acceptodds
Under review as a conference paper at ICLR 2027

Task Geometry: Response Channels Predict Explicit Recovery and Attention Localization

Abstract

In-context learning on fresh low-dimensional tasks requires attention to infer which input directions define task geometry. We study an explicit response-derived route to that geometry. A clipped second-moment response channel yields cross-fitted quadratic-softmax attention with log-free, dimension-explicit subspace recovery and the intrinsic point-query rate when the channel is open; when it is closed, that statistic is non-identifying even though its finite-sample geometry need not be Haar. In a controlled replication with frozen hypotheses on ten synthetic tasks at and , the mean Haar-minus-explicit geometry advantage is larger on six open tasks than on three closed tasks, with one preregistered numerically weak task outside the contrast. On the three closed odd-link tasks, a small task-specific Transformer attains lower prediction risk than both the Haar control and the explicit route. This trained result is predictive rather than geometric: it uses task information beyond the selected second-moment statistic, while an open first-moment channel provides a classical alternative explanation. Together, response channels predict explicit geometry recovery and separate it from broader task-specific prediction.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
attention geometry, Transformer attention, representation geometry, low-rank adaptation, operator identifiability, attention compression

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_BiA5OpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper studies in-context regression on synthetic tasks in which the target depends on a low-dimensional subspace of the input that changes in every context, and models attention as a kernel smoother whose metric should align with that subspace. Building on a geometry–risk result from an anonymous submission, it asks when this subspace can be estimated from the responses of the current context. Using a classical second-moment statistic from sufficient dimension reduction, the authors define a "channel" that is open when the statistic identifies the subspace and closed when it vanishes, for example for odd target functions. They prove recovery and prediction guarantees for an attention rule built on this statistic when the channel is open. They then test the open/closed prediction on ten synthetic tasks, against a random-subspace control and a small task-trained Transformer. Experiments are

Critical strengths
  1. Transparent reporting. The authors are open about the limits of their study. They say which tasks violate the assumptions of the theorem, they report the negative results (the rank-2 trained test fails, the T4 result depends on one seed) and they disclose that the learning rate was changed after seeing the results. Proposition 1 is also a nice small observation: even when the population moment is zero, the estimated subspace is not completely random in finite samples.
Critical weaknesses
  1. The paper is very hard to understand. I could not understand the abstract before reading the whole paper. It uses many terms that are defined only later (e.g. "clipped second-moment response channel", "log-free, dimension-explicit subspace recovery", "finite-sample geometry need not be Haar") and never says in simple words what problem is studied. The introduction is mostly about the protocol (frozen hypotheses, controlled replication, learning-rate correction) and not about the motivation. The main concept, open and closed channels, appears only in Section 4. Overall, it is hard to understand what is actually claimed.

  2. The paper depends on a concurrent anonymous submission that I cannot check. Theorem 1, the geometry–risk frontier on which the whole paper is built, comes from Anonymous (2026). The paper gives the statement and some intuition, but the proof "is not reproduced". The new theory is described as an improvement of that work, and the paper itself says the earlier work already reaches the intrinsic rate in most settings. Since I cannot read the anonymous paper, I cannot verify the main result or understand how big the contribution is.

  3. Limited evidence beyond known results.

  • The main mechanism is classical. The authors acknowledge that the blind spot of this kind of moment on odd functions is already known, so the open/closed result is largely predictable before running the experiment.
  • The experiments do not really test the theory: the implemented method differs from the one analysed in the theorem, several tasks are outside its assumptions, and only one setting is studied.
  • The Transformer comparison is not fair, since the model is trained on thousands of contexts per task while the explicit route only sees the current context. The results are also mixed: on one of the three closed tasks the advantage comes from a single seed, and on the rank-2 task the model is worse than the random control. Despite the title, attention localization is never measured in the trained model.
Questions for authors
  1. What is exactly new with respect to Anonymous (2026)? Can the paper be made self-contained, with the proof of Theorem 1 and a description of the original explicit route and of how its hyperparameters were selected?
Reviewer_6RTEOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This paper attempts to decompose ICL (assuming an x–y pair setting) into a two-stage process: first, suppressing the importance of task-irrelevant features in the input (x) (or, in the paper's terminology, adjusting the mass), and then performing a form of k-NN-like classification using these denoised features (the authors adopt a softer version of k-NN). The authors establish the statistical feasibility of this process and train a toy Transformer to provide a preliminary empirical investigation.

Critical strengths

The mathematical intuition proposed by the authors is interesting and arguably elegant. For a long time, we have tended to think that some well-known ICL mechanisms, such as Induction Heads, directly operate on existing sentence encodings within LLMs without performing additional geometric processing (e.g., removing task-irrelevant features, as discussed in this paper). This paper offers an interesting alternative mathematical perspective.

Critical weaknesses
  1. The proposed theory is difficult to connect to the ICL phenomena observed in real pretrained LLMs, even as I mentioned above, I find the theoretical framework rather elegant. In particular, the idea that ICL suppresses task-irrelevant features, prevents them from influencing the prediction of query outputs, and thereby enables a modified k-NN-like model is a research direction that I consider worth pursuing.

    • Modern interpretability research, especially mechanistic interpretability, increasingly seeks to develop coherent scientific frameworks that explain a broad range of phenomena observed in real models, much like traditional scientific research, rather than restricting the entire discussion to a simulated setting. Regarding the limitations of k-NN-based explanations, I recall a previous paper that attempted to reduce ICL to a k-NN model. If my memory serves me correctly, its reviewers pointed out that such a model could not explain several observed phenomena, such as why ICL tends to copy the label of the demonstration closest to the query.

    • However, the present paper instead constructs a model that is rather far removed from the actual computations performed by attention. The proposed construction is derived from classical statistical tools, and it remains unclear under what conditions, whether mild or restrictive, a parameterized Transformer could actually implement these computations.

    • Furthermore, the underlying mathematical intuition itself is not adequately validated in this paper. For example, what proportion of the hidden-state representation in pretrained LLMs is occupied by task-related features? What is the eccentricity of these representations? Such properties should also be investigated to establish whether the proposed mathematical intuition corresponds to what actually happens in real models.

      • Some earlier studies have already provided evidence that pretrained LLMs do indeed exhibit a process of removing task-irrelevant information from approximately isotropic hidden-state representations. These findings could potentially address the weakness I have raised. However, they would simultaneously diminish the novelty of the present work. Therefore, my overall assessment is that the underlying mathematical intuition is plausible, but it has already been recognized and investigated in prior work. Relevant examples include:
        - In-Context Learning with Transformers: Softmax Attention Adapts to Function Lipschitzness. NeurIPS 2024.
        - An Oracle Property of the Nadaraya–Watson Kernel Estimator for High-dimensional Nonparametric Regression. Scandinavian Journal of Statistics.
        - Pretrained Transformer Efficiently Learns Low-Dimensional Target Functions In-Context. NeurIPS 2024.
        - Mechanism of Task-oriented Information Removal in In-context Learning. ICLR 2026.
    • The connection between the trained toy Transformer and the theoretical model proposed in this paper has not been established. In particular, the authors do not appear to provide evidence at the level of the model's actual computational mechanisms, rather than merely its final outputs. (My understanding is that the authors only evaluate the final prediction results, and I apologize if I have overlooked any relevant analyses.) If the authors cannot demonstrate that the proposed statistical machinery is actually implemented even by a Transformer trained under their highly idealized assumptions, it becomes even more difficult to establish its relevance to real pretrained LLMs. Moreover, I suspect that such a mechanistic validation would be quite challenging. Although the authors acknowledge this limitation in their Limitations section, acknowledging a limitation does not exempt a scientific claim from the need for empirical grounding.

  2. The manuscript appears to be entirely AI-generated, with little evidence of manual refinement, and insufficient attention has been paid to human readability. The paper retains several characteristics commonly associated with AI-generated scientific writing that make it unnecessarily difficult to read. For example:

    • Specific numerical results are repeatedly interspersed with the main prose. Even the presentation of these numbers is somewhat unconventional for a major AI conference (e.g., 0.608 [0.600, 0.617]). This substantially distracts the reader from the main arguments, making much of the paper feel more like an appendix than a coherent scientific narrative.
    • Mathematical notation and equations are overused. Some theoretical details and mathematical derivations could be moved to the appendix, while the main text could focus on the logical arguments and more intuitive explanations. In my view, the mathematical intuition underlying this paper is not particularly complicated, but the authors have presented it in an unnecessarily complex manner.
    • For transparency, I acknowledge that I used an AI agent to assist me in reading this paper, as I found it exceptionally difficult to follow. However, the review comments themselves were written entirely by me and subsequently translated from my native language into English using AI.
  3. The paper cites an inaccessible and unverified source: "Anonymous. Geometry-adaptive attention enables intrinsic-dimensional prediction, 2026. Concurrent submission, under review." This work serves as an important part of the paper's theoretical background. The authors claim that it is a concurrent submission by the same authors, but I was unable to locate any corresponding anonymous manuscript on OpenReview. I consider this a serious concern regarding the foundation of the paper's motivation. Although I am willing to accept the intuition presented in Section 3, it currently lacks sufficient independently verifiable support. I would like to ask the program committee to investigate whether this reference constitutes a hallucinated citation.

Questions for authors

See weakness.

Reviewer_2h4fOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

Y=g(B⊤X)+εY=g(B^\top X)+\varepsilon, where the rank-rr subspace P=BB⊤P=BB^\top is redrawn from the Haar law and gg is bounded and ss-Hölder. An attention head is modeled as a positive normalized kernel row. In quadratic-softmax attention the precision matrix AA acts as the head's "localization geometry", and aligning AA with PP moves the risk from the ambient rate n−2s/(2s+d)n^{-2s/(2s+d)} toward the intrinsic rate n−2s/(2s+r)n^{-2s/(2s+r)}. The geometry–risk frontier that prices this information comes from a concurrent submission (stated as Theorem 1, not proved here). The paper asks which response statistics supply the geometry. Contributions as I understand them: Explicit route (Theorem 2; App. B.1–B.3). For the clipped, pilot-centred second-moment (pHd-type) "channel" Hϕ=BΓϕB⊤H_\phi=B\Gamma_\phi B^\top, the paper gives three results. A deterministic clipping level τ∗\tau_* guarantees sr(Γϕ)≥γ/2s_r(\Gamma_\phi)\ge\gamma/2. The projector error obeys a log-free, dimension-explicit bound EP^−Pop2s≤min⁡1,(Cv2d/(mγ2))sE\hat P-P_{op}^{2s}\le\min1,(Cv^2d/(m\gamma^2))^s. Cross-fitted quadratic-softmax prediction reaches the intrinsic rate when 2s+r≥22s+r\ge 2. Comparison selector (Theorem B.3). A search over an hh-net of subspaces, scored by symmetric pair-smoother comparisons, attains the intrinsic exponent up to a log factor without any eigengap. Its cost is exponential in dr0dr_0. A zero channel does not imply Haar geometry (Proposition 1). For d=2d=2, r=1r=1 and a linear link, the population channel vanishes, but the absolute-leading eigenline of the empirical moment remains biased toward PP. This motivates a matched Haar-projector control. Preregistered study. The study uses ten synthetic tasks (d=16d=16, N=256N=256, 512 contexts each), labeled open, closed or weak before evaluation. Against the matched Haar control, the explicit route's open-minus-closed gain is 0.608 [0.600, 0.617] in geometry and 0.277 [0.258, 0.297] in risk. On the three closed (odd-link) tasks, a task-specific 2-layer Transformer has lower risk than both the Haar control and the explicit route. On the rank-2 interaction T7 it stays at the constant-predictor level.

Critical strengths

S1. The empirical protocol is rigorous and transparent, and the claims are scoped to the evidence. Labels, groups, clipping level and contrasts were frozen before evaluation. The explicit route is compared with a matched Haar-projector control rather than with the nominal 1−r/d1-r/d or an isotropic smoother. Proposition 1 justifies this theoretically. App. H justifies it empirically: on T5, a random 1-D projection already beats isotropic smoothing, so beating isotropic smoothing is not evidence of geometry recovery. Intervals are simultaneous max-t over a context bootstrap. Deviations are disclosed (the learning-rate correction in D.5 and the shared Haar draws), and unflattering results are reported prominently (T7, the T8-driven H3 aggregate, T4's dependence on a single seed). This matters for my assessment because the empirical statements can be trusted. The matched-control lesson is also useful for mechanistic ICL studies beyond this paper.

S2. The theory is clean and explicit, and the null-channel result is subtle and correct. Theorem 2 replaces the concurrent work's assumed clipped gap and log⁡N\sqrt{\log N} clipping level with a deterministic τ∗\tau_*, a derived gap and explicit constants. Proposition 1 makes a point that is easy to get wrong: a zero population moment removes identification but not task-aligned fluctuation, and the bias comes specifically from absolute-value eigen-selection (the algebraically largest eigenline is exactly Haar). In Theorem B.3, the symmetric covariance ΣP,Q\Sigma_{P,Q} keeps likelihood-ratio second moments at order h−rh^{-r} rather than h−dh^{-d}, which is a neat device. The parts examined in detail appear correct: the limiting covariance and sign-regression argument of Proposition 1, the clipping bound behind Proposition B.1, the risk decomposition of Theorem B.2, and the principal-angle identity (B.40). My concerns are therefore about significance rather than correctness.

Critical weaknesses

W1: Limited novelty and significance beyond classical SDR and the concurrent submission. The channel prediction is classical, and H1 is not a risky test. "Open" means that the population target of a pHd-type moment identifies PP (Li, 1992; Cook, 1998). Every closed task in the suite is a parity case, and the same split had already appeared in the authors' earlier benchmark. The non-parity closed link ∣z∣e−z2/2|z|e^{-z^2/2} (Section 4) is the one case where the channel notion says more than parity, and it is not tested. Channel status describes one weak estimator, not the task or trained attention. Theorem B.3 needs no eigengap. In App. H, RMAVE is near-oracle where the fixed clipped moment fails (NR 0.005–0.018 vs. 0.15–1.73). By point estimate, the trained Transformer is worse than E on every open task and better on every closed one. The theoretical delta is incremental. The frontier, the construction and the intrinsic rate for 2s+r>22s+r>2 are in the concurrent submission. New here are a fixed clipping level, a derived gap, explicit constants, and removal of a log factor at the boundary point 2s+r=22s+r=2. Theorem B.3 follows the direction-aggregation and adaptive-index lineage (Gaïffas & Lecué, 2007; Lepski & Serdyukova, 2014), and it cannot be implemented at the paper's scale: (B.17) allows about 1.9×10151.9\times10^{15} candidates for d=16d=16, r=1r=1, h=1/2h=1/2. The closely related Nishikawa et al. (ICML 2025) is not discussed (see Q2). Why it matters: the paper's value would have to come from new understanding of how attention obtains task geometry. The confirmed predictions are classical, and the new theory refines a concurrent result.

W2. The experiments test neither the new theory nor the interpretation of the trained model. No test of the theory. All results come from one operating point (d=16d=16, N=256N=256, m=192m=192). The implementation uses log⁡m\sqrt{\log m} clipping instead of τ∗\tau_* and α=0.25\alpha=0.25 instead of 0, and several "open" tasks fall outside the theorem (T2/T7 unbounded, T6 heteroscedastic, T5a below 2s+r=22s+r=2). The v2d/(mγ2)v^2d/(m\gamma^2) scaling is never examined, and App. F.1 concedes that this needs several pilot sizes. Fixed versus log⁡\sqrt{\log} clipping is not compared either, although E runs in seconds per task. The risk half of H1 rests mainly on T2 (CHE=1.18C_{HE}=1.18); without T2 the point contrast is about 0.10, with no interval. H2 is expected by construction. S is trained per link on up to 24k contexts with noiseless targets. On the odd-link tasks, E[YX]∥PE[YX]\parallel P is available (SIR succeeds on T1 in App. H). Without a first-moment explicit arm, the paper cannot separate its interpretation from the classical one, as it acknowledges. On T4 the win rests on one of three seeds. The negative T7 result comes from runs early-stopped on a plateau (best checkpoints at 300–1,900 steps), and no attention or projector readout is analyzed. Why it matters: the title claims that response channels "predict explicit recovery and attention localization". The evidence covers one explicit estimator at one operating point, plus a trained-model comparison that lacks the control needed to make it informative.

Questions for authors
  1. Could you supplement Appendix I with a precise side-by-side comparison of Theorem 2 and Theorem B.3 against Anonymous (2026), including assumptions, clipping rules, recovery and prediction bounds, constants and boundary cases?

  2. Such a study would provide empirical evidence on trends consistent with the projector bound, rather than validating tight constants. Testing the boundary prediction rate would additionally require varying N and using an appropriate bandwidth.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.