Pith. sign in

REVIEW 4 major objections 5 minor 22 references

Across three decoder-only transformer models, the representations of an ambiguous word's two senses reconverge in late layers while the model's next-token predictions for those senses diverge most sharply at the same depth.

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 ↗

In three decoder-only LLMs, representations of homonym and polyseme senses reconverge in late layers while next-token predictions diverge, and activation patching shows late-layer states still carry the disambiguating signal.

T0 review reviewed 2026-08-04 challenge →

load-bearing objection Good multi-model evidence for mid-layer geometric divergence and late-layer reconvergence; the predictive-divergence headline is selected by the validation filter, so the general claim is conditional, not established. the 4 major comments →

arxiv 2608.01816 v1 pith:E2P7QHWX submitted 2026-08-03 cs.CL

Divergent large language model predictions from convergent representations in ambiguous word pairs

K. Jack Scott , Narun Pat , Veronica Liesaputra This is my paper
classification cs.CL
keywords lexical ambiguityhomonymypolysemytransformer interpretabilityactivation patchinglogit lenscosine similarityembedding geometry
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 reading

This paper reports a consistent layer-wise dissociation in three decoder-only language models spanning 117M to 32B parameters: when the same ambiguous word appears in two different senses, the hidden-state representations of the word become most separated in middle layers, then partly move back together late, at exactly the depth where next-token predictions become most different. Activation patching shows the late-layer states still causally control the output, so the sense distinction is not lost—it is just not visible to cosine similarity over the embeddings. The result holds for both unrelated homonyms and concept-drift polysemes such as "mouse" as rodent versus computer device, and it offers a mechanism for why a model can behave as if it distinguishes senses that its embedding geometry does not separate. If the finding holds, methods that read meaning from late-layer cosine similarity—semantic search, retrieval, clustering—can understate the disambiguation actually happening inside the model.

Core claim

The central claim is that, across two ambiguity types and three model scales, representational separation and predictive separation come apart in a specific, repeatable way. For the same word in two contexts, cosine distance between residual-stream hidden states rises through early and middle layers, peaks, and then partially reconverges in the final layers, while KL divergence between the two contexts' next-token distributions keeps rising and reaches its maximum in those same final layers. Activation patching provides causal evidence that the late-layer representations, despite their reduced geometric separation, carry the disambiguating information: swapping a late-layer activation from o

What carries the argument

The workhorse is the paired-context single-token residual-stream probe. For each ambiguous word, the paper measures at every layer three quantities: cosine distance between the hidden states of the two contexts, cosine distance between their projected logits, and KL divergence between their next-token distributions. It then intervenes with activation patching—replacing one context's residual-stream activation at a layer and measuring how much the final output shifts toward the source—to establish causal sufficiency through the nonlinear remainder of the network. The load-bearing split is between the geometric measure (cosine distance) and the functional measure (KL divergence) at the same de

Load-bearing premise

The pairs tested were kept only when the model already predicted clearly different next tokens, so the late-layer divergence and patch effects could be consequences of that selection rather than a general feature of how ambiguity is resolved.

What would settle it

Re-run the same layer-wise metrics on the full candidate pool of 190 homonym pairs and 97 polyseme pairs before applying the distinctness filters (KL divergence above 0.5 or 1.0, top-k overlap below 0.4). If the late-layer rise in KL divergence and the near-complete activation-patch effects do not reproduce on the unfiltered pool, the central pattern is an artifact of selecting pairs by the outcome being explained.

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

If this is right

  • Late-layer cosine similarity is not a complete readout of a model's sense distinctions; retrieval, clustering, or search built on it can miss distinctions the model is actively using.
  • Middle-layer representations preserve more angular separation than final-layer ones, so extracting embeddings from middle layers could keep more sense information available to similarity measures.
  • Combining embedding similarity with prediction-based signals should recover distinctions that late-layer geometry hides.
  • The gap between behavioural similarity and embedding similarity previously reported for instruction-tuned models can be explained by representational reconvergence occurring alongside predictive divergence.
  • Because single-layer vulnerability differs across architectures, no single layer depth or ablation profile is universal; layer choice must be made per model.

Where Pith is reading between the lines

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

  • Editorial inference: rerunning the same metrics on the unfiltered candidate pool would test whether the late-layer divergence is an artifact of selecting pairs by their already-distinct predictions.
  • Editorial inference: if the pattern transfers to pooled sentence embeddings and fine-tuned embedding models, retrieval systems may be systematically losing sense distinctions; a direct retrieval experiment comparing middle- and final-layer embeddings on sense-disambiguated queries would test this.
  • Editorial inference: a per-layer linear probe would show whether the sense distinction is still linearly decodable even when cosine distance has reconverged—a question the paper explicitly leaves open.
  • Editorial inference: the model-specific ablation profiles suggest the same outcome may be implemented by different circuits in different architectures, which circuit tracing could identify.
Share X Bluesky LinkedIn Reddit HN

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

4 major / 5 minor

Summary. The paper investigates how decoder-only transformer LMs resolve lexical ambiguity by tracking, layer by layer, the geometry of hidden-state representations, projected logits, and next-token prediction distributions for homonym and polyseme pairs in GPT-2-Small, Llama-3.2-3B, and Qwen2.5-32B. It reports that activation-space separation between the two senses peaks in middle layers and partially reconverges in late layers, while KL divergence between next-token distributions continues to increase and peaks late. Activation patching is then used to argue that late-layer representations, despite reduced geometric separation, causally determine outputs, and single-layer ablation is used to compare layer-wise vulnerabilities across models. The authors propose this as a mechanism for the observed weak correspondence between embedding similarity and model behavior, and draw implications for embedding-based retrieval and similarity methods.

Significance. If the central claim holds, the paper makes a useful conceptual contribution: it separates geometric encoding from predictive encoding and provides causal evidence that late-layer representations can remain functionally decisive even when cosine distances shrink. Strengths include the deliberate scaling across three model families and sizes, the inclusion of both homonyms and concept-drift polysemes, a sequence-position control for the activation-distance results, and the public release of stimuli and code. The activation-patching design and the bidirectional consistency of swaps are appropriate tools for the causal claim. However, the generality of the headline finding is currently limited by an outcome-based stimulus filter, a suspected error in the swap-effectiveness definition, post hoc layer-group boundaries, and an internal contradiction between the abstract and the GPT-2 logit-distance result. These issues are fixable, but they are load-bearing for the paper's central generalization.

major comments (4)
  1. [§3.1, Table 1] The stimulus pool is filtered on the dependent variable. Logit-lens pairs are retained only when final-layer next-token D_KL > 0.5; activation-patching pairs only when D_KL > 1.0 and top-10 overlap < 0.4, with Table 1 showing substantial attrition (e.g., Qwen patching concepts: 36/97 candidates). The reported late-layer KL rise and near-complete patch effects are therefore established only for pairs the model already strongly disambiguates at the output. Since the abstract and conclusion generalize to 'ambiguous word pairs' as a class, the unfiltered candidate-pool analyses, or at least threshold sweeps, are needed. The limitations section (§5.1) does not list this selection, which is the main blocker to accepting the general claim.
  2. [§3.3] The Swap Effectiveness formula appears to be reversed or inconsistently notated. As printed, for an A→B insertion the numerator is D_KL(swapped || original A) and the denominator is D_KL(swapped || original B). Complete adoption of source A makes the numerator tend to 0 and the denominator stay positive, so the metric tends to 0, yet §4.2 reports late-layer values of 800–1300 and states that non-finite values represent complete adoption. Either the numerator and denominator are swapped or the notation is mis-defined. This concerns the central causal metric and must be corrected and re-verified.
  3. [§3.6] Layer groups were defined 'from observation of our early logit lens findings' and then used as categorical fixed effects for the confirmatory tests in §4.1. Post hoc boundary selection inflates the apparent significance of layer contrasts, especially with non-parametric rank tests. Report continuous layer-wise analyses as a robustness check, or pre-specify the boundaries, or explicitly label the grouped tests exploratory. This is load-bearing for the layer-trajectory claim.
  4. [§4.1.1 and Abstract] The abstract states that representations become maximally distinct in middle layers and partially reconverge late across all three models, but §4.1.1 reports that GPT-2 logit distance increases monotonically without reconvergence for both homonyms and concept-drift polysemes (e.g., concept drift: 0.101 → 0.290 → 0.410). If the claim is specifically about activation distance, the wording should be restricted accordingly throughout the paper; as written, the blanket statement is internally contradicted by the reported GPT-2 logit-distance trajectory.
minor comments (5)
  1. [Figure 1] The individual panels are not clearly labeled on the face of the figure; the reader must infer which curves are homonyms vs. concept-drift and which are homonyms vs. polysemes. Please add explicit panel labels and legends.
  2. [§3.1] The phrase 'top-ktoken overlap' should specify that k=10, how ties are handled, and whether overlap is computed on tokens or on token IDs.
  3. [§4.2 and Figure 2] The text often reports median swap effectiveness, while the Methods and figure caption describe geometric means with log-space confidence intervals. Clarify which summary is displayed and which is tested.
  4. [§5.1] The anisotropy-baseline limitation is acknowledged, but a supplementary anisotropy-corrected version of the activation-distance curves would make the geometric results directly comparable to Ma et al. and would strengthen the claim that the reconvergence is not an artifact of raw cosine distance.
  5. [Throughout] Several small typos and awkward phrases remain (e.g., 'at the extreme of the divergence being measured'; 'foundation-size' is used without an initial definition). A careful copyedit is recommended.

Circularity Check

1 steps flagged

Outcome-based stimulus filtering partially circularizes the late-layer KL-divergence claim; activation-distance and patching results retain independent support.

specific steps
  1. fitted input called prediction [Section 3.1 (stimulus validation), Table 1; reported in Section 4.1]
    "Context distinctness was quantified via KL divergence between next-token probability distributions, with thresholds of D_KL > 0.5 for logit lens stimuli and D_KL > 1.0 for activation patching stimuli. Additionally, activation patching stimuli required top-k token overlap < 0.4 to ensure strong semantic differentiation."

    The stimulus pool is filtered on the outcome metric itself: logit-lens pairs are kept only when the model's final next-token KL divergence exceeds 0.5, and patching pairs only when D_KL > 1.0 with top-k overlap < 0.4. The paper then reports that KL divergence rises to a late-layer maximum. That late-layer divergence is partly an inclusion criterion, not an independent discovery; the unfiltered candidate pool (190 homonym and 97 polyseme pairs) is not analyzed. This is selection on the dependent variable, so the predictive-divergence claim is conditional on the model already strongly disambiguating the retained pairs. It does not force the layer-wise trajectory, so the circularity is partial rather than definitional.

full rationale

No formal derivation equates a result to its inputs: there are no fitted parameters and no load-bearing self-citations (the only self-reference is to the supplementary stimulus tables). The activation-distance trajectory and the activation-patching results are genuinely independent of the main circularity concern: stimulus selection is based on KL divergence and top-k overlap, not on activation cosine distance, and the sequence-order control plus bidirectional consistency of patch effects provide additional grounding. The circular element is experimental: the headline late-layer predictive-divergence result is measured only on pairs preselected by high final-layer KL divergence (D_KL > 0.5 for logit lens; > 1.0 plus top-k overlap < 0.4 for patching), so the abstract's generalization to ambiguous word pairs as a class is conditional on the outcome. The limitations section (5.1) does not list this selection, nor the post-hoc layer grouping in §3.6, where early/middle/late boundaries were chosen after observing the logit-lens data. Because the representational and patching claims survive the concern and the KL trajectory is not forced by the threshold, the paper is partially circular rather than definitionally circular; an unfiltered re-analysis or a threshold-robustness check would resolve the issue.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 0 invented entities

No new entities are introduced. The paper's 'convergent representations' and 'predictive divergence' are descriptive labels for measured quantities, not new theoretical objects.

free parameters (5)
  • Validation threshold D_KL for logit-lens stimuli = D_KL > 0.5
    Hand-set in Section 3.1; only pairs already producing distinct next-token distributions enter the logit-lens analyses, so the reported late-layer KL rise is measured on an outcome-filtered sample.
  • Validation thresholds for activation-patching stimuli = D_KL > 1.0; top-k overlap < 0.4
    Hand-set in Section 3.1; guarantees baseline output differences and magnifies swap-effectiveness and token-overlap results in Experiment 2.
  • Layer-group boundaries for statistical tests = GPT-2: 0-3/4-8/9-11; Llama: 0-6/7-16/17-27; Qwen: 0-14/15-48/49-63
    Chosen in Section 3.6 from observation of the logit-lens findings, then used as fixed effects; this is a post-hoc division of the data.
  • Top-k for token overlap = 10
    Used in Section 3.2 and 3.3 for both validation and the overlap metric; hand-chosen and minor.
  • KL epsilon smoothing = 1e-10
    Numerical stabilization in Section 3.2; hand-chosen, negligible effect except on zero-probability tokens.
axioms (5)
  • domain assumption Residual stream at the target token position reflects the model's representation of the word sense.
    Used in all three experiments; the paper does not validate this readout against probes or other locations (acknowledged in Section 5.1).
  • ad hoc to paper The KL and token-overlap validation filters define a meaningful subpopulation of ambiguous words.
    The pool of stimuli is constructed using these thresholds in Section 3.1, so all general conclusions about disambiguation are drawn from this filtered subset.
  • ad hoc to paper Post hoc layer groups can serve as categorical fixed effects for confirmatory tests.
    Section 3.6 defines early/middle/late groups after inspecting the data, then tests differences between these groups.
  • domain assumption Activation patching with a full forward pass measures causal influence of a layer's representation.
    Standard interpretability assumption used in Section 3.3; it establishes causal sufficiency under intervention, not the natural computational pathway.
  • standard math Standard definitions of cosine distance, softmax, and KL divergence apply unchanged.
    Equations in Section 3.2 rely on standard definitions, with epsilon smoothing noted.

reviewed 2026-08-04 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Divergent large language model predictions from convergent representations in ambiguous word pairs." pith.science (2026). https://pith.science/paper/E2P7QHWX

@misc{pith2026260801816,
  author       = {Pith},
  title        = {Pith review of: Divergent large language model predictions from convergent representations in ambiguous word pairs},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/E2P7QHWX}},
  note         = {Machine review of arXiv:2608.01816}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

In this work we investigate how decoder-only transformers resolve lexical ambiguity through layer-by-layer analysis of three models spanning three parameter sizes (GPT-2-Small-117M, Llama-3.2-3B, Qwen2.5-32B). For both homonyms and polysemes, we find that representations become maximally distinct in middle layers, then partially reconverge in late layers, while the KL divergence between their next-token predictions reaches its maximum in the final layers. The activation patching experiment provides causal evidence that late-layer representational differences directly determine outputs despite apparent increased similarity in embedding space. Our single-layer ablation experiment indicates that models achieve equivalent disambiguation despite qualitatively different layer-wise vulnerabilities. These findings offer a mechanism for recent observations where models' internal embedding similarities show low correlation with their behavioural outputs despite strong performance. The semantic distinctions therefore remain present but become increasingly invisible to similarity measures over the embeddings, with implications for embedding-based methods such as semantic search, retrieval, and clustering that rely on late-layer cosine similarity.

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

22 extracted references · 3 canonical work pages · 1 internal anchor

  1. [1]

    Evaluating Contextualized Representations of (Spanish) Ambiguous Words: A New Lexical Resource and Empirical Analysis

    https://doi.org/10.48550/arXiv.2406.14678. Skean, Oscar, Md Rifat Arefin, Dan Zhao, Niket Patel, Jalal Naghiyev, Yann LeCun, and Ravid Shwartz-Ziv

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2406.14678

  2. [4]

    Foundation-size effects were evaluated separately for each model, ambiguity type, and degradation metric using Kruskal–Wallis tests

    A foundation size of k indicates that the first k layers were protected and that disruption was applied only at subsequent eligible layers. Foundation-size effects were evaluated separately for each model, ambiguity type, and degradation metric using Kruskal–Wallis tests. These secondary analyses used the raw degradation measures underlying the percentage...

    1932

  3. [5]

    Gage, Philip

    How do multilingual language models remember facts? InFindings of the Association for Computational Linguistics: ACL 2025, pages 16052–16106, Association for Computational Linguistics, Vienna, Austria.https://doi.org/10.18653/v1/2025.findings-acl.827. Gage, Philip

    doi:10.18653/v1/2025.findings-acl.827 2025

  4. [6]

    ArXiv:2407.21783

    The llama 3 herd of models. ArXiv:2407.21783. https://doi.org/10.48550/arXiv.2407.21783. Gurnee, Wes, Nicholas Sofroniew, Adam Pearce, Mateusz Piotrowski, Isaac Kauvar, Runjin Chen, Anna Soligo, Paul Bogdan, Euan Ong, Rowan Wang, Ben Thompson, David Abrahams, Subhash Kantamneni, Emmanuel Ameisen, Joshua Batson, and Jack Lindsey

    Pith/arXiv arXiv doi:10.48550/arxiv.2407.21783

  5. [8]

    ArXiv:2404.15255.https://doi.org/10.48550/arXiv.2404.15255

    How to use and interpret activation patching. ArXiv:2404.15255.https://doi.org/10.48550/arXiv.2404.15255. Jain, Shantanu

    Pith/arXiv arXiv doi:10.48550/arxiv.2404.15255

  6. [10]

    https://doi.org/10.48550/arXiv.2406.19384

    The remarkable robustness of LLMs: Stages of inference? ArXiv:2406.19384. https://doi.org/10.48550/arXiv.2406.19384. Ma, Matthew King-Hang, Xie Chenwei, Wenbo Wang, and William Shiyuan Wang

    Pith/arXiv arXiv doi:10.48550/arxiv.2406.19384

  7. [11]

    InFindings of the Association for Computational Linguistics: ACL 2025, pages 19705–19724, Association for Computational Linguistics, Vienna, Austria

    Exploring layer-wise representations of English and Chinese homonymy in pre-trained language models. InFindings of the Association for Computational Linguistics: ACL 2025, pages 19705–19724, Association for Computational Linguistics, Vienna, Austria. https://doi.org/10.18653/v1/2025.findings-acl.1011. Nanda, Neel and Joseph Bloom

    work page doi:10.18653/v1/2025.findings-acl.1011 2025

  8. [12]

    https://doi.org/10.1109/ACCESS.2023.3266377

    A survey of text representation and embedding techniques in NLP.IEEE Access, 11:36120–36146. https://doi.org/10.1109/ACCESS.2023.3266377. Proietti, Lorenzo, Stefano Perrella, Simone Tedeschi, Giulia Vulpis, Leonardo Lavalle, Andrea Sanchietti, Andrea Ferrari, and Roberto Navigli

    arXiv 2023

  9. [13]

    ArXiv:2510.06477

    Attention sinks and compression valleys in LLMs are two sides of the same coin. ArXiv:2510.06477. https://doi.org/10.48550/arXiv.2510.06477. Radford, Alec, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever

    doi:10.48550/arxiv.2510.06477

  10. [14]

    InFindings of the Association for Computational Linguistics: EACL 2024, pages 868–874, Association for Computational Linguistics, St

    The shape of learning: Anisotropy and intrinsic dimensions in transformer-based models. InFindings of the Association for Computational Linguistics: EACL 2024, pages 868–874, Association for Computational Linguistics, St. Julian’s, Malta. https://doi.org/10.18653/v1/2024.findings-eacl.58. Rivière, Pamela D., Anne L. Beatty-Martínez, and Sean Trott

    doi:10.18653/v1/2024.findings-eacl.58 2024

  11. [15]

    ArXiv:2406.14678, Version

    Bidirectional transformer representations of (spanish) ambiguous words in context: A new lexical resource and empirical analysis. ArXiv:2406.14678, Version

    Pith/arXiv arXiv

  12. [17]

    ArXiv:2502.02013.https://doi.org/10.48550/arXiv.2502.02013

    Layer by layer: Uncovering hidden representations in language models. ArXiv:2502.02013.https://doi.org/10.48550/arXiv.2502.02013. Steck, Harald, Chaitanya Ekanadham, and Nathan Kallus

    Pith/arXiv arXiv doi:10.48550/arxiv.2502.02013

  13. [18]

    Timkey, William and Marten van Schijndel

    Is cosine-similarity of embeddings really about similarity? InCompanion Proceedings of the ACM Web Conference 2024, pages 887–890.https://doi.org/10.1145/3589335.3651526. Timkey, William and Marten van Schijndel

    arXiv 2024

  14. [20]

    Leveraging large language models for word sense disambiguation.Neural Computing and Applications, 37(6):4093–4110.https://doi.org/10.1007/s00521-024-10747-5. Yang, An, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chengyuan Li, Dayiheng Liu, Fei Huang, Guanting Dong, Haoran Wei, Huan Lin, Jian Yang, Jianhong Tu, Jianwei Zhang, Jianxin Yang...

    work page doi:10.1007/s00521-024-10747-5

  15. [21]

    ArXiv:2412.15115

    Qwen2.5 technical report. ArXiv:2412.15115. https://doi.org/10.48550/arXiv.2412.15115. 21 Supplementary Materials Full homonym and concept drift polyseme tables Tables S1 to S4 list the full candidate pools of homonym and concept-drift polyseme pairs from which the per-model stimulus sets in Scott et al. (2026) were drawn. Tables are organised by experime...

    Pith/arXiv arXiv doi:10.48550/arxiv.2412.15115 2026

  16. [2018]

    SentencePiece: A simple and language independent subword tokenizer and detokenizer for neural text processing. InProceedings of the 2018 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, pages 66–71, Association for Computational Linguistics, Brussels, Belgium. https://doi.org/10.18653/v1/D18-2012. 20 Lad, Vedang, Jin ...

    doi:10.18653/v1/d18-2012 2018

  17. [2019]

    How contextual are contextualized word representations? comparing the geometry of BERT, ELMo, and GPT-2 embeddings. InProceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 55–65, Association for Computational Linguistics, Hong K...

    doi:10.18653/v1/d19-1006 2019

  18. [2021]

    ArXiv:2109.04404

    All bark and no bite: Rogue dimensions in transformer language models obscure representational quality. ArXiv:2109.04404. https://doi.org/10.48550/arXiv.2109.04404. Yae, Jung H., Nolan C. Skelly, Neil C. Ranly, and Phillip M. LaCasse

    Pith/arXiv arXiv doi:10.48550/arxiv.2109.04404

  19. [2023]

    ArXiv:2303.08112.https://doi.org/10.48550/arXiv.2303.08112

    Eliciting latent predictions from transformers with the tuned lens. ArXiv:2303.08112.https://doi.org/10.48550/arXiv.2303.08112. Capone, Luca, Serena Auriemma, Martina Miliani, Alessandro Bondielli, and Alessandro Lenci

    Pith/arXiv arXiv doi:10.48550/arxiv.2303.08112

  20. [2024]

    In Proceedings of the 10th Italian Conference on Computational Linguistics (CLiC-it 2024), pages 148–156, CEUR Workshop Proceedings, Pisa, Italy

    Lost in disambiguation: How instruction-tuned LLMs master lexical ambiguity. In Proceedings of the 10th Italian Conference on Computational Linguistics (CLiC-it 2024), pages 148–156, CEUR Workshop Proceedings, Pisa, Italy. https://aclanthology.org/2024.clicit-1.19/. Ethayarajh, Kawin

    2024

  21. [2025]

    Transformer Circuits.https: //transformer-circuits.pub/2025/attribution-graphs/methods.html

    Circuit tracing: Revealing computational graphs in language models. Transformer Circuits.https: //transformer-circuits.pub/2025/attribution-graphs/methods.html. Belrose, Nora, Igor Ostrovsky, Lev McKinney, Zach Furman, Logan Smith, Danny Halawi, Stella Biderman, and Jacob Steinhardt

    2025

  22. [2026]

    Transformer Circuits Thread

    Verbalizable representations form a global workspace in language models. Transformer Circuits Thread. https://transformer-circuits.pub/2026/workspace/index.html. Heimersheim, Stefan and Neel Nanda

    2026

This paper was first reviewed by deepseek-v4-flash on August 4, 2026.