Pith. sign in

REVIEW 4 major objections 5 minor 20 references

A language model’s hidden states mark how a single text is organized: structural value lives in a token’s place in this text, not in its vector alone.

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

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

T0 review · grok-4.5

2026-07-31 07:27 UTC pith:AHUJOCN6

load-bearing objection A real training-free instrument with solid engineered controls and a thinner clinical half; the relational claim is worth taking seriously even if the Lacanian frame is optional. the 4 major comments →

arxiv 2607.28434 v1 pith:AHUJOCN6 submitted 2026-07-30 cs.AI cs.CL

Metaphor Tracer: A Theory-Informed Analysis of Hidden States

Marc Heimann , Roxana Assadi Moghaddam , Olga Brovkina , Mark Pettifor , Lutz Goetzmann This is my paper
classification cs.AI cs.CL
keywords hidden statestransformer geometryaggregator channeldifferentiator channelrelational structuretype transferpsychoanalytic annotationtraining-free interpretability
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.

This paper builds a training-free instrument that, from one forward pass, scores every token position on two geometric properties of the residual stream: an aggregator that consolidates the whole text into a stable configuration around an anchor, and a differentiator that marks transient transport of other tokens into that anchor’s subspace. Constants are frozen on one discovery text; everything else is confirmatory. Across three unrelated models the aggregator follows engineered register boundaries, aligns with a psychoanalyst’s pre-instrument markings of clinical transcripts above lexical and saliency controls, and holds under repetition while surprisal and attention drain. A type-transfer test then shapes the result: models whose anchor structure travels with lexical type read singular discourse worst, and in a matched base/instruct pair tuning raises fidelity without moving type-transfer. The claim is that hidden-state “structural value” is relational—a property of place in this text—operationalizing older theory of language as a testable geometry rather than an essentialist reading of vectors.

Core claim

Structural value in transformer hidden states is a property of a token’s place in this text, not of its vector alone. A frozen, training-free aggregator channel tracks organizing positions against engineered register boundaries (6/6 cells) and a psychoanalyst’s fixed markings (34/36 directional cells) with a graded increment above lexical controls, holds under repetition while surprisal and attention drain, and in the models that read singular discourse best does not transfer with lexical type.

What carries the argument

The Metaphor Tracer: from one teacher-forced pass, every token is treated as a candidate anchor whose core subspace is fixed by positive interaction growth; two channels then score it—the aggregator (final-layer participation ratio / capacity share of the whole-text configuration on that subspace) and the differentiator (pull-gap of depth-displacement energy into the subspace over the ever-recruited basin)—read in both the raw operative view and a quiet, substrate-partialled view.

Load-bearing premise

That one psychoanalyst’s pre-instrument markings of signifiers and ruptures on four clinical excerpts are a valid external criterion for the discourse-organizing structure the aggregator is claimed to measure.

What would settle it

On held-out clinical transcripts with the same frozen constants, either the blue-operative rank-AUC on expert-marked versus unmarked tokens falls to chance after the paper’s own lexical and attention residualizations, or same-type idiosyncratic subspace transfer in the high-fidelity models rises well above the random-anchor scaffold baseline while clinical fidelity stays high—breaking the reported fidelity/transfer inversion.

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

If this is right

  • Cross-text pooling by lexical type averages away the text-individual structure these channels read, especially in instruction-tuned models.
  • Surprisal and attention-rollout are the wrong instruments for organizing signifiers under return: they drain where the aggregator holds.
  • Instruction tuning can raise discourse-reading fidelity without de-essentializing type-carried geometry; lineage fixes transfer, tuning fixes reading skill.
  • Delimiter “pins” are not attention sinks under another name: within the delimiter class the aggregator’s top positions are attention-poor in the tuned models.
  • Dual anisotropy views are complementary maps, not noise variants: discourse mastery sits on the substrate, local binding and lexical web in the quiet dimensions.

Where Pith is reading between the lines

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

  • If structural value is constitutively relational, standard corpus-averaged probes and type-level metaphor classifiers are systematically misaligned with the object this paper isolates—one-text organization—so per-text relational readouts may be required wherever individuality of discourse matters.
  • The stated cross-instrument prediction (elevated turnover of suspended workspace contents at pin positions) offers a direct bridge to causal direction-space methods without sharing methodology.
  • A larger multi-reader clinical corpus could turn the provisional rupture typology (operative friction rim vs quiet-web withdrawal) into a pre-registered segmentation test the channels would have to pass blind.
  • Treating the model as a theory of language rather than only a tool reframes explainability: the target becomes the geometry of a reading, not a classifier decision.

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 introduces Metaphor Tracer, a training-free, single-forward-pass instrument that scores every token position on two channels read from residual-stream geometry: an aggregator (participation ratio / capacity_share on an anchor’s core subspace) and a differentiator (pull_frac_gap on the ever-recruited basin), each in operative and quiet (z-scored) views. Constants are frozen on one discovery text; all other texts are confirmatory. Across three models plus a matched Llama base/instruct twin, the authors report that the aggregator tracks an engineered register across reversed-order boundaries (6/6 cells), aligns directionally with a psychoanalyst’s pre-instrument markings of clinical transcripts (34/36 cells) with a graded residual above lexical/surprisal controls, holds under repetition while surprisal and attention drain, and—in models that track the singular discourse best—does not transfer with lexical type. A matched-lineage follow-up attributes discourse fidelity and channel exclusion to instruction tuning while type-transfer and recruitment order remain lineage-fixed. The central interpretive claim is that structural value is a property of a token’s place in this text, not of its vector alone: a relational rather than essentialist reading of hidden states, offered as an operationalization of prior Lacanian/Saussurean theory.

Significance. If the result holds, the paper supplies a rare training-free, per-text relational readout of hidden-state organization, with a frozen confirmatory battery, public code, deterministic regeneration from cached artifacts, and multiple non-channel controls (lexical baselines, entropy/surprisal, attention rollout, delimiter erasure, permutation nulls, junction-matched minimal pairs, matched base/instruct). The engineered register control, the repetition dissociation (aggregator holds while information and traffic measures drain), the two-view inversion, and the transfer/fidelity split are substantive contributions to interpretability regardless of the psychoanalytic framing. The matched-lineage decomposition—tuning raises clinical fidelity without moving type-transfer—is a clean and citable empirical fact. The theory-informed composition is unusual but made falsifiable at the clause level; that alone is a methodological contribution if the external anchors are treated with the weight they can bear.

major comments (4)
  1. [Abstract; Results §5; Discussion “Ground truth without agreement”] Abstract and §5: the clinical half of the strongest claim is uneven relative to how it is stated. Directional predictions were fixed after transcript H1, so 9/36 cells are in-sample; residualized blue-operative clears both permutation nulls cleanly for phi-4 (and the non-pooled instruct twin), the base clears only the block null, and Qwen is explicitly non-adjudicable after lexical subtraction (0.15–0.22 of rank variance explained by rarity/surprisal). The abstract’s “34/36 cells, with a graded increment above lexical controls” compresses a model-heterogeneous residual into a single success count. Either narrow the abstract/claim to the models and increments the controls support, or add held-out annotated material / a second situated reading so the external clinical anchor is not carried primarily by one reader’s residual.
  2. [Methods “Two channels”; Limitations; Results §3, §5] Methods (Aggregator strength) and Limitations: capacity_share is rank-coupled to core-subspace size k at Spearman 0.85–1.00 (median ≈0.98–0.99 in phi-4). The paper correctly notes that the channel is a participation ratio and runs at a fraction of k, but several headline readouts are rank statistics (clinical AUC, strong-decile pins, boundary density). For those, k and blue are nearly interchangeable. A load-bearing control is missing: report the clinical and boundary results with k (or log k) as the score in place of capacity_share, and/or residualize blue on k within-run. Without that, the claim that the aggregator measures consolidating occupancy rather than subspace extent remains only partly secured for the rank-based findings.
  3. [Results §6; Abstract; Limitations] Results §6 and battery test 7: the type-essence / fidelity inversion is the shape-giving result, but transfer is within-register only (shared types almost exclusively inside register blocks; one residual cross-register pair). The paper flags this in Limitations, yet the anti-essentialist conclusion—“structural value does not transfer with lexical type” in the models that read best—is easy to over-read as general. State the scope condition in the Results §6 paragraph and in the abstract’s transfer sentence, not only in Limitations, and avoid language that implies cross-discourse type-indifference beyond the measured within-register case.
  4. [Results §3; Discussion “The pin function”] Discussion “The pin function” and Results §3: the point-de-capiton / paragraph-binding analogy is operationalized mainly in phi-4 (operative Δcos=+0.022, 14/18 pins, p=0.031; quiet stronger). Qwen produces no paragraph-final pins, so clause (i) is untestable there; the base interleaves strata differently. The functional claim is already hedged as per-model, but the master-signifier vocabulary is used globally in the framing. Either restrict that vocabulary to models where the capiton tests apply, or supply an alternative binding test for lexical pins (e.g., same-subspace binding of co-occurring content spans) so the analogy is not carried by one model’s delimiter crust.
minor comments (5)
  1. [Figure 6; Results §5] Figure 6 / §5: mark in-sample (H1) vs confirmatory transcripts visually in the forest plot so the 34/36 tally’s dependence structure is readable without the prose caveat.
  2. [Methods “Anchor construction”; Reproducibility] Methods: mass_cut μ=0.7 and τ=0.7 are frozen on the discovery text; a brief sensitivity sweep (e.g., μ∈{0.6,0.7,0.8}) on confirmatory boundary direction and clinical AUC would reassure that the frozen constants are not knife-edge, even if not retuned.
  3. [Results §8; Methods “Attention-saliency pass”] Results §8 attention pass runs in bfloat16 with eager attention while the main pipeline is float32; state explicitly that rank statistics only are used and that no frozen battery number depends on this pass (already implied—make it one sentence at the head of the subsection).
  4. [Abstract; throughout] Typographical/formatting: several compounded words lack spaces in the compiled text (e.g., “structuretravelswithlexicaltype”, “base/instructpair”); clean these in production. Also standardize “quiet-dimensions view” vs “quiet view” on first use in Results.
  5. [Discussion “The global workspace”] Related work: the comparison to Gurnee et al. (2026) is useful; if the Jacobian-lens tooling is public as stated, a single quantitative contact (e.g., persistence break at pins on one open model) would turn a Discussion prediction into a small result—optional, not required for revision.

Circularity Check

1 steps flagged

Theory-informed design and self-cited Freudian/Lacanian frame, but confirmatory scores do not reduce to inputs by construction.

specific steps
  1. self citation load bearing [Introduction; Theoretical Framework (Heimann & Hübener 2024, 2025; Heimann 2026a–c)]
    "In prior theoretical work one of the authors argued, on the shared principles of Lacanian theory and transformer architecture, that the “understanding” a language model performs must have a describable structure: points that quilt a discourse... (metaphor)... That argument had no measuring device... The present paper is a first operationalization of these insights... a metric informed by theory, one that is found precisely because we are looking for the metaphoric in Lacan’s sense"

    The interpretive mapping of channels onto master-signifier/quilting and metaphor-as-transport is justified primarily by the authors’ own prior theory papers, not by an external uniqueness result. This is minor and not load-bearing for the numerical claims: engineered boundaries, frozen constants, and pre-instrument annotations still supply independent confirmatory content; the paper disclaims proving the theory and reports geometric tests that could have failed.

full rationale

The paper’s empirical chain is not circular in the analyzer’s sense. Aggregator and differentiator are geometric definitions (participation ratio on an anchor core subspace; basin pull-gap) fixed on one discovery text; engineered boundary directions, clinical annotations, and column declarations are fixed before scoring and are not algebraic restatements of those definitions. The 6/6 register result, type-transfer test, base/instruct split, repetition-vs-surprisal/attention dissociation, and lexical/attention controls are independent measurements, not fitted parameters renamed as predictions. Self-citation (Heimann & Hübener 2024/2025; Heimann 2026a–c) supplies the interpretive vocabulary (quilting, metaphor-as-transport, anti-essentialism) and motivated what to look for, which is ordinary theory-laden instrument design, not a uniqueness theorem that forces the numerical outcomes. The paper explicitly treats the theory as a non-falsifiable perspective and stakes falsifiability on per-text operational clauses. Mild residual concern: clinical marks and channel interpretation share a Lacanian selection grammar, and directional clinical predictions were fixed after H1 (9/36 cells in-sample), but that is design/external-validity risk, not derivation-by-construction. Honest finding: no significant circularity of the load-bearing empirical claims.

Axiom & Free-Parameter Ledger

6 free parameters · 6 axioms · 4 invented entities

The central empirical claim rests on a small set of hand-chosen scan constants frozen on one text, on defining basins and channels in a particular way, and on treating expert marks plus engineered boundaries as ground truth for ‘organization.’ Theoretical Lacanian correspondences are interpretive framing rather than derived necessities; the quantitative claim can be stated without accepting the full theory, but the instrument’s design and the reading of pins as points de capiton assume that framing. No new physical entity is postulated; the invented entities are operational constructs.

free parameters (6)
  • mass_cut μ = 0.7
    Core subspace is the smallest prefix of positive Δ dimensions covering fraction μ of positive mass; frozen at 0.7 on the discovery text and never retuned.
  • operative recruitment threshold τ = 0.7
    Fixed cosine threshold for basin recruitment in the raw view; frozen with scan defaults on the discovery text.
  • STRONG_QUANTILE = 0.90
    Strong anchors defined as top decile union per channel; sets density ratios and pin sets used in battery tests.
  • quiet-view τ_ℓ = 95th percentile per layer
    Per-layer threshold set at the 95th percentile of that layer’s cosine distribution after z-scoring; hand-chosen calibration for the whitened view.
  • consolidation profile window w = max(20, T/15)
    Rolling density window max(20, T/15) used in profiles; a design choice affecting visualization more than headline AUCs.
  • type-transfer filters = top 30 types; >50% scaffold removal; 150 baseline pairs
    Shared-type discovery uses alphabetic tokens ≥5 chars, top 30 by min cross-text frequency, ≤6 occurrences, idiosyncratic dims = dims used by >50% of anchors removed; shapes the transfer percentiles in §6.
axioms (6)
  • ad hoc to paper Ever-recruited basin membership (threshold crossed at any depth) is the right support for measuring processing transport rather than final-layer membership alone.
    Methods explicitly choose ever-recruited basins; channels inherit this processing character by construction.
  • ad hoc to paper Participation ratio of the final-layer configuration on the anchor’s core subspace (capacity_share) measures consolidating organization rather than a bare dimension count, despite high rank-coupling to k.
    Stated in Methods and Limitations; coupling Spearman ~0.85–1.0 is acknowledged.
  • domain assumption Disagreement between raw (operative) and per-layer z-scored (quiet) geometries is scientifically informative rather than a nuisance to correct.
    Contrasts with standard anisotropy-correction practice (Timkey & van Schijndel et al.); dual-view-as-dual-map is a design premise.
  • domain assumption A single psychoanalyst’s marks of signifiers/ruptures, and a domain expert’s column declarations, are appropriate external ground truths for within-text organization (with agreement rejected for the clinical tier).
    Tier-2 design and Discussion “Ground truth without agreement”; load-bearing for the clinical half of the claim.
  • domain assumption Transformer residual-stream geometry on a teacher-forced pass is a legitimate object for reading ‘a model’s reading’ of an individual text.
    Standard mechanistic-interpretability premise; free-running generation may differ (Limitations).
  • standard math Standard linear algebra and rank statistics (cosine on subspaces, participation ratio, Spearman, rank AUC, sign tests) apply as used.
    Ordinary ML/geometry toolkit; no exotic formal system.
invented entities (4)
  • Aggregator channel (blue / capacity_share on anchor core subspace) independent evidence
    purpose: Score positions that consolidate the whole text into a stable high-rank configuration.
    Operational construct composed from known SVD/participation-ratio tools; not an external physical entity. Independent handle is alignment with engineered and expert structure under controls.
  • Differentiator channel (red / pull_frac_gap on ever-recruited basin) no independent evidence
    purpose: Score processing metaphoricity as transient transport into an anchor’s subspace.
    Defined inside the paper’s basin construction; validated mainly via engineered anomaly signatures and exclusion patterns, not a separate external benchmark for ‘metaphor.’
  • Processing metaphoricity vs final-state metaphoricity split no independent evidence
    purpose: Interpret red as transport-in-flight without claiming resting metaphor content.
    Theoretical-operational distinction; paper defers final-state claims to future interventions.
  • Pin / point-de-capiton functional reading of top aggregator anchors no independent evidence
    purpose: Map high-aggregator positions to quilting points that bind segments.
    Analogical entity with stated falsification clauses (paragraph binding, material emptiness, non-transfer); only partly testable (e.g., no paragraph-final pins in Qwen).

pith-pipeline@v1.2.0-daily-grok45 · 38411 in / 4946 out tokens · 103755 ms · 2026-07-31T07:27:22.718921+00:00 · methodology

0 comments
read the original abstract

What do a language model's hidden states say about the organization of a single text? From one forward pass, without training, we score every token position on two properties. The *aggregator* measures whether the position consolidates the whole text into a stable configuration. The *differentiator*, whether other tokens are transiently carried into its subspace as the model reads: metaphor in its root sense, transport. Constants were frozen on one discovery text; every other is confirmatory. The aggregator is not, in the classic sense, an information measure, nor a measure of salience. Across three unrelated models, as a signifier repeats, its surprisal and its attention drain while its aggregator score holds: the channel marks a token's place in the text. That this tracks a reading rests on independent ground truth: an engineered register the aggregator follows across its boundaries (6/6 cells), and a psychoanalyst's marking of clinical transcripts, fixed before the instrument existed, in 34/36 cells, with a graded increment above lexical controls and dissociations no type-level measure reproduces. A transfer test gives the result its shape: the model whose token structure travels with lexical type reads the singular discourse worst, and in a matched base/instruct pair tuning raises fidelity without moving type-transfer. Structural value is a property of a token's place in *this* text, not of its vector alone: a relational rather than essentialist reading of hidden states, operationalizing theory that predated the instrument.

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

20 extracted references · 8 canonical work pages · 4 internal anchors

  1. [5]

    Metaphor Identification Using Large Language Models: A Comparison of RAG, Prompt Engineering, and Fine-Tuning

    “Metaphor Identification Using Large Language Models: A Comparison of RAG, Prompt Engineering, and Fine-Tuning.” Preprint, arXiv.https://doi.org/10.48550/ARXIV.2509.24

    work page doi:10.48550/arxiv.2509.24

  2. [7]

    The Extimate Core of Understanding: Absolute Metaphors, Psychosis and Large Language Models

    “The Extimate Core of Understanding: Absolute Metaphors, Psychosis and Large Language Models.”AI & Society . Advance online publication. https://doi.org/10.1007/s00146-024-01971-7. • Heimann, Marc, and Anne-Friederike Hübener

    work page doi:10.1007/s00146-024-01971-7

  3. [8]

    Circling the Void: Using Heidegger and Lacan to Think about Large Language Models

    “Circling the Void: Using Heidegger and Lacan to Think about Large Language Models.”Cognitive Systems Research 91: 101349. https://doi.org/10.1016/j.cogsys.2025.101349. • Heimann, Marc. 2026a. “Freudian AI? Transformer Models as a Proof of Concept for a Central Hypothesis in Freudian Theory.”Lacunae: APPI International Journal for Lacanian Psychoanalysis, no

    arXiv 2025

  4. [9]

    SIDU-TXT: An XAI Algorithm for NLP with a Holistic Assessment Approach

    “SIDU-TXT: An XAI Algorithm for NLP with a Holistic Assessment Approach.”Natural Language Processing Journal 7 (June): 100078.https://doi.org/10.1016/j.nlp.2024.100078. • Jia, Mumin, and Jairo Diaz-Rodriguez

    arXiv 2024

  5. [10]

    Unsupervised Text Segmentation via Kernel Change-Point Detection on Sentence Embeddings

    “Unsupervised Text Segmentation via Kernel Change-Point Detection on Sentence Embeddings.” Preprint, arXiv.https://doi.org/10.485 50/ARXIV.2601.18788. • Kazmierczak, Rémi, Steve Azzolin, Eloïse Berthier, et al

    doi:10.48550/arxiv.2601.18788

  6. [11]

    Benchmarking XAI Explanations with Human-Aligned Evaluations

    “Benchmarking XAI Expla- nations with Human-Aligned Evaluations.” Preprint, arXiv.https://doi.org/10.48550/ARX IV.2411.02470. • Kramer, Oliver

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2411.02470

  7. [12]

    Conceptual Metaphor Theory as a Prompting Paradigm for Large Language Models

    “Conceptual Metaphor Theory as a Prompting Paradigm for Large Language Models.” Preprint, arXiv.https://doi.org/10.48550/ARXIV.2502.01901. • Lacan, Jacques

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2502.01901

  8. [13]

    How Attention Sinks Emerge in Large Language Models: An Interpretability Perspective

    “How Attention Sinks Emerge in Large Language Models: An Interpretability Perspective.” Preprint, arXiv.https://doi.org/10.48550/ARXIV.2603.06591. • Plank, Barbara

    doi:10.48550/arxiv.2603.06591

  9. [14]

    The ‘Problem’ of Human Label Variation: On Ground Truth in Data, Modeling and Evaluation

    “The ‘Problem’ of Human Label Variation: On Ground Truth in Data, Modeling and Evaluation.”Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, 10671–82. https://aclanthology.org/2022.emnlp-main.731/. • Rudman, William, Catherine Chen, and Carsten Eickhoff

    2022

  10. [16]

    Stable Anisotropic Regularization

    “Stable Anisotropic Regularization.” Preprint, arXiv.https://doi.org/10.48550/ARXIV.2305.19358. • Solbiati, Alessandro, Kevin Heffernan, Georgios Damaskinos, Shivani Poddar, Shubham Modi, and Jacques Cali

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2305.19358

  11. [18]

    All Bark and No Bite: Rogue Dimensions in Transformer Language Models Obscure Representational Quality

    “All Bark and No Bite: Rogue Dimensions in Transformer Language Models Obscure Representational Quality.” Preprint, arXiv.https: //doi.org/10.48550/ARXIV.2109.04404. • Uma, Alexandra N., Tommaso Fornaciari, Dirk Hovy, Silviu Paun, Barbara Plank, and Mas- simoPoesio

    Pith/arXiv arXiv doi:10.48550/arxiv.2109.04404

  12. [19]

    The Geometry of Tokens in Internal Representations of Large Language Models

    “The Geometry of Tokens in Internal Representations of Large Language Models.” Preprint, arXiv.https://doi.org/10.48550/ARXIV.2501.10573. • Yan, Yu, Sheng Sun, Zenghao Duan, et al

    Pith/arXiv arXiv doi:10.48550/arxiv.2501.10573

  13. [20]

    From Benign Import Toxic: Jailbreaking the Language Model via Adversarial Metaphors

    “From Benign Import Toxic: Jailbreaking the Language Model via Adversarial Metaphors.” Preprint, arXiv.https://doi.org/10.48550 /ARXIV.2503.00038. 39

    doi:10.48550/arxiv.2503.00038

  14. [2020]

    Quantifying Attention Flow in Transformers

    “Quantifying Attention Flow in Transformers.” In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics , 4190–4197. https://doi.org/10.18653/v1/2020.acl-main.385. • Aghazadeh, Ehsan, Mohsen Fayyaz, and Yadollah Yaghoobzadeh

    doi:10.18653/v1/2020.acl-main.385 2020

  15. [2021]

    Unsupervised Topic Segmentation of Meetings with BERT Embeddings

    “Unsupervised Topic Segmentation of Meetings with BERT Embeddings.” Preprint, arXiv.https://doi.org/10.48550/ARXIV.2106.12978. • Timkey, William, and Marten van Schijndel

    Pith/arXiv arXiv doi:10.48550/arxiv.2106.12978

  16. [2022]

    Metaphors in Pre-Trained Language Models: Probing and Generalization Across Datasets and Languages

    “Metaphors in Pre-Trained Language Models: Probing and Generalization Across Datasets and Languages.” Preprint, arXiv.https://doi.org/10.48550/ARXIV.2203.14139. • Aroyo, Lora, and Chris Welty

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2203.14139

  17. [2023]

    Outlier Dimensions Encode Task Specific Knowledge

    “Outlier Dimensions Encode Task Specific Knowledge.” Proceedings of the 2023 Conference on Empirical Methods in 38 Natural Language Processing, 14596–605. https://doi.org/10.18653/v1/2023.emnlp-main.9

    work page doi:10.18653/v1/2023.emnlp-main.9 2023

  18. [2024]

    SepLLM: Accelerate Large Language Models by Compressing One Segment into One Separator

    “SepLLM: Accelerate Large Language Models by Compressing One Segment into One Separator.” Preprint, arXiv.https://doi.org/10.485 50/ARXIV.2412.12094. • Fuoli, Matteo, Weihang Huang, Jeannette Littlemore, Sarah Turner, and Ellen Wilding

    Pith/arXiv arXiv doi:10.48550/arxiv.2412.12094

  19. [2025]

    Leveraging Large Language Models to Estimate Clinically Relevant Psychological Constructs in Psychotherapy Transcripts

    “Leveraging Large Language Models to Estimate Clinically Relevant Psychological Constructs in Psychotherapy Transcripts.”Computational Psychiatry 9 (1): 187–209. https: //doi.org/10.5334/cpsy.141. • Abnar, Samira, and Willem Zuidema

    work page doi:10.5334/cpsy.141

  20. [2026]

    Verbalizable Representations Form a Global Workspace in Language Models

    “Verbalizable Representations Form a Global Workspace in Language Models.”Transformer Circuits Thread. https://tran sformer-circuits.pub/2026/workspace/. 37 • Heimann, Marc, and Anne-Friederike Hübener

    2026