acceptodds
Withdrawn from ICLR 2027

Entanglement Is All You Don’t Need

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

This paper primarily focuses on mechanistic investigation, conducting controlled experimentation at a 1B parameter scale to effectively isolate scaling effects, rather than large language model optimization aimed solely at achieving higher benchmark scores. Operationally, we quantify feature entanglement through three measurable properties: (1) decreased intrinsic dimensionality relative to physical dimensionality, (2) increased overlap of activated neurons across different inputs, and (3) lack of hierarchical specialization across layers. The intrinsic dimensionality collapse observed in DeepSeek-V2-Lite-Chat, Gemma, and Qwen indicates premature feature fusion, leading to feature entanglement, which limits the model's representational capacity. To address this problem, we propose a principled architectural design called EPMORE (Explainable Process Mixture-of-Experts) based on the core intuition that the reasoning process is a conditionally dependent feature (from abstract to concrete) orthogonal accumulation process. EPMORE consists of four major components. SPE (Simple Position Encoding) achieves disentanglement between position and content. EPL (Expansive Processing Logic) increases feature capacity by dynamically expanding from extremely low dimensions, making feature disentanglement possible. PIL (Parameter Independence Loss) in the architecture decouples the factor entanglement in weight matrices, thereby achieving feature disentanglement. The MOR (Middle Output Reuse) mechanism allows tokens whose intrinsic dimensionality increases during inference to align with tokenizers at different levels of abstraction; during training, intermediate ground truth labels (not manually annotated) aligned in a hierarchical manner based on the token's level of abstraction (i.e., intrinsic dimensionality) provide intermediate losses that shorten backpropagation paths, accelerating training and reducing computational cost. Ultimately, it enables feature disentanglement between layers. Experiments demonstrate that the EPMORE architecture alleviates feature entanglement, making it possible to provide preliminary evidence toward process-level interpretability. The example of “the capital of France” in the supplementary materials clearly shows how the model derives the result step by step according to the semantic hierarchy.

Primary area:
interpretability and explainable AI
Keywords:
EPMORE, SPE, EPL, PIL, MOR

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 3 reviews, above 0% of the others.
Reviewer_W5c2OpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The authors of the paper propose EPMORE (Explainable Process Mixture-of-Experts), a Transformer-based architecture that is designed to reduce feature entanglement through progressively expanding hidden dimensions, sparse activations, and intermediate supervision. Instead of maintaining a fixed hidden dimension throughout the network, the authors propose to use a hierarchy of representations;

Critical strengths

The authors pose a generally interesting question about feature entanglement in language models and whether architectural modifications can alleviate it.

Critical weaknesses
  1. Poor presentation and organization. The paper is extremely difficult to read and follow. Its structure deviates substantially from conventional research-paper organization, with excessive use of bold text, inconsistent formatting, unclear explanations, and poorly connected sections. Overall, the presentation significantly undermines the paper's clarity.

  2. Weak experimental methodology. The authors report training approximately 1B-parameter models from scratch using GSM8K, which contains only around 7.5K training examples, without clearly reporting any additional pretraining corpus. It is unclear what meaningful conclusions about language-model architectures can be drawn from such a limited training setup. Furthermore, the paper lacks convincing evaluations across diverse benchmarks and properly trained, competitive baselines.

  3. Questionable definition and measurement of entanglement. The proposed operationalization of feature entanglement is insufficiently justified. In particular, higher intrinsic dimensionality and reduced neuron activation overlap do not necessarily imply semantic feature disentanglement. The proposed parameter independence metric also has mathematical problems (see Question 1), making the central empirical conclusions difficult to interpret.

  4. Insufficient interpretability evidence. The interpretability claims are mainly supported by qualitative evidence, primarily a single simple factual example.

  5. Poor figure quality. Many figures are extremely difficult to read, with very small text and insufficient explanations. Several provide limited useful information, making it difficult to understand the results or independently assess the authors' conclusions.

  6. Unconvincing comparisons and theoretical justification. The baseline comparisons do not adequately control for architectural differences, activation functions, or training convergence.

The manuscript feels rushed and insufficiently developed--many strong claims concerning disentanglement, efficiency, and interpretability are not supported by the evidence; the presentation is below conference standards;

Questions for authors

See Weaknesses

Q1: The proposed parameter independence metric is mathematically questionable. By the inclusion-exclusion principle, Equation (2) -- PIL -- simply computes the normalized cardinality of the union of activated neurons, rather than their independence. As a result, the metric can assign higher scores to disjoint activation sets but also to fully overlapping dense activations, making its interpretation ambiguous. Could the authors elaborate on the choice of such a loss term?

Reviewer_RCA1OpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper introduces EPMORE, an interpretability focused LLM architecture that combines SPE, EPL, PIL, MOR. The work shows that increasing the dimensionality of the hidden representation, and their other components SPE, PIL, MOR, increases the intrinsic dimensionality across depth, while baseline models do not.

Critical strengths

The idea to increase dimensionality to decrease entanglement is interesting.

Critical weaknesses

At a high level, there is insufficient justification of the central premise. While the work argues that reductions in the intrinsic dimensionality equate to greater entanglement, that relationship is never sufficiently established. For example, lower dimensionality could mean convergence toward the final token, or specialization based on context.

The experimental evidence is also unconvincing. First, the baselines that are compared against (DeepSeek-V2-Lite-Chat, Gemma4-E4B, Qwen2.5) are all trained from scratch to match the 1B parameter mark. In Table 4, the perplexities for these models have extraordinarily poor performance, and it is unclear whether these baselines have learned to model language at all and thus whether this training setup is a valid comparison in the first place.

The evidence for interpretability itself is also weak, being limited generally to discussions around the “capital of France” example. The relationship between the other metrics, such as activation sparsity and overlap, is not clear.

Finally, the presentation of the paper is lacking. The writing generally relies heavily upon intuition rather than rigorous explanations, without the necessary ablations to support the claims, such as the “fact level MoE layer”. The formatting of the writing, in particular the results, is also highly unconventional. Other inconsistencies include a disagreement between Figure 3, which suggests four layers for EPMORE, going from “city” -> “political” -> “Paris” explicitly across these layers, and Figure 1, which suggests 16 layers.

Overall, I feel that the paper would benefit substantially from a revision, both in terms of its prose and its experimental justification.

Questions for authors

No additional questions.

Reviewer_6Bc1OpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper argues that standard LLMs suffer from "intrinsic dimension (ID) collapse": the PCA dimension of hidden states shrinks in deeper layers. The authors call this "feature entanglement." They propose EPMORE, with four parts:

  • SPE: position is stored in a separate dimension.
  • EPL: layer width grows layer by layer (for example 128 -> 640 -> 1152 -> ...) and uses ReLU.
  • PIL: a loss/metric for overlap between activated neurons.
  • MOR: intermediate layers predict tokens and receive extra losses during training. They train EPMORE and "1B versions" of DeepSeek-V2-Lite, Gemma4 and Qwen2.5 from scratch and report that EPMORE's ID goes up across layers while the baselines' ID goes down. They also show one qualitative example ("The capital of France is" -> city -> political center -> Paris).
Critical strengths
  • Interesting and relevant question. The paper tries to build interpretability into the architecture instead of explaining a model after training. Looking at layer-wise representation geometry (ID per block) is a useful lens.

  • Simple, easy-to-try ideas. The changes are small: growing width, ReLU sparsity, a separate position dimension, and intermediate prediction heads. The core Transformer operators stay the same, so others can test these ideas easily.

  • Good effort on reproducibility and controls. Code is released. All baselines are trained from scratch with the same setup, with no pretrained weights. The authors also run ablations (without EPL, without MOR) and statistical tests.

  • Honest scoping in many places. The authors clearly mark the complex-number analogy, the Discussion section and the "capital of France" case as intuition or illustration, not proof. They state that the results are limited to the 1B scale.

Critical weaknesses
  • In Table 4, Qwen2.5 has loss 12.34 and perplexity 228,355. With a ~152k vocabulary, random guessing gives a loss of about 11.9, so this baseline did not learn at all.

  • All baselines are trained for only 1,000 steps. Large MoE-style models shrunk to 1B likely need different settings.

  • The models do not look matched. EPMORE uses 8.7 GFLOPs and DeepSeek uses 58.5 GFLOPs (Table 5), and parameter counts are never reported per model.

  • EPMORE's layer width grows with depth, so its PCA dimension can grow simply because each layer has more dimensions. Baselines have a fixed width.

  • EPMORE uses ReLU. The baselines use SwiGLU/GeLU-style activations, which are almost never exactly zero. So "activation sparsity = 1.0000" and "overlap = 1.0000" for the baselines reflect the activation function, not entanglement. A fair comparison needs a ReLU baseline of the same width.

  • D.1: the claim that "J = W^T * diag(sigma') is a non-negative matrix" is false, because W has negative entries. The gradient-alignment argument therefore fails. There is also a dimension mismatch: a d_K vector is projected to d_{K-1} and then compared.

  • B.4: "implicit regularization makes the rank non-increasing" is assumed, not proven. Also, "grad_theta L -> 0 implies grad_h L -> 0" uses d(theta)/d(h), which is the wrong direction for the chain rule.

Typos

  • "Equation equation 11" appears twice in D.1 (an \eqref issue).
  • "Gemma4-1B" (Table 4) vs. "Gemma4-E4B" elsewhere.
  • "We have the EPMORE of architecture's hierarchical..." (D.1) is unclear written
Questions for authors
  • "Expansive Processing Logic" (Sec. 2.4) vs. "Expansion Processing Logic" (Sec. 3.2.2): are they different?
  • "Simple Position Encoding" vs. "Simple Positional Encoding"
  • "rotational positional embeddings" should be "rotary positional embeddings"

See Weaknesses

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

Related papers

Discussion (0)

Sign in to comment.