acceptodds
Under review as a conference paper at ICLR 2027

A Learned Domain-Discriminative Subspace of the Residual Stream

Abstract

Do trained language models represent domain identity, i.e., how the text they are processing departs from their default distribution, in structured form, and is that structure learned rather than an artifact of how we probe for it? We define , a per-layer subspace of the residual stream obtained by joint diagonalization of per-domain generalized-eigenvalue discriminants against generic text, and recover the same compact object in five model families spanning three token-mixing mechanisms. It occupies -% of the hidden dimension, -% in the mixture-of-experts read one sub-layer earlier, sized by the number of calibrated domains rather than by stream width. Because such a construction returns a basis from anything, the depth claim is made at fixed rank and against a floor drawn from its own candidates - an arbitrary subspace of equal rank, which has a rising gate of its own. Against that floor, discriminability along rises with depth in every trained checkpoint, decays in random-initialization controls, and clears it by a near-constant - in all five families. Pseudo-concepts stay an order of magnitude below at every depth and rank. Removing degrades even the least affected calibration corpus - more than the generic reference, whereas a matched-rank variance-only subspace damages domain and generic text indiscriminately. Yet nothing we can measure reads the code, and its causal necessity does not extend beyond the corpora it was calibrated on, even though its discriminative format does. What survives is a learned, compact, corpus-selective code.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Learned subspace, Residual stream, Representation geometry, Representation learning, Mechanistic interpretability

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_kdn9OpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

The paper identifies a small set of directions within LLMs that distinguish specialized text, such as medical content, from general text. This signal becomes stronger in deeper layers of trained models but not in untrained models. Further studies show that removing these directions harms performance on domain-specific text.

Critical strengths
  1. The controls are comprehensive, including random models with pretrained embeddings, matched-rank random and PCA subspaces, and arbitrary-partition and style controls. Fixed-rank comparisons help rule out dimensionality as a confound in the depth-wise trends

  2. The analysis spans five model families with diverse architectures. Subspace ablations also show more selective damage to the corpora used for subspace construction than matched-rank PCA ablations, providing evidence beyond classification or visualization.

Critical weaknesses
  1. The core idea of VjointV_{\text{joint}} is introduced without clear motivation. It is unclear how it connects to the initial research question or why this layer-wise quantity is relevant to understanding how models “store domain identity.”

  2. Some terminology in the paper is confusing. For example, “calibration” appears to refer simply to subspace estimation using what seems to be an evaluation/test set, rather than calibration as commonly understood in the LLM literature. The paper also does not clearly explain why key metrics such as Fisher S/N provide appropriate evidence for the claimed second-moment code.

  3. The ablation experiments rely on in-sample evaluation, reusing text used to construct the subspace. Consequently, the selective performance degradation does not establish that the same causal effect holds on unseen documents.

  4. The paper is poorly presented, making its arguments and logical progression difficult to follow. As a human reader, I found most of the logic difficult to understand despite my best efforts.

Questions for authors

Q1: The paper states that, of the 10 initial claims, 4 were withdrawn, 4 were narrowed, and 2 were retained. I am curious whether the authors randomly selected (with AI) the initial 10 claims and then removed those that were unsupported. If not, how were the initial 10 claims formulated?

Reviewer_5utBOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

The paper investigates whether domain-specific information is encoded in compact subspaces of the hidden representations of LLMs. The authors propose a covariance-based approach to identify directions that discriminate between domain-specific and generic text. They analyze these subspaces across several model architectures and layers, using different controls and ablation experiments to assess their properties and relevance.

Critical strengths
  • Interesting research question. Understanding how LLMs internally represent information from different domains is an interesting and potentially valuable direction for interpretability research.

  • Extensive experimental analysis. The authors investigate several model architectures and include different controls, such as randomly initialized models, random subspaces, and PCA-based comparisons.

Critical weaknesses
  • Major lack of clarity. This is my main concern. The paper is unnecessarily complex and difficult to follow. Throughout the manuscript, the authors introduce numerous technical concepts without sufficient intuition or clear motivation (e.g., residual stream, GEV discriminants, contrastive codes, depth gates, etc.). Although the authors eventually define some concepts, their meaning and role in the proposed approach remain difficult to understand. Even at a high level, the paper does not clearly communicate the central contribution or its significance. Despite careful reading, I found it extremely difficult to follow the reasoning and reliably assess the technical content.

  • Limited novelty and unclear significance. Identifying discriminative subspaces in neural representations is not new. In particular, Dasgupta et al. (2026) already investigate related subspaces using a covariance-based generalized eigenvalue approach. Although the authors acknowledge this work, their additional contributions are not sufficiently novel or significant to me.

Questions for authors
  1. Could the authors provide a substantially clearer and more intuitive explanation of the proposed method, ideally with a simple example or diagram illustrating how the domain-discriminative subspaces are identified and analyzed?

  2. What is the main novel scientific finding beyond existing work on discriminative subspaces, particularly Dasgupta et al. (2026)?

Unfortunately, the lack of clarity prevents me from providing deeper insights (despite my best efforts).

Reviewer_v635OpenReview ↗
Rating
2
Confidence
1
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents a method to investigate whether the residual stream of trained language models maintains "domain identity" and whether it is learned rather than an artifact of probing. The paper defines a per-layer subspace obtained from generalized eigenvalue discriminants and shows its compactness and effectiveness through various experiments.

Critical strengths
  • The obtained per-layer subspace, VjointV_{joint}, is validated from multiple viewpoints and proved to be compact and effective.
  • The limitations, including that the subspace does not appear to be read or steered, are clearly described.
Critical weaknesses
  • The motivation of focusing on domain identity and the residual stream is not clearly described. Although domain identity is introduced as one of concrete features of the representational mechanism, the importance of domain identity should be described if it exists. The reason for using the residual stream is partially mentioned in section 3.1, it should be described earlier.
  • "Learned code" is stronger than the functional evidence. The title and abstract use the phrase "learned domain-discriminative subspace", and the paper often refers to VjointV_{joint} as a code. However, the functional evidence does not show that any component explicitly reads this code. Therefore, the term "code" sounds too strong.
Questions for authors
  • The advantage over matched-rank floors is modest: 1.49-1.73x in final level (line 319). This margin is consistent and valuable but not large. Why is this margin sufficient evidence for a learned object rather than merely a favorable selection effect?
  • What is the definition of "D" in line 045? Later, it is defined as the number domains and also as the set of domains.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.