- 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
-
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.
-
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
-
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.
-
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.
-
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.
-
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?
-
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?
-
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