The Anatomy of a Truth Direction: Knowledge-Dependent Dimensionality, a Relational Law, and a Convergent Category Geometry in Small Language Models

Francesco Karim Vicidomini

REVIEW 3 major objections 4 minor 17 references

Truth in small language models is a single knowledge-gated direction that attention builds and the feed-forward network overwrites.

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

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

T0 review · deepseek-v4-flash

2026-08-01 20:04 UTC pith:GXFNXPXE

load-bearing objection A serious, self-correcting empirical study of truth directions in small LMs — worth refereeing, but the headline 'propagate vs. overwrite' law is conditional on a pre-frame control that leaves an admitted indirect-path confound. the 3 major comments →

arxiv 2607.16741 v2 pith:GXFNXPXE submitted 2026-07-18 cs.LG

The Anatomy of a Truth Direction: Knowledge-Dependent Dimensionality, a Relational Law, and a Convergent Category Geometry in Small Language Models

Francesco Karim Vicidomini This is my paper

classification cs.LG

keywords truth directionminimal pairssingular value decompositionknowledge-dependent dimensionalityattention and feed-forward networksSwiGLU value streamcross-model convergenceinterpretability

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 tries to establish that truth in small language models is a low-dimensional, knowledge-dependent representational object with a measurable mechanism. Using a training-free axis derived from the SVD of hidden-state differences over true/false minimal pairs, it shows that on facts a model knows the truth signal concentrates on a single direction, while on facts it does not know the signal diffuses and a single axis recovers less of it. Mechanistically, the paper argues that attention propagates truth frames it did not write, while the feed-forward network opposes the current frame and deposits the material of later ones, with the post-peak decay carried by the value stream of the SwiGLU activation. Across five models from three families, per-category truth axes form a semantically signed arrangement that converges across families, but only for categories the models share knowledge of; the convergence is knowledge-gated and quantified as classical attenuation. If correct, the work gives a cheap structural probe for where truth is built and degraded in a transformer, and suggests the category geometry of truth belongs to the knowledge domain, not the training recipe.

Core claim

The central claim: a single SVD axis over true/false minimal-pair differences captures most of the truth signal on facts a model knows, and its failure on unknown or heterogeneous material is itself the signal—the dimensionality of truth is knowledge-dependent. Polarity's second dimension exists but is recoverable without polarity labels. Inside the residual stream, attention propagates truth frames it did not write, the feed-forward network opposes the current frame, and the post-peak erosion is causally carried by the SwiGLU value stream. Per-category truth axes form a semantically signed arrangement that converges across model families, knowledge-gated and quantified as classical attenuat

What carries the argument

The central object is the truth axis: the top right-singular vector of the SVD of hidden-state differences over true/false minimal pairs, read at the last token and identified without labels up to one global sign (Lemma 1). It costs two dot products per token and needs no training. Carrying the argument are three further instruments: an exact additive decomposition of residual states into attention and FFN contributions, gated by an identity check; a pre-frame versus post-frame control that removes the containment circularity and exposes the relational law; and an exact SwiGLU split into gate and value terms (plus a norm term for the sandwich-normalized architecture) that attributes post-pea

Load-bearing premise

The pre-frame control is assumed to remove all effects of the FFN contribution being measured, but indirect paths—accumulated earlier FFN writes that shape the pre frame—are not removed, and Empirical Law 1's attribution depends on those indirect paths being negligible.

What would settle it

Ablate the FFN in the layers before a frame block and re-fit the pre frame: if the axis moves materially, the measured 'opposition' of the FFN is an artifact of earlier writes rather than a genuine law; if the axis is stable, indirect contamination is negligible. The paper's own seven falsifications targeted the one-dimensionality claim, so a decisive test of the relational law should target the contamination path directly.

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

If this is right

A training-free, two-dot-product axis can trace where truth is built (attention in middle layers), where it is degraded (FFN value stream after the peak), and why the single-axis reading fails (unknown or heterogeneous material).
On facts a model knows, adding a second dimension to the truth axis adds no separation; the one-dimensional reading survives a seven-experiment falsification battery.
The same axis has two readings: its sign classifies true vs false, and its separability measures how well the model knows the material—a geometric, label-free confidence estimate.
Per-category truth axes form a semantically signed arrangement that converges across model families; the convergence is knowledge-gated, so the arrangement is a property of the knowledge domain, not the training recipe.
A measured dictionary of per-category axes and FFN write directions is exported, with a four-criterion contract for any sparse autoencoder claiming to do better.

Where Pith is reading between the lines

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

If the knowledge gate is real, the same mechanism predicts where hallucination is geometrically invisible: facts a model does not know produce overlapping true/false representations that no single direction can separate, suggesting detection must shift to per-atom expansion (e.g., learned dictionaries) in the low-knowledge regime.
The propagate/overwrite law suggests a candidate microscopic writer for macroscopic rotational-dynamics observations in factual-constraint processing; connecting the two could turn a correlational anatomy into a circuit-level mechanism.
The private geometry observed in one model family's dense bundle, if confirmed, implies that training diet (e.g., code and mathematics corpora) leaves a measurable residue in how truth is represented, beyond what shared knowledge explains; a fourth family with a known training pedigree would stress this.
The consensus sign gauge may generalize beyond truth axes: any SVD-based analysis with per-component orientations identified only up to sign can use the leading eigenvector of the cosine matrix to stabilize cross-condition comparisons.

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

A serious, self-correcting empirical study of truth directions in small LMs — worth refereeing, but the headline 'propagate vs. overwrite' law is conditional on a pre-frame control that leaves an admitted indirect-path confound.

read the letter

This paper is worth your time, but know that the headline 'Empirical Law 1' — attention propagates truth frames, the FFN opposes them — is the most novel claim and also the least settled. The basic result, a training-free SVD axis that reads truth on known facts, is credible and well-tested. The law is conditional on a control that has not yet been run.

What's genuinely new: a knowledge-dependent dimensionality claim (one axis on known facts, diffuse on unknown), a careful seven-experiment falsification battery, an exact attention/FFN decomposition attributing post-peak erosion to the SwiGLU value stream across five models, and a signed per-category geometry that converges across families in a knowledge-gated way. The paper also corrects its own earlier interpretation: the apparent 'flip at peak+1' is shown to be an artifact of the post-frame containing its own FFN write. That self-correction, together with the released code and one-command reproducibility, is real evidence of disciplined work.

Soft spots, in order of importance. First, the pre-frame control excludes the current block's FFN write but not earlier FFN writes that shape the axis. The paper admits this in Sec. 26. If later FFN writes are autocorrelated with earlier ones, the measured 'opposition' could be a projection artifact. The synthetic control in Sec. 16.6 did not model that specific contamination. This can be tested and is not a rejection of the study, but it does mean the law is not yet on solid ground. Second, the law-level and geometry claims rest on small datasets (250 CounterFact pairs, 60 pairs per category), single template per relation, and the continent anti-alignment is flagged as a possible template artifact. Third, the category sign gauge is fit to the same data whose convergence it demonstrates; the consensus gauge helps, but the 'private geometry' result depends on an unverified one-factor reliability model.

These are real caveats, but they are in proportion: the paper names most of them itself, which is rare. The core signal-on-known-facts result, the value-stream erosion, and the cross-family geometry are all solid enough to warrant referee time.

For you: if you work on interpretability or hallucination detection, this is a useful read and a good example of how to run a self-falsifying research process. I would cite it once the indirect-path control is published, but I'd be cautious about citing the law as settled. Send it to peer review — a serious referee will push for the control and for larger datasets, and that is exactly what this work needs.

Referee Report

3 major / 4 minor

Summary. The paper extends Bürger et al. (2024) in three directions: the knowledge-dependence of truth-direction dimensionality, the architectural component that builds and degrades the truth direction, and the semantic mixture underlying it. Part I, on Qwen2.5-1.5B, constructs a training-free axis from the SVD of hidden-state minimal pairs, reports held-out AUC 0.938 on 45 curated pairs (permutation p≈0.005, full-probe ceiling 0.989), shows the axis-to-probe gap grows on less-known and heterogeneous material, recovers the supervised polarity direction t_P without polarity labels, and reports that seven falsification attempts fail to overturn one-dimensionality on known facts. An exact attention/FFN decomposition and causal ablations attribute signal construction to attention in middle layers and post-peak erosion to the FFN. Part II extends the framework to five models across three families, introducing truth frames (pre and post), a fixed-frame gap statistic, and a SwiGLU split. The headline claims are: Empirical Law 1 (attention propagates frames it did not write; the FFN opposes them), the causal attribution of post-peak erosion to the SwiGLU value stream, a semantically signed per-category geometry that converges across families under a spectral consensus sign gauge, a knowledge-gated interpretation of that convergence as classical attenuation, and a model-private Qwen geometry at high sample density. The paper includes extensive registered predictions, negative results, tooling

Significance. If the central claims hold, the paper makes a substantial contribution: a cheap, training-free probe competitive with trained probes on known facts; a mechanistic, cross-family law about attention and FFN contributions; and a quantitative, knowledge-gated account of when single-direction truth reading fails. The manuscript's methodological discipline is a genuine strength: a one-command-per-number reproducibility contract, held-out protocols with permutation nulls that price layer selection, pre-registered predictions with failures reported as results, exact decomposition identity checks, and a synthetic axis-provenance control. The paper is also unusually honest, explicitly flagging confounds (Secs. 12 and 26) and correcting its own earlier interpretations. However, the most novel claim — Empirical Law 1 — rests on a pre-frame control whose admitted residual contamination has not been ruled out, and the knowledge-dependent dimensionality claim of Part I remains partially confounded with dataset and template identity. These are correctness risks on load-bearing points, not merely presentation issues.

major comments (3)

[Secs. 16.1, 17.3, 26; Table 9] Empirical Law 1 depends on the pre-frame control. The pre frame F_pre(b) is fit on h_b + a_b, which contains all FFN writes from earlier blocks. Section 26 admits: 'The pre frame removes the direct containment circularity but not indirect paths (accumulated earlier writes shape the pre frame).' If the pre-frame axis is substantially aligned with the subspace of these earlier FFN writes, then the measured negative d'(f_L; F_pre(b)) for L > b could be autocorrelation between successive FFN writes rather than a general 'propagate vs. overwrite' regularity. The synthetic axis-provenance control (Sec. 16.6) validated the pre-frame collapse on planted circular writers, but it did not model this specific indirect-path contamination. I suggest a direct test: regress the pre-frame axis onto the span of FFN writes from blocks < b, and re-measure the sub-diagonal of Table 9 after abating earlier FF
[Sec. 5.2, Table 1, Sec. 12] The knowledge-dependent dimensionality claim compares 45 curated pairs (behaviorally verified, one template set) with 250 CounterFact pairs (know-rate 30%, many templates) and a CounterFact/TruthfulQA mix. The axis-to-probe gap differences (0.051, 0.062, 0.203) are therefore confounded with dataset identity, template syntax, and topic heterogeneity. The paper itself notes in Sec. 12 that the 'unknown-fact' reading is 'confounded with sentence-format effects' and that filtering by a behavioral label is future work. Since the statement 'when the model knows a fact, its truth content is concentrated on a single direction' is one of the paper's two central contributions, the claim needs a within-dataset, within-template comparison of behaviorally known vs. unknown facts. As it stands, the evidence supports a monotone relation between know-rate and axis gap at the dataset level, but not the s
[Secs. 21, 22, 26; Table 13] The arrangement law is promoted to a law at the registered protocol, but the underlying category geometry has no bootstrap error bars (admitted in Sec. 26: 'no bootstrap error bars'), and the law is stated only after a sign-gauge repair. The sign instability documented in Sec. 22 is substantial: three of six Qwen seed re-samplings anti-correlate with the canonical matrix under the per-category gauge, and the dense cross-family comparison fails under that gauge. The consensus gauge repairs this, but it is selected post hoc from the same data. The authors argue convincingly that the consensus gauge is computed independently per model and that relative signs are gauge-invariant, but the report should include bootstrap confidence intervals on the Mantel statistics and on the per-category cosines, and a sensitivity analysis showing the law's conclusions are stable across reasonable alternativ

minor comments (4)

[Throughout] The manuscript alternates between 'hidden level' and 'block' indexing (Part I peak layer 16 for Qwen2.5-1.5B; Table 6 peak p=15). This is consistent but easy to misread; a single explicit conversion sentence in Sec. 4 or 15 would help.
[Sec. 4.2] 'Know-rate' is defined as a string-match lower bound on 200 prompts, and Sec. 22 later shows its per-relation version correlates only weakly with adherence (Spearman +0.21). This limitation is disclosed, but the narrative sometimes leans on the word 'knows' as if it were an established behavioral label; rephrasing to 'behavioral lower-bound know-rate' throughout would reduce over-reading.
[Sec. 22, Eq. (8)] The spectral consensus gauge is introduced via the leading eigenvector of the cosine matrix. It would help to state explicitly that the eigenvector is defined only up to sign and that the final sign is fixed by majority, as is done in the text, but the equation itself could carry a comment to that effect.
[Sec. 23] The private-geometry section is presented as a finding, though it rests on one dense bundle per family and its three predictions are registered but untested. The text does mark it as 'beyond the gate' and the mechanism as 'declared, not established,' which is appropriate; consider labeling it Tier C given the absence of a tested prediction.

Circularity Check

0 steps flagged

No significant circularity: the one admitted containment artifact is corrected, and the final laws rest on pre-frame controls plus independent held-out evidence.

full rationale

The paper's derived chain is self-contained rather than circular. The only place the word 'circularity' appears is in the admitted limitation that the pre-frame 'removes the direct containment circularity but not indirect paths' (Sec. 26), and in the correction of the earlier post-frame flip (Sec. 17.2), where the axis fit on h_{b+1} literally contained f_b and therefore dragged itself toward its own ingredient. That measurement is explicitly exposed as an artifact and is not load-bearing for the final claims: Empirical Law 1 is validated on the sub-diagonal of the pre-frame matrices, where neither a_L nor f_L is contained in the frame, and the erosion attribution uses frame-free ablations and an exact algebraic SwiGLU split (Eq. 5). The central dimensionality claim is evaluated held-out with permutation nulls against external datasets (CounterFact, TruthfulQA) and a full-probe ceiling, not fitted into the conclusion. The consensus gauge is computed independently per model, and the paper explicitly notes that two independent gauges cannot collude. The residual indirect-path contamination is a real, declared validity threat to the law's mechanistic reading, but it is a correctness risk, not a definitional reduction of the law to its inputs. No self-citation is load-bearing, and no fitted parameter is renamed as a prediction in the final derivation chain.

Axiom & Free-Parameter Ledger

6 free parameters · 7 axioms · 3 invented entities

The central claims rest on a small set of hand-chosen constants (peak layers, stability floors, category protocols, knowledge thresholds) and on a sign gauge fit to the same data that is later used to measure convergence. Several axioms are admitted by the author: minimal-pair isolation of truth, the pre-frame's imperfection, and the one-factor attenuation model for private geometry.

free parameters (6)

Peak layer p per model = 15, 16, 7, 9, 11 (Qwen-1.5B, Qwen-3B, Llama-1B, Llama-3B, Gemma)
Chosen as the block maximizing held-out truth AUC on training folds; all frame-relative laws (flip at p+1, erosion band) depend on this selection.
Consensus sign gauge s* = leading eigenvector of C (Eq. 8)
Fit to the per-category cosine matrix to fix unidentifiable per-axis signs; the cross-family convergence law is reported under this fitted gauge.
Knowledge-restriction threshold tau = 0.6 (within-category held-out AUC on both families)
The law's strength (+0.744 at K=33) is measured on the filtered set C_tau = {i : W_i ≥ 0.6} on both families; the threshold is hand-set.
Stability floor and significance margins = 2 sigma and 0.03 AUC/d'
A value is declared 'stable' only if it exceeds both 2σ across 5 seeds and a floor of 0.03; these constants gate which measurements enter the laws.
Category protocol K and pairs per relation = K=8 (top relations) with 60 pairs; K=33 with 60 and 888 pairs
The arrangement law is stated at one protocol (K=8, 60 pairs) and re-derived at K=33; category count, category set, and pair density are researcher-chosen.
Registered magnitude windows for confirmations = e.g., R1 cosine window [+0.15, +0.45]
The R1 window is described as registered before Llama measurements but was set after inspecting Qwen data, and was missed on the high side; such windows are fitted to earlier observations.

axioms (7)

domain assumption Minimal-pair hidden-state differences isolate truth from topic and surface form
The entire axis construction assumes that flipping one content word cancels shared context and leaves a truth component; declared lexical confounds and controls in Sec. 4.2 reduce but do not eliminate this risk.
domain assumption Truth is linearly readable in the residual stream
The axis/probe analysis presupposes linear decodability of sentence truth; supported by [2,5] but not derived here.
domain assumption Greedy-completion know-rate measures behavioral knowledge
The knowledge-dependence claims group facts by whether a greedy 8-token completion contains the target; this is a noisy lower bound (Sec. 4.2, Sec. 26) and correlates only weakly with cross-family adherence (Sec. 22).
standard math Residual stream decomposes additively into attention and FFN contributions
By construction with forward hooks; identity check verified at ~1e-7 in Sec. 4.7, so this is exact in the model's forward pass.
domain assumption The pre-frame excludes all effects of the measured FFN write
Sec. 26 explicitly admits indirect paths remain (accumulated earlier writes shape the pre frame); Empirical Law 1 depends on these being negligible.
ad hoc to paper Sign comparability of per-category axes across models
The consensus gauge (Eq. 8) is fit per model to fix the unidentifiable global sign of each category axis; cross-model agreement under it is then read as a law, which is the claim under test.
ad hoc to paper One-factor latent arrangement for attenuation/reliability
The private-geometry estimate rel(Qwen,60) = 0.537 uses a one-factor reading of the cross-model agreement triangle (Sec. 23); a fourth family is registered but untested.

invented entities (3)

Truth frame (pre/post) no independent evidence
purpose: Oriented SVD axis at a block; the object relative to which attention/FFN alignment is measured.
A measurement construct, not externally observable; Empirical Law 1 is defined entirely in its terms.
Knowledge gate / classical attenuation independent evidence
purpose: Explains why cross-family agreement weakens for low-knowledge categories.
Falsifiable prediction (restricting to known categories raises agreement) was tested and confirmed with the +0.424→0.744 gradient; though thresholds are chosen post hoc.
Private geometry (Qwen dense bundle) no independent evidence
purpose: Accounts for Qwen@888 structure not explained by the shared arrangement.
Pre-registered discriminations (OLMo-2, Qwen2.5-Coder, fourth distilled generalist) are untested; current support is a one-factor reliability model.

pith-pipeline@v1.3.0-alltime-deepseek · 32284 in / 14498 out tokens · 146180 ms · 2026-08-01T20:04:16.692642+00:00 · methodology

0 comments

read the original abstract

B\"urger et al. (2024) demonstrated that truth representations in large language models are universal across statement polarity but reside within a multidimensional subspace. We extend this framework along three questions: how the dimensionality of the subspace depends on the model's knowledge, which architectural component builds the truth direction, and what the direction is a mixture of. In Part I, a training-free directional probe derived from the SVD of hidden-state minimal pairs shows that the dimensionality of truth is knowledge-dependent: the signal concentrates on a single axis for known facts and diffuses as knowledge decreases. In Part II, a relational law emerges across multiple model families: attention propagates truth frames, the feed-forward network opposes the current block's frame, and post-peak decay is causally attributed to the SwiGLU value stream. Furthermore, per-category truth axes form a semantically signed arrangement that converges across families. Stress tests expose a sign instability in this orientation, which we repair with a spectral consensus gauge to sharpen the convergence into a knowledge-gated law. Finally, a replication campaign on Gemma-2-2b, extending our decomposition tools to accommodate its sandwich normalization, confirms these laws and attributions. We quantify the knowledge gate as classical attenuation and isolate a stable, model-specific private geometry.

Figures

Figures reproduced from arXiv: 2607.16741 by Francesco Karim Vicidomini.

**Figure 2.** Figure 2: Experiment 7, CounterFact at layer 16. Left: the 2D plane; the true (blue) and false (red) centroids (diamonds) are close relative to the cloud spread (centroid-separation ratio 0.55, overlapping). Right: the three-dimensional (v1, v2, v3) scatter shows a gradient along v1 (position) but no distinct clumps. The separation is confined to the horizontal position axis; the extra dimensions are noise. 7.1 Per-… view at source ↗

**Figure 3.** Figure 3: Experiment 7, TruthfulQA at layer 16. The same picture on more ambiguous facts: centroid-separation ratio 0.51 (overlapping), no three-dimensional clumps. Lower knowledge, lower separability, identical one-dimensional structure. margin +0.284, pˆ = 0.0099; CounterFact 0.816 vs. 0.561, margin +0.254, pˆ = 0.0099). But relative to the compressed residual at the same layer, the expanded space gains essentiall… view at source ↗

**Figure 4.** Figure 4: The FFN’s frame writing (d ′ ) across depth, aligned at the truth peak, four models. Right: normalized by the flip amplitude. 17.3 The relational law [PITH_FULL_IMAGE:figures/full_fig_p023_4.png] view at source ↗

**Figure 5.** Figure 5: Propagate vs. overwrite on four models: held-out [PITH_FULL_IMAGE:figures/full_fig_p024_5.png] view at source ↗

**Figure 6.** Figure 6: The category geometry: early vs. peak signed-cosine matrices, and decoding vs. the lexical [PITH_FULL_IMAGE:figures/full_fig_p028_6.png] view at source ↗

**Figure 7.** Figure 7: The arrangement law (R5). Left: the two K=8 peak signed-cosine matrices. Center: their off-diagonal cells against each other. Right: cross-model agreement at the peak and at the embedding level, and the within-model reorganization between them. recomputes r on Cτ = { i : WA i ≥ τ ∧ WB i ≥ τ } with Wi the within-category held-out AUC and τ = 0.6. Per-category agreement is the adherence adhi = Pearson Oi(MA)… view at source ↗

discussion (0)

Reference graph

Works this paper leans on

17 extracted references · 8 linked inside Pith

[1]

Burns, H

C. Burns, H. Ye, D. Klein, J. Steinhardt.Discovering Latent Knowledge in Language Models Without Supervision.arXiv:2212.03827, 2022 (ICLR 2023)

Pith/arXiv arXiv 2022
[2]

Marks, M

S. Marks, M. Tegmark.The Geometry of Truth: Emergent Linear Structure in Large Language Model Representations of True/False Datasets.arXiv:2310.06824, 2023 (COLM 2024)

Pith/arXiv arXiv 2023
[3]

K. Li, O. Patel, F. Viégas, H. Pfister, M. Wattenberg.Inference-Time Intervention: Eliciting Truthful Answers from a Language Model.NeurIPS, 2023

2023
[4]

Zou et al.Representation Engineering: A Top-Down Approach to AI Transparency

A. Zou et al.Representation Engineering: A Top-Down Approach to AI Transparency. arXiv:2310.01405, 2023

Pith/arXiv arXiv 2023
[5]

Bürger, F

L. Bürger, F. A. Hamprecht, B. Nadler.Truth is Universal: Robust Detection of Lies in LLMs. arXiv:2407.12831, 2024 (NeurIPS 2024)

Pith/arXiv arXiv 2024
[6]

Cunningham, A

H. Cunningham, A. Ewart, L. Riggs, R. Huben, L. Sharkey.Sparse Autoencoders Find Highly Interpretable Features in Language Models.arXiv:2309.08600, 2023

Pith/arXiv arXiv 2023
[7]

Bricken et al.Towards Monosemanticity: Decomposing Language Models With Dictionary Learning.Transformer Circuits Thread, Anthropic, 2023

T. Bricken et al.Towards Monosemanticity: Decomposing Language Models With Dictionary Learning.Transformer Circuits Thread, Anthropic, 2023

2023
[8]

Elhage et al.Toy Models of Superposition.Transformer Circuits Thread, Anthropic, 2022 (arXiv:2209.10652)

N. Elhage et al.Toy Models of Superposition.Transformer Circuits Thread, Anthropic, 2022 (arXiv:2209.10652)

Pith/arXiv arXiv 2022
[9]

Elhage et al.A Mathematical Framework for Transformer Circuits.Transformer Circuits Thread, Anthropic, 2021

N. Elhage et al.A Mathematical Framework for Transformer Circuits.Transformer Circuits Thread, Anthropic, 2021

2021
[10]

Bar-Shalom, F

G. Bar-Shalom, F. Frasca, Y. Galron, Y. Ziser, H. Maron.Beyond Token Probes: Hallucination Detection via Activation Tensors with ACT-ViT.arXiv:2510.00296, 2025

arXiv 2025
[11]

K. Meng, D. Bau, A. Andonian, Y. Belinkov.Locating and Editing Factual Associations in GPT.NeurIPS, 2022. Dataset mirror:NeelNanda/counterfact-tracing

2022
[12]

arXiv:2507.16488, 2025

ICR Probe: Tracking Hidden State Dynamics for Reliable Hallucination Detection in LLMs. arXiv:2507.16488, 2025

Pith/arXiv arXiv 2025
[13]

LLM Hallucination Detection: A Fast Fourier Transform Method Based on Hidden Layer Temporal Signals (HSAD).arXiv:2509.13154, 2025

arXiv 2025
[14]

How Transformers Reject Wrong Answers: Rotational Dynamics of Factual Constraint Process- ing.Preprint, February 2026

2026
[15]

The HeadGenome Taxonomy: A Structural and Causal Map of Attention Head Specialization in Transformers.Preprint, OpenReview iG4i2pc4wg
[16]

M. Huh, B. Cheung, T. Wang, P. Isola.The Platonic Representation Hypothesis.ICML 2024. arXiv:2405.07987

Pith/arXiv arXiv 2024
[17]

S. Lin, J. Hilton, O. Evans.TruthfulQA: Measuring How Models Mimic Human Falsehoods. ACL, 2022. 41

2022

[1] [1]

Burns, H

C. Burns, H. Ye, D. Klein, J. Steinhardt.Discovering Latent Knowledge in Language Models Without Supervision.arXiv:2212.03827, 2022 (ICLR 2023)

Pith/arXiv arXiv 2022

[2] [2]

Marks, M

S. Marks, M. Tegmark.The Geometry of Truth: Emergent Linear Structure in Large Language Model Representations of True/False Datasets.arXiv:2310.06824, 2023 (COLM 2024)

Pith/arXiv arXiv 2023

[3] [3]

K. Li, O. Patel, F. Viégas, H. Pfister, M. Wattenberg.Inference-Time Intervention: Eliciting Truthful Answers from a Language Model.NeurIPS, 2023

2023

[4] [4]

Zou et al.Representation Engineering: A Top-Down Approach to AI Transparency

A. Zou et al.Representation Engineering: A Top-Down Approach to AI Transparency. arXiv:2310.01405, 2023

Pith/arXiv arXiv 2023

[5] [5]

Bürger, F

L. Bürger, F. A. Hamprecht, B. Nadler.Truth is Universal: Robust Detection of Lies in LLMs. arXiv:2407.12831, 2024 (NeurIPS 2024)

Pith/arXiv arXiv 2024

[6] [6]

Cunningham, A

H. Cunningham, A. Ewart, L. Riggs, R. Huben, L. Sharkey.Sparse Autoencoders Find Highly Interpretable Features in Language Models.arXiv:2309.08600, 2023

Pith/arXiv arXiv 2023

[7] [7]

Bricken et al.Towards Monosemanticity: Decomposing Language Models With Dictionary Learning.Transformer Circuits Thread, Anthropic, 2023

T. Bricken et al.Towards Monosemanticity: Decomposing Language Models With Dictionary Learning.Transformer Circuits Thread, Anthropic, 2023

2023

[8] [8]

Elhage et al.Toy Models of Superposition.Transformer Circuits Thread, Anthropic, 2022 (arXiv:2209.10652)

N. Elhage et al.Toy Models of Superposition.Transformer Circuits Thread, Anthropic, 2022 (arXiv:2209.10652)

Pith/arXiv arXiv 2022

[9] [9]

Elhage et al.A Mathematical Framework for Transformer Circuits.Transformer Circuits Thread, Anthropic, 2021

N. Elhage et al.A Mathematical Framework for Transformer Circuits.Transformer Circuits Thread, Anthropic, 2021

2021

[10] [10]

Bar-Shalom, F

G. Bar-Shalom, F. Frasca, Y. Galron, Y. Ziser, H. Maron.Beyond Token Probes: Hallucination Detection via Activation Tensors with ACT-ViT.arXiv:2510.00296, 2025

arXiv 2025

[11] [11]

K. Meng, D. Bau, A. Andonian, Y. Belinkov.Locating and Editing Factual Associations in GPT.NeurIPS, 2022. Dataset mirror:NeelNanda/counterfact-tracing

2022

[12] [12]

arXiv:2507.16488, 2025

ICR Probe: Tracking Hidden State Dynamics for Reliable Hallucination Detection in LLMs. arXiv:2507.16488, 2025

Pith/arXiv arXiv 2025

[13] [13]

LLM Hallucination Detection: A Fast Fourier Transform Method Based on Hidden Layer Temporal Signals (HSAD).arXiv:2509.13154, 2025

arXiv 2025

[14] [14]

How Transformers Reject Wrong Answers: Rotational Dynamics of Factual Constraint Process- ing.Preprint, February 2026

2026

[15] [15]

The HeadGenome Taxonomy: A Structural and Causal Map of Attention Head Specialization in Transformers.Preprint, OpenReview iG4i2pc4wg

[16] [16]

M. Huh, B. Cheung, T. Wang, P. Isola.The Platonic Representation Hypothesis.ICML 2024. arXiv:2405.07987

Pith/arXiv arXiv 2024

[17] [17]

S. Lin, J. Hilton, O. Evans.TruthfulQA: Measuring How Models Mimic Human Falsehoods. ACL, 2022. 41

2022