acceptodds
Withdrawn from ICLR 2027

A Dual Scalar Field on Manifold for Recursive H-Neurons Detection and Mitigation in Large Language Models

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Hallucination-Associated Neurons (H-Neurons) refer to a pro-hallucinatory subset of separable neurons. Current methods identify this subset within the activation space by deploying linear probes to impose external statistical boundaries. Such an external empirical estimator lacks geometric interpretability, fundamentally struggling to yield a unified hallucination mitigation framework, while the exhaustive fitting of probes across successive layers incurs prohibitive cross-layer computational costs. Accordingly, we dually derive a scalar field on the manifold and recursively propagate it through the network layers, with its geometric properties intrinsically dictating H-Neurons detection and the mitigation direction. Specifically, we theoretically prove the equivalence between the tangent vector of the scalar field and that of the geodesic on the support connecting hallucinated states to truthful states. This fundamental equivalence enables both the identification of H-Neurons and the comparison of hallucination degrees: (1) decomposing the tangent vector onto activation axes allows component polarities to identify H-Neurons by revealing their pro-hallucinatory dynamical roles and (2) differentiating along the geodesic yields a non-negative derivative, establishing that the scalar field is monotonically non-decreasing along the geodesic trajectory toward truthful states. Subsequently, guided by tangential equivalence and monotonicity of the scalar field, we formalize hallucination mitigation as a path integral process governed by a first-order ordinary differential equation. This motivates a local trajectory along which hallucinated activations follow the ascent direction of the scalar field, thereby steering them toward regions associated with truthfulness. Finally, to alleviate cross-layer computational overhead, we introduce a manifold pullback mechanism to recursively derive high-quality scalar field candidates with a bounded approximation error, thereby reducing the need for exhaustive layer-wise fitting. Experiments demonstrate that our framework can accurately detect H-Neurons and significantly mitigate hallucinations in large language models.

Primary area:
foundation or frontier models, including LLMs
Keywords:
H-Neurons Detection and Mitigation, Scalar Field, Large Language Models

Reviews

Among:
Shape:
1234this paper, Bottom 25%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.75 over 4 reviews, above 25% of the others.
Reviewer_G9BXOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes a geometric framework for identifying hallucination-associated neurons and mitigating hallucinations in large language models. It models activation distributions as Riemannian manifolds and learns an approximate Kantorovich–Rubinstein potential between truthful and hallucinated activations using a two-layer MLP. The Riemannian gradient is used both to identify neurons through its component signs and to steer activations through discretized gradient flow. Normalized pullback constructs candidate potentials for preceding layers to reduce repeated training. The paper evaluates the method on TruthfulQA and PDTB using several model families, including LLaMA-3-8B and LLaMA models up to 65B, and reports neuron interventions, component ablations, cross-layer analyses, and efficiency measurements.

Critical strengths
  1. A common scalar potential provides an interesting connection between representation analysis and intervention. Using the Fisher geometry to define a steering direction, while reusing the potential across layers, is a coherent research direction. If validated, this could offer a useful way to connect neuron analysis with an actionable intervention and reduce the cost of fitting separate models at every layer.

  2. The paper makes a substantial effort to evaluate its individual components. Beyond aggregate benchmark results, it includes neuron interventions, component ablations, comparisons between independently learned and pulled-back fields, and runtime/memory measurements. These analyses are relevant to the claimed benefits and provide useful starting points for assessing the approach, although the concerns below limit the conclusions that can currently be drawn.

Critical weaknesses
  1. The reliability and comparability of the reported empirical gains are not sufficiently established. Table 4 reports TRUE = 57.1 and True×Info = 59.5 for LLaMA-7B with the proposed method. Under the definitions in E.1, the joint or product score should not exceed truthfulness, and averaging across runs cannot explain this inequality. If TRUE denotes a different metric, that metric needs to be explicitly defined. Table 1 also labels the DeepSeek MC1 increase from 51.0 to 56.2 as 9.3%, whereas the relative increase is approximately 10.2%. The composite-score aggregation in Table 1 and the baseline/units in Table 6 need clarification; differences between a product of means and a mean of products should be explicitly distinguished.

    Baseline provenance is also unclear. The non-Ours numerical rows in Table 1 match Table 1 of SPACE v2 (p. 7), whose experimental section describes a mixture of published results and local experiments. The Base/ACT rows in Table 4 match Table 7 of ACT, which evaluates models with few-shot prompting. ACT uses 40 training questions in its few-shot data setting, whereas this submission states a 20% training split. Although E.3 asserts matched inference settings and averaging over independent runs, it does not identify imported versus rerun results or establish matching splits, prompts, and evaluators. Numerical agreement alone does not establish whether the baselines were rerun, but this provenance and protocol ambiguity, together with the metric inconsistencies, prevents a reliable assessment of the claimed gains.

  2. The geometric results do not establish the claimed neuron-level causal interpretation or the quality of the recursive fields. Even granting Theorem 1 on exact optimal-transport paths, a negative coordinate of the joint direction v = g⁻¹∇f does not establish the effect of independently suppressing that neuron, as required by Definition 3. For example, on ℝ² with a constant positive-definite metric, let g⁻¹ = [[1, −2], [−2, 5]] and f(x) = (x₁ + x₂)/√2. Then v = (−1, 3)/√2 has unit Riemannian norm. For a source point mass at y = (2, 2) and a target point mass at x = y + v, f is an optimal KR potential with exact geodesic saturation. Nevertheless, decreasing x₁ alone decreases f, despite v₁ < 0. Thus the component sign does not justify the claimed isolated pro-hallucinatory role. The full-vector update in Algorithm 1/Eq. (162) also needs to be reconciled with the selected-neuron versus random-neuron comparison in Table 5.

    The recursive guarantee has a separate gap: Eqs. (144)–(147) use Euclidean spectral norms to bound mappings between Fisher-metric spaces without establishing the necessary metric conversion factors. Moreover, Eqs. (155)–(158) concern a dual-objective suboptimality gap; they do not establish that this gap is small or control potential values, gradients, or neuron-sign stability. These issues affect the paper's central claims of causal identification and high-quality recursive fields, and require substantive clarification or revision beyond correcting the reported numbers.

Questions for authors

The following requests concern the validity and interpretation of the submitted results; their answers could materially change my assessment.

  1. Evaluation clarification: Which entries in Tables 1, 3, and 4 were rerun, and which were imported? Please specify the splits, training budgets, prompts, and truthfulness/informativeness evaluators used for the proposed method and the relevant baselines. Please reconcile Table 4's True×Info > TRUE, explain the composite-score aggregation, and provide corrected figures with per-example or per-run evidence and uncertainty estimates.

  2. Causal-interpretation clarification: How does Corollary 1 establish the isolated intervention effect in Definition 3 for a non-diagonal metric, including in the counterexample above? What neuron mask and matched intervention budget were used in Table 5, given that Algorithm 1/Eq. (162) updates the full vector?

  3. Theoretical clarification: Which lower-layer metric is implemented, and what justifies the Riemannian Lipschitz bound in Eq. (147)? Beyond dual-objective suboptimality, what result or independent evidence supports the claimed quality of the pulled-back gradients used for neuron identification and steering?

  4. Inference-protocol clarification: E.5 selects the lowest-scoring 15% of tokens after a response has been generated. How does this selection produce the final response that is evaluated? Please specify whether generation uses multiple passes, how regeneration/token alignment and the KV cache are handled, and whether the reported runtime includes the entire procedure

Reviewer_2MxHOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper studies hallucination-associated neurons (H-Neurons) and inference-time hallucination mitigation in large language models. It learns a scalar field over truthful and hallucinated activations using a regularized Kantorovich–Rubinstein objective and a Fisher-based metric. The resulting field is used to identify H-Neurons through tangent-vector components and to steer activations through multi-step updates. A cross-layer pullback is introduced to reduce the cost of fitting separate fields at multiple layers. The method is evaluated on TruthfulQA and PDTB across several language models.

Critical strengths
  1. The paper addresses an important problem and offers a unified perspective on neuron identification and activation-based hallucination mitigation.
  2. The evaluation covers multiple model families and includes competitive baselines, intervention experiments, ablations, and an analysis of pullback efficiency. The reported benchmark improvements are encouraging.
Critical weaknesses

Weaknesses

  1. The theoretical analysis is not convincingly connected to the practical intervention. Theorem 1 establishes gradient–geodesic alignment for an optimal Kantorovich potential along geodesics associated with transport pairs satisfying duality saturation. The practical method instead learns an approximate field with a soft gradient-norm penalty and follows its gradient through Euler updates (Appendix E.4–E.5). The paper does not establish that these updates approximate the theoretically characterized transport directions at the states encountered during intervention. Although Fig. 2(c) and Appendix F.8/F.10 provide geometric analyses, they do not fully bridge this gap. As a result, the theory does not yet provide a convincing explanation for why the implemented updates should reliably move activations toward truthful states.

  2. The role of H-Neuron identification in the final method remains unclear. Corollary 1 labels neurons with negative tangent-vector components as H-Neurons, but the sign of a local steering direction alone does not establish a neuron's causal role in hallucination. Figs. 3–4 and the random-neuron ablation in Appendix Table 5 provide useful intervention evidence. However, the final mitigation procedure updates the full activation vector of selected tokens, rather than only the identified H-Neurons (Appendix E.5). It is therefore unclear how much of the reported improvement depends on the proposed neuron identification criterion. A matched comparison between editing the identified neurons and editing the full vector would be particularly informative.

  3. The benefit of the Kantorovich formulation over simpler nonlinear steering is not isolated. In practice, the method learns a nonlinear score that distinguishes truthful from hallucinated activations and uses its metric-adjusted gradient for intervention. While the paper compares against several existing methods and ablates important components, these experiments do not show whether the optimal-transport objective itself is necessary. A more direct control would be a similarly sized supervised nonlinear scorer with the same Fisher metric, token selection, intervention layers, and update steps. Without this comparison, it is difficult to separate the proposed geometric formulation from the benefits of nonlinear scoring and iterative steering.

Minor Concerns

  1. The cross-layer mapping in Appendix B depends on the residual-stream state as well as the FFN activation, while the practical Lipschitz constants are estimated from sampled perturbations. The scope of the pullback approximation guarantee should be stated more carefully.
  2. The mechanism analyses are conducted on different subsets of models and largely focus on TruthfulQA. More consistent cross-model statistics and representative examples of corrected and uncorrected responses would help substantiate the interpretation.
  3. Some improvements over strong baselines are small. Reporting variation across runs would make these comparisons easier to assess.
Questions for authors
  1. Is there independent evidence that the learned gradient aligns with an optimal-transport direction, or moves held-out activations closer to the truthful distribution during intervention?
  2. How does H-Neuron-only editing compare with full-vector editing under the same intervention budget?
  3. How does a capacity-matched nonlinear truthfulness scorer perform when paired with the same Fisher-adjusted, multi-step editing procedure?
Reviewer_69xROpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The paper trains a WGAN-style critic f_L (Kantorovich–Rubinstein dual of W1 under a Fisher metric) between truthful and hallucinated activations. It labels neurons with a negative component of g^{-1}∇f_L as H-Neurons and mitigates hallucination by a few normalized gradient-ascent steps on f_L. A pullback f_L∘φ/K is used to obtain fields for earlier layers. Experiments on TruthfulQA and PDTB report gains over ITI, TruthX, SPACE, ACT and others.

Critical strengths

Using a nonlinear Lipschitz critic instead of a linear probe, and steering along its gradient, is a reasonable and simple idea. The empirical gains on two benchmarks are consistent across several 7B models.

Critical weaknesses
  1. The theory is known, and its validation is circular. Theorem 1 (the gradient of the optimal W1 potential has unit norm and points along transport paths) is Proposition 1 of WGAN-GP (Gulrajani et al., 2017), which is not cited. The two corollaries restate it. It holds only for the exact optimal potential on supp(π*). The trained critic is a 2-layer MLP with a soft penalty at sample points, and geodesics under the Fisher metric are never computed. Fig. 8 "verifies" Theorem 1 by the cosine between the gradient and the mitigation trajectory, but that trajectory is generated by following the same gradient, so the high cosine holds by construction.
  2. The claimed unification of detection and mitigation is not reflected in the algorithm. In practice the method is gradient steering with a learned critic, which is close to PPLM (Dathathri et al., 2020) and is not discussed. Algorithm 1 edits the full hidden state and never uses the detected H-Neurons. The pullback theory has no practical effect: the Euler step is normalized and token selection is top-k, so the 1/K factor that Props. 1–2 are about cancels out.
Questions for authors
  1. How are the geodesics in Fig. 1(a) and d_g(x, y) in Fig. 6(b) computed under the Fisher metric?
  2. The rule "v^μ < 0 means H-Neuron" labels about half of all neurons, while H-Neurons are defined as an extremely sparse subset. What threshold is used in practice, and how are H-Neurons used in mitigation?
  3. How is the intervention applied during open-ended generation? App. E.5 selects tokens "given the generated response". Is the response regenerated after editing?
  4. Was the step size α re-tuned for the "w/o Fisher Metric" ablation? Without re-tuning, the Fisher-normalized step and the Euclidean step have very different sizes.
Reviewer_qfStOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper presented a geometric framework for identifying Hallucination-Associated Neurons, i.e., H-Neurons, and mitigating hallucinations in LLMs by learning a Kantorovich-Rubinstein dual scalar field on an activation manifold equipped with the Fisher Information Metric. The authors formulated inference-time activation steering as the 1st-order ordinary differential equation along the scalar field gradient and introduced a normalized manifold pullback operator to propagate scalar fields across adjacent layers without independent retraining. Experiments on TruthfulQA and PDTB across several open-weight LLMs are presented to validate the proposed detection and mitigation framework.

Critical strengths
  1. The paper explored an interesting geometric perspective that connects optimal transport duality and the Fisher Information Metric to internal representation steering in LLMs.

  2. The submission employed a low-rank Monte Carlo score approximation together with the Woodbury matrix identity in Appendix B.1, which provided a clean numerical way to apply the inverse Fisher metric without instantiating a full d-by-d matrix.

Critical weaknesses
  1. The mathematical definition of neuron contribution and multiplicative activation scaling in Appendix B (Definition 3, Equations 16 to 19) was taken directly from Section 6.1.2 and Section 6.2.1 without attribution, while the multiplicative activation scaling formulation directly contradicted the additive ODE intervention that was actually used in Section 3.2.

  2. The experimental section contains the material for the main experiments along with an incorrect description regarding how the evaluation was done. For example, Table 1 (same as the descriptions in Section 5.1 and Appendix E.1) is copied from the SPACE publication (Wang et al., 2025a, arXiv:2506.11088), and even the incorrect swapping of the Targeted and Counterfactual columns in the results for LLaMA2-7B-Chat appears to have been duplicated from that publication as well. Table 4 has been copied from Table 7 of the ACT publication (Wang et al., 2025b, arXiv:2406.00034), and 50-shot prompting results have been relabeled as Base scores, although Appendix E.3 states all baselines have been reevaluated using identical conditions across multiple runs.

  3. Theoretical equivalence between Theorem 1 and our proposed method depends upon two conditions: namely, that the function must be a truly global 1-Lipschitz continuous function, and that the Kantorovich-Rubinstein duality must be saturated as we move along the minimizing geodesics of the Fisher metric. Unfortunately, in Appendix E.4, we do not apply the Kantorovich-Rubinstein duality but instead only apply a weak one-sided penalty to samples taken at discrete points. In addition, Figure 8 demonstrates that the empirical trajectories were simply normalized onto a Euclidean hypersphere and therefore did not follow the geodesics of the Fisher metric.

  4. There is a fundamental disconnect between H-Neuron identification and the mitigation algorithm: Corollary 1 defines H-Neurons locally via state-dependent negative tangent, whereas Algorithm 1 and Equation 162 steer all hidden dimensions densely without masking or selecting a sparse subset of H-Neurons.

  5. The cross-layer pullback in Section 4 and the inference rule in Appendix E.5 are mathematically and causally flawed: Definition 2 treats the cross-layer Transformer update as a single-token map despite multi-token attention dependencies, Proposition 2 bounds the Lipschitz constant with Euclidean spectral norms that contradict the isometric pullback metric of Lemma 3, and selecting the lowest 15% of tokens via TopK over the full response of length T cannot be done causally during single-pass autoregressive generation.

Questions for authors
  1. What is the reason behind the fact that Definition 3 reiterates the CETT norm-ratio and the multiplicative scaling model from Section 6.1.2 and the multiplicative scaling model from Section 6.2.1 as found in Gao et al. (2025) sans acknowledgement, and what is the relationship between the additive ODE structure from Equation (162) you have proposed and Definition 3?

  2. Would you explain how the numerical values in Table 1 of your document are identical to those found in Table 1 of SPACE (Wang et al., 2025a), and how the numerical values in Table 4 of your document are affected?

  3. How were the reference Fisher geodesics and geodesic distances in Figures 1(a), 2(c), and 6(b) computed in high-dimensional activation space, and how many backward passes through the language model head are required per token to estimate the online Fisher matrix during Euler integration?

  4. How do you aggregate the state-dependent tangent vector components across tokens and inputs to obtain a fixed global subset of H-Neurons in Figures 3(a) and 6(a), and did the benchmark runs in Tables 1, 3, and 4 intervene only on sparse H-Neurons or on the full dense hidden vector?

  5. How can the token selection rule in Appendix E.5 be executed during real-time autoregressive decoding before all T tokens have been generated?

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

Related papers

Discussion (0)

Sign in to comment.