Pith. sign in

REVIEW 3 major objections 6 minor 47 references

This paper argues that in media bias detection, an explanation's agreement with expert rationales (plausibility) and its reflection of the model's actual computation (faithfulness) are distinct properties that must be measured separately.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-01 11:11 UTC pith:FBSUP6DX

load-bearing objection A solid, well-caveated empirical study with a genuinely new encoder-focused circuit analysis, but the faithfulness axis rests on an unvalidated linear probe that needs strengthening before the mechanistic claims fully land. the 3 major comments →

arxiv 2607.19954 v1 pith:FBSUP6DX submitted 2026-07-22 cs.CL

A Multi-Dimensional Evaluation of Explainability in Media Bias Detection

Ting Chen , Raina Zhang , Benjamin M. Ampel , Sagar Samtani This is my paper
classification cs.CL
keywords media bias detectionexplainabilityplausibilitymechanistic faithfulnessattention supervisioncircuit discoveryBERTRoBERTa
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

Media bias classifiers must justify their predictions, but a plausible explanation—one that highlights the same tokens an expert would—may not be a faithful one, since the model might rely on different cues. This paper studies BERT and RoBERTa models (base and large) on the BABE benchmark and evaluates them along three separate axes: classification accuracy, token-level agreement with expert rationales, and mechanistic faithfulness measured by how well compact sets of attention heads recover the model's predictions under counterfactual masking. By using attention-supervised finetuning as a controlled intervention that nudges attention toward expert-highlighted tokens, the authors show that plausibility and faithfulness move independently across architectures. The central message is that explanation quality is not a single number: predictive performance, plausibility, and faithfulness characterize different behavioral aspects and should be reported separately.

Core claim

The paper's central claim is that predictive performance, attribution plausibility, and mechanistic faithfulness are non-interchangeable properties of an explanation in media bias detection. Using the BABE dataset, the authors fine-tune BERT and RoBERTa (base and large) under standard supervision and under attention-supervised finetuning, which adds a loss that pushes the CLS attention distribution toward expert-highlighted tokens. They measure plausibility as AUROC between three attention-based attribution scores (CLS attention, attention rollout, gradient-weighted attention) and expert rationales, and they measure faithfulness by adapting activation-patching circuit discovery to encoders:

What carries the argument

Attention-supervised finetuning: an auxiliary MSE loss that encourages the final-layer CLS attention distribution to match binary expert rationale masks, used as a controlled intervention on plausibility. Activation-patching circuit discovery adapted from ACDC: for each attention head, the causal effect is the absolute change in expected probe probability when clean head activations replace corrupted ones; top-k heads form a candidate circuit evaluated in elimination (none) and isolation (circuit_only) modes. Retention = (probability gap of isolated circuit)/(absolute base gap) measures sufficiency; Rescue = gap(circuit_only) − gap(none) measures recovered signal beyond ablation. A logistic

Load-bearing premise

The faithfulness measurements are computed against a lightweight logistic-regression probe over the final [CLS] representation rather than the model's actual classification head; if this probe does not faithfully reflect the model's real decision procedure, Retention and Rescue measure the probe, not the model.

What would settle it

Re-run activation-patching circuit discovery using each model's original BERT/RoBERTa classification head as the causal readout (replacing the linear probe) on the same counterfactual pairs; if retention and rescue scores change substantially under the actual head—or if the ranking of heads flips—the paper's faithfulness conclusions do not hold for the original models.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • If plausibility and faithfulness are independent, benchmark suites for explainable NLP should report both token-level agreement with human rationales and circuit-level recoverability, not just one.
  • Attention-supervised finetuning does not guarantee more faithful circuits; its effect is architecture-dependent, so rationale supervision should not be advertised as a faithfulness intervention.
  • Model scale does not determine compressibility: RoBERTa-large resists compact circuit recovery, indicating that larger encoders may distribute bias signal more diffusely.
  • Positive rescue only at larger circuit sizes suggests bias judgments rely on coordinated multi-head computation, guiding future circuit-search efforts toward multi-component explanations.
  • Since attention-based attributions achieve only moderate AUROC against expert rationales, relying on raw attention heatmaps for bias explanations is insufficient.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The probe-based faithfulness results are the paper's own; a testable extension would repeat circuit discovery using the original classification heads to see whether retention/rescue scores match those obtained under the logistic probe, since the paper's faithfulness claims refer to 'the model' while the probe is a stand-in.
  • The counterfactual pipeline masks expert-annotated biased keywords; unbiased sentences lack rationales, so the circuit analysis may effectively cover only biased examples. An extension would construct counterfactuals for unbiased sentences (e.g., inserting neutral alternatives) to check whether circuit behavior generalizes across classes.
  • The moderate plausibility numbers (most AUROCs near 0.5–0.6) suggest that token-level rationales with only 6% coverage are a hard target; an inference is that attribution methods tuned for sparse rationales, such as supervised rationale extraction, may close part of the gap.
  • If the dimensionality claim generalizes beyond BABE, evaluation of explainability in other high-stakes text classification (e.g., hate speech, misinformation) should adopt the same three-axis separation.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 6 minor

Summary. The paper presents a multi-dimensional evaluation of explainability for encoder-based media bias detection on the BABE dataset. It fine-tunes BERT and RoBERTa (base and large) under standard and attention-supervised training, then measures (i) predictive performance via Macro F1, (ii) explanation plausibility via AUROC between three attention-based attribution methods and expert rationales, and (iii) mechanistic faithfulness via activation-patching circuit discovery, quantified by new Retention and Rescue metrics. The central claim, stated in the abstract and conclusion, is that predictive performance, attribution plausibility, and mechanistic faithfulness are distinct dimensions of model behavior and should be evaluated separately. In particular, the paper reports no monotonic relationship between plausibility and circuit retention (Figure 3), suggesting that expert-aligned rationales do not imply that the model relies on those tokens. The paper contributes an adaptation of ACDC to encoder classification and an intervention on plausibility via attention supervision, but several methodological issues affect the validity of the faithfulness measurements.

Significance. If the methodological concerns are resolved, the paper would make a useful contribution: it is among the first to jointly evaluate plausibility and mechanistic faithfulness for encoder-based text classification, and it provides a concrete setup for doing so on a socially relevant task. The empirical observation that attribution plausibility and circuit retention do not align—if supported by sound metrics—would reinforce the growing consensus that human-grounded plausibility and model-grounded faithfulness must be assessed independently. The paper also makes its results transparent through detailed tables (e.g., Table E1) and reports per-class and per-seed results, which strengthens the empirical basis. However, the faithfulness axis is currently built on a linear probe that replaces the models' own classification heads, and the probe is not validated against the original classifiers. As a result, the headline claim about whether the model actually relies on highlighted tokens is not directly supported.

major comments (3)
  1. [§3.5, Eq. (1); §4.2, Table E1] The circuit faithfulness metrics (Retention and Rescue) are computed with a logistic-regression probe over final [CLS] hidden states, explicitly replacing the architecture-specific classification heads of BERT and RoBERTa. The probe is trained only on clean representations, yet activation patching feeds corrupted and patched representations into the probe, which may lie outside its training distribution. The paper reports no accuracy, calibration, or agreement between the probe and the original classifier on clean or corrupted inputs. If the probe's decisions differ from the model's, the circuits characterize the probe, not the BERT/RoBERTa model whose explanations are being evaluated. This directly undermines the abstract's claim that the results speak to 'whether the model actually relies on the highlighted tokens.' The authors justify the probe as a 'uniform diagnostic objective,' but
  2. [§3.5, §3.6; Table E1] The definition of ΔP(base), ΔP(none), and ΔP(circuit_only) is under-specified: the text refers only to a 'predicted probability gap' but does not state exactly which examples are included (e.g., mean probability for biased minus unbiased sentences, or a per-pair difference) or how the counterfactual pairs are aggregated. This is especially problematic because unbiased sentences have no expert rationales (K_i empty), so the corruption function replaces no tokens and the 'corrupted' input is identical to the clean input. The paper does not state whether unbiased examples are excluded from circuit analysis or how identity counterfactuals are handled. Without this clarity, the Retention and Rescue scores cannot be interpreted, and the coverage of the faithfulness analysis is unclear. Please specify the exact computation of the probability gap and the treatment of unbiased sentences.
  3. [§4.2, Figure 3] The central claim that plausibility and mechanistic faithfulness are distinct rests on the assertion that 'no consistent monotonic relationship' is observed between best AUROC and circuit retention (k=30). This is based on visual inspection of eight points. With such a small sample, absence of a visible trend is weak evidence for the claimed lack of association. Please provide a quantitative analysis (e.g., Spearman correlation with confidence intervals, or a permutation test) and justify the choice of best AUROC and k=30. If the relationship is non-monotonic or noisy, the claim should be phrased with appropriate statistical uncertainty.
minor comments (6)
  1. [§3.6] Typo: 'explanation plausibilityplausibility' should be 'explanation plausibility.'
  2. [Figure 2a caption] The caption states that attention-supervised finetuning 'generally improves circuit retention and rescue scores,' but the text in §4.2 describes mixed effects (e.g., weaker retention for BERT-large and RoBERTa-large). Please align the caption with the reported results.
  3. [§3.5] The citation for BABE should be Spinde et al. (2021a), not Spinde et al. (2024), which refers to a taxonomy paper. Please correct.
  4. [§3.3] For unbiased sentences, the expert rationale mask is all zeros. The attention-supervision loss L_attn then encourages all normalized attention weights to zero, which is degenerate. Please specify how L_attn is handled for unbiased sentences (e.g., applied only to biased ones) or provide a non-degenerate formulation.
  5. [§3.4] In the Attention Rollout description, the matrix \bar{A}^{(l)} is not defined (presumably head-averaged attention). Please define it. Also, the reference 'Hao et al.' is cited without a year in the text and with an incomplete bibliographic entry.
  6. [References] The reference 'Naim and Asher' lacks a year in the citation; the entry appears incomplete. Please ensure all references are complete and consistently formatted.

Circularity Check

0 steps flagged

No circularity: faithfulness and plausibility are measured independently; the probe-based readout is a validity limitation, not a definitional reduction.

full rationale

The paper's derivation chain is self-contained. Plausibility is operationalized as AUROC between attention-based attributions and BABE expert rationales (Section 3.6), while mechanistic faithfulness is operationalized through activation-patching retention and rescue metrics (Sections 3.5-3.6). The central empirical claim that these dimensions are distinct rests on the observed lack of a consistent monotonic relationship in Figure 3, which is not forced by the definitions of the metrics. The retention/rescue equations are standard definitions adapted from ERASER and ACDC, and the circuit-selection procedure does not fit parameters to the target conclusion. The paper introduces no load-bearing self-citations: its methodological anchors (ACDC, ERASER, BABE, attention-supervision losses) are external prior work. The main caveat is the logistic probe used as the causal readout in Section 3.5, replacing the original classification heads; the paper does not report probe agreement with the original classifiers, so the faithfulness results may describe the encoder-plus-probe system rather than the original model. This is a construct-validity and generalizability limitation, not a circular reduction: the probe is trained to predict labels, not to produce a particular retention/rescue outcome, and no equation reduces the model's behavior to the probe's parameters by construction. No self-definitional, fitted-input-as-prediction, or self-citation circularity is present.

Axiom & Free-Parameter Ledger

2 free parameters · 4 axioms · 2 invented entities

The paper's empirical conclusions stand on the BABE annotation ground truth, a probe-based causal target, a deterministic keyword-masking counterfactual, and author-defined retention/rescue metrics. None of these load-bearing constructs is externally validated, and one (the probe) changes the object whose faithfulness is being measured.

free parameters (2)
  • λ (attention supervision weight) = 0.5
    Selected by a validation sweep on BERT-base (Appendix C, λ∈{0,...,5.0}) and then fixed for all architectures; it is a fitted hyperparameter that controls the intervention, not a constant of a derivation.
  • Linear probe parameters (w,b) = learned logistic regression weights
    Trained on final-layer [CLS] representations to predict the bias label (Eq. 1, Section 3.5). This probe is the causal readout for all retention/rescue circuit metrics, so the central faithfulness numbers depend on these fitted parameters.
axioms (4)
  • domain assumption BABE expert bias labels and token rationales are valid ground truth for bias and explanation plausibility.
    Used throughout as supervision signal and plausibility targets; the paper itself notes bias is subjective and contested.
  • ad hoc to paper The linear probe over final hidden states is a valid proxy for the model's bias classification decision.
    Section 3.5 replaces architecture-specific heads with a logistic probe; if the probe diverges from the model's head, faithfulness conclusions don't transfer to the actual classifier.
  • domain assumption Masking expert-annotated keywords with [MASK] produces valid counterfactual inputs.
    Section 3.5 defines the corruption function; replacing a biased keyword with [MASK] is assumed to remove the bias signal without creating other artifacts. Unbiased sentences with empty rationales are not addressed.
  • ad hoc to paper Top-k attention heads selected by activation-patching score approximate the model's computational circuit.
    Section 3.5 sweeps k∈{1,...,30} and the Limitations section states the result is approximate; this is a stated approximation, not a proven circuit decomposition.
invented entities (2)
  • Retention metric no independent evidence
    purpose: Quantifies the fraction of the full model (probe) probability gap reproduced by an isolated top-k circuit
    Newly defined in Section 3.6; not validated against any external benchmark and can be inflated when the base gap is small.
  • Rescue metric no independent evidence
    purpose: Measures signed recovered predictive margin beyond the fully ablated baseline
    Newly defined in Section 3.6; no external validation, but it complements retention by avoiding near-zero denominator inflation.

pith-pipeline@v1.3.0-alltime-deepseek · 12563 in / 14700 out tokens · 135846 ms · 2026-08-01T11:11:27.653899+00:00 · methodology

0 comments
read the original abstract

Detecting media bias automatically is difficult because biased framing is often subtle, yet in domains such as news analysis, accurate predictions alone are insufficient without explanations that reflect the model's underlying reasoning. We present a multi-dimensional evaluation of explainability in encoder-based media bias detection using the Bias Annotations By Experts (BABE) dataset. Specifically, we study BERT and RoBERTa as classifiers (base and large variants) along three complementary axes: predictive performance, explanation plausibility (token-level alignment with expert rationales), and mechanistic faithfulness (whether compact sets of attention heads recover predictive signal under counterfactual rationale masking). To induce variation in plausibility, we additionally investigate attention-supervised finetuning, which incorporates expert rationale annotations as an auxiliary training signal. Attention supervision serves as an intervention on attribution plausibility, while the effectiveness of attribution methods varies substantially across architectures. Circuit analysis further reveals substantial variation in mechanistic recoverability across architectures, suggesting that model scale alone does not determine circuit compressibility. Taken together, our findings suggest that predictive performance, attribution plausibility, and mechanistic faithfulness characterize different aspects of model behavior and should be evaluated separately when studying explainability in media bias detection.

Figures

Figures reproduced from arXiv: 2607.19954 by Benjamin M. Ampel, Raina Zhang, Sagar Samtani, Ting Chen.

Figure 1
Figure 1. Figure 1: Flowchart of the media bias analysis pipeline. [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Circuit faithfulness analysis across encoder-based bias detection models. Attention-supervised finetuning [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Relationship between attribution plausibility [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

47 extracted references · 10 canonical work pages · 4 internal anchors

  1. [1]

    Findings of the B lackbox NLP 2025 Shared Task: Localizing Circuits and Causal Variables in Language Models

    Arad, Dana and Belinkov, Yonatan and Chen, Hanjie and Kim, Najoung and Mohebbi, Hosein and Mueller, Aaron and Sarti, Gabriele and Tutek, Martin. Findings of the B lackbox NLP 2025 Shared Task: Localizing Circuits and Causal Variables in Language Models. Proceedings of the 8th BlackboxNLP Workshop: Analyzing and Interpreting Neural Networks for NLP. 2025. ...

    work page doi:10.18653/v1/2025.blackboxnlp-1.32 2025

  2. [2]

    Attribution Patching Outperforms Automated Circuit Discovery

    Syed, Aaquib and Rager, Can and Conmy, Arthur. Attribution Patching Outperforms Automated Circuit Discovery. Proceedings of the 7th BlackboxNLP Workshop: Analyzing and Interpreting Neural Networks for NLP. 2024. doi:10.18653/v1/2024.blackboxnlp-1.25

    doi:10.18653/v1/2024.blackboxnlp-1.25 2024

  3. [3]

    2026 , eprint=

    When Bigger Isn't Better: A Comprehensive Fairness Evaluation of Political Bias in Multi-News Summarisation , author=. 2026 , eprint=

    2026

  4. [4]

    2024 , eprint=

    LLMs left, right, and center: Assessing GPT's capabilities to label political bias from web domains , author=. 2024 , eprint=

    2024

  5. [5]

    Automated identification of bias inducing words in news articles using linguistic and context-oriented features , journal =

    Timo Spinde and Lada Rudnitckaia and Jelena Mitrović and Felix Hamborg and Michael Granitzer and Bela Gipp and Karsten Donnay , keywords =. Automated identification of bias inducing words in news articles using linguistic and context-oriented features , journal =. 2021 , issn =. doi:https://doi.org/10.1016/j.ipm.2021.102505 , url =

    arXiv 2021

  6. [6]

    Mandrekar , keywords =

    Jayawant N. Mandrekar , keywords =. Receiver Operating Characteristic Curve in Diagnostic Test Assessment , journal =. 2010 , issn =. doi:https://doi.org/10.1097/JTO.0b013e3181ec173d , url =

    work page doi:10.1097/jto.0b013e3181ec173d 2010

  7. [7]

    2022 , eprint=

    In-context Learning and Induction Heads , author=. 2022 , eprint=

    2022

  8. [8]

    2021 , journal=

    A Mathematical Framework for Transformer Circuits , author=. 2021 , journal=

    2021

  9. [9]

    2023 , eprint=

    Chain-of-Thought Prompting Elicits Reasoning in Large Language Models , author=. 2023 , eprint=

    2023

  10. [10]

    2019 , eprint=

    Attention is not not Explanation , author=. 2019 , eprint=

    2019

  11. [11]

    2023 , eprint=

    Towards Automated Circuit Discovery for Mechanistic Interpretability , author=. 2023 , eprint=

    2023

  12. [12]

    Mechanistic Fine-tuning for In-context Learning

    Cho, Hakaze and Luo, Peng and Kato, Mariko and Kaenbyou, Rin and Inoue, Naoya. Mechanistic Fine-tuning for In-context Learning. Proceedings of the 8th BlackboxNLP Workshop: Analyzing and Interpreting Neural Networks for NLP. 2025. doi:10.18653/v1/2025.blackboxnlp-1.21

    work page doi:10.18653/v1/2025.blackboxnlp-1.21 2025

  13. [13]

    2019 , eprint=

    What Does BERT Look At? An Analysis of BERT's Attention , author=. 2019 , eprint=

    2019

  14. [14]

    Computational Linguistics , author =

    Towards Faithful Model Explanation in. Computational Linguistics , author =. 2024 , date =. doi:10.1162/coli_a_00511 , shorttitle =

    doi:10.1162/coli_a_00511 2024

  15. [15]

    Improving Attention-Based Interpretability of Text Classification Transformers , url =

    Mylonas, Nikolaos and Mollas, Ioannis and Tsoumakas, Grigorios , year =. Improving Attention-Based Interpretability of Text Classification Transformers , url =. doi:10.1007/978-3-031-17849-8_15 , booktitle =

    work page doi:10.1007/978-3-031-17849-8_15

  16. [16]

    Where Americans get their news: New data from

    Carter, Allison , year =. Where Americans get their news: New data from

  17. [17]

    Evaluating self-attention interpretability through human-grounded experimental protocol

    Bhan, Milan and Achache, Nina and Legrand, Victor and Blangero, Annabelle and Chesneau, Nicolas , urldate =. Evaluating self-attention interpretability through human-grounded experimental protocol , url =. 2023 , date =. doi:10.48550/arXiv.2303.15190 , abstract =. 2303.15190 [cs] , keywords =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2303.15190 2023

  18. [18]

    On Explaining with Attention Matrices , url =

    Naim, Omar and Asher, Nicholas , urldate =. On Explaining with Attention Matrices , url =. 2024 , date =. doi:10.3233/FAIA240594 , eprinttype =. 2410.18541 [cs] , keywords =

    Pith/arXiv arXiv 2024

  19. [19]

    Self-Attention Attribution: Interpreting Information Interactions Inside Transformer , url =

    Hao, Yaru and Dong, Li and Wei, Furu and Xu, Ke , urldate =. Self-Attention Attribution: Interpreting Information Interactions Inside Transformer , url =. 2021 , date =. doi:10.48550/arXiv.2004.11207 , shorttitle =. 2004.11207 [cs] , keywords =

    Pith/arXiv arXiv doi:10.48550/arxiv.2004.11207 2021

  20. [20]

    Towards Machine Learning Explainability in Text Classification for Fake News Detection , url =

    Kurasinski, Lukas and Mihailescu, Radu-Casian , urldate =. Towards Machine Learning Explainability in Text Classification for Fake News Detection , url =. 2020 19th. doi:10.1109/ICMLA51294.2020.00127 , abstract =

    arXiv 2020

  21. [21]

    Generalizing Backpropagation for Gradient-Based Interpretability , url =

    Du, Kevin and Hennigen, Lucas Torroba and Stoehr, Niklas and Warstadt, Alexander and Cotterell, Ryan , urldate =. Generalizing Backpropagation for Gradient-Based Interpretability , url =. 2023 , date =. doi:10.48550/arXiv.2307.03056 , abstract =. 2307.03056 [cs] , keywords =

    Pith/arXiv arXiv doi:10.48550/arxiv.2307.03056 2023

  22. [22]

    The Devil is in the Neurons: Interpreting and Mitigating Social Biases in Pre-trained Language Models

    Liu, Yan and Liu, Yu and Chen, Xiaokang and Chen, Pin-Yu and Zan, Daoguang and Kan, Min-Yen and Ho, Tsung-Yi , urldate =. The Devil is in the Neurons: Interpreting and Mitigating Social Biases in Pre-trained Language Models , url =. 2024 , date =. doi:10.48550/arXiv.2406.10130 , shorttitle =. 2406.10130 [cs] , keywords =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2406.10130 2024

  23. [23]

    Chain-of-Thought Prompting Elicits Reasoning in Large Language Models , url =

    Wei, Jason and Wang, Xuezhi and Schuurmans, Dale and Bosma, Maarten and Ichter, Brian and Xia, Fei and Chi, Ed and Le, Quoc and Zhou, Denny , urldate =. Chain-of-Thought Prompting Elicits Reasoning in Large Language Models , url =. 2023 , date =. doi:10.48550/arXiv.2201.11903 , abstract =. 2201.11903 [cs] , keywords =

    Pith/arXiv arXiv doi:10.48550/arxiv.2201.11903 2023

  24. [24]

    How Interpretable are Reasoning Explanations from Prompting Large Language Models? , url =

    Yeo, Wei Jie and Satapathy, Ranjan and Goh, Rick Siow Mong and Cambria, Erik , urldate =. How Interpretable are Reasoning Explanations from Prompting Large Language Models? , url =. Findings of the Association for Computational Linguistics:. 2024 , date =. doi:10.18653/v1/2024.findings-naacl.138 , abstract =. 2402.11863 [cs] , keywords =

    Pith/arXiv arXiv 2024

  25. [25]

    Understanding Interpretability: Explainable

    Gr. Understanding Interpretability: Explainable. 2023 , publisher =. doi:10.1007/978-3-031-44070-0_3 , booktitle =

    work page doi:10.1007/978-3-031-44070-0_3 2023

  26. [26]

    Five Key Insights Into Americans' Views of the News Media , url =

    Brenan, Megan and Saad, Lydia , year =. Five Key Insights Into Americans' Views of the News Media , url =

  27. [27]

    2023 , note =

    Dealing with Propaganda, Misinformation and Fake News , url =. 2023 , note =

    2023

  28. [29]

    The Media Bias Taxonomy: A Systematic Literature Review on the Forms and Automated Detection of Media Bias , url =

    Spinde, Timo and Hinterreiter, Smi and Haak, Fabian and Ruas, Terry and Giese, Helge and Meuschke, Norman and Gipp, Bela , urldate =. The Media Bias Taxonomy: A Systematic Literature Review on the Forms and Automated Detection of Media Bias , url =. 2024 , date =. doi:10.48550/arXiv.2312.16148 , shorttitle =. 2312.16148 [cs] , keywords =

    Pith/arXiv arXiv doi:10.48550/arxiv.2312.16148 2024

  29. [30]

    A Theoretical Framework for

    Giordano, Davide and Kavasidis, Isaak and Spampinato, Concetto , year =. A Theoretical Framework for. 2212.14447 , archivePrefix =

    Pith/arXiv arXiv

  30. [31]

    Attention heads of large language models , volume =

    Zheng, Zifan and Wang, Yezhaohui and Huang, Yuxin and Song, Shichao and Yang, Mingchuan and Tang, Bo and Xiong, Feiyu and Li, Zhiyu , year =. Attention heads of large language models , volume =. Patterns , shortjournal =. doi:10.1016/j.patter.2025.101176 , abstract =

    arXiv 2025

  31. [32]

    Adaptable and Reliable Text Classification using Large Language Models

    Wang, Zhiqiang and Pang, Yiran and Lin, Yanbin and Zhu, Xingquan , urldate =. Adaptable and Reliable Text Classification using Large Language Models , url =. 2024 , date =. doi:10.48550/arXiv.2405.10523 , abstract =. 2405.10523 [cs] , note =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2405.10523 2024

  32. [33]

    Neural Media Bias Detection Using Distant Supervision With

    Spinde, Timo and Plank, Manuel and Krieger, Jan-David and Ruas, Terry and Gipp, Bela and Aizawa, Akiko , urldate =. Neural Media Bias Detection Using Distant Supervision With. Findings of the Association for Computational Linguistics:. 2021 , date =. doi:10.18653/v1/2021.findings-emnlp.101 , abstract =. 2209.14557 [cs] , keywords =

    Pith/arXiv arXiv 2021

  33. [34]

    doi:10.48550/arXiv.1911.03429 , shorttitle =

    2020 , date =. doi:10.48550/arXiv.1911.03429 , shorttitle =. 1911.03429 [cs] , keywords =

    Pith/arXiv arXiv doi:10.48550/arxiv.1911.03429 2020

  34. [35]

    Data-Centric Human Preference with Rationales for Direct Preference Alignment

    Just, Hoang Anh and Jin, Ming and Sahu, Anit and Phan, Huy and Jia, Ruoxi , urldate =. Data-Centric Human Preference with Rationales for Direct Preference Alignment , url =. 2025 , date =. doi:10.48550/arXiv.2407.14477 , abstract =. 2407.14477 [cs] , keywords =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2407.14477 2025

  35. [36]

    Tursunalieva, Ainura and Alexander, David L. J. and Dunne, Rob and Li, Jiaming and Riera, Luis and Zhao, Yanchang , urldate =. Making Sense of Machine Learning: A Review of Interpretation Techniques and Their Applications , volume =. Applied Sciences , shortjournal =. 2024 , date =. doi:10.3390/app14020496 , shorttitle =

    work page doi:10.3390/app14020496 2024

  36. [37]

    From Understanding to Utilization: A Survey on Explainability for Large Language Models , url =

    Luo, Haoyan and Specia, Lucia , urldate =. From Understanding to Utilization: A Survey on Explainability for Large Language Models , url =. 2024 , date =. doi:10.48550/arXiv.2401.12874 , shorttitle =. 2401.12874 [cs] , keywords =

    Pith/arXiv arXiv doi:10.48550/arxiv.2401.12874 2024

  37. [38]

    Attention is not not Explanation , url =

    Wiegreffe, Sarah and Pinter, Yuval , urldate =. Attention is not not Explanation , url =. 2019 , date =. doi:10.48550/arXiv.1908.04626 , abstract =. 1908.04626 [cs] , keywords =

    Pith/arXiv arXiv doi:10.48550/arxiv.1908.04626 2019

  38. [39]

    Automated identification of media bias in news articles: an interdisciplinary literature review , volume =

    Hamborg, Felix and Donnay, Karsten and Gipp, Bela , urldate =. Automated identification of media bias in news articles: an interdisciplinary literature review , volume =. International Journal on Digital Libraries , shortjournal =. 2019 , date =. doi:10.1007/s00799-018-0261-y , shorttitle =

    doi:10.1007/s00799-018-0261-y 2019

  39. [40]

    Where Americans Get Their News: New Data from

    Carter, Allison , year =. Where Americans Get Their News: New Data from

  40. [41]

    Attitudes to How Bias in News Sources Affects Ability to Sort Out the Facts in the

  41. [42]

    2019 , eprint=

    BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding , author=. 2019 , eprint=

    2019

  42. [43]

    2019 , eprint=

    RoBERTa: A Robustly Optimized BERT Pretraining Approach , author=. 2019 , eprint=

    2019

  43. [44]

    2023 , eprint=

    Qwen Technical Report , author=. 2023 , eprint=

    2023

  44. [45]

    2020 , eprint=

    Quantifying Attention Flow in Transformers , author=. 2020 , eprint=

    2020

  45. [46]

    2021 , eprint=

    Self-Attention Attribution: Interpreting Information Interactions Inside Transformer , author=. 2021 , eprint=

    2021

  46. [47]

    2023 , eprint=

    Attention Is All You Need , author=. 2023 , eprint=

    2023

  47. [48]

    2019 , eprint=

    Attention is not Explanation , author=. 2019 , eprint=

    2019