Pith. sign in

REVIEW 5 major objections 4 minor 13 references

Transformers separate right from wrong by rotating answers in representation space, not by scaling them

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 →

T0 review · deepseek-v4-flash

2026-08-02 20:57 UTC pith:I7EP7MZ3

load-bearing objection Novel trajectory measurement undermined by missing control and abstract-body mismatch—worth a referee's time, not acceptance as-is. the 5 major comments →

arxiv 2603.13259 v3 pith:I7EP7MZ3 submitted 2026-02-25 cs.CL cs.AI

How Transformers Reject Wrong Answers: Rotational Dynamics of Factual Constraint Processing

classification cs.CL cs.AI
keywords rotational divergenceactive suppressionlogit lensforced-completion probingfactual constraint processingcommitment ratiohallucination detectionscale threshold
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.

Correct and incorrect forced answers to factual queries trace paths through a transformer's hidden states that keep the same length but rotate away from each other, with the angular gap peaking at mid-depth and partially closing later. In models above about 1.6 billion parameters, the incorrect path pushes the logit-lens commitment to the correct answer down to about 0.08—far below the 0.5 equal-probability baseline—which the paper reads as the model actively suppressing the right token rather than passively failing. The pattern appears across decoder-only transformers from roughly 1B to 13B parameters and is absent in a 1.5B model, suggesting a scale threshold for this factual-processing mechanism. If correct, the finding implies that factual correctness shows up in direction on an approximate hypersphere, not in representation magnitude, which would redirect hallucination detection toward angular measurements.

Core claim

The paper's core discovery is that factual constraint processing in decoder-only transformers is geometric and rotational: when a model is forced to process a correct versus an incorrect single-token continuation of the same query, the displacement from the query-only representation has nearly identical length for both continuations (norm ratio within 3% of 1) while their angular separation grows through mid-depth (cosine similarity drops to roughly 0.65–0.69 in the larger models), then partially reconverges. Superimposed on this, the logit-lens commitment to the correct token collapses in the incorrect run to about 0.08 in the 7B and 13B models, meaning the unembedding readout assigns about

What carries the argument

The central machinery is the displacement field (Definition 1) and the derived measures of rotational divergence—cosine separation of displacement vectors with a norm ratio near unity—and the commitment ratio (Definition 3), which applies the final layer norm and unembedding to each intermediate hidden state to read out the model's intermediate preference between the correct and incorrect token. These two devices let the author separate angular from radial divergence and quantify 'active suppression' as κ < 0.5. The paper also uses linear probes and an attention allocation ratio as supporting measurements.

Load-bearing premise

The load-bearing premise is that the normalized logit-lens readout restricted to the two forced tokens faithfully reflects the model's internal preference, so that a commitment ratio below 0.5 demonstrates 'active suppression' rather than the model simply reading the token it was forced to carry; the paper explicitly concedes that only a random-token control can settle this.

What would settle it

Force a random unrelated token, such as 'apple', as a third continuation of the same factual queries and measure the commitment ratio between the correct token and the random token across layers; if κ also collapses below 0.5, the suppression reading is an artifact of the unembedding reading out the forced token, and the active-suppression claim collapses.

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

Share X Bluesky LinkedIn Reddit HN

If this is right

  • Hallucination detectors that compare embedding norms or magnitudes will miss the signal; the discriminative information lies in angular position, so cosine-based or projection-based methods are the ones with a geometric basis.
  • The commitment collapse offers a potential label-free internal flag because it appears before the final layer and does not require knowing the true answer if contrastive continuations are constructed, though the paper does not bridge to open-ended autoregressive generation.
  • The scale threshold implies that models below roughly 1.6B parameters lack the internal circuitry for this kind of conflict resolution, so mitigation strategies that work at larger scales may have nothing to engage with at smaller scales.
  • The positive correlation between linear separability and commitment gap in larger models suggests that representation and generation become coupled with scale, making intermediate layers the place to look for factual commitments.

Where Pith is reading between the lines

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

  • A direct test of whether rotational divergence is specific to factual constraints or a general binary-distinction mechanism would run the same protocol on non-factual contrasts (e.g., syntactic or stylistic continuations); if the same isometric divergence appears, the finding is about transformer computation generally, not hallucination specifically.
  • The paper's own caveat implies a decisive control: adding a random-token condition would separate 'active suppression' from 'conformity to the forced token'; the authors did not run it, so the suppression claim should be treated as provisional.
  • The claimed 1.5B/1.6B threshold could be tested within a single model family by evaluating the same queries at several nearby scales, which would distinguish a sharp phase transition from a gradual improvement artifact.
  • If angular divergence is the real substrate, then activation patching or steering should aim to rotate representations along the correct-answer direction rather than rescale them; that is a testable prediction for intervention methods.

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

5 major / 4 minor

Summary. The paper introduces 'forced-completion probing,' a method for tracking five geometric quantities across transformer layers when a model processes correct versus incorrect single-token continuations of factual queries. The central empirical claims are: (1) correct and incorrect answer paths separate primarily by rotation rather than by scaling, with near-constant displacement norms and falling cosine similarity; (2) in models above roughly 1.6B parameters, the incorrect path drives a two-token logit-lens commitment ratio far below 0.5, interpreted as 'active suppression' of the correct token; and (3) this behavior is absent in the smallest tested model, suggesting a scale threshold. The paper reports results for four models (LLaMA-2 13B, Mistral 7B, StableLM-2 1.6B, Qwen2 1.5B) across 300 queries in three domains, with additional linear-probe and attention-allocation analyses.

Significance. If the rotational-divergence result holds, it would be a useful, non-obvious observation about how transformer hidden states encode factual correctness, with potential implications for hallucination detection and interpretability. The paper also demonstrates good methodological instincts: it includes control query categories, Mann-Whitney tests with FDR correction, token-span matching for multi-token keywords, and it explicitly acknowledges several limitations and a missing control. However, the significance is currently undercut by major inconsistencies between the abstract and the body (model counts, activation-patching experiments, scale-threshold claims) and by the lack of a random-token control that the authors themselves identify as necessary to interpret the 'active suppression' result. The geometric results are interesting but the headline mechanistic claim is not yet supported.

major comments (5)
  1. [Abstract vs. §4.2, Table 1, Conclusion] The abstract claims 'six decoder-only transformers spanning four architecture families' plus a seventh flat model, and the final sentence of §8 claims 'seven models.' The body evaluates only four models (LLaMA-2 13B, Mistral 7B, StableLM-2 1.6B, Qwen2 1.5B); no Gemma or other Llama/Mistral variants appear. This directly undermines the 'holds across six models/four families' and 'phase transition' statements. Either add the missing models or restrict all claims to the four evaluated.
  2. [§5.2, Definition 3–4] The central 'active suppression' claim is not identifiable without a random-token control, as the abstract itself concedes. κ(ℓ) is a two-token softmax over exactly t+ and t−; when the forced token t− is processed, the residual stream at the last position contains the embedding/context of that token, so κ_min=0.08 may simply reflect the unembedding reading out the forced token, not active suppression. The paper presents no control where an unrelated random token is appended and κ is computed. Without it, Definition 4's interpretation collapses to forced-token conformity.
  3. [Abstract vs. body] The abstract promises activation-patching results: 'Where single-layer activation patching is cleanly interpretable (LLaMA-2 13B, Mistral 7B) it yields no layer band of consistent recovery; a third model ... we diagnose as a code-path artefact and exclude.' No section in the body reports any activation-patching experiment. This is a missing empirical result that is load-bearing for the 'distributed-by-trajectory' conclusion. Either add the experiment or remove the claim from the abstract.
  4. [Table 3] The t-statistics in Table 3 are computed at the layer ℓ*/L where κ_min occurs, after scanning all layers. This post-selection inference is not corrected for multiple comparisons, so reported p<10^-100 is inflated. Pairwise comparisons in §4.5 use FDR across layers, but the one-sample t-tests do not. Report the full κ profile and apply FDR or a hold-out layer selection procedure before claiming significance.
  5. [§1, §5.4, §7.5, Abstract] The scale-threshold claim is stated inconsistently. The abstract and §1 say 'phase transition' and 'emerge at 1.6B'; §5.4 and Limitation 5 say 'cannot be determined from four models alone' and 'could reflect architectural differences rather than scale alone.' The abstract also calls Qwen2 1.5B flat 'plausibly a tokenizer artefact' while §5.4 says it is 'evidence of a genuine absence.' These contradictions must be resolved and the claims carefully hedged.
minor comments (4)
  1. [Table 2] Norm ratios η are reported as point means without variance or confidence intervals. Given the claim of isometric divergence, report SEM or a distribution; otherwise the near-unity values cannot be assessed.
  2. [§5.4] The phrase 'Qwen2 achieves only 0.50 at its single valid layer' is unexplained. Which layer is 'valid' and why are other layers invalid? Clarify.
  3. [References] The StableLM-2 technical report reference has a malformed author list: 'Reshinth Adithyan Sauer, et al.' Should be corrected.
  4. [Abstract] The sentence 'It holds across six decoder-only transformers ... a seventh (Qwen2 1.5B) is flat' is ambiguous: is Qwen2 one of the six or an additional seventh model? Clarify the count.

Circularity Check

1 steps flagged

Active-suppression finding is Definition 4 restated: κ<0.5 is labeled 'active suppression,' then the observed κmin=0.08 is reported as independent evidence, despite the abstract conceding the forced-token conformity alternative and omitting the random-token control.

specific steps
  1. self definitional [Definition 4 (Section 3, M4); Section 5.2; Section 8 Conclusions]
    "Definition 4 (Active suppression). A model shows active suppression at layer ℓ if, when processing the incorrect answer, κ(ℓ) < 0.5 − ϵ ... This means the intermediate representation actively favors the incorrect token over the correct one—the model is not only uncertain but is driving probability away from the right answer. ... Section 5.2: κ does not only stagnate—it collapses below 0.5, reaching κmin = 0.08 ... meaning the model actively drives probability away from the correct token."

    Def. 4 makes 'active suppression' the name for κ<0.5 in the incorrect run. Section 5.2 treats κmin=0.08 as showing the model 'actively drives probability away from the correct token,' and the conclusion repeats it. That is the defining condition restated as a discovery, not a tested mechanism. Additionally, κ is a two-token softmax over exactly t+ and t− taken after t− has been appended; the residual stream carries the forced token, so low κ can be the forced token's own readout. The abstract concedes this ('could equally reflect the model conforming to the token it is forced to carry, which only a random-token control can settle'), and no such control is supplied. The suppression result is thus equivalent to Definition 4's criterion, not independent evidence.

full rationale

The rotational-divergence measurements (M1–M3) are self-contained: η≈1 and ρ are computed from matched correct/incorrect/query runs and checked against neutral and control categories, so those findings are not circular. No self-citation chain or imported uniqueness theorem is used; all references are external. The one load-bearing circularity is the active-suppression claim: Definition 4 stipulates that κ<0.5 during the incorrect run counts as 'active suppression,' and then Sections 5.2 and 8 report the measured κmin=0.08 as evidence for the mechanism. Since κ is a two-token ratio over exactly the correct and incorrect tokens after the incorrect token has been appended, the sub-0.5 value is plausibly the forced-token input being read out through the unembedding; the abstract itself flags this and calls for a random-token control that is never run. This makes the suppression result constructed by definition (partial circularity), though the paper's caveat and its independent geometric findings prevent it from being wholly circular; score 6.

Axiom & Free-Parameter Ledger

0 free parameters · 5 axioms · 0 invented entities

No numeric constants are fitted to data; the central parameters are design choices (query set, correct/incorrect pairs, keyword annotations, logit-lens normalization). The important load-bearing axioms concern what the measurements mean: the normalized logit lens is treated as a faithful readout of internal preference; last-token residual states are treated as the locus of answer processing; neutral queries are assumed to control for forced-completion artifacts without a random-token baseline; the Qwen2 vs StableLM comparison is treated as evidence about scale; attention weights are treated as mechanistic.

axioms (5)
  • domain assumption Normalized logit lens (final RMSNorm/LayerNorm + lm_head applied to intermediate hidden states) faithfully estimates the model's per-token preference.
    Invoked in Section 4.3 for commitment ratio (M4); if this projection is contaminated by the forced input token, κ_min < 0.5 does not imply active suppression.
  • domain assumption The hidden state at the last token position is the correct locus for studying answer processing at every layer.
    All geometric measurements (M1-M5) use only the last-token residual stream; the paper provides no evidence that other positions carry the relevant factual constraint signal.
  • domain assumption Neutral query controls are sufficient to distinguish factual processing from forced-completion artifacts, and no random-token control is needed.
    Section 4.1 uses a neutral category for this purpose, but the abstract explicitly states that a random-token control is required to settle the suppression interpretation; that control is never run.
  • domain assumption The absence of the phenomena in Qwen2 1.5B can be attributed to parameter scale rather than tokenizer or architecture differences.
    Section 5.4 and the abstract interpret Qwen2 flatness as a scale threshold, while the limitations section concedes architecture and tokenizer confounds cannot be disentangled with four models.
  • domain assumption Attention weights from the last token to manually annotated keyword spans are a valid mechanistic correlate of factual engagement.
    M5 (Definition 5) treats attention allocation as supporting evidence for commitment collapse; the paper does not establish causal direction from attention to suppression.

pith-pipeline@v1.3.0-alltime-deepseek · 11047 in / 14899 out tokens · 138635 ms · 2026-08-02T20:57:09.839903+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of How Transformers Reject Wrong Answers: Rotational Dynamics of Factual Constraint Processing." pith.science (2026). https://pith.science/paper/I7EP7MZ3

@misc{pith2026260313259,
  author       = {Pith},
  title        = {Pith review of: How Transformers Reject Wrong Answers: Rotational Dynamics of Factual Constraint Processing},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/I7EP7MZ3}},
  note         = {Machine review of arXiv:2603.13259}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

When a decoder-only transformer is forced to process matched correct and incorrect single-token continuations of a factual query, the two pathways through hidden-state space diverge: displacement vectors from the query-only representation keep near-equal magnitude but rotate apart, with angular separation growing through mid-depth before late layers resolve an asymmetric outcome. A logit-lens preference in the incorrect run falls far below the equal-probability prior (roughly 11.5x more mass on the incorrect token than the correct one). We read this pattern, rotational divergence then late-layer asymmetric commitment, as the geometric signature of the model externally appearing to reject a wrong continuation, while staying explicit that it is observational, not causal: the incorrect run could equally reflect the model conforming to the token it is forced to carry, which only a random-token control can settle. It holds across six decoder-only transformers spanning four architecture families (Llama, Mistral, Gemma, StableLM) from 1B to 13B parameters; a seventh (Qwen2 1.5B) is flat under our protocol, plausibly a tokenizer artefact, leaving an emergence threshold open. Linear probes recover the distinction at intermediate depth, and cross-domain transfer is structurally asymmetric, a financial-medical corridor transferring far better than transport pairs. Where single-layer activation patching is cleanly interpretable (LLaMA-2 13B, Mistral 7B) it yields no layer band of consistent recovery; a third model (StableLM-2 1.6B) recovers uniformly above the ceiling, which we diagnose as a code-path artefact and exclude. Under this scoped null the late-layer asymmetry is not localized to a single component, fitting a distributed-by-trajectory account rather than single-layer localized recall. We document this with forced-completion probing across seven models, three domains, and 300 queries.

Figures

Figures reproduced from arXiv: 2603.13259 by Javier Mar\'in.

Figure 1
Figure 1. Figure 1: shows trajectory similarity τ (ℓ) across normalized depth for three models (all Qwen2 1.5B values exceed 0.99, consistent with the scale threshold). In LLaMA-2 13B, τ (ℓ) drops from 0.99 at the embedding layer to a minimum of 0.49 at ℓ/L = 0.35 before reconverging to 0.94 at the final layer. Mistral 7B shows a similar pattern (τmin = 0.51 at ℓ/L = 0.31), and StableLM-2 a shallower version (τmin = 0.79 at ℓ… view at source ↗
Figure 2
Figure 2. Figure 2: Commitment ratio κ(ℓ) across normalized depth. Solid lines: correct answers by category type. Dashed gray: incorrect answers (all categories pooled). In LLaMA-2 and Mistral, κ for incorrect answers collapses below 0.10, indicating active suppression. StableLM-2 shows a weaker effect; Qwen2 1.5B (not shown) remains at 0.50 throughout. The larger models do something qualitatively different: they drive κ to 0… view at source ↗
Figure 3
Figure 3. Figure 3: Left: Probe accuracy (5-fold CV) across normalized depth. All three models peak at intermediate layers, with accuracy declining toward the final layer. Circles mark peaks. Right: Cross-domain transfer AUROC. Within-domain (blue) is near-perfect; cross-domain (red) degrades, particularly for StableLM-2. Qwen2 1.5B (not shown) achieves 0.50 throughout. 5.4 Scale Threshold and Linear Probing [PITH_FULL_IMAGE… view at source ↗

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

13 extracted references · 8 linked inside Pith

  1. [2]

    Jump to conclusions: Short- cutting transformers with linear transformations.arXiv preprint arXiv:2303.09435,

    Alexander Yom Din, Noga Tamir, Idan Szpektor, and Yoav Goldberg. Jump to conclusions: Short- cutting transformers with linear transformations.arXiv preprint arXiv:2303.09435,

  2. [5]

    Dissecting recall of factual as- sociations in auto-regressive language models

    Mor Geva, Jasmijn Bastings, Katja Filippova, and Amir Globerson. Dissecting recall of factual as- sociations in auto-regressive language models. InProceedings of the 2023 Conference on Empirical Methods in Natural Language Processing,

  3. [7]

    A survey on hallucination in large language models: Principles, taxonomy, challenges, and open questions.arXiv preprint arXiv:2311.05232,

    Lei Huang, Weijiang Yu, Weitao Ma, Weihong Zhong, Zhangyin Feng, Haotian Wang, Qianglong Chen, Weihua Peng, Xiaocheng Feng, Bing Qin, et al. A survey on hallucination in large language models: Principles, taxonomy, challenges, and open questions.arXiv preprint arXiv:2311.05232,

  4. [8]

    Mistral 7B.arXiv preprint arXiv:2310.06825,

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

  5. [9]

    The geometry of truth: Emergent linear structure in large language model representations of true/false datasets.arXiv preprint arXiv:2310.06824,

    Samuel Marks and Max Tegmark. The geometry of truth: Emergent linear structure in large language model representations of true/false datasets.arXiv preprint arXiv:2310.06824,

  6. [10]

    LLaMA: Open and efficient foundation language models.arXiv preprint arXiv:2302.13971,

    14 Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. LLaMA: Open and efficient foundation language models.arXiv preprint arXiv:2302.13971,

  7. [13]

    Representation engineering: A top-down approach to AI transparency.arXiv preprint arXiv:2310.01405,

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

  8. [2019]

    Transformer feed-forward layers are key-value memories

    Mor Geva, Roei Schuster, Jonathan Berant, and Omer Levy. Transformer feed-forward layers are key-value memories. InProceedings of the 2021 Conference on Empirical Methods in Natural Language Processing,

  9. [2020]

    Qwen2 technical report.arXiv preprint arXiv:2407.10671,

    An Yang, Baosong Yang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Zhou, et al. Qwen2 technical report.arXiv preprint arXiv:2407.10671,

  10. [2021]

    How contextual are contextualized word representations? comparing the ge- ometry of BERT, ELMo, and GPT-2 representations

    Kawin Ethayarajh. How contextual are contextualized word representations? comparing the ge- ometry of BERT, ELMo, and GPT-2 representations. InProceedings of the 2019 Conference on Empirical Methods in Natural Language Processing,

  11. [2022]

    Transformers: State-of-the-art natural language processing

    Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumond, Clement Delangue, Anthony Moi, Pierric Cistac, Tim Rault, Remi Louf, Morgan Funtowicz, et al. Transformers: State-of-the-art natural language processing. InProceedings of the 2020 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, pages 38–45,

  12. [2023]

    Stable LM 2 1.6B technical report.arXiv preprint arXiv:2402.17834,

    Marco Bellagente, Jonathan Tow, Dakota Mahan, Duy Phung, Maksym Zhuravinskyi, Reshinth Adithyan Sauer, et al. Stable LM 2 1.6B technical report.arXiv preprint arXiv:2402.17834,

  13. [2024]

    Designing and interpreting probes with control tasks

    John Hewitt and Percy Liang. Designing and interpreting probes with control tasks. InProceedings of the 2019 Conference on Empirical Methods in Natural Language Processing,