Curved Inference: Concern-Sensitive Geometry in Large Language Model Residual Streams

Rob Manson

REVIEW 4 major objections 4 minor 1 cited by

Curved Inference: Concern-Sensitive Geometry in Large Language Model Residual Streams

T0 review · 4 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read Concern-shifted wording measurably bends the residual-stream trajectories of transformer language models, and in LLaMA the effect scales with concern intensity.

desk verdict A clean, reproducible geometric probe whose central claim is unconfirmed for lack of a synonym-swap control—worth referee time if the fix is made. read the letter →

arxiv 2507.21107 v1 pith:MJ6V76AH submitted 2025-07-08 cs.CL cs.AI

Rob Manson This is my paper · ORCID

classification cs.CLcs.AI

keywords residualstreamcurvaturesaliencesemanticmetricpullbacklargelanguagemodelinterpretabilitypromptengineeringconcern-shiftedprompts

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

Curved Inference claims that the way a large language model's internal state bends as it reads a prompt is a measurable, semantics-sensitive quantity. Across 20 matched prompt scaffolds, the paper finds that shifting one word toward emotional, moral, or logical concern changes the residual-stream trajectory in both models, and in LLaMA3.2-3b the change in curvature and salience grows systematically and significantly as concern intensity increases. If correct, this gives a geometric, model-grounded signal for when and how strongly a model is reorienting its representation, with practical uses in alignment monitoring and architectural comparison.

What carries the argument

The carrier of the argument is the pullback semantic metric $G = U^{\top} U$, built from the unembedding matrix $U$ that maps residual vectors to logits; it redefines distances and angles in residual space so they are aligned with the model's output vocabulary. Under this metric, salience is the first-order step norm $S(t) = \|x_{t+1} - x_t\|_G$, and curvature is the second-order reorientation $\kappa_i = \sqrt{\|a_i\|_G^2 \|v_i\|_G^2 - \langle a_i,v_i\rangle_G^2} / \|v_i\|_G^3$, estimated by 3-point finite differences of residual vectors across layers. This metric makes curvature coordinate-invariant and ties bending to token-level semantic change; the discrete curvature formula plus matched prompt scaffolds (control versus one-word concern-shifted variants) is what lets the paper attribute curvature deltas to semantic concern.

What would settle it

Replace each concern word in the 20 scaffolds with a frequency- and length-matched neutral synonym; if curvature and salience deltas of the same size appear, the concern-specific claim is false. Alternatively, redo the pullback metric with randomly permuted unembedding rows: if the LLaMA scaling survives, the signal is not semantic.

Watch

Extended reading notes

Core claim

Concern-shifted prompts do not merely change model output; they deform the internal trajectory. Measuring per-token, per-layer curvature $\kappa_i$ and salience $S(t)$ in native residual space under the pullback metric $G = U^{\top} U$, the paper reports that LLaMA3.2-3b shows early, statistically significant scaling of both metrics with concern strength, while Gemma3-1b responds to concern but differentiates moderate from strong variants only weakly. The paper interprets this as evidence for two interacting geometric layers: a static latent conceptual structure in the embedding and unembedding matrices, and a contextual trajectory in the residual stream that bends according to prompt-specific inference. It also treats attention and MLP blocks as semantic lenses whose integrated effect is visible as curvature.

Load-bearing premise

The load-bearing premise is that swapping a single word in a matched scaffold changes only the intended semantic concern, and that the output-projection geometry built from the unembedding matrix turns residual-stream bending into a measure of meaning-change rather than a generic reaction to any token substitution.

Editorial extensions

If this is right

Concern-induced bends appear at the same token-layer positions whether the wording is moderate or strong, but the bends grow larger with stronger concern, so curvature tracks graded semantic pressure rather than being a generic response to lexical change.
LLaMA3.2-3b and Gemma3-1b show different geometric styles: LLaMA bends early, strongly, and persistently, while Gemma bends shallowly near layers 5-6, giving a quantitative handle on cross-model comparison.
Because salience and curvature are strongly anticorrelated in LLaMA ($r=-0.89$), the paper argues for a representational trade-off: high reorientation tends to come with shorter total movement, suggesting a kind of internal representational budget.
Low-dimensional projections distort the curvature signal below the paper's fidelity threshold, so native-space residual metrics are required to observe the effect.
Curvature spikes on morally or emotionally charged prompts could serve as real-time alignment telemetry, flagging risky internal states before they surface in output.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

A direct extension would add null baselines such as synonym swaps, scrambled word order, and random-weight forward passes; if curvature deltas survive these controls unchanged, the concern-specific reading would need revision.
The two-layer geometry suggests tracking tokens that are singular in embedding space through varied prompts to see which contexts amplify or neutralize their latent irregularity.
Early-layer curvature differences between the two models could partly reflect tokenization or positional-encoding differences, so testing under a shared tokenizer would separate semantic effect from preprocessing.
If the effect is genuinely semantic rather than token-level, translated versions of the same prompts should show similar curvature shapes, extending the paper's cross-lingual speculation.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

read the letter

Quick take: this is a competent, reproducible pilot that is over-interpreted. The core observation—changing one word changes residual-stream geometry—is trivially true; the meaningful question is whether the change tracks 'concern' specifically rather than any token edit. That question is left open because the only control is the neutral scaffold. No synonym-swap or structure-preserving non-concern substitutions were run, and Section 6 concedes this. Until that null baseline exists, 'concern-sensitive geometry' is indistinguishable from 'any word edit bends the residual stream.'

What's genuinely new: the specific empirical pattern—discrete curvature and salience under the pullback metric G=U^TU scaling with hand-coded concern intensity in LLaMA positive polarity (p=0.006, p=0.016)—isn't in the cited literature. The framing itself is standard differential geometry; the contribution is the matched-pair extension and the fact that code and data are available. The UMAP audit is a nice piece of negative method work.

Soft spots, in order of severity. First, the missing null control is load-bearing, not peripheral. Second, residual stream was selected as the only interesting site after looking at three sites, then significance was reported without multiple-comparison correction—so the p-values are optimistic. Third, the salience–curvature anticorrelation (r=-0.89 for LLaMA) is largely a definitional artifact: curvature divides by ||v||^3, salience is ||v||, so speed and turniness trade off mechanically. The 'representational budget' story doesn't address that. Fourth, Appendix C's 'core result'—a non-colinear shift induces curvature—is just the definition of curvature, dressed up as a criterion. It doesn't add information.

The paper is honest about its limitations, which I respect. The math is elementary but correct, and the code/data availability is real evidence. For a reader in interpretability, this is a useful cautionary example of how a legitimate geometric measurement can be over-attributed to a semantic variable. It is not a demonstrated result.

Recommendation: I would not accept this as-is, but I would not desk-reject it either. The missing control is a fixable flaw, not a conceptual one. Send it to review with a clear demand for synonym-swap and random-substitution baselines, correction for multiple comparisons, and a re-derivation of the salience–curvature relationship. If the authors do that, the paper could become a credible pilot. As it stands, it's a well-executed first draft of an experiment.

Referee Report

4 major / 4 minor

Summary. The paper proposes 'Curved Inference,' a geometric-interpretability framework that measures curvature and salience of the residual-stream trajectory of transformer LLMs under a pullback metric G = U^T U induced by the unembedding matrix. Using 20 hand-curated matched prompt pairs across seven semantic domains, the authors compare neutral controls with 'concern-shifted' variants in Gemma3-1b and LLaMA3.2-3b. They report that concern shifts alter residual-stream curvature and salience in both models, with LLaMA showing statistically significant scaling between moderate and strong positive-polarity variants. The paper also presents correlation analyses between metrics, a 'Semantic Lens' interpretation of attention/MLP layers, and a discussion of limitations including the absence of null controls.

Significance. The idea of treating the residual stream as a curve and measuring semantic reorientation via curvature is potentially interesting and complements existing interpretability tools. The authors include useful details on finite-difference curvature estimation, acknowledge the need for native-space (rather than projected) analysis, and make their code and prompts available. However, the headline claim—that curvature and salience scale with 'concern intensity'—is not currently established, because the empirical design lacks the decisive control conditions and statistical corrections needed to rule out generic token-edit sensitivity. If the authors can supply those controls and temper the claims accordingly, the framework could become a useful diagnostic; as it stands, the evidence is suggestive but not convincing.

major comments (4)

[§3.2, §4.3, Table 5, §6] The central claim that curvature tracks 'concern intensity' is confounded with generic lexical-substitution effects. The concern-shifted prompts differ from controls by replacing a single word (e.g., 'repeatedly' → 'desperately'), but no synonym-swap, structure-preserving random substitution, or other non-concern lexical baseline is reported. Any token substitution generically moves residual activations and changes both salience S(t) and the finite-difference curvature κ_i, so the observed LLaMA scaling (Table 5: positive curvature p = 0.006, salience p = 0.016) could arise from non-semantic sensitivity to any word edit. The paper itself concedes in §6: 'No null baselining... no scrambled, synonym-swapped, and random-weight prompts were tested.' This is not a peripheral limitation; it is the construct-validity condition for 'concern-sensitive geometry.' Without a null distribution of Δκ and ΔS under matched non-concern edits, the central claim is indistinguishable from 'any word edit bends the residual stream.'
[§4.3, Table 5, Abstract] The statistical support for the scaling claim is limited to one polarity in one model. Of the eight model–metric–polarity combinations in Table 5, only LLaMA positive-polarity curvature (p = 0.006) and salience (p = 0.016) reach conventional significance; all Gemma rows and both LLaMA negative-polarity rows are non-significant (p ≥ 0.127). The Abstract's statement that 'LLaMA exhibiting consistent, statistically significant scaling' overstates this pattern, and no multiple-comparison correction is applied across the eight tests. Moreover, the residual-stream site and the G = U^T U metric were selected after examining the data (as described in §4: 'This was not assumed a priori—it emerged through a comparative analysis'), so the reported p-values do not account for this selection. The authors should report corrected p-values, disclose the selection procedure, and soften the headline claim accordingly.
[Appendix C] The 'core result' in Appendix C is tautological rather than a substantive geometric criterion. Stating that a shift-induced step Δv not colinear with the control step v produces positive curvature is just the definition of curvature for a polygonal path bent into a two-dimensional plane; it does not establish that such bending reflects 'semantic divergence' or 'concern'. Similarly, the converse statement (colinear shifts produce zero curvature) is a geometric identity. The appendix is therefore not evidence for the empirical claims, and invoking it as support for κ_i legitimizes a circular argument. The authors should reframe this appendix as a sanity check of the discretization, not as a derivation that concern-induced semantic divergence produces curvature.
[§4.2, Table 3, Table 4] Cross-model curvature magnitudes are not comparable as reported because the pullback metric G = U^T U scales with the norm of U, which differs across models. Table 3 shows mean curvatures of ~3.7e-5 for Gemma versus ~0.074 for LLaMA—a factor of roughly 2000—yet the paper interprets this as LLaMA exhibiting 'early, strong, sustained semantic curvature' while Gemma shows 'weak, shallow curvature.' Without normalizing by a model-specific scale (e.g., a per-model null distribution of κ under random or neutral inputs, or reporting standardized effect sizes), these cross-model differences may simply reflect differences in unembedding norms or residual-stream step sizes. The 'representational budget' interpretation of the salience–curvature anticorrelation (r = −0.89 for LLaMA) is likewise unfalsifiable as stated, since no alternative mechanism is specified or tested.

minor comments (4)

[§5.2] There is a typo in the sentence 'sharp spikes in response to certain concernt-shifted prompts'—'concernt' should be 'concern'.
[Abstract and §1] The paper has several missing spaces and formatting artifacts, e.g., 'We proposeCurved Inference' and 'Curved Inferencereveals' in the Abstract; these should be corrected throughout.
[Table 5] The table would be clearer if it reported the number of prompts per polarity and the exact test used (e.g., Wilcoxon signed-rank), as well as effect sizes, rather than only p-values and ratios.
[§6] The Limitations section is admirably candid, but several of its concessions (e.g., 'curvature magnitudes remain relatively interpreted') directly contradict the strength of the claims in the Abstract and Conclusions; the authors should align the framing of the main text with these limitations.

Circularity Check

2 steps flagged · score 4.0 of 10

Partial circularity: 'concern' is operationalized using the very geometric divergence it is said to explain, and Appendix C's curvature 'criterion' restates the definition of curvature; the LLaMA scaling test itself is an empirical comparison and not forced.

self definitional [Section 1 (Introduction); operational definition of 'semantic concern']
"We define semantic concern as a latent dimension of meaning (such as emotional tone, moral framing, or identity signalling) that affects how the model integrates information. Concern-shifted prompts induce bends in the model’s internal trajectory even when surface tokens remain similar. While concern is defined operationally here (via prompt-class manipulations and the resulting geometric divergence), the findings suggest a layered perspective:"

The operational definition of 'concern' bundles the independent variable (prompt-class manipulation) with the dependent variable ('the resulting geometric divergence'). The paper's central conclusion that concern-shifted prompts bend the residual stream is therefore partly true by definition: a divergence is part of what the construct is stipulated to be. The residual empirical content is only that these particular word substitutions move activations relative to the neutral control; without a non-concern lexical-substitution baseline (which Section 6 admits was not run), 'concern sensitivity' and 'any token edit bends the trajectory' are not separated.
other [Appendix C (Discrete Curvature: Geometric Criterion for Semantic Divergence), C.1 Core Result and C.2 Interpretation]
"If there exists any layer ℓ∗ such that ∆vℓ∗ is not colinear with vctrlℓ∗, then the concern-shifted trajectory exhibits strictly positive curvature at that layer. That is, the path bends in a two-dimensional plane spanned by {vctrlℓ∗, ∆vℓ∗}."

This 'criterion' asserts that a direction-changing perturbation produces positive curvature, which is exactly the definition of curvature as non-colinearity of adjacent/derivative vectors used in Section 3.4. It is a mathematical identity, not an empirical or derived property of concern shifts: any non-colinear perturbation of any curve would satisfy it. Presenting it as a result that 'confirms' concern-induced bending and 'supports the use of κi' adds no information beyond the metric definition, and the paper uses it to make the empirical heatmaps look theoretically forced.

full rationale

The LLaMA scaling result (Table 5: positive curvature p=0.006, salience p=0.016) is a genuine comparison of fixed model activations between moderate and strong concern-shifted prompts; it is not produced by fitting a parameter, by a self-citation chain, or by an equation that defines the outcome. Self-citations [12] and [13] are descriptive and not load-bearing. The post-hoc selection of the residual-stream site (Section 4: 'only the residual stream exhibited...') is a multiple-comparisons concern but does not by construction force the reported deltas. However, the construct 'semantic concern' is operationally defined in Section 1 partly as 'the resulting geometric divergence', so the conclusion that concern causes divergence is partially stipulative. Appendix C is a tautological restatement of curvature as direction change, used as if it were a geometric proof of concern effects. The paper's own Section 6 limitation ('no scrambled, synonym-swapped, and random-weight prompts were tested... curvature magnitudes remain relatively interpreted') is a serious confound for the 'concern-specific' interpretation, but that is a validity threat rather than a circular reduction. Overall score 4: partial framing-level circularity exists, while the central quantitative comparisons retain independent empirical content.

Assumptions & free parameters 1 free parameters · 3 assumptions · 2 invented entities

The central claim rests on the assumption that residual-stream trajectories can be meaningfully treated as curves under a pullback metric, and that the prompt manipulation isolates a single semantic dimension. No free numerical parameters are fitted to data, but the ordinal concern levels and the choice of the unembedding pullback are modeling decisions. No new entities are introduced beyond the latent 'concern' construct and the 'semantic lens' analogy.

free parameters (1)

Concern intensity levels = ordinal (moderate, strong)
The scaling claim depends on a hand-chosen ordinal coding of prompt variants; the intensity axis is not quantitatively calibrated.

assumptions (3)

domain assumption Residual stream activations form a smooth trajectory in R^d such that 3-point finite differences estimate its curvature.
With only 26-28 layers, the discrete samples are sparse; the central metric assumes the underlying trajectory is smooth enough for these differences to be meaningful (Section 3.4).
domain assumption The pullback metric G = U^T U defines a semantically meaningful geometry.
The paper asserts that distances and angles under this metric reflect token-output semantics, but this is a modeling choice not validated against external benchmarks (Appendix A.4).
ad hoc to paper Replacing one word in a scaffold changes only 'semantic concern' and not other lexical properties.
The experimental design in Section 3.2 assumes the control and concern-shifted prompts differ semantically but not structurally; tokenization and word frequency still differ, and the paper acknowledges token-count variation between models.

invented entities (2)

Semantic concern
purpose: Latent dimension hypothesized to drive curvature changes in the residual stream when prompts carry emotional, moral, or identity framing.
Defined operationally via prompt classes and never measured independently; the paper itself states 'concern is defined operationally here' (Introduction, Appendix B).
Semantic Lens model
purpose: Metaphor for attention and MLP layers acting as lenses that bend token trajectories.
A descriptive analogy in Appendix A.6, not a tested mechanism.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Curved Inference: Concern-Sensitive Geometry in Large Language Model Residual Streams." pith.science (2026). https://pith.science/paper/MJ6V76AH

@misc{pith2026250721107,
  author       = {Pith},
  title        = {Pith review of: Curved Inference: Concern-Sensitive Geometry in Large Language Model Residual Streams},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/MJ6V76AH}},
  note         = {Machine review of arXiv:2507.21107}
}

read the original abstract

We propose Curved Inference - a geometric Interpretability framework that tracks how the residual stream trajectory of a large language model bends in response to shifts in semantic concern. Across 20 matched prompts spanning emotional, moral, perspective, logical, identity, environmental, and nonsense domains, we analyse Gemma3-1b and LLaMA3.2-3b using five native-space metrics, with a primary focus on curvature (\k{appa}_i) and salience (S(t)). These metrics are computed under a pullback semantic metric derived from the unembedding matrix, ensuring that all measurements reflect token-aligned geometry rather than raw coordinate structure. We find that concern-shifted prompts reliably alter internal activation trajectories in both models - with LLaMA exhibiting consistent, statistically significant scaling in both curvature and salience as concern intensity increases. Gemma also responds to concern but shows weaker differentiation between moderate and strong variants. Our results support a two-layer view of LLM geometry - a latent conceptual structure encoded in the embedding space, and a contextual trajectory shaped by prompt-specific inference. Curved Inference reveals how models navigate, reorient, or reinforce semantic meaning over depth, offering a principled method for diagnosing alignment, abstraction, and emergent inference dynamics. These findings offer fresh insight into semantic abstraction and model alignment through the lens of Curved Inference.

Figures

Figures reproduced from arXiv: 2507.21107 by the authors.

**Figure 2.** Neutral, Concern-Shift and Delta heatmaps for one variant of a single prompt using Gemma3-1b [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗

**Figure 3.** Neutral and Delta Curvature heatmaps for each variant of a single prompt using Gemma3-1b [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗

Figures from the paper (5 more)

**Figure 4.** Figure 4: Neutral and Delta Curvature heatmaps for each variant of a single prompt using LLaMA3.2-3b [PITH_FULL_IMAGE:figures/full_fig_p010_4.png]

**Figure 5.** Figure 5: Neutral and Delta Salience heatmaps for each variant of a single prompt using Gemma3-1b [PITH_FULL_IMAGE:figures/full_fig_p010_5.png]

**Figure 6.** Figure 6: Neutral and Delta Salience heatmaps for each variant of a single prompt using LLaMA3.2-3b [PITH_FULL_IMAGE:figures/full_fig_p011_6.png]

**Figure 7.** Figure 7: Mean Curvature Delta plots by prompt using Gemma3-1b [PITH_FULL_IMAGE:figures/full_fig_p014_7.png]

**Figure 8.** Figure 8: Mean Salience Delta plots by prompt using LLaMA3.2-3b [PITH_FULL_IMAGE:figures/full_fig_p015_8.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

Riemannian Geometry for Pre-trained Language Model Embeddings
cs.CL 2026-07 conditional novelty 6.0 of 10

Aggregating per-token pullback metrics via the Fréchet mean on the SPD manifold outperforms Euclidean mean pooling for sentence classification, with most of the gain attributable to geometric aggregation rather than l...

Reference graph

Works this paper leans on

1 extracted references · 1 canonical work pages · cited by 1 Pith paper

[1]

Exploring the Residual Stream of Transformers

1 -Yu, Z., et al.(2023) “Exploring the Residual Stream of Transformers”arXiv 2 -Jain, S. & Wallace, B.(2019) “Attention is not Explanation”arXiv 3 -Conneau, A., et al.(2018) “What you can cram into a single vector: Probing sentence embeddings for linguistic properties” arXiv 4 -Olah, C., et al.(2020) “Zoom In: An Introduction to Circuits”Distill 5 -Elhage...

work page 2023

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.

[1] [1]

Exploring the Residual Stream of Transformers

1 -Yu, Z., et al.(2023) “Exploring the Residual Stream of Transformers”arXiv 2 -Jain, S. & Wallace, B.(2019) “Attention is not Explanation”arXiv 3 -Conneau, A., et al.(2018) “What you can cram into a single vector: Probing sentence embeddings for linguistic properties” arXiv 4 -Olah, C., et al.(2020) “Zoom In: An Introduction to Circuits”Distill 5 -Elhage...

work page 2023