acceptodds
Under review as a conference paper at ICLR 2027

Semantic Fusion and Subspace Separation in Language Models: A Hypothesis and Its Predictions

Abstract

We propose a hypothesis of semantic fusion and subspace separation in language models. Early information exchange may build shared representations of semantic units, reducing independent constituent access, while explicit boundaries may favor separately readable subspaces. The hypothesis predicts first-constituent substitutions in predecessor retrieval and their reduction with separators. Across three instruction-tuned language models in English and Chinese, semantic pairs elicit more such errors than random triples, and commas reduce them on paired items. To study the proposed mechanism, we construct a minimal arithmetic-and-retrieval task that permits explicit control over pair meanings and retrieval exposure. A small Transformer exhibits the predicted substitutions for operator pairs assigned arithmetic meanings but withheld from retrieval training. Blocking the early first-to-second-token attention pathway improves retrieval while sharply impairing arithmetic. These findings support the hypothesis's predictions; the proposed subspace organization remains a mechanistic interpretation.

Primary area:
interpretability and explainable AI
Keywords:
LLM, transformer, semantic fusion

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_u3dHOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes the hypothesis of semantic fusion and subspace separation in language models, suggesting that semantically related adjacent tokens may form shared representations,thereby hindering the retrieval of individual tokens,and that explicit delimiters can mitigate this interference. The authors validate this hypothesis through controlled experiments using toy Transformers, as well as experiments in both Chinese and English across three large language models. The results indicate that semantically related token pairs are more prone to first-component substitution errors, which can be reduced by the insertion of commas. Furthermore, while blocking early-stage attention pathways improves retrieval performance, it simultaneously impairs arithmetic reasoning capabilities.

Critical strengths

1.The paper proposes an interesting hypothesis connecting semantic composition with constituent-level retrieval interference, offering clear and testable predictions. 2.The experiments across three language models and two languages (English and Chinese) show consistent behavioral patterns, providing empirical support for the proposed predictions. 3.The toy Transformer setup allows explicit control over semantic associations and retrieval exposure, helping investigate the relationship between semantic composition and constituent accessibility.

Critical weaknesses

1.Although the experimental results support the behavioral predictions, the paper does not directly demonstrate the existence of semantic fusion or subspace separation in the models' internal representations. 2.The semantic and random conditions are not fully controlled for factors such as word frequency, token co-occurrence, and lexical familiarity, which may contribute to the observed performance differences. 3.The semantic operator pairs are excluded from retrieval training, making it difficult to distinguish interference caused by semantic fusion from out-of-distribution generalization failures. 4.The use of a post-hoc selected checkpoint for one random seed raises concerns about evaluation consistency. Additionally, the architectural comparison does not isolate the contributions of convolution and linear attention.

Questions for authors

Questions For Authors

  1. Direct evidence for the proposed mechanism: Could the authors provide additional analyses, such as linear probing or activation patching, to directly examine whether semantic fusion reduces the accessibility of individual token representations?

  2. Potential confounding factors: Would the observed differences between semantic and random pairs remain significant after controlling for word frequency, token co-occurrence, and lexical familiarity?

  3. Out-of-distribution generalization: Would similar retrieval errors occur for token pairs that are excluded from retrieval training but are not assigned arithmetic meanings? This would help distinguish semantic interference from generalization failures.

  4. Hybrid-model evaluation: Could the authors report results using a consistent checkpoint selection strategy across all random seeds and further isolate the effects of convolution and linear attention?

Reviewer_oDqfOpenReview ↗
Rating
2(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper hypothesizes that semantic composition can impair access to individual constituents, while separators preserve access. For example, when asked for the word before "lake" in "ice cream lake," a model may answer "ice" or "ice cream" rather than "cream." It reports such errors across three LLMs and two languages, reduced errors with commas, and a toy-model attention intervention that improves retrieval while impairing arithmetic.

Critical strengths
  • The toy intervention improves retrieval for every pair and seed. For pair (0, 1), mean accuracy rises from 14.55% to 76.59%, while overall arithmetic accuracy falls from 99.73% to 10.78%. This is a useful demonstration of competing task demands on the same pathway.
  • The authors distinguish measured behavior from the proposed subspace interpretation and disclose the hybrid model's checkpoint-selection issue.
Critical weaknesses
  • My main concern is the limited explanatory contribution. Section 6 reports an interesting behavioral pattern, but the preceding sections do not establish why it occurs. The proposed subspaces are not measured, the toy experiment confounds arithmetic meaning with retrieval-training exposure, and its connection to the real-LLM results remains untested. Consequently, the paper provides evidence for a narrow behavioral observation, but insufficient evidence for the broader account that motivates the work. Although the authors appropriately describe that account as a hypothesis, the experiments do not discriminate it sufficiently from alternatives to make the explanation convincing.
  • In Section 4.3, arithmetic pairs are excluded from retrieval training. Their failures could reflect generalization to excluded adjacencies or interference from arithmetic training, rather than semantic fusion. The answer-token controls do not distinguish these possibilities. Equally held-out pairs without arithmetic meanings, with token identities and exposure controlled, would provide a more informative baseline.
Questions for authors
  • Section 4.1 devotes substantial space to standard Transformer computations (Eqs. 5–8). I suggest retaining the notation needed to define the attention-pathway intervention and moving the remaining architectural details to the appendix. The space would be better used for the literal evaluation prompts or the hybrid model’s checkpoint-selection caveat for better flow of the paper.
  • There are a few relevant papers that maybe the authors should read:
    • Wu et al. (ICLR 2025). The Semantic Hub Hypothesis: Language Models Share Semantic Representations Across Languages and Modalities.
    • Dankers et al. (ACL 2022). Can Transformer be Too Compositional? Analysing Idiom Processing in Neural Machine Translation.
    • Chen et al. (ICML 2025). SepLLM: Accelerate Large Language Models by Compressing One Segment into One Separator.
    • Li and Subramani (ACL 2026). Model Internal Sleuthing: Finding Lexical Identity and Inflectional Features in Modern Language Models.
    • Arora et al. (ICLR 2024). Zoology: Measuring and Improving Recall in Efficient Language Models.
Reviewer_sJk1OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The paper proposes a hypothesis, there is a link between the semantic fusion, from words that are familiar pairs such as "ice cream", with subspace separation, the presence of a comma or a punctuation mark should favorate the retrieval of words. While the semantic fusion make it dificult to retrive the previous word in the case there is a presence of familiar pairs, a delimiter like comma and a toy-model of attention reduce this dificulty of retrive words.

Critical strengths
  1. The proposed hypothesis of semantic fusion and subspace separation, shows a interesting relationship between the arithmetic connection of words and the retrieval capacity of the model to treat each word as an individual token and retrieve the aisolated information.

  2. The experiment shown consistent patterns with the hypothesis and the use of an isolated transformer allows to control the experiment.

Critical weaknesses
  1. Even the tiny transformer experiment shows a clear dilema between arithmetic connection and retrieval capacity, it is not clear for the LLMs models. There is no statistical test to ensure a significant diference between the LLMs models results.

  2. No experiment show estandar desviation and in the experiment in table 5 shows high variability between each seed. More experiment should be tested for the tiny transformer experiment.

Questions for authors

Compound words needs to be separated? May be like human mind the compound words only shows that some concepts or representations needs to be treated as one word.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.