acceptodds
Under review as a conference paper at ICLR 2027

Identifying Token and Task-Relevant Representational Subspaces in Large Language Models

Abstract

Understanding how language models predict particular tokens and perform specific tasks requires identifying representational subspaces that selectively support these behaviors across contexts. We develop a unified framework based on gradients of token scores and behavioral losses with respect to hidden states. Averaging gradient outer products yields a sensitivity matrix whose leading eigenvectors capture directions of greatest local sensitivity. At the token level, we separate the context-averaged gradient direction from a residual subspace that captures context-dependent sensitivity orthogonal to it. We extend this framework to task selectivity by contrasting sensitivity matrices for target and comparison tasks, yielding Contrastive Task Directions. We find that token-relevant residual subspaces in Qwen and Llama generalize across independent contexts, with sensitivity increasingly concentrated in later layers. On held-out multihop reasoning prompts, ablating subspaces associated with an intermediate concept selectively reduces the probability of the correct downstream answer, while counterfactual interventions can redirect generation toward the answer implied by a substituted concept. Extending this analysis to task-level behavior, we find that ablating contrastive subspaces across seven Qwen and Llama models produces selective effects on syntactic, world-knowledge, and commonsense-reasoning benchmarks, with less off-target disruption on average than non-contrastive subspaces. Together, these results connect local gradient geometry to selective causal effects at both token and task scales.

Primary area:
foundation or frontier models, including LLMs
Keywords:
language models, representations, multihop reasoning, functional selectivity, mechanistic interpretability, causal interventions

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_SRGTOpenReview ↗
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
  1. 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.

  1. 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
  1. 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.

  1. 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.

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

This paper extends J-Lens by considering the second moment of token-score gradients, introducing residual-F subspaces to capture context-dependent sensitivity beyond the mean gradient direction. It further proposes Contrastive Task Directions (CTDs) to identify task-selective subspaces.

Critical strengths
  1. The method is simple, well-motivated, and provides a unified framework for token- and task-level analysis.
  2. The experiments cover multiple model families, with held-out evaluations and causal interventions.
Critical weaknesses
  1. While Figure 3 demonstrates counterfactual patching, the interventions rely on query-matched donor activations. Could the authors evaluate test-time donor-free steering, where a fixed steering vector is estimated from independent calibration data and applied to unseen prompts without constructing example-specific donors? Comparing J-Lens, residual-F, and J+F under this setting would help establish whether the identified subspaces support reusable steering beyond filtering example-specific activation differences.
  2. Although CTDs improve average selectivity, the gains over raw-F are modest or inconsistent for some models, particularly on BLiMP. It would be helpful to clarify when and why the contrastive formulation offers an advantage.
Questions for authors

See Critical Weaknesses.

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

The paper uses gradient second moments to identify token-residual and task-contrastive subspaces. Held-out analyses, multihop ablations and counterfactual patches support selective token effects in Qwen and Llama. Task ablations across seven common models reduce off-target disruption on average. The framework is useful, but its explanatory contribution and the robustness of its comparative advantages remain my main concerns.

Critical strengths
  1. Evidence beyond correlation. Independent contexts, held-out task examples, wrong-concept and donor-mismatch controls, and rank- and magnitude-matched patches support selective intervention effects (§§3–4). Residual subspaces contribute information beyond the mean token direction.
Critical weaknesses
  1. The explanatory advance is modest. Gradient second moments have established mathematical precedent (Constantine et al.), although the residual and contrastive applications here are distinct. The token and task studies share a useful tool, but do not establish how these subspaces interact or are constructed and read out (§5). In my view, the contribution is stronger as an intervention method than as an explanation of representational organization.
  2. Comparative gains need better qualification. Matched controls and rank sweeps strengthen the evidence, but the selection of layer bands, main-comparison ranks and patch strength is unexplained (§3.3). Figures 2–3 also omit the bootstrap sampling unit despite multiple prompts per case. These gaps make the reported advantages harder to assess; they do not establish biased selection or invalid intervals.
Questions for authors
  1. How were layer bands, emphasized ranks and patch strength selected, and were these choices independent of final evaluation outcomes?
  2. Did the bootstrap in Figures 2–3 resample underlying cases while preserving associated prompts and paired method comparisons?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.