Pith. sign in

REVIEW 5 major objections 5 minor 2 cited by

LLMs encode cognitive difficulty as a linearly readable internal variable, reaching roughly 95% probe accuracy by layer five.

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

2026-08-02 22:17 UTC pith:YKSXVIV4

load-bearing objection Bloom labels are linearly decodable from residual streams, but the paper's 'cognitive complexity' interpretation rests on a lexical confound the controls don't kill. the 5 major comments →

arxiv 2602.17229 v2 pith:YKSXVIV4 submitted 2026-02-19 cs.AI cs.CL

Mechanistic Interpretability of Cognitive Complexity in LLMs via Linear Probing using Bloom's Taxonomy

Bianca Raimondi , Maurizio Gabbrielli This is my paper
classification cs.AI cs.CL
keywords Bloom's Taxonomylinear probingresidual streamcognitive complexitylarge language modelsmechanistic interpretabilityrepresentation geometryCognitive Separability Onset
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.

The paper attempts to show that Bloom's Taxonomy is not only a human grading scheme but also a structure the model actually builds inside itself. The authors extract activation vectors from the residual streams of four open-weights decoder-only LLMs and fit simple logistic-regression probes to classify each prompt's Bloom level. They report roughly 95% mean accuracy, with the useful signal appearing by around layer five and remaining stable deeper in the network. Probe errors are predominantly between adjacent Bloom levels, and the same geometric separation appears through centroid distances even without the probe. If the claim holds, cognitive complexity becomes a measurable, internal property of LLM representations rather than a label imposed from outside.

Core claim

The paper claims that the six Bloom levels (Remember, Understand, Apply, Analyze, Evaluate, Create) are linearly decodable from the final-token residual-stream activation of a Transformer. A multiclass logistic-regression probe trained per layer on a balanced set of 1,128 educational questions reaches approximately 95% mean accuracy across all levels and across all four tested models. The authors define the Cognitive Separability Onset (CSO) as the first layer where probe accuracy reaches 90%, which occurs near layer 5. They further show that pairwise Euclidean distances between adjacent-level centroids increase sharply around the same point, and that misclassifications stay mostly between n

What carries the argument

Layer-wise residual-stream activations x_l at the final token position, probed with a capacity-limited multiclass logistic-regression probe at every layer. The Cognitive Separability Onset (CSO) is defined as l* = min{l : A_l >= 0.90}, the earliest layer where probe accuracy is high enough to declare the representations linearly separable; centroid-based Euclidean distances D_l(k) = ||mu_{l,k+1} - mu_{l,k}|| provide a probe-independent check on the same geometric claim.

Load-bearing premise

The central claim rests on the assumption that the high probe accuracy comes from an internal representation of cognitive complexity rather than from surface cues like the specific instruction verbs in the prompts.

What would settle it

Train the same probes on a dataset where each prompt's Bloom label is held fixed but the governing verb is replaced with a verb from another level (e.g., 'Define' becomes 'Design'); if accuracy falls toward chance after the swap, the encoding was anchored to surface lexical form rather than to an abstract difficulty dimension. A second check is to test on a new, independently written set of Bloom-annotated prompts and see whether the ~95% accuracy and adjacent-error pattern persist.

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

If this is right

  • Probe accuracy can serve as a representation-level audit of task difficulty without fine-tuning, complementing behavioral benchmarks.
  • Because separability appears by roughly layer five and persists, downstream computation may consume an already-formed complexity estimate rather than recomputing it.
  • Adjacent-level confusion patterns suggest the learned representation respects the ordinal structure of the taxonomy, mirroring known ambiguity in human assessment.
  • Shallow text features alone (TF-IDF, sentence embeddings) reach lower accuracy, so the paper argues the signal is not merely lexical.

Where Pith is reading between the lines

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

  • The paper leaves open whether linear separability is causally used; a natural next step is to ablate the probe's most predictive dimensions and test whether generation performance degrades in a level-specific way.
  • If Bloom-level separability generalizes beyond educational questions, the same technique could be used to anticipate when an agent is about to engage in shallow recall instead of deeper reasoning.
  • The dataset's prompts likely share instruction verbs such as 'define' or 'design'; a verb-swap or cross-dataset test would decide whether the encoding is of abstract complexity or of surface phrasing.

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

5 major / 5 minor

Summary. The paper proposes to probe whether Bloom's Taxonomy cognitive levels (Remember through Create) are linearly decodable from the residual-stream activations of four open-weight LLMs. The authors extract final-token hidden states at every layer, train a multiclass logistic-regression probe per layer, and report that probe accuracy reaches ~95%, that the onset of high separability ('Cognitive Separability Onset', CSO) occurs around layer 5, and that probe errors are concentrated between adjacent Bloom levels. Control experiments with TF-IDF and sentence embeddings are used to argue that the signal is not purely lexical. The paper claims that cognitive complexity is encoded in a linearly accessible, geometrically ordered subspace of the model's representations.

Significance. If the central claim were fully supported, this would be a useful contribution: it would directly connect an established educational-psychology hierarchy to internal LLM representations and would provide a concrete, reproducible probing workflow. The paper also has strengths: it evaluates four architectures, uses a balanced dataset, and performs probe-independent centroid analyses. However, the current evidence does not yet establish that the probe is reading abstract cognitive complexity rather than instruction-verb surface cues, and several quantitative claims are not consistent with the paper's own exhibited data. These issues are load-bearing for the headline interpretation, so the paper cannot be accepted in its present form, but it is a reasonable basis for a major revision if the confounds are addressed experimentally.

major comments (5)
  1. [Abstract, §4, Fig. 2 vs. Fig. 4/8] The headline claim of '~95% mean accuracy' is not consistent with the confusion matrices shown. For Llama-3.1 at layer 5 (Fig. 4), the mean diagonal is 90.9%; for gemma-3-4b it is 90.9%; for DeepSeek-R1-Distill it is 93.0%; for Qwen3-4B it is 97.1%. The average across the four reported matrices is ~93%, not ~95%. No error bars, confidence intervals, or repeated-seed results are reported for any accuracy curve. This discrepancy is load-bearing because the ~95% figure is the paper's central quantitative claim. Please report means and standard deviations over multiple probe initializations/data splits and correct the abstract and text to match the actual numbers.
  2. [§4.2 (Control Experiments) and §5/§6] The controls do not rule out the lexical/instruction-verb confound. The dataset is constructed from prompts whose main discriminative feature is the instruction verb itself (Define, Explain, Use, Compare, Justify, Design). TF-IDF reaches 73% and sentence embeddings 61%—far above the 16.7% chance level—so a strong surface signal exists. The higher probe accuracy (~90-97%) could simply mean that the LLM's residual stream encodes these same lexical/semantic cues more richly, not that it represents an abstract cognitive-complexity dimension. The paper's own Discussion (Sec. 5) concedes that 'residual-stream separability could still partially reflect training-induced regularities in prompt formulation,' but the Conclusion (Sec. 6) asserts the effect is 'not merely a linguistic artifact.' These statements are in tension. A verb-scrambling control (e.g., swapping instruction verbs across levels
  3. [§4.1, Eq. (2)] The 'Cognitive Separability Onset' is defined as the first layer at which probe accuracy exceeds an arbitrarily chosen threshold τ = 0.90. The 'early onset' claim is therefore partly a construction of the threshold, not an independent discovery. If τ were set to 0.85 or 0.95, the reported CSO layer could shift. No sensitivity analysis over τ is provided, no confidence intervals for l* are given, and the same threshold is applied to all models without accounting for model-specific accuracy variability. To support RQ2, please report the full accuracy curve with uncertainty and characterize the onset via a model-agnostic criterion (e.g., derivative, comparison to a baseline), or show that the conclusion is robust across a range of thresholds.
  4. [§4.1.1, Eqs. (4)-(5), Fig. 5/9] The centroid-distance analysis is presented as 'probe-independent' geometric evidence, but the L2 distance between class centroids is not normalized. Residual-stream norms generally grow with depth across all inputs, so the observed monotonic increase in inter-centroid distance could simply reflect a global norm inflation rather than specific cognitive separation. The analysis should be controlled for overall norm growth, for example by dividing by the average within-class spread or by comparing distances to a null distribution obtained from random label assignments. Without such normalization, Figure 5/9 does not establish that Bloom levels are being 'progressively disentangled' in a meaningful geometric sense.
  5. [§4.1, Eq. (3)] The claim that the expected absolute error E[|ŷ−y|] ≈ 1 is not substantiated by the presented confusion matrices. For the Llama-3.1 matrix in Figure 4, the mean absolute error over all samples is approximately 0.12, and conditional on a misclassification it is approximately 1.3, not 1. Similar calculations can be made for the appendix matrices, where some rows show substantial non-adjacent errors (e.g., Create confused with Apply in multiple models). No statistical test is reported against a random-error baseline (which would give an expected distance of ~2.33 conditional on error). The 'adjacency effect' is a central piece of evidence for the ordinal-structure claim, so it needs to be computed explicitly and tested.
minor comments (5)
  1. [Figure 5 caption] The caption reads 'Figure 5: Figure 5: Layer-wise Euclidean distances...'—duplicate 'Figure 5' should be removed.
  2. [§3.3 Dataset] For EduQG, it is not specified whether the prompt includes the multiple-choice answer options. If options are included, the final-token representation and the probe could be influenced by the wording of the options. Please clarify the exact prompt format used for activation extraction.
  3. [Table 3] Table 3 shows a few arbitrary activation values and does not convey meaningful information. Consider replacing it with a description of the tensor shape and storage format, or removing it.
  4. [§3.5 Linear Probing] The logistic-regression details are underspecified: the 'default ℓ2 regularization' strength (C) and the feature-normalization procedure should be reported for reproducibility.
  5. [Throughout] The phrase '~95%' is used in the abstract, introduction, and results. If the actual mean accuracy is ~93% (or model-dependent), the wording should be updated consistently.

Circularity Check

0 steps flagged

No significant circularity: held-out linear probing and independent centroid analysis carry the central claim; only a minor background self-citation and threshold-defined CSO terminology warrant a low score.

full rationale

The paper's central result is an empirical supervised-probing claim: a logistic-regression probe is trained on residual-stream activations and evaluated on a stratified held-out test split, with labels taken from external educational datasets annotated with Bloom's Taxonomy. The target quantity is not fitted into the probe and then relabeled as a prediction, so the main claim is not circular in the fitted-input sense. The Cognitive Separability Onset (CSO, Eq. 2) is explicitly defined as the first layer with probe accuracy A_l >= 0.90; calling the resulting layer a 'discovery' is a threshold-dependent operational summary of the measured accuracy curve, not a derivation of a separate result from its own input. The centroid-distance analysis in Sec. 4.1.1 is computed independently of the linear classifier, so it does not reduce to the probe. The control experiments (TF-IDF 73%, sentence embeddings 61%) are genuine baseline comparisons, though they do not rule out the instruction-verb lexical confound; that is a correctness/validity limitation, not circularity. The Discussion even concedes that 'residual-stream separability could still partially reflect training-induced regularities in prompt formulation,' while the Conclusion asserts the opposite—an inconsistency, but not a circular reduction. The only self-citation is Raimondi et al. 2025 in Related Work, used as background support that MI can probe high-level constructs; it is not load-bearing for the reported experiments. Therefore no specific circular step can be exhibited, and the low score reflects only the minor background self-citation and the threshold-defined CSO terminology.

Axiom & Free-Parameter Ledger

6 free parameters · 5 axioms · 1 invented entities

The paper's central claim rests on the annotations of the source datasets, the architectural assumption about the residual stream, and the interpretational leap from linear decodability to a cognitive-complexity construct. The free parameters are mostly standard experimental choices, but the CSO threshold and the uniform l*=5 are load-bearing for the 'early onset' narrative. The only invented entity is the CSO name, which lacks independent evidence.

free parameters (6)
  • CSO threshold τ = 0.90
    Chosen by hand (Section 4.1) to define the layer at which probe accuracy is considered converged; setting τ directly affects the reported l*=5 and thus the 'early onset' finding.
  • CSO layer l* = 5
    The paper reports all analyses at layer 5 for all four models, even though the definition Eq. (2) would produce model-specific layers; this choice affects the headline comparison (Figures 3, 4, 8).
  • Downsampled class size = 188
    Controlled downsampling (Section 3.3) balances the dataset but the selection criteria from the two source datasets are not specified.
  • Train/test split ratio = 80/20 stratified
    Standard choice, but the random seed is not reported, affecting reproducibility.
  • Logistic regression regularization = default ℓ2
    Section 3.5 says 'default ℓ2 regularization with feature normalization' but the strength (C) is not given; this can affect probe accuracy.
  • Position for activation extraction = final token
    Section 3.4 chooses the final token; mean pooling is deferred to future work, so the result is conditional on this design choice.
axioms (5)
  • domain assumption Bloom labels in the aggregated dataset are correct and correspond to genuine cognitive demand
    Section 3.3 inherits annotations from Zaman et al. (2024) and EduQG; the paper provides no independent validation of these labels.
  • domain assumption The final token's residual stream state at layer l fully encodes the prompt's context before generation
    Section 3.4 justifies this via attention over the whole prompt, but it is an architectural assumption, not a proven equivalence.
  • standard math The residual stream update x_{l+1}=x_l+Attn(x_l)+MLP(x_l) is a sufficient model of a decoder-only transformer layer
    Section 3.4 presents this as a 'simplified formulation'; it is standard but omits layer norm and other details that may affect representation geometry.
  • domain assumption High linear probe accuracy implies the model has a representation of cognitive complexity rather than of the prompt's surface form
    This is the interpretational leap from Section 4 and the abstract; the control experiments do not exhaust the alternative (lexical cue) explanation.
  • domain assumption The two source datasets use a consistent Bloom scale and the aggregation does not introduce label noise
    Section 3.3 aggregates datasets with different question formats; no inter-annotator agreement or label-adjudication process is reported.
invented entities (1)
  • Cognitive Separability Onset (CSO) no independent evidence
    purpose: A named layer l* where probe accuracy first exceeds τ=0.90, used to claim that cognitive complexity emerges early in the forward pass
    CSO is defined entirely by the authors' threshold on their own probe accuracy curves; it has no external validation and is not shown to be invariant to threshold choice or dataset.

pith-pipeline@v1.3.0-alltime-deepseek · 11739 in / 17892 out tokens · 149162 ms · 2026-08-02T22:17:10.613519+00:00 · methodology

0 comments
read the original abstract

The black-box nature of Large Language Models necessitates novel evaluation frameworks that transcend surface-level performance metrics. This study investigates the internal neural representations of cognitive complexity using Bloom's Taxonomy as a hierarchical lens. By analyzing high-dimensional activation vectors from different LLMs, we probe whether different cognitive levels, ranging from basic recall (Remember) to abstract synthesis (Create), are linearly separable within the model's residual streams. Our results demonstrate that linear classifiers achieve approximately 95% mean accuracy across all Bloom levels, providing strong evidence that cognitive level is encoded in a linearly accessible subspace of the model's representations. These findings provide evidence that the model resolves the cognitive difficulty of a prompt early in the forward pass, with representations becoming increasingly separable across layers.

Figures

Figures reproduced from arXiv: 2602.17229 by Bianca Raimondi, Maurizio Gabbrielli.

Figure 1
Figure 1. Figure 1: Overview of the experimental pipeline, from dataset construction and activation extraction to layer-wise [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Layer-wise probe accuracy across all evaluated models. [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 4
Figure 4. Figure 4: Representative confusion matrix of the linear [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Figure 5: Layer-wise Euclidean distances [PITH_FULL_IMAGE:figures/full_fig_p006_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Confusion matrices for the control experiments. [PITH_FULL_IMAGE:figures/full_fig_p007_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Layer-wise linear probe accuracy for each Bloom level. [PITH_FULL_IMAGE:figures/full_fig_p011_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Confusion matrices of the linear probe for four representative models at the CSO layer [PITH_FULL_IMAGE:figures/full_fig_p011_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Layer-wise Euclidean distances between adjacent Bloom-level centroids for all evaluated models. [PITH_FULL_IMAGE:figures/full_fig_p011_9.png] view at source ↗

discussion (0)

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

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. From Execution to Education: A Bloom-Aligned Framework for Measuring Educational Control in LLMs

    cs.CL 2026-07 conditional novelty 6.5

    On 2,520 programming tasks, matched Qwen general and coder models reliably raise Bloom cognitive demand but fail to lower it, so execution skill does not imply educational control.

  2. UniSteer: Text-Guided Flow Matching in Activation Space for Versatile LLM Steering

    cs.CL 2026-05 unverdicted novelty 5.0

    UniSteer trains a conditional flow matching model on LLM residual-stream activations to enable text-conditioned steering and classification across multiple behavioral tasks.

Reference graph

Works this paper leans on

38 extracted references · 13 linked inside Pith · cited by 2 Pith papers

  1. [1]

    Lorin W Anderson and David R Krathwohl. 2001. A taxonomy for learning, teaching, and assessing: A revision of Bloom's taxonomy of educational objectives: complete edition. Addison Wesley Longman, Inc

    2001

  2. [2]

    Yonatan Belinkov. 2022. Probing classifiers: Promises, shortcomings, and advances. Computational Linguistics, 48(1):207--219

    2022

  3. [3]

    \'E lo \" se Benito-Rodriguez, Einar Urdshals, Jasmina Nasufi, and Nicky Pochinkov. 2025. Beyond tokens in language models: Interpreting activations through text genre chunks. arXiv preprint arXiv:2511.16540

    arXiv 2025

  4. [4]

    Leonard Bereska and Efstratios Gavves. 2024. Mechanistic interpretability for ai safety--a review. arXiv preprint arXiv:2404.14082

    Pith/arXiv arXiv 2024

  5. [5]

    Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, and 1 others. 2020. Language models are few-shot learners. Advances in neural information processing systems, 33:1877--1901

    2020

  6. [6]

    Devichand Budagam, Ashutosh Kumar, Mahsa Khoshnoodi, Sankalp KJ, Vinija Jain, and Aman Chadha. 2024. Hierarchical prompting taxonomy: A universal evaluation framework for large language models aligned with human cognitive principles. arXiv preprint arXiv:2406.12644

    Pith/arXiv arXiv 2024

  7. [7]

    Bartosz Cywi \'n ski, Emil Ryd, Senthooran Rajamanoharan, and Neel Nanda. 2025. Towards eliciting latent knowledge from llms with mechanistic interpretability. arXiv preprint arXiv:2505.14352

    Pith/arXiv arXiv 2025

  8. [8]

    Nelson Elhage, Neel Nanda, Catherine Olsson, Tom Henighan, Nicholas Joseph, Ben Mann, Amanda Askell, Yuntao Bai, Anna Chen, Tom Conerly, and 1 others. 2021. A mathematical framework for transformer circuits. Transformer Circuits Thread, 1(1):12

    2021

  9. [9]

    Sabina Elkins, Ekaterina Kochmar, Jackie CK Cheung, and Iulian Serban. 2024. How teachers can use large language models and bloom’s taxonomy to create educational quizzes. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 23084--23091

    2024

  10. [10]

    Sandeep Reddy Gantla. 2025. Exploring mechanistic interpretability in large language models: Challenges, approaches, and insights. In 2025 International Conference on Data Science, Agents & Artificial Intelligence (ICDSAAI), pages 1--8. IEEE

    2025

  11. [11]

    Amir Hadifar, Semere Kiros Bitew, Johannes Deleu, Chris Develder, and Thomas Demeester. 2023. Eduqg: A multi-format multiple-choice dataset for the educational domain. Ieee Access, 11:20885--20896

    2023

  12. [12]

    Dan Hendrycks, Collin Burns, Steven Basart, Andy Zou, Mantas Mazeika, Dawn Song, and Jacob Steinhardt. 2020. Measuring massive multitask language understanding. arXiv preprint arXiv:2009.03300

    Pith/arXiv arXiv 2020

  13. [13]

    Anne Herrmann-Werner, Teresa Festl-Wietek, Friederike Holderried, Lea Herschbach, Jan Griewatz, Ken Masters, Stephan Zipfel, and Moritz Mahling. 2024. Assessing chatgpt’s mastery of bloom’s taxonomy using psychosomatic medicine exam questions: mixed-methods study. Journal of medical Internet research, 26:e52113

    2024

  14. [14]

    John Hewitt and Percy Liang. 2019. Designing and interpreting probes with control tasks. arXiv preprint arXiv:1909.03368

    Pith/arXiv arXiv 2019

  15. [15]

    Mohammad Hmoud and Shaqour Ali. 2024. Aied bloom’s taxonomy: A proposed model for enhancing educational efficiency and effectiveness in the artificial intelligence era. The International Journal of Technologies in Learning, 31(2):111

    2024

  16. [16]

    Thomas Huber and Christina Niklaus. 2025. Llms meet bloom’s taxonomy: A cognitive view on large language model evaluations. In Proceedings of the 31st International Conference on Computational Linguistics, pages 5211--5246

    2025

  17. [17]

    Robert Jankowski, Filippo Radicchi, M Serrano, Mari \'a n Bogu \ n \'a , and Santo Fortunato. 2025. Task complexity shapes internal representations and robustness in neural networks. arXiv preprint arXiv:2508.05463

    Pith/arXiv arXiv 2025

  18. [18]

    Mingyu Jin, Qinkai Yu, Jingyuan Huang, Qingcheng Zeng, Zhenting Wang, Wenyue Hua, Haiyan Zhao, Kai Mei, Yanda Meng, Kaize Ding, and 1 others. 2025. Exploring concept depth: How large language models acquire knowledge and concept at different layers? In Proceedings of the 31st international conference on computational linguistics, pages 558--573

    2025

  19. [19]

    Been Kim, Martin Wattenberg, Justin Gilmer, Carrie Cai, James Wexler, Fernanda Viegas, and 1 others. 2018. Interpretability beyond feature attribution: Quantitative testing with concept activation vectors (tcav). In International conference on machine learning, pages 2668--2677. PMLR

    2018

  20. [20]

    Junsol Kim, James Evans, and Aaron Schein. 2025. Linear representations of political perspective emerge in large language models. arXiv preprint arXiv:2503.02080

    Pith/arXiv arXiv 2025

  21. [21]

    David R Krathwohl. 2002. A revision of bloom's taxonomy: An overview. Theory into practice, 41(4):212--218

    2002

  22. [22]

    Ramya Kumar, Dhruv Gulwani, and Sonit Singh. 2025. Automated analysis of learning outcomes and exam questions based on bloom's taxonomy. arXiv preprint arXiv:2511.10903

    arXiv 2025

  23. [23]

    Yifan Luo, Zhennan Zhou, and Bin Dong. 2025 a . Inversescope: Scalable activation inversion for interpreting large language models. arXiv preprint arXiv:2506.07406

    Pith/arXiv arXiv 2025

  24. [24]

    Yiming Luo, Ting Liu, Patrick Cheong-Iao Pang, Dana McKay, Ziqi Chen, George Buchanan, and Shanton Chang. 2025 b . Enhanced bloom's educational taxonomy for fostering information literacy in the era of large language models. arXiv preprint arXiv:2503.19434

    Pith/arXiv arXiv 2025

  25. [25]

    Chris Olah, Nick Cammarata, Ludwig Schubert, Gabriel Goh, Michael Petrov, and Shan Carter. 2020. Zoom in: An introduction to circuits. Distill, 5(3):e00024--001

    2020

  26. [26]

    Shahzad Qaiser and Ramsha Ali. 2018. Text mining: use of tf-idf to examine the relevance of words to documents. International journal of computer applications, 181(1):25--29

    2018

  27. [27]

    Daking Rai, Yilun Zhou, Shi Feng, Abulhair Saparov, and Ziyu Yao. 2024. A practical review of mechanistic interpretability for transformer-based language models. arXiv preprint arXiv:2407.02646

    arXiv 2024

  28. [28]

    Bianca Raimondi, Daniela Dalbagno, and Maurizio Gabbrielli. 2025. Analysing moral bias in finetuned llms through mechanistic interpretability. arXiv preprint arXiv:2510.12229

    Pith/arXiv arXiv 2025

  29. [29]

    Nils Reimers and Iryna Gurevych. 2019. https://arxiv.org/abs/1908.10084 Sentence-bert: Sentence embeddings using siamese bert-networks . In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing. Association for Computational Linguistics

    Pith/arXiv arXiv 2019

  30. [30]

    Dong Shu, Xuansheng Wu, Haiyan Zhao, Daking Rai, Ziyu Yao, Ninghao Liu, and Mengnan Du. 2025. A survey on sparse autoencoders: Interpreting the internal mechanisms of large language models. arXiv preprint arXiv:2503.05613

    arXiv 2025

  31. [31]

    Katharina Simbeck and Mariam Mahran. 2025. Mechanistic interpretability with saes: Probing religion, violence, and geography in large language models. arXiv preprint arXiv:2509.17665

    arXiv 2025

  32. [32]

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, ukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. Advances in neural information processing systems, 30

    2017

  33. [33]

    Yongan Yu, Mengqian Wu, Yiran Lin, and Nikki G Lobczowski. 2025. Think: Can large language models think-aloud? arXiv preprint arXiv:2505.20184

    Pith/arXiv arXiv 2025

  34. [34]

    Khandoker Ashik Uz Zaman, Ashraful Islam, Yusuf Mahbubul Islam, and Md Abu Sayed. 2024. Dataset of computer science course queries from students: Categorized and scored according to bloom's taxonomy. Data in Brief, 53:110109

    2024

  35. [35]

    Haiyan Zhao, Fan Yang, Bo Shen, Himabindu Lakkaraju, and Mengnan Du. 2024. Towards uncovering how large language model works: An explainability perspective. arXiv preprint arXiv:2402.10688

    Pith/arXiv arXiv 2024

  36. [36]

    Maria-Eleni Zoumpoulidi, Georgios Paraskevopoulos, and Alexandros Potamianos. 2025. Bloomwise: enhancing problem-solving capabilities of large language models using bloom’s-taxonomy-inspired prompts. In Proceedings of The 3rd Workshop on Mathematical Natural Language Processing (MathNLP 2025), pages 34--49

    2025

  37. [37]

    online" 'onlinestring :=

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint eprinttype howpublished institution journal key month note number organization pages publisher school series title type volume year doi pubmed url lastchecked label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block STRING...

  38. [38]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...