acceptodds
Under review as a conference paper at ICLR 2027

Logical subspace in LLMs

Abstract

Recent work has identified a human brain network specialized for abstract formal reasoning (Kean et al., 2025). Does the same hold true in language models? To answer this question, we introduce the minimal viable subspace (MVS) method, which searches for the lowest-rank activation subspace at a layer that preserves task performance when retained alone. Using MVS, we demonstrate low-rank subspaces supporting logical inference on Gemma and Qwen models. Furthermore, these subspaces exhibit dissociation from other tasks: retaining late logic subspaces preserves inference while impairing factual knowledge, working memory, cognitive control, and arithmetic; ablating them reduces logical inference accuracy to chance while largely sparing these other capacities. Our results suggest a functionally-localized core machinery for logic akin to that in the human brain.

Primary area:
applications to neuroscience & cognitive science
Keywords:
interpretability, functional localization, reasoning, minimal viable subspaces, human-AI parallels, pruning

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_1ouJOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper introduces Minimal Viable Subspaces (MVS), a method for identifying low-dimensional activation subspaces that are sufficient to preserve performance on a target “logical inference” task.

Using complementary retention and ablation interventions on Qwen3-8B and Gemma-4-12B, the authors identify subspaces associated with logical inference that exhibit functional dissociations from several other model capabilities.

Their contribution in this reviewer’s opinion appears two-fold:

  1. They propose a method to extract task relevant subspaces using an unsupervised learning approach, where the subspace is learned by minimizing the KL loss between the output distributions of original and intervened models.
  2. They apply their method to logical inference, design a synthetic dataset to evaluate this capability, and demonstrate the causal relevance of the identified subspace to the target behavior through retention and ablation interventions.
Critical strengths
  1. They sufficiently and clearly motivate the problem and discuss the parallels with human cognition that is understandable to readers in both cognitive science and computer science. Also the positioning relative to previous studies seems well-justified.

  2. The paper is generally well-written and understandable with clear explanations and coherent structure. The formulations are clear and the figures are intuitive, clean and understandable and help the paper comprehension tremendously.

  3. Retention and ablation experiments provide stronger evidence for functional localization than correlational probing alone (Section 5.2, Figure 4). The effects generalize and their diverse control tasks support functional dissociation (Figures 2 and 4) supporting their narrative. The paper experiments across two model families, which shows the generalizability of the mechanism beyond a single model.

Critical weaknesses
  1. Novelty seems moderate; their causal subspace approach is not particularly novel. Their approach’s novelty compared to DAS (Geiger et al) and Yuan & Su is not sufficiently established.
  2. The method’s presentation (Section 3) seems somewhat rushed and would benefit from more detailed preliminaries and explanations. In particular, the authors could better explain the intuition behind the projection operation, and how the learned subspace is optimized and the minimum sufficient rank is selected.
  3. While the paper mentions their advantage of DAS is in the “unsupervised”ness of their method, the paper lacks direct comparison with the alternatives (especially DAS) in terms of performance. If we ignore the lack of labels problem, does DAS outperform MVS otherwise? This was never sufficiently discussed.
  4. Despite using multiple seeds during optimization, the paper lacks uncertainty estimates for its main accuracy results and causal intervention effects, making it difficult to assess their statistical robustness.
  5. Figure 6 is not explicitly referenced or adequately discussed in the main text. They mention in Figures caption that “the activation for Qwen3-8B are linearly separable” but this needs more elaboration, stating what it signifies in the context of the paper.
  6. Limited validation of logical inference; The primary evaluation relies on a single synthetic benchmark involving one-step, binary conditional inference, limiting evidence that the identified subspace captures logical reasoning more generally. The held-out tests vary vocabulary and logical formulations but do not sufficiently extend reasoning complexity (Section 4.1). This ties closely to their mixed results on PrOntoQA. Notably, both retaining and ablating the identified subspace impair PrOntoQA performance (Figure 4), yet this important observation is insufficiently discussed. As a result, the findings may establish localization of task-specific representations rather than a general logical inference capability. As a final point, it appears that the extracted subspace corresponds specifically to their synthetic dataset of simple, single-step inferences. Extending their experiments beyond this limited setting would significantly strengthen their claims about isolating logical inference capabilities. As of now, this seems like an overclaim. The performance drop on PrOntoQA when ablating (S) (Figure 4) potentially suggests some support for general logical inference behavior in the subspace, but this needs to be investigated further.
Questions for authors
  1. Beyond its unsupervised nature, what distinguishes MVS from DAS and Yuan & Su (2026)? Could the authors provide direct experimental comparisons?

  2. Could the authors clarify the intuition behind the projection operation, the choice of neutral-corpus mean, and how the minimum sufficient rank is determined?

  3. How consistent are the main results and intervention effects across optimization seeds? Could the authors report uncertainty estimates?

  4. Can the authors evaluate MVS on more diverse logical reasoning benchmarks, particularly those involving multi-step inference and nonbinary outputs?

  5. Why do both retention and ablation impair PrOntoQA performance (Figure 4), and what does this imply about the generality of the identified logic subspace?

Reviewer_UuHBOpenReview ↗
Rating
3
Confidence
4
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes Minimal Viable Subspace(MVS), a unsupervised method:(1) the method searches for the lowest-rank activation subspace that preserves a model's performance on a target task when all orthogonal activity is removed; (2) Identify MVSs across layers by one-step conditional inferences expressed using both ordinary and meaningless vocabulary; (3) Test whether subspace is selectively necessary for inference through complementary intervention (retain subspace & ablate it). Interventions on language, working memory, factual knowledge, arithmetic, and a separate rule-based reasoning benchmark show that retaining subspace keeps near-ceiling logical performance while impairing tasks while ablating it reduces logic to near chance.

Critical strengths
  1. Rather than localizing task-attributed units or a domain-general workspace, this paper explore whether a capability can be isolated within a lowest-dimensional activation subspace. The MVS method is unsupervised without requiring gold labels or prespecified causal variables.

  2. The mean activation μl calculated on a neutral corpus disjoint from evaluated tasks, reflecting that it does not carry any task-specific information.

  3. The idea is interesting.

Critical weaknesses
  1. The main logic task is one-step conditional syllogism. However, LLM logical reasoning also includes multi-step deduction, quantifier reasoning, set reasoning, abduction and induction.
  2. PrOntoQA as a multi-step deduction benchmark is selected to the control task. From Figure 4, in the purple PrOntoQA bar, although the Intact accuracy is the highest, the Retain S accuracy does not seem to outweigh the Ablate S accuracy. The authors acknowledge that it is "affected by both interventions" and will "return to this in the Discussion", but there without any deeper analysis about it. This may suggests that the "logic subspace" label is too strong.
  3. Working memory, MSIT and Multiplication are all at ceiling (1.000) when Intact, and they still keep ceiling when Ablate S, but whether this effect is due to the task being too simple is unknown. And for Qwen column is not interpretable: Intact accuracy 0.540 vs. chance 0.500. So it may not cautiously support the main conclusion that "control tasks are largely preserved" . And the control tasks may need to be reconsidered.
  4. Some key implementation details are missing, which may affect reproducibility. In MVS method, “μl the mean activation over a neutral corpus disjoint from all evaluated tasks”, but which benchmarks or tasks are used to calculate μl is not clear. At section 5.1, linear probes are not clear define. By the way, the Figure 3 and 6 are not cited anywhere in the main context.
Questions for authors

None

Reviewer_PMVqOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper introduces the minimal viable subspace (MVS) method, which trains a low rank bottleneck into a language model on a narrow task distribution. This subspace bottleneck is trained such that its inclusion has minimal impact on e2e performance on the dataset, as measured by KL divergence. By doing followup experimentation on the subspaces learned, they show that this subspace is also necessary - when you ablate it, performance on the task drops to near-random.

The authors use this technique to try to isolate a subspace wherein abstract formal reasoning takes place in 2 models - Qwen3-8b and Gemma4-12b. They hypothesise that this might exist, given an analogous structure has been found in the human brain.

They procedurally generate a dataset of logical statements, each of which is either "valid" or "invalid" - roughly, whether the statement is logically consistent. the statements are three sentence syllogisms.

They use MVS to find low rank, linear subspaces which preserve performance on the syllogisms task when applied. They do this at various model layers, but focus primarily layer 21 in Gemma and 22 in Qwen. They also sweep subspace rank.

They find that performance on the syllogisms task can be retained even under very low rank (1 or 2) MVS subspace projection. Under the same projections they find performance on other tasks is destroyed.

They then train linear probes on raw model activations on the syllogisms task and find that various logical aspects of the prompt, including the validity (which is in this case, the "label"), are already linearly decodable by layer 21 in Qwen and layer 24 in Gemma.

They also evaluate the MVS subspaces under a held-out distribution of syllogism data, and find that the subspace still retains performance here.

Critical strengths

The most interesting part of the paper is where the authors show that here they can reduce performance on the task to chance by ablating their found subspace. This is a good validation that the subspace is causally important, and is mildly surprising, as I would naively expect some sort of hydra effect to at least retain some of the performance, by routing information through "backup" representations. That said, the ablation happens late enough into the model (roughly halfway) that I suppose it's not surprising that any hydra effects might be entirely localised to before it.

Critical weaknesses

I thought overall the paper, while containing some interesting results, does not sufficiently validate its main hypothesis, while admitting a far simpler explanation of the results. As I understand it, the claim is that the logical reasoning happens inside the subspace that they learn. I don't believe the authors show this. In fact the authors provide compelling evidence against this claim.

Fundamentally, I would argue the authors have failed to recognise that, while they have indeed found a "sufficient and necessary" subspace, they only do so for a single, and particularly late, point in the model's computational graph. To make a trivial illustrating example: the directions in the final residual stream corresponding to the tokens "valid" and "invliad" are also "sufficient and necessary", but are obviously not where the logical reasoning happens.

I would argue there is a simple explanation for the observations: MVS has found, at the highlighted layers, a subspace in which the already-computed answer lies.

The authors provide evidence for this. in Figure 6, they show activations projected onto the 2d subspace they isolate. The activations are, at this point, arranged into an antipolar representation, with "valid" and "invalid" points at opposite ends. This representation seems to already contains the model's computed answer. The authors also show that linearly probing for the answer is possible around/earlier than the point of subspace projection in Figure 3, but I would argue that while this might be cause enough for suspicion, the MVS projection evidence is even clearer evidence.

Questions for authors

I think the goal of isolating a subspace in which formal, abstract reasoning happens is very interesting. However, more experiments are needed in order to show that this has been found, and the current evidence admits the explanation that they have simply found a subspace containing the output of the reasoning.

I would like to see evidence for a subspace with the same empirical correspondence with reasoning (eg. ablating it harms abstract reasoning but not knowledge recall) at layers where the answer hasn't obviously already been computed.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.