Pith. sign in

REVIEW 5 major objections 6 minor 42 references

Functional Abstraction of Knowledge Recall in Large Language Models

T0 review · 5 major / 6 minor · reviewed 2026-08-16 · deepseek-v4-flash

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

Pith's one-line read Knowledge recall in LLMs is a function call in activation space.

desk verdict A clean functional framing of knowledge recall that mostly restages known results; the editing trick works, but the independence claim is under-supported without controls. read the letter →

arxiv 2504.14496 v1 pith:3TDPXFSX submitted 2025-04-20 cs.CL

Zijian Wang , Chang Xu This is my paper · ORCID
classification cs.CL
keywords knowledgerecallactivationpatchingfunctionalabstractionrepresentationalinterpretabilityeditinginterchangeinterventionfactualassociationtransformer
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

This paper tries to establish that when a transformer LLM recalls a factual association, its hidden activations implement a function call: the subject token's activation is the input argument, the relation token's activation is the function body, and the final-token activation is the return value. Using a patching-based knowledge-scoring algorithm, the authors locate activation vectors that carry subject, relation, and object information and show that they are localized along specific token positions and layers. Counter-knowledge interchange tests then show that replacing a single identified activation vector with the corresponding vector from another query flips the model's answer to the other query's object most of the time, with roughly 92-100% interchange accuracy depending on the component. The result matters because it offers a concrete functional abstraction of knowledge recall that can be applied to knowledge editing: patching the object vector of a new fact into a context prompt raises the edited fact's success from near zero to about 65-69% on their tests.

What carries the argument

The load-bearing machinery is activation patching used as both a probe and an intervention. For probing, the paper defines knowledge-encoding scores for subject, relation, and object by corrupting the relevant token embeddings with Gaussian noise, then restoring each corrupted activation one at a time and measuring the recovered object probability; high scores mark the activation vectors that carry each piece of knowledge. For verification, the located vectors are interchanged between a source query and a reference query, and the predicted object is compared with the object that would follow from the corresponding textual swap. The functional abstraction itself, (subject, relation) mapping to object, is the identity that organizes these operations and gives the components their names.

What would settle it

Take a subject-only interchange pair and patch only a random half of the subject token positions; if the reference object still wins as often as with full subject patching, the 'subject vector' is not the local carrier claimed. Conversely, patch a late-layer final-token vector the scorer marks as low-OES; if it flips the predicted object at nearly the same rate as the high-OES vector, the scoring step is not identifying a specific return value.

Watch

Extended reading notes

Core claim

The paper asserts that knowledge recall in transformer LLMs is, at the level of activations, a function execution with three separable components: subject-token activations in early layers act as the input argument, relation-token activations in early layers act as the function body, and final-token activations in middle-to-late layers act as the return value. The evidence is a knowledge-scoring step that localizes these vectors, followed by interchange interventions: swapping only the subject vector from another query changes the prediction to the other query's object with about 93-97% accuracy; swapping the relation vector gives about 92-95%; swapping the object vector gives about 99-100%; swapping both subject and relation gives about 89-93% on the filtered dataset. The authors conclude that these activation vectors function as distinct, separable components in the knowledge recall process rather than as a distributed tangle.

Load-bearing premise

The argument assumes that the knowledge-scoring and interchange procedures isolate a single activation vector that completely and independently carries the subject, relation, or object of a fact; if those representations are distributed across many interacting vectors, high interchange accuracy would not by itself prove the claimed independence.

Editorial extensions

If this is right

  • Factual knowledge can be edited without weight updates by patching the object vector into a context prompt: the paper reports the edited fact's efficacy score rising from about 0.14-0.19 without patching to about 0.66-0.69 with patching.
  • The same component localization appears under two prompt templates with different subject-relation order, implying the functional roles are tied to token content rather than to template position.
  • Dual interchange composes a subject from one fact and a relation from another, so if the paper is right, factual recall is modular enough to support compositional knowledge operations at the activation level.
  • The knowledge-scoring algorithm gives an interpretability tool that locates factual associations in comparable transformers without training probes or parameter access.

Reading between the lines

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

  • Editorial inference: Because object-only interchange is nearly perfect while subject- and relation-only interchange are slightly lower, the return value may be more cleanly localized than the input argument and function body; a natural test is whether the subject vector's residual error shrinks as more subject-token positions are patched.
  • Editorial inference: The paper tests six relation types on two 7-billion-parameter models; a stress test for the framework would be a multi-hop or compositional fact, where a relation function must be applied twice, to see whether the same component boundaries hold.
  • Editorial inference: The knowledge-editing results are short-term and measured on immediate queries; an untested extension is whether repeated patched prompting can consolidate the new object into long-term recall.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

5 major / 6 minor

Summary. This paper proposes that knowledge recall in transformer LLMs can be abstracted as a function execution in activation space, with subject, relation, and object activation vectors acting as independent functional components (input argument, function body, and return value). The authors introduce a patching-based knowledge-scoring algorithm (SES/RES/OES) to localize these vectors, validate them via counter-knowledge interchange experiments, and apply the resulting decomposition to a contextual knowledge editing method. Experiments on Llama2-7b and Mistral-7b use filtered relational knowledge data and two prompt templates; the main evidence for the functional decomposition is the high interchange accuracy reported in Table 1, and the editing results are reported in Table 2.

Significance. The proposed abstraction is attractive: if validated, it would give a sparse, human-readable functional decomposition of factual recall and a cheap, non-parametric knowledge editing intervention. The paper has some strengths: it tests two 7B models, uses two prompt templates with different subject-relation orders, reports score heatmaps and layer/token locality analyses, and provides detailed appendix descriptions of data filtering and interchange construction. The central claim is falsifiable in principle, and the interchange protocol is a standard causal-intervention tool. However, the current evidence does not establish the independence claim because several controls are missing and one of the interventions is, by construction, close to the model's output head. The significance of the paper therefore depends on whether the authors can add the missing controls and address the selection/evaluation overlap.

major comments (5)
  1. [§3.1, §4.1.3, Table 1] The object-only interchange result in Table 1 (99–100%) does not provide independent evidence for a separable 'return value' component. Section 4.1.3 defines the object vector as the final-token activation in layers 15–31, and Section 3.1 states that the output distribution is Softmax(phi(h_{L-1,n-1})), so replacing exactly that vector with the reference vector necessarily overwrites the primary input to the classification head. A random vector of matched norm at the same position could plausibly produce a similarly large effect. I ask for two controls: (i) patch the object position with a non-reference activation vector, such as random noise or a vector from an unrelated example, and report the resulting accuracy; and (ii) patch only a subset of the object layer range, for example layers 15–20 and 21–31 separately, to show that the effect is not simply due to overwriting the head input.
  2. [§4.1.3, §4.2, Appendix A.1.1] The layer ranges for subject (0–14) and relation (0–10) vectors are derived from the SES/RES/OES heatmaps in Figures 4–6, which are computed on the same filtered dataset that is later used for the interchange evaluation in Table 1. Because the component definitions and the evaluation labels come from the same facts, the high interchange accuracy is partly a consequence of circular selection. Please provide a held-out evaluation, for example selecting components on a development split and evaluating on a disjoint test split; report bootstrap or split-level variance for Table 1; and include a random-vector baseline to quantify how much of the interchange effect is due to the locality of computation rather than to the specific semantic content of the vector.
  3. [Table 1, §4.2] The dual interchange accuracies (88.5–92.7%) are consistently 5–10 points below the object-only accuracies (99–100%) and also lower than the subject-only and relation-only accuracies for most rows. Under the proposed independence hypothesis, replacing both input arguments should not degrade performance below the single-component interventions unless the two vectors interact. This gap is direct evidence against the 'distinct, separable components' claim as stated, and it is not discussed. Please analyze the failure cases, report whether the gap is statistically significant, and compare the observed dual accuracy with the accuracy predicted under an independence model, such as the product of single-intervention accuracies.
  4. [§5.1] The knowledge editing method is described inconsistently. The first paragraph says to patch 'the late layer activation vectors of the previous token of objects, e.g., the "is" token' with the new object mean activation vector, while the next paragraph says to extract 'the mean activation vector at the early layers (0–14) from the object tokens of new knowledge prompting.' This contradicts Section 4.1.3, where object vectors are defined in layers 15–31 at the final token, and it is unclear whether the patched location is the token before the object, an object token, or the final token. Please clarify the exact token positions and layer ranges used for both extraction and patching, and justify why the editing setup differs from the object-vector definition used in the scoring and interchange experiments.
  5. [Appendix A.1.1, Table 3] All experiments are run on a filtered subset containing only facts that the model already predicts correctly with both templates, for example 155–547 of 836 examples for Landmark-Country. This filter may select for facts whose recall is unusually easy and localized, so the high interchange accuracies may not generalize to the full knowledge distribution. Please report the filtering rate for each relation and, if feasible, run the scoring and at least the subject-only and relation-only interchanges on the unfiltered data or on a random sample stratified by model confidence, to bound the selection effect.
minor comments (6)
  1. [Appendix A.3] There are typos in the appendix headings and text: 'Activation V ectors Interchange' (A.3.2) and 'a the relation' (A.3.1) should be corrected.
  2. [Table 1, Table 3] The relation name is inconsistent: Table 1 uses 'Fruit-Color' while Table 3 lists 'Food-Color'; please reconcile.
  3. [Section 4, Section 4.1.3, Figure 4] The template in Section 4 ('Given <subject>, it's <relation> is') differs from the template in Section 4.1.3 and Figure 4 ('Given <subject>, its <relation> is'); standardize the phrasing.
  4. [Section 5.2, Appendix A.4] Section 5.2 says 3000 new triplets are constructed, while Appendix A.4 describes replacing objects for each factual triple; please specify whether 3000 is the total or per relation and report the number of examples behind Table 2.
  5. [General] The paper does not state whether code or data will be released; for reproducibility, please include a link or an availability statement.
  6. [Figures 4–6] The captions of Figures 4–6 do not define the color mapping for SES, RES, and OES; please add a legend or describe the color scheme in each caption.

Circularity Check

2 steps flagged · score 6.0 of 10

Object-only interchange is an architectural tautology, and component layer ranges are selected on the same filtered facts used for validation.

  1. self definitional [Section 3.1 (output definition), Section 4.1.3 (object vector definition), Section 3.3 (object-only intervention), Table 1]
    "P (o|I(s,r )) =Softmax(ϕ(hL−1,n−1)) ... Object Vectors: Activation vectors located in late layers (15-31) in the final token, serving as return values. ... 3. Object-only Intervention. Only replace the source object vector with the reference one to get object probability Po(˜o|I(s1,r 1), do(βO1 = βO2))"

    The model's output distribution is defined as Softmax(phi(h_{L-1,n-1})), i.e., a function of the final-token activation at the last layer. The 'object vector' component is then defined as the activation vector in the final token at late layers (15-31). An object-only interchange simply replaces that very vector with the reference run's final-token vector, so the classification head receives the reference representation by construction. Near-100% object-only interchange accuracy is therefore expected from the architecture and cannot serve as evidence that the vector is an independent 'return value' component; the test is a definitional identity, not a controlled causal decomposition.

  2. fitted input called prediction [Section 4.1.3 (component formulation), Section 4.2 (counter-knowledge testing), Appendix A.1.1 (filtered dataset)]
    "Based on the above findings, we define activation vectors that can be treated as components of a knowledge recall function. • Subject Vectors: Activation vectors located in early layers (0-14) within the subject tokens range ... • Relation Vectors: Activation vectors located in early layers (0-10) within the relation tokens range ... Our subsequent knowledge scoring and counter-knowledge test experiments are all conducted on this filtered dataset."

    The subject and relation component locations are chosen from the SES/RES/OES score heatmaps computed on the same filtered knowledge triples that are then used to compute the interchange accuracies in Table 1. There is no held-out split or separate validation set. Thus the 'prediction' that subject and relation vectors in those ranges independently drive object recall is partly a re-description of the selection criterion: the layers and token positions were fitted to the same labels used for evaluation. The dual-interchange accuracy (88-93%) also falls several points below object-only accuracy, indicating unaddressed interaction, so the claim of cleanly separable components is not independently established.

full rationale

The paper's central claim is that subject, relation, and object activation vectors behave as independent functional components in knowledge recall. The strongest evidence, object-only interchange, is circular by construction because the object vector is defined as the final-token late-layer activation, which is exactly the vector the classification head consumes. Subject-only and relation-only interchanges are not strictly tautological, but their component definitions are derived from the same filtered dataset on which they are evaluated, with no held-out data or random-vector controls, so those results are partially fitted rather than predictive. The paper cites external methods such as causal mediation analysis and interchange interventions, and those citations are not load-bearing self-citations; the same authors' prior work is cited only in related work, not as the justification for the functional decomposition. The editing application is downstream and does not by itself create circularity. Overall, the paper contains real empirical work, but one of its four headline validations reduces by construction, and the other two are selected on the evaluation data, warranting a moderate circularity score of 6.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

The central claim rests primarily on domain assumptions about how to interpret activation patching scores as knowledge content and how to treat token-position vectors as independent function components. The layer ranges and thresholds are hand-chosen based on the same data used for evaluation, which raises concerns about post hoc selection. No fundamentally new entities are postulated; the 'input argument', 'function body', and 'return value' are labels applied to existing activation vectors.

free parameters (4)
  • High-score threshold = 0.05
    Used in Section 4.1 (Figure 5) to classify 'high-scoring' activation vectors. This threshold is arbitrary; changing it changes which vectors are considered components.
  • Layer ranges for functional components = subject 0-14, relation 0-10, object 15-31
    Defined in Section 4.1.3 by inspecting score locality plots. These boundaries are fit to the observed data and then used in the counter-knowledge tests, creating post hoc selection.
  • Noise variance for embedding corruption = 5 times standard deviation of token embeddings
    Set in Appendix A.2.1. This noise level is chosen by hand and affects which activations are considered corrupted.
  • Mean activation vector layer range for editing = early layers 0-14
    Used in Section 5.1 to extract the new object vector. This range is hand-chosen and not justified independently of the results.
assumptions (4)
  • domain assumption Knowledge recall is equivalent to a function execution where the relation is a transformation between entities.
    This is the central hypothesis stated in Section 1 and Section 3. It is not derived from data; it is a modeling choice. The paper cites Bordes et al. (2013) to argue relationships are transformations, but that is not evidence about LLM internals.
  • domain assumption The causal mediation score (difference between restored and corrupted probability) measures the knowledge content encoded in an activation vector.
    Used in Section 3.2. This is a common assumption in activation patching literature, but it is an assumption that the score is a faithful proxy for knowledge encoding.
  • domain assumption Interchange accuracy, defined as the match between text intervention and vector intervention, indicates that the vector faithfully encodes the knowledge.
    Section 3.3.1. This assumes the model's output is determined by these vectors in a way that matches the causal effect of the text.
  • domain assumption The next-token probability of the first object token is a sufficient measure of knowledge recall.
    Section A.2.2 and Section 3.1. The model may generate multi-token objects; here only the first token is used, which is a simplification that may miss part of the recall behavior.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Functional Abstraction of Knowledge Recall in Large Language Models." pith.science (2026). https://pith.science/paper/3TDPXFSX

@misc{pith2026250414496,
  author       = {Pith},
  title        = {Pith review of: Functional Abstraction of Knowledge Recall in Large Language Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/3TDPXFSX}},
  note         = {Machine review of arXiv:2504.14496}
}
read the original abstract

Pre-trained transformer large language models (LLMs) demonstrate strong knowledge recall capabilities. This paper investigates the knowledge recall mechanism in LLMs by abstracting it into a functional structure. We propose that during knowledge recall, the model's hidden activation space implicitly entails a function execution process where specific activation vectors align with functional components (Input argument, Function body, and Return values). Specifically, activation vectors of relation-related tokens define a mapping function from subjects to objects, with subject-related token activations serving as input arguments and object-related token activations as return values. For experimental verification, we first design a patching-based knowledge-scoring algorithm to identify knowledge-aware activation vectors as independent functional components. Then, we conduct counter-knowledge testing to examine the independent functional effects of each component on knowledge recall outcomes. From this functional perspective, we improve the contextual knowledge editing approach augmented by activation patching. By rewriting incoherent activations in context, we enable improved short-term memory retention for new knowledge prompting.

Figures

Figures reproduced from arXiv: 2504.14496 by the authors.

Figure 1
Figure 1. Illustration of our abstraction frame￾work. In a knowledge recall process, i.e., (subject, relation)→object, we find that knowledge-related rep￾resentations are locally distributed, and they are well￾aligned with functional components. model. The mainstream is module-based mechani￾cal approaches, which typically attribute the stor￾age and retrieval of knowledge to specific mod￾ules by identifying important model par… view at source ↗
Figure 2
Figure 2. Illustration of two activation patching opera [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Illustration of the counter-knowledge testing experiments. If the prediction of the object after the [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: Scores heat map visualization. The left col [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 5
Figure 5. Figure 5: Locality along token positions. This bar graph [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]
Figure 6
Figure 6. Figure 6: Locality along layer positions. The plots of [PITH_FULL_IMAGE:figures/full_fig_p006_6.png]
Figure 7
Figure 7. Figure 7: The plots of intervention effects of textual and [PITH_FULL_IMAGE:figures/full_fig_p007_7.png]
Figure 8
Figure 8. Figure 8: Figure (a) shows the knowledge conflict in [PITH_FULL_IMAGE:figures/full_fig_p008_8.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

42 extracted references · 3 canonical work pages

  1. [1]

    Antoine Bordes, Nicolas Usunier, Alberto Garcia-Duran, Jason Weston, and Oksana Yakhnenko. 2013. Translating embeddings for modeling multi-relational data. Advances in neural information processing systems, 26

    2013

  2. [2]

    Collin Burns, Haotian Ye, Dan Klein, and Jacob Steinhardt. 2022. Discovering latent knowledge in language models without supervision. arXiv preprint arXiv:2212.03827

    arXiv 2022

  3. [3]

    Roi Cohen, Eden Biran, Ori Yoran, Amir Globerson, and Mor Geva. 2024. Evaluating the ripple effects of knowledge editing in language models. Transactions of the Association for Computational Linguistics, 12:283--298

    2024

  4. [4]

    Arthur Conmy, Augustine Mavor-Parker, Aengus Lynch, Stefan Heimersheim, and Adri \`a Garriga-Alonso. 2023. Towards automated circuit discovery for mechanistic interpretability. Advances in Neural Information Processing Systems, 36:16318--16352

    2023

  5. [5]

    Damai Dai, Li Dong, Yaru Hao, Zhifang Sui, Baobao Chang, and Furu Wei. 2021. Knowledge neurons in pretrained transformers. arXiv preprint arXiv:2104.08696

    arXiv 2021

  6. [6]

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

    2021

  7. [7]

    Atticus Geiger, Hanson Lu, Thomas Icard, and Christopher Potts. 2021. Causal abstractions of neural networks. Advances in Neural Information Processing Systems, 34:9574--9586

    2021

  8. [8]

    Atticus Geiger, Zhengxuan Wu, Christopher Potts, Thomas Icard, and Noah Goodman. 2024. Finding alignments between interpretable causal variables and distributed neural representations. In Causal Learning and Reasoning, pages 160--187. PMLR

    2024

Show all 42 references
  1. [9]

    Mor Geva, Jasmijn Bastings, Katja Filippova, and Amir Globerson. 2023. Dissecting recall of factual associations in auto-regressive language models. arXiv preprint arXiv:2304.14767

    2023 arXiv

  2. [10]

    Asma Ghandeharioun, Avi Caciularu, Adam Pearce, Lucas Dixon, and Mor Geva. 2024. Patchscope: A unifying framework for inspecting hidden representations of language models. arXiv preprint arXiv:2401.06102

    2024 arXiv

  3. [11]

    Michael Hanna, Ollie Liu, and Alexandre Variengien. 2024. How does gpt-2 compute greater-than?: Interpreting mathematical abilities in a pre-trained language model. Advances in Neural Information Processing Systems, 36

    2024

  4. [12]

    Peter Hase, Mohit Bansal, Been Kim, and Asma Ghandeharioun. 2024. Does localization inform editing? surprising differences in causality-based localization vs. knowledge editing in language models. Advances in Neural Information Processing Systems, 36

    2024

  5. [13]

    Roee Hendel, Mor Geva, and Amir Globerson. 2023. In-context learning creates task vectors. arXiv preprint arXiv:2310.15916

    2023 arXiv

  6. [14]

    Dan Hendrycks and Mantas Mazeika. 2022. X-risk analysis for ai research. arXiv preprint arXiv:2206.05862

    2022 arXiv

  7. [15]

    Evan Hernandez, Arnab Sen Sharma, Tal Haklay, Kevin Meng, Martin Wattenberg, Jacob Andreas, Yonatan Belinkov, and David Bau. 2023. Linearity of relation decoding in transformer language models. arXiv preprint arXiv:2308.09124

    2023 arXiv

  8. [16]

    Albert Q Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lucile Saulnier, et al. 2023. Mistral 7b. arXiv preprint arXiv:2310.06825

    2023 arXiv

  9. [17]

    Zhengbao Jiang, Frank F Xu, Jun Araki, and Graham Neubig. 2020. How can we know what language models know? Transactions of the Association for Computational Linguistics, 8:423--438

    2020

  10. [18]

    Nora Kassner, Oyvind Tafjord, Ashish Sabharwal, Kyle Richardson, Hinrich Schuetze, and Peter Clark. 2023. https://doi.org/10.18653/v1/2023.emnlp-main.877 Language models with rationality . In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processin...

    2023 doi

  11. [19]

    Kenneth Li, Oam Patel, Fernanda Vi \'e gas, Hanspeter Pfister, and Martin Wattenberg. 2024. Inference-time intervention: Eliciting truthful answers from a language model. Advances in Neural Information Processing Systems, 36

    2024

  12. [20]

    Tom Lieberum, Matthew Rahtz, J \'a nos Kram \'a r, Neel Nanda, Geoffrey Irving, Rohin Shah, and Vladimir Mikulik. 2023. Does circuit analysis interpretability scale? evidence from multiple choice capabilities in chinchilla. arXiv preprint arXiv:2307.09458

    2023 arXiv

  13. [21]

    Aman Madaan, Niket Tandon, Peter Clark, and Yiming Yang. 2022. Memory-assisted prompt editing to improve gpt-3 after deployment. arXiv preprint arXiv:2201.06009

    2022 arXiv

  14. [22]

    Kevin Meng, David Bau, Alex Andonian, and Yonatan Belinkov. 2022 a . Locating and editing factual associations in gpt. Advances in Neural Information Processing Systems, 35:17359--17372

    2022

  15. [23]

    Kevin Meng, Arnab Sen Sharma, Alex Andonian, Yonatan Belinkov, and David Bau. 2022 b . Mass-editing memory in a transformer. arXiv preprint arXiv:2210.07229

    2022 arXiv

  16. [24]

    Jack Merullo, Carsten Eickhoff, and Ellie Pavlick. 2023. A mechanism for solving relational tasks in transformer language models

    2023

  17. [25]

    Judea Pearl. 2022. Direct and indirect effects. In Probabilistic and causal inference: the works of Judea Pearl, pages 373--392

    2022

  18. [26]

    Fabio Petroni, Tim Rockt \"a schel, Patrick Lewis, Anton Bakhtin, Yuxiang Wu, Alexander H Miller, and Sebastian Riedel. 2019. Language models as knowledge bases? arXiv preprint arXiv:1909.01066

    2019 arXiv

  19. [27]

    Adam Roberts, Colin Raffel, and Noam Shazeer. 2020. How much knowledge can you pack into the parameters of a language model? arXiv preprint arXiv:2002.08910

    2020 arXiv

  20. [28]

    Eric Todd, Millicent L Li, Arnab Sen Sharma, Aaron Mueller, Byron C Wallace, and David Bau. 2023. Function vectors in large language models. arXiv preprint arXiv:2310.15213

    2023 arXiv

  21. [29]

    Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. 2023. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288

    2023 arXiv

  22. [30]

    Jesse Vig, Sebastian Gehrmann, Yonatan Belinkov, Sharon Qian, Daniel Nevo, Yaron Singer, and Stuart Shieber. 2020. Investigating gender bias in language models using causal mediation analysis. Advances in neural information processing systems, 33:12388--12401

    2020

  23. [31]

    Dimitri von R \"u tte, Sotiris Anagnostidis, Gregor Bachmann, and Thomas Hofmann. 2024. A language model's guide through latent space. arXiv preprint arXiv:2402.14433

    2024 arXiv

  24. [32]

    Cunxiang Wang, Xiaoze Liu, Yuanhao Yue, Xiangru Tang, Tianhang Zhang, Cheng Jiayang, Yunzhi Yao, Wenyang Gao, Xuming Hu, Zehan Qi, et al. 2023. Survey on factuality in large language models: Knowledge, retrieval and domain-specificity. arXiv preprint arXiv:2310.07521

    2023 arXiv

  25. [33]

    Kevin Wang, Alexandre Variengien, Arthur Conmy, Buck Shlegeris, and Jacob Steinhardt. 2022. Interpretability in the wild: a circuit for indirect object identification in gpt-2 small, 2022. URL https://arxiv. org/abs/2211.00593

    2022 arXiv

  26. [34]

    Zijian Wang, Britney White, and Chang Xu. 2024. Locating and extracting relational concepts in large language models. arXiv preprint arXiv:2406.13184

    2024 arXiv

  27. [35]

    Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumond, Clement Delangue, Anthony Moi, Pierric Cistac, Tim Rault, R \'e mi Louf, Morgan Funtowicz, et al. 2020. Transformers: State-of-the-art natural language processing. In Proceedings of the 2020 conference on empirical met...

    2020

  28. [36]

    Yunzhi Yao, Ningyu Zhang, Zekun Xi, Mengru Wang, Ziwen Xu, Shumin Deng, and Huajun Chen. 2024. Knowledge circuits in pretrained transformers. arXiv preprint arXiv:2405.17969

    2024 arXiv

  29. [37]

    Xunjian Yin, Jin Jiang, Liming Yang, and Xiaojun Wan. 2024. History matters: Temporal knowledge editing in large language model. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 19413--19421

    2024

  30. [38]

    Fred Zhang and Neel Nanda. 2023. Towards best practices of activation patching in language models: Metrics and methods. arXiv preprint arXiv:2309.16042

    2023 arXiv

  31. [39]

    Ce Zheng, Lei Li, Qingxiu Dong, Yuxuan Fan, Zhiyong Wu, Jingjing Xu, and Baobao Chang. 2023. Can we edit factual knowledge by in-context learning? arXiv preprint arXiv:2305.12740

    2023 arXiv

  32. [40]

    Andy Zou, Long Phan, Sarah Chen, James Campbell, Phillip Guo, Richard Ren, Alexander Pan, Xuwang Yin, Mantas Mazeika, Ann-Kathrin Dombrowski, et al. 2023. Representation engineering: A top-down approach to ai transparency. arXiv preprint arXiv:2310.01405

    2023 arXiv

  33. [41]

    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...

  34. [42]

    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 gl...

Pith tools

Reviewed August 16, 2026 · model on record in the stance chip above.