl, Causality

    J. Pearl, Causality. Cambridge university press, 2009

    2009

  26. [32]

    Intrinsic dimensionality explains the effectiveness of language model fine-tuning,

    A. Aghajanyan, S. Gupta, and L. Zettlemoyer, “Intrinsic dimensionality explains the effectiveness of language model fine-tuning,” in Proceed- ings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural...

    2021

  27. [33]

    Parameter-efficient fine-tuning of large- scale pre-trained language models,

    N. Ding, Y . Qin, G. Yang, F. Wei, Z. Yang, Y . Su, S. Hu, Y . Chen, C.-M. Chan, W. Chen et al. , “Parameter-efficient fine-tuning of large- scale pre-trained language models,” Nature Machine Intelligence, vol. 5, no. 3, pp. 220–235, 2023

    2023

  28. [36]

    Active learning principles for in-context learning with large language models,

    K. Margatina, T. Schick, N. Aletras, and J. Dwivedi-Yu, “Active learning principles for in-context learning with large language models,” in Find- ings of the Association for Computational Linguistics: EMNLP 2023 , 2023, pp. 5011–5034

    2023

  29. [37]

    Semeval-2016 task 5: Aspect based sentiment analysis,

    M. Pontiki, D. Galanis, H. Papageorgiou, I. Androutsopoulos, S. Man- andhar, M. Al-Smadi, M. Al-Ayyoub, Y . Zhao, B. Qin, O. De Clercq et al. , “Semeval-2016 task 5: Aspect based sentiment analysis,” in International workshop on semantic evaluation , 2016, pp. 19–30

    2016

  30. [38]

    Language models are few-shot learners,

    T. B. Brown, B. Mann, N. Ryder, M. Subbiah, J. Kaplan, P. Dhariwal, A. Neelakantan, P. Shyam, G. Sastry, A. Askellet al., “Language models are few-shot learners,” in Advances in Neural Information Processing Systems, 2020

    2020

  31. [39]

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

    J. Wei, X. Wang, D. Schuurmans, M. Bosma, B. Ichter, F. Xia, E. H. Chi, Q. V . Le, and D. Zhou, “Chain-of-thought prompting elicits reasoning in large language models,” in Advances in Neural Information Processing Systems, 2022

    2022

  32. [40]

    Trusting your evidence: Hallucinate less with context-aware decoding,

    W. Shi, X. Han, M. Lewis, Y . Tsvetkov, L. Zettlemoyer, and W. Yih, “Trusting your evidence: Hallucinate less with context-aware decoding,” in Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Tech...

    2024

  33. [41]

    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,” CoRR, vol. abs/2307.09288, 2023

    2023 arXiv

  34. [42]

    Llms with chain- of-thought are non-causal reasoners,

    G. Bao, H. Zhang, L. Yang, C. Wang, and Y . Zhang, “Llms with chain- of-thought are non-causal reasoners,” CoRR, vol. abs/2402.16048, 2024

    2024 arXiv

  35. [43]

    Language models are few- shot learners,

    T. B. Brown, B. Mann, N. Ryder, M. Subbiah, J. Kaplan, P. Dhariwal, A. Neelakantan, P. Shyam, G. Sastry et al., “Language models are few- shot learners,” in Advances in Neural Information Processing Systems , 2020

    2020

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.