pith. machine review for the scientific record. sign in

arxiv: 2604.17761 · v1 · submitted 2026-04-20 · 💻 cs.AI · cs.CL

Recognition: unknown

Contrastive Attribution in the Wild: An Interpretability Analysis of LLM Failures on Realistic Benchmarks

Rongyuan Tan , Jue Zhang , Zhuozhao Li , Qingwei Lin , Saravan Rajmohan , Dongmei Zhang
Authors on Pith no claims yet

Pith reviewed 2026-05-10 05:08 UTC · model grok-4.3

classification 💻 cs.AI cs.CL
keywords LLM interpretabilitycontrastive attributionLRPfailure analysisbenchmarkstoken attributionmodel debugging
0
0 comments X

The pith

Token-level contrastive attribution using LRP yields informative signals for some LLM failures on realistic benchmarks but is not universally applicable.

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

The paper tests whether contrastive attribution can explain why large language models produce wrong outputs on standard benchmarks instead of toy problems. It does this by tracing the logit gap between an incorrect token and a correct alternative back through the model using layer-wise relevance propagation. An efficient extension allows building cross-layer graphs even for long inputs. Results from comparisons across datasets, model sizes, and checkpoints show the approach highlights useful patterns in certain errors but leaves many others unexplained. Readers would care because it sets practical limits on one popular interpretability tool for real LLM debugging.

Core claim

We formulate failure analysis as contrastive attribution, attributing the logit difference between an incorrect output token and a correct alternative to input tokens and internal model states, and introduce an efficient extension that enables construction of cross-layer attribution graphs for long-context inputs. Our systematic empirical study across benchmarks shows that this token-level contrastive attribution can yield informative signals in some failure cases, but is not universally applicable.

What carries the argument

Contrastive attribution, which traces the logit difference between a wrong output token and a correct alternative back to input tokens and states via LRP rules, extended to cross-layer graphs for long sequences.

If this is right

  • Attribution patterns differ systematically across datasets, model sizes, and training checkpoints.
  • In applicable failure cases the method can isolate specific input tokens or internal states driving the error.
  • The approach has clear limits so it cannot replace broader suites of diagnostic tools for LLM analysis.
  • Efficient cross-layer graph construction makes the technique feasible for realistic long-context benchmarks.

Where Pith is reading between the lines

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

  • Developers could track how attribution quality evolves across training checkpoints to decide when interpretability tools become reliable.
  • Combining contrastive attribution with other methods might cover the failure cases where LRP signals stay weak.
  • The observed variability suggests benchmark design should include failure subsets where attribution is known to work well.

Load-bearing premise

The contrastive logit difference and LRP propagation rules accurately reflect the model's causal decision process rather than method-specific artifacts or correlations.

What would settle it

Compare attribution scores to results from causal interventions such as ablating the highest-scoring input tokens and checking whether the model's output flips as the scores would predict.

Figures

Figures reproduced from arXiv: 2604.17761 by Dongmei Zhang, Jue Zhang, Qingwei Lin, Rongyuan Tan, Saravan Rajmohan, Zhuozhao Li.

Figure 1
Figure 1. Figure 1: Distribution of attribution outcomes (top) and failure patterns (bottom) across benchmarks. [PITH_FULL_IMAGE:figures/full_fig_p008_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Examples of input attribution heatmaps. (a) URT: Qwen3-0.6B underweights the [PITH_FULL_IMAGE:figures/full_fig_p009_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Sample ablated attribution graph for an NC-IA + M-AG case; the expanded attribution [PITH_FULL_IMAGE:figures/full_fig_p011_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Logit difference comparisons between Qwen3-0.6B and larger models (1.7B and 4B) on [PITH_FULL_IMAGE:figures/full_fig_p012_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Input attribution relevance score breakdown across prompt segments. [PITH_FULL_IMAGE:figures/full_fig_p012_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Evolution of logit differences across training checkpoints for Olmo-3-7B-Think on IFEval [PITH_FULL_IMAGE:figures/full_fig_p013_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Input attribution relevance score breakdown across training checkpoints by prompt [PITH_FULL_IMAGE:figures/full_fig_p014_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Expanded attribution graph for case in Figure 3. [PITH_FULL_IMAGE:figures/full_fig_p035_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Expanded attribution graph for an example case from MATH, with biases embedded in [PITH_FULL_IMAGE:figures/full_fig_p037_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Normalized relevance profiles of the prediction token across all failure cases, colored by [PITH_FULL_IMAGE:figures/full_fig_p038_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Clustered relevance profiles (k=3). Each panel shows individual traces (thin lines) and the cluster mean (thick line). 2.0 1.5 1.0 0.5 0.0 0.5 1.0 PC1 (88.1%) 0.4 0.2 0.0 0.2 0.4 0.6 PC2 (5.8%) Cluster 0 Cluster 1 Cluster 2 IFEval MATH EvalPlus [PITH_FULL_IMAGE:figures/full_fig_p039_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: PCA 2D projection of the 29-dimensional normalized relevance profiles, colored by [PITH_FULL_IMAGE:figures/full_fig_p039_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Composition space: SB fraction vs. total magnitude (SB+OC), colored by relevance [PITH_FULL_IMAGE:figures/full_fig_p041_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: Distribution of composition features by layer segment (Early/Mid/Late). Diamonds [PITH_FULL_IMAGE:figures/full_fig_p041_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: Layer-wise |∆| heatmaps for SB, BOS, and OC across all traces (rows). Black boxes mark the peak transition layer for each trace. Darker colors indicate larger magnitude of change. 43 [PITH_FULL_IMAGE:figures/full_fig_p043_15.png] view at source ↗
Figure 16
Figure 16. Figure 16: Cross-model attribution decomposition comparison for a representative failure case. Each [PITH_FULL_IMAGE:figures/full_fig_p044_16.png] view at source ↗
read the original abstract

Interpretability tools are increasingly used to analyze failures of Large Language Models (LLMs), yet prior work largely focuses on short prompts or toy settings, leaving their behavior on commonly used benchmarks underexplored. To address this gap, we study contrastive, LRP-based attribution as a practical tool for analyzing LLM failures in realistic settings. We formulate failure analysis as \textit{contrastive attribution}, attributing the logit difference between an incorrect output token and a correct alternative to input tokens and internal model states, and introduce an efficient extension that enables construction of cross-layer attribution graphs for long-context inputs. Using this framework, we conduct a systematic empirical study across benchmarks, comparing attribution patterns across datasets, model sizes, and training checkpoints. Our results show that this token-level contrastive attribution can yield informative signals in some failure cases, but is not universally applicable, highlighting both its utility and its limitations for realistic LLM failure analysis. Our code is available at: https://aka.ms/Debug-XAI.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The paper introduces contrastive LRP-based attribution to analyze LLM failures on realistic benchmarks. It formulates the task as attributing the logit difference between an incorrect output token and a correct alternative, extends LRP with an efficient cross-layer mechanism for long-context inputs, and reports a systematic empirical comparison of attribution patterns across datasets, model sizes, and training checkpoints. The central conclusion is that token-level contrastive attribution produces informative signals in some failure cases but is not universally applicable.

Significance. If the attributions are faithful to causal token contributions, the work would supply a practical interpretability tool for debugging LLMs on standard benchmarks rather than toy settings, with the multi-model, multi-dataset design helping to delineate the method's scope and limits.

major comments (2)
  1. [§4] §4 (Empirical evaluation): the paper reports observed attribution patterns but provides no quantitative definition or metric for what constitutes an 'informative signal' (e.g., no correlation with perturbation effects on the incorrect-vs-correct logit gap), leaving the strength of the utility claim only partially supported.
  2. [§3.2] §3.2 (LRP extension): no intervention or faithfulness tests (token ablation, activation patching, or logit-difference sensitivity) are described to confirm that the contrastive LRP scores track causal contributions rather than propagation artifacts from LayerNorm, attention, or residual handling; this is load-bearing for interpreting the patterns as diagnostic of failure modes.
minor comments (2)
  1. The abstract would be improved by naming the specific benchmarks and model families used, rather than referring only to 'realistic benchmarks.'
  2. [Figures] Figure captions for the cross-layer graphs should explicitly define the sign and magnitude encoding of the attribution edges.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive and detailed feedback. We address each major comment below, clarifying our approach and outlining planned revisions to strengthen the manuscript.

read point-by-point responses

  1. Referee: [§4] §4 (Empirical evaluation): the paper reports observed attribution patterns but provides no quantitative definition or metric for what constitutes an 'informative signal' (e.g., no correlation with perturbation effects on the incorrect-vs-correct logit gap), leaving the strength of the utility claim only partially supported.

    Authors: We agree that a quantitative metric would make the notion of 'informative signal' more precise and would better support the utility claims. In the current manuscript, we use the term to describe attribution patterns that highlight input tokens whose removal or perturbation would be expected to affect the incorrect-versus-correct logit difference, based on the systematic visual and comparative analysis across datasets, model sizes, and checkpoints. To address the concern, the revised version will introduce an explicit quantitative metric: the Spearman rank correlation between token attribution scores and the change in logit gap after ablating the highest-attributed tokens. We will report these correlations separately for the subsets of cases where patterns appeared informative versus those where they did not, thereby providing a clearer, data-driven delineation of the method's scope. revision: yes

  2. Referee: [§3.2] §3.2 (LRP extension): no intervention or faithfulness tests (token ablation, activation patching, or logit-difference sensitivity) are described to confirm that the contrastive LRP scores track causal contributions rather than propagation artifacts from LayerNorm, attention, or residual handling; this is load-bearing for interpreting the patterns as diagnostic of failure modes.

    Authors: We acknowledge that explicit faithfulness validation is important for any attribution method, especially when extending LRP to contrastive logit differences and long contexts. The cross-layer mechanism we introduce follows the standard LRP propagation rules for attention, residuals, and LayerNorm that have been validated in prior transformer work; our contribution is the efficient aggregation across layers for long sequences. Because the paper's primary goal was to apply the method to realistic benchmarks and document observed patterns (including where they fail to be informative), we did not include new intervention experiments. In the revision we will add a dedicated limitations subsection that explicitly discusses potential propagation artifacts from LayerNorm and residual connections, notes the absence of direct causal tests, and frames the multi-model, multi-dataset consistency as indirect empirical support rather than definitive proof of causality. This will allow readers to interpret the diagnostic value of the patterns with appropriate caution. revision: partial

Circularity Check

0 steps flagged

No circularity: purely empirical comparison of attribution patterns

full rationale

The paper defines contrastive attribution as the attribution of logit differences between incorrect and correct output tokens using LRP propagation, introduces a cross-layer extension for long contexts, and reports observed patterns across benchmarks, model sizes, and checkpoints. No derivations, predictions, or first-principles results are claimed; conclusions rest on direct empirical comparisons without parameter fitting to target outcomes, self-definitional reductions, or load-bearing self-citations. The analysis is self-contained and falsifiable via replication on the stated benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The central claim rests on the assumption that LRP rules remain valid when applied contrastively to transformer logits and that the chosen benchmarks are representative of realistic LLM usage.

axioms (1)
  • domain assumption LRP attribution rules can be applied to the logit difference between incorrect and correct tokens in transformer models
    Invoked when formulating failure analysis as contrastive attribution.

pith-pipeline@v0.9.0 · 5487 in / 1251 out tokens · 60300 ms · 2026-05-10T05:08:58.626889+00:00 · methodology

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

83 extracted references · 38 canonical work pages · 8 internal anchors

  1. [1]

    Achtibat, R., Hatefi, S. M. V., Dreyer, M., Jain, A., Wiegand, T., Lapuschkin, S., and Samek, W. Attnlrp: Attention-aware layer-wise relevance propagation for transformers. InForty-first International Conference on Machine Learning, ICML 2024, Vienna, Austria, July 21-27, 2024. OpenReview.net, 2024. URLhttps://openreview.net/forum?id=emtXYlBrNF

    2024

  2. [2]

    XAI for transformers: Better explanations through conservative propagation

    Ali, A., Schnake, T., Eberle, O., Montavon, G., Müller, K., and Wolf, L. XAI for transformers: Better explanations through conservative propagation. In Chaudhuri, K., Jegelka, S., Song, L., Szepesvári, C., Niu, G., and Sabato, S. (eds.),International Conference on Machine Learning, ICML 2022, 17-23 July 2022, Baltimore, Maryland, USA, volume 162 ofProceed...

    2022

  3. [3]

    Ameisen, E., Lindsey, J., Pearce, A., Gurnee, W., Turner, N. L., Chen, B., Citro, C., Abrahams, D., Carter, S., Hosmer, B., Marcus, J., Sklar, M., Templeton, A., Bricken, T., McDougall, C., Cunningham, H., Henighan, T., Jermyn, A., Jones, A., Persic, A., Qi, Z., Ben Thompson, T., Zimmerman, S., Rivoire, K., Conerly, T., Olah, C., and Batson, J. Circuit tr...

    2025

  4. [4]

    Y.Gan,C.Li,J.Xie,L.Wen,M.Purver,andM.Poesio

    Andrews, P., Benhalloum, A., Bertran, G. M.-T., Bettini, M., Budhiraja, A., Cabral, R. S., Do, V., Froger, R., Garreau, E., Gaya, J.-B., Laurençon, H., Lecanu, M., Malkan, K., Mekala, D., Ménard, P., Mialon, G., Piterbarg, U., Plekhanov, M., Rita, M., Rusakov, A., Scialom, T., Vorotilov, V., Wang, M., and Yu, I. Are: Scaling up agent environments and eval...

    work page arXiv 2025

  5. [5]

    A close look at decomposition-based xai-methods for transformer language models, 2025

    Arras, L., Puri, B., Kahardipraja, P., Lapuschkin, S., and Samek, W. A close look at decomposition-based xai-methods for transformer language models, 2025. URL https:// arxiv.org/abs/2502.15886

    work page arXiv 2025

  6. [6]

    ErrorMap and ErrorAtlas: Charting the failure landscape of large language models.arXiv preprint arXiv:2601.15812, 2026

    Ashury-Tahan, S., Mai, Y., Bandel, E., Shmueli-Scheuer, M., and Choshen, L. Errormap and erroratlas: Charting the failure landscape of large language models, 2026. URLhttps: //arxiv.org/abs/2601.15812

    work page arXiv 2026

  7. [7]

    On pixel-wise explanations for non-linear classifier decisions by layer-wise relevance propagation.PloS one, 10(7):e0130140, 2015

    Bach, S., Binder, A., Montavon, G., Klauschen, F., Müller, K.-R., and Samek, W. On pixel-wise explanations for non-linear classifier decisions by layer-wise relevance propagation.PloS one, 10(7):e0130140, 2015

    2015

  8. [8]

    Monitoring reasoning models for misbehavior.arXiv preprint arXiv:2503.11926,

    Baker, B., Huizinga, J., Gao, L., Dou, Z., Guan, M. Y., Madry, A., Zaremba, W., Pachocki, J., and Farhi, D. Monitoring reasoning models for misbehavior and the risks of promoting obfuscation, 2025. URLhttps://arxiv.org/abs/2503.11926

    work page arXiv 2025

  9. [9]

    doi:10.48550/arXiv.2504.17550 , abstract =

    Bang, Y., Ji, Z., Schelten, A., Hartshorn, A., Fowler, T., Zhang, C., Cancedda, N., and Fung, P. Hallulens: Llm hallucination benchmark, 2025. URLhttps://arxiv.org/abs/2504.17550

    work page arXiv 2025

  10. [10]

    Spectral filters, dark signals, and attention sinks

    Cancedda, N. Spectral filters, dark signals, and attention sinks. InProceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pp. 4792–4808, 2024. 16

    2024

  11. [11]

    Why Do Multi-Agent LLM Systems Fail?

    Cemri, M., Pan, M. Z., Yang, S., Agrawal, L. A., Chopra, B., Tiwari, R., Keutzer, K., Parameswaran, A., Klein, D., Ramchandran, K., Zaharia, M., Gonzalez, J. E., and Stoica, I. Why do multi-agent llm systems fail?, 2025. URLhttps://arxiv.org/abs/2503.13657

    work page internal anchor Pith review arXiv 2025

  12. [12]

    , Xie X: A survey on evaluation of large language models

    Chang, Y., Wang, X., Wang, J., Wu, Y., Yang, L., Zhu, K., Chen, H., Yi, X., Wang, C., Wang, Y., Ye, W., Zhang, Y., Chang, Y., Yu, P. S., Yang, Q., and Xie, X. A survey on evaluation of large language models.ACM Trans. Intell. Syst. Technol., 15(3), March 2024. ISSN 2157-6904. doi: 10.1145/3641289. URLhttps://doi.org/10.1145/3641289

    work page doi:10.1145/3641289 2024

  13. [13]

    M., and Lee, S

    Covert, I., Lundberg, S. M., and Lee, S. Explaining by removing: A unified framework for model explanation.J. Mach. Learn. Res., 22:209:1–209:90, 2021. URLhttps://jmlr.org/ papers/v22/20-1316.html

    2021

  14. [14]

    URL https://doi.org/10.18653/v1/2022.acl -long.581

    Dai, D., Dong, L., Hao, Y., Sui, Z., Chang, B., and Wei, F. Knowledge neurons in pretrained transformers. In Muresan, S., Nakov, P., and Villavicencio, A. (eds.),Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pp. 8493–8502, Dublin, Ireland, May 2022. Association for Computational Linguistic...

    work page doi:10.18653/v1/2022.acl-long.581 2022

  15. [15]

    Graphghost: Tracing structures behind large language models, 2025

    Dai, X., Guo, K., Lo, C.-H., Zeng, S., Ding, J., Luo, D., Mukherjee, S., and Tang, J. Graphghost: Tracing structures behind large language models, 2025. URLhttps://arxiv.org/abs/2510. 08613

    2025

  16. [16]

    DeepSeek-V3.2: Pushing the Frontier of Open Large Language Models

    DeepSeek-AI, Liu, A., Mei, A., Lin, B., Xue, B., Wang, B., Xu, B., Wu, B., Zhang, B., Lin, C., Dong, C., et al. Deepseek-v3.2: Pushing the frontier of open large language models, 2025. URL https://arxiv.org/abs/2512.02556

    work page internal anchor Pith review arXiv 2025

  17. [17]

    Extraction of salient sentences from labelled documents, 2015

    Denil, M., Demiraj, A., and de Freitas, N. Extraction of salient sentences from labelled documents, 2015. URLhttps://arxiv.org/abs/1412.6815

    work page arXiv 2015

  18. [18]

    Transcoders find interpretable LLM feature circuits

    Dunefsky, J., Chlenski, P., and Nanda, N. Transcoders find interpretable LLM feature circuits. In Globersons, A., Mackey, L., Belgrave, D., Fan, A., Paquet, U., Tomczak, J. M., and Zhang, C. (eds.),Advances in Neural Information Processing Systems 38: Annual Conference on Neural Information Processing Systems 2024, NeurIPS 2024, Vancouver, BC, Canada, Dec...

    2024

  19. [19]

    and Voita, E

    Ferrando, J. and Voita, E. Information flow routes: Automatically interpreting language models at scale. In Al-Onaizan, Y., Bansal, M., and Chen, Y. (eds.),Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, EMNLP 2024, Miami, FL, USA, November 12-16, 2024, pp. 17432–17445. Association for Computational Linguistics,

    2024

  20. [20]

    Information Flow Routes: Automatically Interpreting Language Models at Scale , booktitle =

    doi: 10.18653/V1/2024.EMNLP-MAIN.965. URL https://doi.org/10.18653/v1/ 2024.emnlp-main.965

    work page doi:10.18653/v1/2024.emnlp-main.965 2024

  21. [21]

    I., Tsiamas, I., and Costa-jussà, M

    Ferrando, J., Gállego, G. I., Tsiamas, I., and Costa-jussà, M. R. Explaining how transformers use context to build predictions. In Rogers, A., Boyd-Graber, J. L., and Okazaki, N. (eds.), Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), ACL 2023, Toronto, Canada, July 9-14, 2023, pp. 5486–5513...

    work page doi:10.18653/v1/2023.acl-long.301 2023

  22. [22]

    Ferrando, J., Sarti, G., Bisazza, A., and Costa-jussà, M. R. A primer on the inner workings of transformer-based language models, 2024. URLhttps://arxiv.org/abs/2405.00208. 17

    work page arXiv 2024

  23. [23]

    Fong, R. C. and Vedaldi, A. Interpretable explanations of black boxes by meaningful perturba- tion. InIEEE International Conference on Computer Vision, ICCV 2017, Venice, Italy, October 22-29, 2017, pp. 3449–3457. IEEE Computer Society, 2017. doi: 10.1109/ICCV.2017.371. URL https://doi.org/10.1109/ICCV.2017.371

    work page doi:10.1109/iccv.2017.371 2017

  24. [24]

    The llm evaluation guidebook, 2025

    Fourrier, C., Frere, T., Penedo, G., and Wolf, T. The llm evaluation guidebook, 2025. URL https://huggingface.co/spaces/OpenEvals/evaluation-guidebook#recommendations

    2025

  25. [25]

    Boyi Deng, Yu Wan, Baosong Yang, Yidan Zhang, and Fuli Feng

    Galichin, A., Dontsov, A., Druzhinina, P., Razzhigaev, A., Rogov, O. Y., Tutubalina, E., and Oseledets, I. I have covered all the bases here: Interpreting reasoning features in large language models via sparse autoencoders, 2025. URLhttps://arxiv.org/abs/2503.18878

    work page arXiv 2025

  26. [26]

    Transformer Feed-Forward Layers Are Key-Value Memories

    Geva, M., Schuster, R., Berant, J., and Levy, O. Transformer feed-forward layers are key-value memories. In Moens, M.-F., Huang, X., Specia, L., and Yih, S. W.-t. (eds.),Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pp. 5484–5495, Online and Punta Cana, Dominican Republic, November 2021. Association for Computatio...

    work page internal anchor Pith review doi:10.18653/v1/2021.emnlp-main.446 2021

  27. [27]

    2023 , archivePrefix=

    Goldowsky-Dill, N., MacLeod, C., Sato, L., and Arora, A. Localizing model behavior with path patching, 2023. URLhttps://arxiv.org/abs/2304.05969

    work page arXiv 2023

  28. [28]

    Circuit-tracer: A new library for finding feature circuits

    Hanna, M., Piotrowski, M., Lindsey, J., and Ameisen, E. Circuit-tracer: A new library for finding feature circuits. In Belinkov, Y., Mueller, A., Kim, N., Mohebbi, H., Chen, H., Arad, D., and Sarti, G. (eds.),Proceedings of the 8th BlackboxNLP Workshop: Analyzing and Interpreting Neural Networks for NLP, pp. 239–249, Suzhou, China, November 2025. Associat...

    work page doi:10.18653/v1/2025.blackboxnlp-1.14 2025

  29. [29]

    How to use and interpret activation patching.arXiv preprint arXiv:2404.15255,

    Heimersheim, S. and Nanda, N. How to use and interpret activation patching, 2024. URL https://arxiv.org/abs/2404.15255

    work page arXiv 2024

  30. [30]

    Measuring mathematical problem solving with the MATH dataset

    Hendrycks, D., Burns, C., Kadavath, S., Arora, A., Basart, S., Tang, E., Song, D., and Steinhardt, J. Measuring mathematical problem solving with the MATH dataset. InThirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track (Round 2), 2021. URLhttps://openreview.net/forum?id=7Bywt2mQsCe

    2021

  31. [31]

    A Survey on Hallucination in Large Language Models: Principles, Taxonomy, Challenges, and Open Questions , volume=

    Huang, L., Yu, W., Ma, W., Zhong, W., Feng, Z., Wang, H., Chen, Q., Peng, W., Feng, X., Qin, B., and Liu, T. A survey on hallucination in large language models: Principles, taxonomy, challenges, and open questions.ACM Trans. Inf. Syst., 43(2), January 2025. ISSN 1046-8188. doi: 10.1145/3703155. URLhttps://doi.org/10.1145/3703155

    work page doi:10.1145/3703155 2025

  32. [32]

    J., Madotto, A., and Fung, P

    Ji, Z., Lee, N., Frieske, R., Yu, T., Su, D., Xu, Y., Ishii, E., Bang, Y. J., Madotto, A., and Fung, P. Survey of hallucination in natural language generation.ACM Comput. Surv., 55(12), March

  33. [33]

    ACM Comput

    ISSN 0360-0300. doi: 10.1145/3571730. URLhttps://doi.org/10.1145/3571730

    work page doi:10.1145/3571730

  34. [34]

    In: Merlo, P., Tiedemann, J., Tsarfaty, R

    Kobayashi, G., Kuribayashi, T., Yokoi, S., and Inui, K. Incorporating residual and normalization layers into analysis of masked language models. In Moens, M., Huang, X., Specia, L., and Yih, S. W. (eds.),Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, EMNLP 2021, Virtual Event / Punta Cana, Dominican Republic, 7-11 ...

    work page doi:10.18653/v1/2021 2021

  35. [35]

    control bars

    Kramár, J., Lieberum, T., Shah, R., and Nanda, N. Atp*: An efficient and scalable method for localizing LLM behaviour to components.CoRR, abs/2403.00745, 2024. doi: 10.48550/ARXIV. 2403.00745. URLhttps://doi.org/10.48550/arXiv.2403.00745

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

  36. [36]

    HaluEval: A large-scale hallucination evaluation benchmark for large language models

    Li, J., Cheng, X., Zhao, X., Nie, J.-Y., and Wen, J.-R. HaluEval: A large-scale hallucination evaluation benchmark for large language models. In Bouamor, H., Pino, J., and Bali, K. (eds.), Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pp. 6449–6464, Singapore, December 2023. Association for Computational Linguisti...

    2023

  37. [37]

    From System 1 to System 2: A Survey of Reasoning Large Language Models

    Li, Z.-Z., Zhang, D., Zhang, M.-L., Zhang, J., Liu, Z., Yao, Y., Xu, H., Zheng, J., Wang, P.-J., Chen, X., Zhang, Y., Yin, F., Dong, J., Li, Z., Bi, B.-L., Mei, L.-R., Fang, J., Liang, X., Guo, Z., Song, L., and Liu, C.-L. From system 1 to system 2: A survey of reasoning large language models, 2025. URLhttps://arxiv.org/abs/2502.17419

    work page internal anchor Pith review arXiv 2025

  38. [38]

    12 Distributional Open-Ended Evaluation of LLM Cultural Value Alignment Based on Value Codebook Masoud, R., Liu, Z., Ferianc, M., Treleaven, P

    Lin, S., Hilton, J., and Evans, O. TruthfulQA: Measuring how models mimic human falsehoods. In Muresan, S., Nakov, P., and Villavicencio, A. (eds.),Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pp. 3214–3252, Dublin, Ireland, May 2022. Association for Computational Linguistics. doi: 10.186...

    work page doi:10.18653/v1/2022 2022

  39. [39]

    Lindsey, J., Gurnee, W., Ameisen, E., Chen, B., Pearce, A., Turner, N. L., Citro, C., Abrahams, D., Carter, S., Hosmer, B., Marcus, J., Sklar, M., Templeton, A., Bricken, T., McDougall, C., Cunningham, H., Henighan, T., Jermyn, A., Jones, A., Persic, A., Qi, Z., Thompson, T. B., Zimmerman, S., Rivoire, K., Conerly, T., Olah, C., and Batson, J. On the biol...

    2025

  40. [40]

    Attribot: A bag of tricks for efficiently approximating leave-one-out context attribution

    Liu, F., Kandpal, N., and Raffel, C. Attribot: A bag of tricks for efficiently approximating leave-one-out context attribution. InThe Thirteenth International Conference on Learning Representations, ICLR 2025, Singapore, April 24-28, 2025. OpenReview.net, 2025. URL https://openreview.net/forum?id=9kJperA2a4

    2025

  41. [41]

    S., Wang, Y., and Zhang, L

    Liu, J., Xia, C. S., Wang, Y., and Zhang, L. Is your code generated by chatGPT really correct? rigorous evaluation of large language models for code generation. InThirty-seventh Conference on Neural Information Processing Systems, 2023. URLhttps://openreview.net/forum?id= 1qvx610Cu7

    2023

  42. [42]

    Evaluating language models for efficient code generation

    Liu, J., Xie, S., Wang, J., Wei, Y., Ding, Y., and Zhang, L. Evaluating language models for efficient code generation. InFirst Conference on Language Modeling, 2024. URLhttps: //openreview.net/forum?id=IBCBMeAhmC

    2024

  43. [43]

    Lundberg, S. M. and Lee, S. A unified approach to interpreting model predictions. In Guyon, I., von Luxburg, U., Bengio, S., Wallach, H. M., Fergus, R., Vishwanathan, S. V. N., and Garnett, R. (eds.),Advances in Neural Information Processing Systems 30: Annual Conference on Neural Information Processing Systems 2017, December 4-9, 2017, Long Beach, CA, US...

    2017

  44. [44]

    From understanding to utilization: A survey on explainability for large language models, 2024

    Luo, H. and Specia, L. From understanding to utilization: A survey on explainability for large language models, 2024. URLhttps://arxiv.org/abs/2401.12874. 19

    work page arXiv 2024

  45. [45]

    DoVer : Intervention-driven auto debugging for LLM multi-agent systems

    Ma, M., Zhang, J., Yang, F., Kang, Y., Lin, Q., Yang, T., Rajmohan, S., and Zhang, D. Dover: Intervention-driven auto debugging for llm multi-agent systems, 2025. URLhttps: //arxiv.org/abs/2512.06749

    work page arXiv 2025

  46. [47]

    Locating and editing factual associations in gpt

    Meng, K., Bau, D., Andonian, A., and Belinkov, Y. Locating and editing factual associations in gpt. In Koyejo, S., Mohamed, S., Agarwal, A., Belgrave, D., Cho, K., and Oh, A. (eds.), Advances in Neural Information Processing Systems, volume 35, pp. 17359–17372. Curran Associates, Inc., 2022

    2022

  47. [48]

    S., Andonian, A

    Meng, K., Sharma, A. S., Andonian, A. J., Belinkov, Y., and Bau, D. Mass-editing memory in a transformer. InThe Eleventh International Conference on Learning Representations, 2023. URLhttps://openreview.net/forum?id=MkbcAHIYgyS

    2023

  48. [49]

    Attribution patching: Activation patching at industrial scale

    Nanda, N. Attribution patching: Activation patching at industrial scale. 2023. URLhttps: //www.neelnanda.io/mechanistic-interpretability/attribution-patching

    2023

  49. [50]

    and Bloom, J

    Nanda, N. and Bloom, J. Transformerlens. https://github.com/TransformerLensOrg/ TransformerLens, 2022

    2022

  50. [51]

    Olmo, T., :, Ettinger, A., Bertsch, A., Kuehl, B., Graham, D., Heineman, D., Groeneveld, D., Brahman, F., Timbers, F., Ivison, H., Morrison, J., Poznanski, J., Lo, K., Soldaini, L., Jordan, M., Chen, M., Noukhovitch, M., Lambert, N., Walsh, P., Dasigi, P., Berry, R., Malik, S., Shah, S., Geng, S., Arora, S., Gupta, S., Anderson, T., Xiao, T., Murray, T., ...

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  51. [52]

    gpt-oss-120b & gpt-oss-20b model card, 2025

    OpenAI. gpt-oss-120b & gpt-oss-20b model card, 2025. URLhttps://arxiv.org/abs/2508. 10925

    2025

  52. [53]

    Introducing gpt-5, 2025

    OpenAI. Introducing gpt-5, 2025. URLhttps://openai.com/index/introducing-gpt-5/

    2025

  53. [54]

    D., Ermon, S., and Finn, C

    Rafailov, R., Sharma, A., Mitchell, E., Manning, C. D., Ermon, S., and Finn, C. Direct preference optimization: Your language model is secretly a reward model.Advances in neural information processing systems, 36:53728–53741, 2023

    2023

  54. [55]

    Rosser, J., García, J. L. R., Penha, G., Palla, K., and Bouchard, H. Stream: Scaling up mechanistic interpretability to long context in llms via sparse attention, 2025. URL https://arxiv.org/abs/2510.19875

    work page arXiv 2025

  55. [56]

    Deep Inside Convolutional Networks: Visualising Image Classification Models and Saliency Maps

    Simonyan, K., Vedaldi, A., and Zisserman, A. Deep inside convolutional networks: Visualising image classification models and saliency maps. In Bengio, Y. and LeCun, Y. (eds.),2nd International Conference on Learning Representations, ICLR 2014, Banff, AB, Canada, April 14-16, 2014, Workshop Track Proceedings, 2014. URLhttp://arxiv.org/abs/1312.6034. 20

    work page Pith review arXiv 2014

  56. [57]

    SmoothGrad: removing noise by adding noise

    Smilkov, D., Thorat, N., Kim, B., Viégas, F. B., and Wattenberg, M. Smoothgrad: removing noise by adding noise.CoRR, abs/1706.03825, 2017. URL http://arxiv.org/abs/1706. 03825

    work page Pith review arXiv 2017

  57. [58]

    A survey on large language model reasoning failures

    Song, P., Han, P., and Goodman, N. A survey on large language model reasoning failures. In 2nd AI for Math Workshop @ ICML 2025, 2025. URLhttps://openreview.net/forum?id= hsgMn4KBFG

    2025

  58. [59]

    Axiomatic attribution for deep networks

    Sundararajan, M., Taly, A., and Yan, Q. Axiomatic attribution for deep networks. In Precup, D. and Teh, Y. W. (eds.),Proceedings of the 34th International Conference on Machine Learning, ICML 2017, Sydney, NSW, Australia, 6-11 August 2017, volume 70 ofProceedings of Machine Learning Research, pp. 3319–3328. PMLR, 2017. URLhttp://proceedings.mlr.press/v70/...

    2017

  59. [60]

    Kimi K2: Open Agentic Intelligence

    Team, K., Bai, Y., Bao, Y., Chen, G., Chen, J., Chen, N., Chen, R., Chen, Y., Chen, Y., Chen, Y., et al. Kimi k2: Open agentic intelligence, 2025. URLhttps://arxiv.org/abs/2507.20534

    work page internal anchor Pith review arXiv 2025

  60. [61]

    N., Kaiser, L

    Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N., Kaiser, L. u., and Polosukhin, I. Attention is all you need. In Guyon, I., Luxburg, U. V., Bengio, S., Wallach, H., Fergus, R., Vishwanathan, S., and Garnett, R. (eds.),Advances in Neural Information Processing Systems, volume 30. Curran Associates, Inc., 2017

    2017

  61. [62]

    R., Variengien, A., Conmy, A., Shlegeris, B., and Steinhardt, J

    Wang, K. R., Variengien, A., Conmy, A., Shlegeris, B., and Steinhardt, J. Interpretability in the wild: a circuit for indirect object identification in GPT-2 small. InThe Eleventh International Conference on Learning Representations, ICLR 2023, Kigali, Rwanda, May 1-5,

    2023

  62. [63]

    URLhttps://openreview.net/forum?id=NpsVSN6o4ul

    OpenReview.net, 2023. URLhttps://openreview.net/forum?id=NpsVSN6o4ul

    2023

  63. [64]

    Efficient streaming language models with attention sinks

    Xiao, G., Tian, Y., Chen, B., Han, S., and Lewis, M. Efficient streaming language models with attention sinks. InThe Twelfth International Conference on Learning Representations, 2024. URLhttps://openreview.net/forum?id=NG7sS51zVF

    2024

  64. [65]

    Diagnosing failures in large language models’ answers: Integrating error attribution into evaluation framework

    Xu, Z., Xie, S., Lv, Q., Xiao, S., Song, L., Wenjuan, S., and Lin, F. Diagnosing failures in large language models’ answers: Integrating error attribution into evaluation framework. In Findings of the Association for Computational Linguistics: ACL 2025, pp. 21148–21165, 2025

    2025

  65. [66]

    Qwen3 Technical Report

    Yang, A., Li, A., Yang, B., Zhang, B., Hui, B., Zheng, B., Yu, B., Gao, C., Huang, C., Lv, C., Zheng, C., Liu, D., Zhou, F., Huang, F., Hu, F., Ge, H., Wei, H., Lin, H., Tang, J., Yang, J., Tu, J., Zhang, J., Yang, J., Yang, J., Zhou, J., Zhou, J., Lin, J., Dang, K., Bao, K., Yang, K., Yu, L., Deng, L., Li, M., Xue, M., Li, M., Zhang, P., Wang, P., Zhu, Q...

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  66. [67]

    Token-importance guided direct preference optimization, 2025

    Yang, N., Lin, H., Liu, Y., Tian, B., Liu, G., and Zhang, H. Token-importance guided direct preference optimization, 2025. URLhttps://arxiv.org/abs/2505.19653

    work page arXiv 2025

  67. [68]

    Knowledge circuits in pretrained transformers

    Yao, Y., Zhang, N., Xi, Z., Wang, M., Xu, Z., Deng, S., and Chen, H. Knowledge circuits in pretrained transformers. In Globerson, A., Mackey, L., Belgrave, D., Fan, A., Paquet, U., Tomczak, J., and Zhang, C. (eds.),Advances in Neural Information Processing Systems, volume 37, pp. 118571–118602. Curran Associates, Inc., 2024. doi: 10.52202/079017-3765. 21

    work page doi:10.52202/079017-3765 2024

  68. [70]

    and Nanda, N

    Zhang, F. and Nanda, N. Towards best practices of activation patching in language mod- els: Metrics and methods. InThe Twelfth International Conference on Learning Repre- sentations, ICLR 2024, Vienna, Austria, May 7-11, 2024. OpenReview.net, 2024. URL https://openreview.net/forum?id=Hf17y6u9BC

    2024

  69. [71]

    From reasoning to answer: Empirical, attention- based and mechanistic insights into distilled DeepSeek r1 models

    Zhang, J., Lin, Q., Rajmohan, S., and Zhang, D. From reasoning to answer: Empirical, attention- based and mechanistic insights into distilled DeepSeek r1 models. In Christodoulopoulos, C., Chakraborty, T., Rose, C., and Peng, V. (eds.),Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, pp. 3985–4002, Suzhou, China, November

    2025

  70. [72]

    InProceedings of the 2025 Conference on Empirical Methods in Natural Language Processing

    Association for Computational Linguistics. ISBN 979-8-89176-332-6. doi: 10.18653/v1/ 2025.emnlp-main.198. URLhttps://aclanthology.org/2025.emnlp-main.198/

    work page doi:10.18653/v1/ 2025

  71. [73]

    Which agent causes task failures and when? on automated failure attribution of LLM multi-agent systems

    Zhang, S., Yin, M., Zhang, J., Liu, J., Han, Z., Zhang, J., Li, B., Wang, C., Wang, H., Chen, Y., and Wu, Q. Which agent causes task failures and when? on automated failure attribution of LLM multi-agent systems. InForty-second International Conference on Machine Learning,

  72. [74]

    URLhttps://openreview.net/forum?id=GazlTYxZss

  73. [75]

    S iren ' s Song in the AI Ocean: A Survey on Hallucination in Large Language Models

    Zhang, Y., Li, Y., Cui, L., Cai, D., Liu, L., Fu, T., Huang, X., Zhao, E., Zhang, Y., Chen, Y., Wang, L., Luu, A. T., Bi, W., Shi, F., and Shi, S. Siren’s song in the ai ocean: A survey on hallucination in large language models.Computational Linguistics, 51(4):1373–1418, 12 2025. ISSN 0891-2017. doi: 10.1162/COLI.a.16. URLhttps://doi.org/10.1162/COLI.a.16

    work page doi:10.1162/coli.a.16 2025

  74. [76]

    Verifyingchain-of-thought reasoning via its computational graph.CoRR, abs/2510.09312, 2025

    Zhao, Z., Koishekenov, Y., Yang, X., Murray, N., and Cancedda, N. Verifying chain-of-thought reasoning via its computational graph, 2025. URLhttps://arxiv.org/abs/2510.09312

    work page arXiv 2025

  75. [77]

    latency (seconds) / memory (MB)

    Zhou, J., Lu, T., Mishra, S., Brahma, S., Basu, S., Luan, Y., Zhou, D., and Hou, L. Instruction- following evaluation for large language models, 2023. URLhttps://arxiv.org/abs/2311. 07911. 22 A Batch-Packed Multi-Target Backpropagation for Attribution Graph Construction This appendix presents an efficient procedure for constructing attribution-graph edges...

    2023

  76. [78]

    add all source nodes(s,i)such that|R(s) i |>0

  77. [79]

    add all target nodes(t,j)such that|R(t) j |>0

  78. [80]

    Dataset Description

    add directed edges(s,i)→(t,j)for selected entriesA (s→t) j,i after pruning, with edge weight w(s,i)→(t,j)=A (s→t) j,i .(18) Pruning objectives.The interaction matricesA (s→t)are generally dense, making direct graph construction impractical. We therefore prune edges using magnitude-based criteria applied to |A(s→t)|. Pruning is performedindependently for e...

  79. [81]

    , "ground_truth

    Lyon[13] .[14]", "ground_truth": "Paris, the capital of France." } - The "indexed_completion" field provides a tokenized version of the completion, where each token is followed by its index in square brackets. This will help you localize the position of tokens in the completion. - Both the "prompt" and "completion" fields may contain special tokens, e.g.,...

  80. [82]

    Assuming the target token is the model’s top-1 output token at the failure position (i.e., the token with the highest logit)

Showing first 80 references.