acceptodds
Under review as a conference paper at ICLR 2027

Hyperspherical Steering with Optimal Transport Weighted Dual Prototypes for LLMs

Abstract

Training-free hyperspherical activation steering enables lightweight inference correction for large language models (LLMs) via norm-preserving rotation, yet its geometric framework has inherent limitations: uniformly aggregated centroids are vulnerable to boundary noise; the single-difference prototype implies an antipodal constraint that causes systematic bias in rotation direction; global parameters fail to balance the demands of discriminative and generative tasks. This paper proposes an optimal transport-weighted dual-prototype hyperspherical steering method. In the offline pre-computation stage, we first construct prototypes via hyperspherical optimal transport weighting to suppress boundary noise and enhance inter-class geodesic separation. We then build an independent dual-prototype paradigm to lift the enforced antipodal constraint, using truthful prototype as correction anchor to eliminate rotation direction bias at its geometric root. During online inference, a task-aware von Mises-Fisher gating mechanism is adopted to differentially regulate intervention strength. Experiments on LLaMA-3.1-8B and Qwen2.5-7B across six benchmark datasets show that the proposed method introduces negligible extra computational overhead, and achieves average improvements of 5.33% and 10.14% over SOTA baselines on multiple-choice and open-ended generation tasks respectively. Rank collapse analysis verifies that our method induces milder representation collapse and preserves the LLM’s native representation space more completely. This work improves the theoretical framework of hyperspherical steering from its geometric fundamentals, providing a more self-consistent technical path for training-free lightweight hallucination correction.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Large language models, Hyperspherical Optimal Transport, Steering

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_MWV6OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper extends spherical activation steering with truthful and hallucinatory prototypes obtained from an entropy-regularized optimal-transport coupling. Task-specific gating controls the intervention, which is evaluated on two language models using TruthfulQA and five additional multiple-choice benchmarks.

Critical strengths

The improvements over the task-tuned single-prototype control are worth taking seriously. Compared with Variant 3, the full method raises LLaMA's generation TRUE×INFO\mathrm{TRUE}\times\mathrm{INFO} from 60.28 to 73.63 percent and Qwen's multiple-choice average from 56.50 to 60.62 percent. The gain on LLaMA multiple-choice is smaller, from 61.42 to 62.74 percent (Tables 2 and 6). Tables 3 and 7 also show improvements over spherical steering across the six benchmarks.

The method is inexpensive to use: two stored directions, less than 0.14 MB of prototype storage, and no backbone updates. The rotation preserves activation norms. Table 4 puts LLaMA generation latency at about 1.29 seconds per prompt for both methods, with little added cost over spherical steering.

Critical weaknesses
  1. Theorem 2 is too strong as stated. Take unit vectors on a circle at positive-class angles (−8∘,5∘,6∘)(-8^\circ,5^\circ,6^\circ) and negative-class angles (−1∘,0∘,1∘)(-1^\circ,0^\circ,1^\circ). With the intended pairwise geodesic cost, uniform marginals, and ε=0.01\varepsilon=0.01, applying Equations (5) to (8) reduces prototype separation from about 1.01∘1.01^\circ to 0.09∘0.09^\circ. This small construction should be straightforward to check directly. The proof in Appendix C.4 assumes centroid-norm growth that the weighting does not guarantee. A conditional theorem would need explicit, justified assumptions; otherwise the separation claim should be treated as an empirical hypothesis.

  2. I cannot reconcile one ablation with the stated gate. Table 5 gives Qwen Variant 6 generation α=0\alpha=0 and β=−0.3\beta=-0.3. Algorithm 2 uses t=clip⁡(ατ−β1−β,0,1)t=\operatorname{clip}\left(\alpha\frac{\tau-\beta}{1-\beta},0,1\right) when τ>β\tau > \beta, and zero otherwise, so this setting leaves activations unchanged. Yet Table 6 gives TRUE×INFO\mathrm{TRUE}\times\mathrm{INFO} of 77.36 percent, compared with 74.40 percent for the unmodified model. Appendix A also leaves the common tuning grid and selection objective unspecified.

Questions for authors

Is Theorem 2 meant to assume additional conditions on the activation distributions? If so, what rules out the example above, and what evidence supports those conditions in the studied models?

Reviewer_KtoKOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper addresses the geometric limitations of existing training-free activation steering methods for Large Language Models (LLMs). The authors argue that prior geometry-aware rotation frameworks (such as vanilla hyperspherical steering) suffer from three critical flaws: (1) vulnerability to boundary noise due to uniform centroid aggregation; (2) an alleged "antipodal constraint" caused by single-difference prototypes; (3) a lack of task-specific adaptation in global gating parameters. To tackle these issues, the paper proposes a framework comprising: (1) using Sinkhorn iterations on the unit hypersphere to compute robust weighted centroids that suppress ambiguous boundary samples; (2) establishing independent positive and negative prototypes to replace single difference vectors; (3) introducing task-specific preset hyperparameters (α,β)(\alpha, \beta) to adjust intervention strength across multiple-choice discrimination and open-ended generation tasks.

Critical strengths
  1. Evaluation across six diverse benchmarks (TruthfulQA, COPA, StoryCloze, MMLU, WinoGrande, BoolQ) and two open-source LLMs provides empirical evidence.

  2. The inclusion of representation rank collapse analysis offers explanations on why hyperspherical rotation preserves generation fluency and informativeness better than linear additive steering methods (ITI, CAA).

Critical weaknesses
  1. The authors claim in the main contributions that they “adopt a task-aware gating mechanism that tailors rotation magnitudes and intervention thresholds to different tasks”. However, Table 5 reveals that α\alpha and β\beta are manually tuned grid-search hyperparameters specific to each task category. This is not an automated or algorithmic "task-aware" adaptation module. It is manual hyperparameter tuning. If deployed on a novel and user-defined open-ended task where ground-truth validation sets are unavailable, a practitioner has no principled way to set α\alpha and β\beta. This significantly limits practical deployment.

  2. In Algorithm 1, line 4, the index for the negative sample incorrectly uses ii (⟨zi+(l),zi−(l)⟩\langle z_i^{+(l)}, z_i^{-(l)}\rangle) instead of cross-pairing ii and jj (⟨zi+(l),zj−(l)⟩\langle z_i^{+(l)}, z_j^{-(l)}\rangle).

  3. The paper introduces Sinkhorn optimal transport to compute sample weights (wi+,wi−w_i^+, w_i^-) for the weighted dual prototypes. However, it lacks an ablation comparing Optimal Transport (OT) against computationally cheaper and intuitive robust weighting alternatives (e.g., distance-to-centroid exponential weighting, trimming top kk\\% outlier samples, or density-based weighting).

  4. All evaluations are restricted to 7B/8B parameter models (LLaMA-3.1-8B-Instruct and Qwen-2.5-7B-Instruct). However, testing whether the method maintains its gains, incurs representation collapse, or scales effectively on larger models (such as LLaMA-3.1-70B or larger instruction-tuned checkpoints) is essential to prove generalizability.

Questions for authors

see Weaknesses

Reviewer_kBvBOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes a training-free hyperspherical steering method that reduces hallucinations in large language models by rotating hidden activations toward a truthful direction while preserving their norm. Building on spherical steering, it computes class centroids with optimal transport weighting and uses the truthful prototype itself as the rotation target, aiming to improve steering accuracy and reduce directional bias. A task-aware confidence gating mechanism further adjusts the intervention strength for multiple-choice and open-ended generation, and two theorems are provided to support this design. Experiments on LLaMA-3.1-8B and Qwen-2.5-7B show improved factual accuracy and generation quality over spherical steering, with minimal computational overhead and milder representation collapse.

Critical strengths
  1. The proposed method yields consistent improvements with negligible computational overhead. It achieves the highest average performance among all compared methods on TruthfulQA and on the five additional multiple-choice benchmarks. Relative to spherical steering, the additional inference latency is limited to 2.15% for multiple-choice and 0.56% for open-ended generation.

  2. The ablation is well organized. Table 1 defines variants for the combinations of the three components, and Tables 2 and 6 report all of 6 variants, so the effect of each component can be checked directly.

  3. Figure 1 is clear and intuitive. It shows how the prototypes are built with OT on the hypersphere, and how activations are rotated toward the truthful prototype at inference.

  4. The rank-collapse analysis in Fig. 2 links performance to representation degradation.

  5. The paper is clearly written and easy to follow

Critical weaknesses

The source of the gains is unclear, and the evidence does not support the paper's main claims. The paper makes two main claims: (i) uniform centroids are sensitive to boundary noise, so OT-weighted centroids should be used and (ii) the single difference-based prototype imposes an antipodal structure and is therefore a poor rotation target, so the truthful prototype should serve as the anchor.

(1) OT-weighted centroids. The paper does not explain why this particular weighting is needed or how it relates to performance. In Eq. 6, the weight of a sample grows with its transport cost, so samples far from the other class, including outliers, receive larger weights. The main effect may therefore be to push m+m^+ and m−m^- apart. Theorem 2 shows only this separation. It does not show that the weighted m+m^+ is a better estimate of the truthful distribution. Since OT weighting is compared only with uniform averaging, it is not possible to tell whether it is better than other centroid estimators, let alone optimal.

The ablation also suggests a different explanation. Variant 3 keeps uniform centroids and the single prototype and only adds task-specific (α,β)(\alpha, \beta). Yet it raises the LLaMA TruthfulQA MC average from 53.17 to 61.42, while the full method reaches 62.74. OT weighting and dual prototypes together therefore add only 1.32 points on LLaMA. At least on LLaMA, most of the improvement comes from tuning (α,β)(\alpha, \beta) per task, not from the geometric components that the paper emphasizes.

(2) Dual prototypes. The paper attributes a directional bias to the antipodal assumption of the single difference-based prototype, but provides no empirical analysis supporting this claim beyond end-task scores. For example, it does not report how far the two prototypes are from antipodal in practice, how much the difference-based target deviates from the truthful prototype, or whether this deviation actually degrades steering.

(3) Task-aware gating. The task-aware component distinguishes only two task types, multiple-choice and open-ended generation, and the latter is evaluated only on TruthfulQA. With so few task types, the per-task (α,β)(\alpha, \beta) values in Table 5, obtained by grid search, appear to reflect a largely heuristic choice rather than a general mechanism. It would also help to clarify how the task discrimination module identifies the task type, as it is currently mentioned only briefly.

Questions for authors
  1. Could the authors clarify how different the uniform and OT-weighted prototypes are in practice? Theorem 2 only provides an inequality. It would therefore be helpful to see the angle between the two prototypes and the distribution of the weights wiw_i.

  2. The task-aware gating is defined for only two task types. Could the authors show whether the method can be extended to task types beyond multiple-choice and open-ended generation?

  3. Could the authors clarify how the rank drop in Fig. 2 is defined?

  4. Theorem 1 shows that the difference-based target is fully aligned with the truthful prototype only when the two prototypes are antipodal. It would be helpful to understand how the truthful and hallucinatory activations are actually distributed on the hypersphere at the intervention layer. Could the authors report the angle between the two prototypes and the degree of overlap between the two clusters near the boundary?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.