acceptodds
Under review as a conference paper at ICLR 2027

Geometry of Large Language Model Controllability

Abstract

Prompt engineering remains a brittle, heuristic practice governed by a fundamental *Controllability Dilemma*: extreme vulnerability to minor lexical variations (*Paraphrase Sensitivity*) alongside a persistent propensity to ignore explicit structural instructions (*Instruction Neglect*). In this paper, we reveal that this dilemma originates from a severe topological mismatch between the superficial prompt text space and the model's true operational Hilbert output space, invalidating the implicit isometric assumption of current prompt optimizers. To resolve this, we formalize prompting as the dynamic selection of a function within a continuous Hilbert space, equipping the prompt manifold with a pullback Riemannian metric tensor induced by functional output divergence over task distributions. Through this geometric lens, we mathematically prove that multi-head attention saturation induces a topological partition of the prompt manifold, driving Instruction Neglect via rank-deficient -null foliations and Paraphrase Sensitivity via unbounded transversal sectional curvature across deep Transformer layers. Operationalizing these theoretical insights, we introduce **Pullback-Guided Discrete Exploration (PGDE)**, a gradient-free algorithm that leverages empirical pullback distances as a zero-order geometric filter to dynamically enforce distant functional jumps and local variance bounds. Empirical evaluations across 7B to 70B parameter models demonstrate that PGDE significantly outperforms state-of-the-art baselines on strict instruction following (IFEval, +8.6% absolute accuracy gain on Llama-3-8B) and mathematical reasoning (GSM8K), while reducing search wall-clock latency by over 50% by early-rejecting functionally inert proposals and suppressing output variance under paraphrasing by 82%.

Primary area:
foundation or frontier models, including LLMs
Keywords:
large language model, Riemannian geometry, prompt engineering, controllability

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_yHtJOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

Paper Summary

The paper attributes two prompt-engineering failure modes, "Instruction Neglect" and "Paraphrase Sensitivity," to a mismatch between distances in prompt-text space and distances in the model's output ("functional") space.

Theory.

  • It models a prompt as a point zz in continuous prompt-embedding space and defines Φ(z)(x)=Ey[ψ(y)]\Phi(z)(x) = \mathbb E_{y}[\psi(y)] as a map into H=L2(Ω,μ;E)\mathcal{H} = L^2(\Omega, \mu; E). The prompt space is equipped with the pullback metric gμg^\mu (Eq. 2).
  • It states three theorems:
    • attention saturation partitions the prompt manifold into ZnullZ_{\text{null}} and ZregZ_{\text{reg}} (Thm. 1);
    • ϵ\epsilon-null foliations cause Instruction Neglect (Thm. 2);
    • unbounded transversal sectional curvature causes Paraphrase Sensitivity (Thm. 3).

Algorithm. PGDE computes an "empirical pullback distance" dpbd_{pb}: the RMS output distance, over N=100N = 100 calibration inputs, between greedy outputs of the base prompt and of a candidate prompt. It discards candidates with dpbd_{pb} below a percentile threshold and picks the best remaining candidate by task loss.

Validation. On Mistral-7B and Llama-3-8B/70B, the paper reports:

  • gains over APE/OPRO/EvoPrompt on IFEval and GSM8K (e.g., "+8.6%" on Llama-3-8B IFEval);
  • 80–90% lower output variance under paraphrasing (Table 3);
  • roughly 50% lower search wall-clock time.
Critical strengths

None. Measuring prompt candidates by output-space divergence rather than text similarity is a sensible practical heuristic. However, it is not substantiated as the source of the gains (see Weakness 2), and it does not by itself support acceptance.

Critical weaknesses

Critical Weaknesses

  1. The central theory does not hold, and the algorithm does not use it. Theorem 1's "partition" is the trichotomy of real numbers applied to σmin⁡(Jz)\sigma_{\min}(J_z) and a threshold. With the paper's own setup (k=768k = 768, N=100N = 100), gμg^\mu has rank at most 76,800, so for prompts longer than about 18 tokens σmin⁡=0\sigma_{\min} = 0 identically and Theorem 3 (stated on ZregZ_{\text{reg}}) is vacuous. Theorem 3 also contradicts Definition 5, which requires dΦz(U)=0d\Phi_z(U) = 0 where dΦzd\Phi_z is injective. PGDE itself uses greedy decoding and none of the theoretical quantities; dpbd_{pb} is simply an output-divergence heuristic.

  2. The empirical claims are unreliable. The headline "+8.6" is computed against EvoPrompt rather than the strongest baseline in Table 2 (Random Proposal Search), against which the gain is +5.6. The claimed compute savings conflict with the algorithm, since dpbd_{pb} already requires generating all calibration outputs. Llama-3-70B is reportedly served in FP16 on 4× L4 GPUs (96 GB), which cannot hold its ~141 GB of weights. The key ablation, PGDE's proposal generator without the filter, is missing.

Questions for authors

All questions are clarifications that could change my assessment.

  1. Are Figs. 12 and 13 plotted from measured data? App. F.3 says Fig. 12b "includes realistic transition noise points … [t]o reflect authentic experimental behavior."
  2. How was Llama-3-70B served on 4× L4 (96 GB) in FP16, and where do the IFEval and GSM8K "validation" calibration sets come from?
  3. What is the performance of PGDE's proposal generator without the dpbd_{pb} filter, under the same budget?
Reviewer_QKpFOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This manuscript explores the issues of Instruction Neglect and Paraphrase Sensitivity in prompt engineering for LLMs through the framework of differential geometry. The authors attribute these phenomena to metric degeneracy and unbounded curvature on the prompt manifold. Based on this analysis, they propose a method called Pullback-Guided Discrete Exploration (PGDE), which operates without gradients and selects candidate prompts by evaluating the distance between their resulting outputs. Experimental results across three different LLMs demonstrate that this approach outperforms existing baselines in terms of instruction following, mathematical reasoning, output stability, and search efficiency.

Critical strengths
  1. Practical significance of the research problem. This manuscript addresses instruction neglect and paraphrase sensitivity, two important challenges in LLM prompt optimization. Guiding prompt search through model output differences rather than textual similarity offers a meaningful perspective.
  2. Consistent empirical improvements across tasks and models. PGDE outperforms the evaluated baselines on IFEval and GSM8K across three LLMs, while also demonstrating improvements in output stability and search efficiency. Additional sensitivity analyses further support the empirical findings.
Critical weaknesses
  1. Theoretical claims require further justification. The proofs of the main theorems appear to contain several gaps. In particular, Theorem 2 does not fully establish the conditions required for Frobenius integrability. Furthermore, the unbounded functional output distances derived in Eq. (162) of Theorem 3 appear inconsistent with the boundedness assumption in Appendix B.1. These issues raise concerns about the validity of the theoretical claims.

  2. Insufficient connection between the geometric theory and the proposed algorithm. The theoretical framework defines a pullback Riemannian metric through derivatives of expected model outputs, whereas PGDE relies on empirical distances between discrete outputs. The relationship between these quantities remains unclear, particularly whether the proposed filtering mechanism actually captures metric degeneracy and curvature. This weakens the theoretical justification for PGDE.

Questions for authors
  1. Insufficient justification of the proposed geometric mechanisms. Theorem 1 attributes the partition of the prompt manifold to attention saturation, but the partition follows directly from thresholding Jacobian singular values. Although Figure 4 in Appendix D.2 shows attention saturation and singular value decay, it does not directly link these observations to instruction-following failures through comparisons between successful and failed cases. For Paraphrase Sensitivity, the reported reduction in output variance may simply result from filtering behaviorally similar prompts, rather than demonstrating improved robustness to unseen paraphrases. Similarly, Theorem 2 links instruction neglect to metric degeneracy without sufficiently establishing the claimed geometric structure. The theoretical basis for these connections requires stronger justification.

  2. Unclear connection between the theoretical framework and PGDE. The theoretical pullback metric is defined through derivatives of expected model outputs, whereas PGDE relies on empirical distances between discrete outputs. The paper does not establish how these distances approximate the theoretical metric, leaving the connection between the proposed geometry and the algorithm insufficiently supported.

  3. Limited generalization to contemporary LLMs. The evaluation relies exclusively on models released in 2023–2024. Given the rapid evolution of LLMs, experiments on more recent instruction-tuned and reasoning-oriented models are needed to substantiate the broader claims about LLM controllability.

  4. Mathematical inconsistencies in the appendix. Appendix B.1 establishes bounded expected outputs, whereas Eq. (162) derives unbounded functional output distances. Additionally, Eq. (160) appears to equate the norm of an integral with the integral of the norm, which is generally invalid. These inconsistencies require clarification to establish the validity of Theorem 3.

  5. Unjustified functional distance filtering. PGDE retains candidates with large functional distances, but behavioral divergence does not necessarily imply performance improvement. This may discard promising prompts that selectively correct errors while preserving correct predictions, yet retain prompts that introduce more errors. The authors should further justify the rationale for this filtering strategy and present comparative experiments with alternative filtering approaches.

1 comment

Since I have not found the authors' responses to my concerns, I will keep my rating.

Reviewer_iGH7OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper identifies two common issues in prompt engineering: models are overly sensitive to minor changes in wording, and they often ignore explicit structural instructions. The authors formalize prompts as manifolds in a continuous space, define a pullback metric based on differences in model outputs, and geometrically explain the relationship between attention saturation and these problems. Based on this theory, they propose a gradient-free search method called Pullback-Guided Discrete Exploration (PGDE), which filters candidate prompts using empirical pullback distances. Experiments cover tasks such as instruction following and mathematical reasoning, comparing PGDE with baseline prompt optimization methods across models of various scales.

Critical strengths
  1. Using the same geometric framework to uniformly explain the two types of prompt questions is more organized than scattered techniques, and this framework is verifiable. The paper shows that attention saturation splits the prompt space into two parts, and explains one of the two problems in each part. If the two problems really come from the same mechanism, one fix could handle both, instead of fixing them one by one.

  2. The experiments cover several models and several tasks. Appendix E also has ablations on calibration size, metric weight, thresholds and encoder, and Appendix C has the implementation details. All methods use the same search budget, so it wil not feeling the gain only comes from more compute. The ablations also show the gain is not from one lucky setting.

Critical weaknesses
  1. In the summarization experiment (Table 3), the paper only reports the output variance and the worst-case drop. It does not report the average summary quality. A small variance may also only mean the output is stable but plain. So it is difficult to determine whether the stability comes at the cost of quality.

  2. The theoretical distance is defined in a continuous space, but the algorithm actually uses the distance between several discrete texts without explaining how the two correspond. In this way, it is difficult to determine whether the good experimental effect is due to geometric theory or screening techniques.

Questions for authors
  1. Could the most crucial step in the proof of Theorem 3 (from the unbounded curvature to the inevitable divergence of the result) be written out in a clearer and more easily verifiable way?

  2. Could you provide the error range between the theoretical distance and the actual distance used in the algorithm?

  3. In Appendix E of the paper, the influences of calibration set size, mixed weights, thresholds, and encoders were measured respectively. However, each factor was adjusted individually. When these factors change together, will the results still be stable?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.