- Rating
- 1
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
This paper investigates whether token predictions and task-specific behaviors in large language models rely on identifiable low-dimensional subspaces of hidden representations. The authors propose a unified framework based on the first and second moments of gradients with respect to hidden states. At the token level, they extend the J-Lens direction by identifying residual sensitivity subspaces (Residual-F) that capture context-dependent gradient variations beyond the mean direction. At the task level, they introduce Contrastive Task Directions (CTDs), obtained through a generalized eigenvalue formulation that maximizes sensitivity to a target task relative to comparison tasks. Experiments on Qwen and Llama models demonstrate that token-level sensitivity becomes increasingly concentrated and reproducible in later layers. Ablation and counterfactual patching experiments on multihop reasoning tasks provide evidence that these subspaces influence downstream predictions. Task-level experiments on BLiMP, EWoK, PIQA, and SocialIQA further suggest that CTDs can produce selective behavioral impairments while reducing off-target disruption compared with non-contrastive sensitivity subspaces.
Critical strengths
- Causal intervention experiments provide meaningful evidence beyond correlational representation analysis.
A major strength is that the paper goes beyond identifying gradient-sensitive directions and evaluates their functional relevance through activation ablation and counterfactual patching. In particular, the intermediate-concept experiments demonstrate that manipulating selected subspaces can alter downstream predictions in multihop reasoning tasks. These interventions strengthen the evidence that the identified subspaces are behaviorally relevant, rather than merely correlated with token predictions.
- Evaluation across multiple tasks, model families, and held-out settings.
The experimental evaluation spans token-level and task-level behaviors, including syntax, world knowledge, and commonsense reasoning. The use of held-out contexts, concept pairs, and task subdomains provides useful evidence that the identified directions can generalize beyond the specific examples used for their estimation. The inclusion of multiple Qwen and Llama models further improves the breadth of the empirical analysis.
Critical weaknesses
- Limited methodological novelty and insufficient comparison with alternative subspace identification methods.
The main methodological components appear to be relatively straightforward applications of established techniques. Residual-F is constructed by removing the mean gradient direction and performing eigendecomposition on the residual gradient second-moment matrix, while CTDs use a generalized eigenvalue formulation to identify directions with high relative sensitivity to a target task. Although their application to token- and task-level interpretability is interesting, the paper does not sufficiently establish what is fundamentally new about these constructions compared with existing gradient-based sensitivity analysis and contrastive subspace identification approaches.
Moreover, the experimental comparisons primarily involve J-Lens, raw-F, activation PCA, and random subspaces. While these are informative controls, they do not fully establish the advantages of the proposed framework over alternative representation analysis or intervention methods. This limits the strength of the methodological contribution.
- The claimed improvement in task selectivity is not sufficiently robust or consistently demonstrated.
Although CTDs achieve higher average selectivity than raw-F across the evaluated benchmarks, the improvements are not consistently observed across models. For example, Table 1 shows that on BLiMP, CTDs achieve an average selectivity of 10.31 percentage points compared with 9.54 for raw-F, a relatively modest improvement. Furthermore, CTDs outperform raw-F on only two of the seven evaluated models for this benchmark.
More importantly, the reduction in off-target damage may be partially associated with weaker overall interventions. In the BLiMP experiments, CTDs produce smaller average target-task accuracy drops than raw-F (11.1 versus 16.5 percentage points). While reduced off-target interference is desirable, comparing methods only at matched subspace ranks does not establish that CTDs achieve a better trade-off between target-task impairment and off-target disruption. A comparison under matched target-task damage would provide more convincing evidence of their selective advantages.
The token-level counterfactual results also exhibit substantial model dependence. For example, combining J-Lens and Residual-F is most effective for Qwen-7B, whereas standalone Residual-F performs better for Llama-8B. In addition, the counterfactual conversion rates are evaluated on relatively small donor-capable subsets. These observations raise questions about the robustness and generalizability of the reported intervention effects.
Overall, the experiments demonstrate promising selective intervention effects but do not yet provide sufficiently strong evidence that the proposed methods offer a consistently superior or broadly generalizable approach to identifying behaviorally relevant subspaces.
Questions for authors
Q1. Selectivity under matched intervention effects.
Can the authors compare CTDs and raw-F under matched target-task accuracy degradation, rather than only matching subspace rank? For example, reporting target-task damage versus off-target damage curves across intervention strengths would help determine whether CTDs genuinely achieve a more favorable selectivity trade-off. This is particularly important for BLiMP, where CTDs cause smaller target-task degradation on average.
Q2. Robustness and statistical significance across models.
Given that CTDs do not consistently outperform raw-F across models, particularly on BLiMP, can the authors provide statistical uncertainty or significance analyses for the differences between these methods? How sensitive are the conclusions to the selected layers, subspace ranks, and regularization parameter?
Q3. Generalizability of the counterfactual intervention results.
The counterfactual experiments are evaluated on donor-capable subsets, and the most effective intervention differs substantially between Qwen and Llama. Could the authors clarify how sensitive the conversion results are to this subset selection and whether uncertainty estimates account for multiple prompts originating from the same underlying MQuAKE case? This would help assess the strength of the causal evidence and its generalizability.