acceptodds
Under review as a conference paper at ICLR 2027

Shared Semantic Distinctions and Model-Dependent Expression in Text Encoders

Abstract

Text embeddings support semantic decisions, yet downstream success does not explain which distinctions their vectors encode or how those distinctions are organized. Recent work suggests that independently trained models converge geometrically, raising the question of what meaning they encode in common. We study seven frozen text encoders using controlled sentences that vary six distinctions: success versus failure (outcome), affirmation versus negation (polarity), which event happens first (order), who acts on whom (role), what contains what (containment), and who has which attribute (binding). We compare recovery across constructions and vocabulary, correspondence in responses to semantic changes, and decision-rule transfer as a diagnostic of representational compatibility. Outcome and polarity remain readily recoverable, while relational distinctions show model- and expression-dependent generalization. Binding reveals a clear dissociation: nonlinear readouts improve recovery across all seven encoders on held-out constructions, yet corresponding response geometry coexists with a substantial transfer deficit through the tested linear maps relative to matched within-model readouts. Useful alignment constraints also differ by distinction: matching full responses benefits order, whereas matching their averages performs better for containment. These findings show how apparent semantic deficits can depend on the readout and why geometric correspondence alone is insufficient to establish compatible decision rules. They provide a controlled account of semantic content and organization, clarifying which conclusions about shared meaning persist across linguistic expressions and ways of accessing embeddings.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Text Embeddings, Representation Learning, Latent Space Alignment, Semantic Accessibility

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

The paper studies semantic recoverability and cross-model compatibility across seven modern text embedding models across six distinctions (outcome, polarity, order, role, containment, binding) under four shift conditions. The authors evaluate native linear vs. nonlinear (128-unit GELU) probes, geometric response correlation (Δ\Delta vectors), and linear rule transfer. They find that relational binding is latently recoverable via nonlinear heads (0.667→0.9380.667 \to 0.938 AUC on construction shifts) and shows cross-model response correlation (ρ≈0.84\rho \approx 0.84), yet fails linear rule transfer (0.711→0.5340.711 \to 0.534 AUC).

Critical strengths

Solid Empirical Suite: Comprehensive evaluation across seven modern encoder families covering both native recoverability and 42 cross-model transfer directions.

Controlled Generalization Splits: Separating construction (syntax) from lexical (vocabulary) shifts prevents probes from relying on superficial co-occurrence heuristics.

Geometry vs. Transfer Disconnect: Effectively demonstrates that geometric response correlation between models does not guarantee functional rule transfer via linear interfaces.

Critical weaknesses

Mismatched Platonic Framing: Framing this as a refutation of the Platonic Representation Hypothesis (Huh et al., 2024) is an ad-hoc stretch. PRH addresses macro-level conceptual convergence across natural data distributions, not whether single pooled sentence vectors linearly preserve combinatorial, XOR-style binding rules through linear ridge stitching.

Missing Foundational Literature & Uncontrolled Probe Capacity: The paper omits foundational Edge Probing work (Tenney et al., ICLR 2019; ACL 2019), which already showed that relational semantics (SRL, coreference) require nonlinear composition heads over linear projections. Without these, it remains unproven whether the 128-unit GELU probe is reading out an intrinsic manifold or simply memorizing synthetic template regularities.

Ecological Validity: The findings rely almost entirely on synthetic, hand-crafted templates. Real-world retrieval/ranking (e.g., Amazon ESCI) involves heavy noise, vocabulary mismatch, and intent ambiguity interacting simultaneously. It is unverified whether these synthetic boundaries reflect real retrieval failure modes.

Descriptive Cataloguing Without Practical or Theoretical Depth: No Mechanistic Theory: The paper offers no geometric or linguistic theory explaining why nonlinear readouts help Binding across shifts but degrade Role under construction shifts (ΔAUC=−0.010\Delta\text{AUC} = -0.010).

Questions for authors

Real-World Confounders: How do your template-derived findings transfer to noisy retrieval/ranking benchmarks (e.g., ESCI or multi-hop QA) where syntax, vocabulary shifts, and intent ambiguity interact simultaneously? Or even consider additional noises (unrelated attributes) in the sentences? What would be the right real-world benchmark for your paper?

Taxonomy Motivation: What formal semantic framework guided selecting these six distinctions over other relational structures.

Mechanistic Divergence: What underlying representational theory explains why GELU non-linearities substantially benefit Binding across all shifts, yet degrade out-of-distribution construction generalization for Role?

1 comment

Thank you for the review!

We agree that success or failure of a ridge map on pooled sentence vectors does not adjudicate the Platonic Representation Hypothesis. We now state this explicitly in the introduction. Our question is narrower, for the same change of meaning, does agreement in how two encoders' vectors change imply that one encoder's decision rule works on the other's vectors through a particular map? The six questions are a controlled sample, not a formal taxonomy. Outcome and polarity admit overt lexical marking. Temporal order, thematic role, spatial containment, and attribute assignment let us vary relations, often while preserving the words. This contrast motivates the selection without claiming coverage of semantics in general.

We now situate the study relative to Tenney et al.'s edge probing (ICLR 2019; ACL 2019), which examines linguistic structure in contextualized token and span representations. Our readout sees one pooled sentence vector, without access to identified argument spans. Matching probe width and training isolates the activation comparison but does not identify an encoder-internal mechanism.

There is a concrete explanation available for much of the original binding gain. Person mention order and color mention order are each linearly readable and their interaction determines the original label exactly. A classifier using the two decoded scores and their product gets .926 AUC against .938 for GELU on held constructions. A GELU advantage of .248 is already present at width eight. With randomized labels assigned within scenes, high training performance is possible, but final AUC stays near chance on disjoint scenes. That control distinguishes memorizing scene labels from learning a regularity that generalizes. The positional interaction is precisely such a regularity. Breaking it in new sentences separates badge contrasts (frozen GELU .900) from “latter/former” and “second/first” reference (.245 across that family). We report the failure as prominently as the gain.

Role is different and mention order is tied to the label within a template, and its direction can reverse across constructions. We tested whether disagreement between training and held-out mention-order cues predicts the GELU penalty. It does not account for the role results. The descriptive association across held templates has the opposite sign to our prediction. The activation effect is small and shift-specific (−.010 on construction, +.058 on vocabulary shift), unlike the large interaction effect on the original binding task. We can give a sufficient readout-level account for much of the latter but we cannot attribute the former to the same mechanism.

We tested your suggestion of irrelevant context directly, keeping the trained heads frozen. Prepending an unrelated sentence, identical for all four versions of a scene, reduces held-construction affine role AUC from .859 to .647 and GELU binding from .938 to .768. Appending it instead gives .712 and .840. The label is unchanged, but its ranking is sensitive to surrounding text. Because the distractor varies across scenes, this test includes between-scene score shifts as well as changes to each label contrast.

For a natural-text benchmark, PAWS is a closer fit to our readout question than product relevance since sentence pairs have high word overlap but can differ in meaning when words are reordered. On 8,000 test pairs, matched GELU beats linear on identical pair features (.671 vs .645 AUC; difference +.026, 95% shared-sentence-group interval [.022, .030]). Plain cosine does better than either (.701). This is a separate paraphrase task, not transfer of our synthetic binding head. These encoders' PAWS outcomes were also previously explored. We did not evaluate ESCI or multi-hop QA. We use PAWS, the existing MultiNLI check, and the distractor result to delimit the claim that readout choice matters beyond the templates, but the controlled scores do not predict retrieval quality in noisy applications.

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

This paper investigates which semantic distinctions text encoders share and tests seven encoders on six binary distinctions (outcome, polarity, order, role, containment, binding). The authors use three measurements. 1. They train probes on each model (affine, linear & nonlinear) to test whether a distinction can be classified from its output embeddings. 2.They compute, within each model, the cosine similarities among contrast vectors (the embedding difference when one label flips), and then rank-correlate these similarity patterns across models. 3. They learn a linear map from one model to another and test whether a probe trained on the first model still works on the second. Outcome and polarity are near ceiling in recoverability and transfer. The relational distinctions lose accuracy on held-out constructions. The authors' main claim concerns binding: nonlinear probes recover it much better than linear probes, and its responses correspond across models, yet linear-map transfer of an affine probe stays far below the target model's own probe

Critical strengths
  1. Distinguishing the three measurements recoverability, geometric correspondence and rule transfer is clear and well motivated. Measuring all three on the same scenes is a useful design. Holding out construction families is a good test of generalization. The lexical controls (Table 16) show that this removes most n-gram shortcuts for the relational distinctions, though not for outcome and polarity.
  2. Reporting is detailed and open: per-model and per-family tables, all intervals, the analysis chronology, and explicit limitations.
Critical weaknesses
  1. Terminology. The term “distinction” refers to three related things: the six categories, the specific binary question associated with each category, and the labels assigned to the two outcomes (“Each label identifies a specific binary distinction,” Appendix B.1). The term “recover” appears in the abstract and introduction before being formally defined, where it is used alongside “encode” and “make accessible.” A concise table early in the main text would improve clarity by listing each binary question. More broadly, the main text is quite dense, with numerous qualifiers and abbreviations that make the central results difficult to follow.
  2. The scope is rather narrow. Each distinction uses fixed focal terms, and prompts differ across encoders (Qwen and SFR receive a "word or phrase" instruction for sentences). A prompt ablation would make the model comparisons more reliable.
  3. Context-condition scores. For order, role and containment, word 1–2gram controls reach 0.83, 0.94 and 1.00 AUC in the context condition, close to the encoders (0.999, 0.998, 0.991). The main text should state that these scores carry little information about semantic encoding, and that the construction and joint conditions are the informative tests.
  4. For binding, the background assignment is mention order. "The doctor has the red badge and the lawyer has the blue badge" and "The lawyer has the blue badge and the doctor has the red badge" both have label 1. The label is then close to an XOR of two surface features: who is mentioned first and which color is mentioned first. A linear probe would fail on such a label even if the encoder represents both features, so the gain from nonlinear probes may reflect the stimulus design rather than binding itself. The paper does not state what the background assignment varies for order, role and containment.
  5. Below chance results: I noticed that there are several below chance AUCs, which are not further analyzed. E.g. Table 14 (p. 17) and Table 17 (p.19).
Questions for authors
  1. For each distinction, what does the background assignment vary? Is mention order crossed with the label inside each template?
  2. What explains the below-chance AUC ?
  3. How sensitive are the results to the encoding prompts? In particular, do Qwen and SFR give similar results without the "word or phrase" instruction?
  4. Could you define "recover" in the introduction and state whether "encode", "make accessible" and "recover" refer to the same property? A table in the main text with the binary question, the meaning of label 1 and the background assignment for each distinction would also help.
1 comment

Thank you, your questions point to a distinction we should have made on page one: that is “recover” means "predict the answer to a specified question from a frozen vector on held-out examples, using a specified readout". We will reserve “distinction” for the property, give its binary question explicitly, and use “recoverable” for the measurement. Recovery does not establish how the encoder computes or uses the answer.

For a fixed site and filler category, each template crosses the answer (yes/no) with one background choice. We have put this table in the main text:

DistinctionBinary question (yes = label 1)Background choice at fixed label
OutcomeDid the event succeed?Which event
PolarityIs the device active?Which device
OrderDid the focal event occur after the briefing?Which focal event
RoleIs the doctor the agent?Which action
ContainmentIs the box the inner object?Which other container
BindingDoes the doctor have the red badge?Doctor or partner mentioned first

Only for binding is participant mention order explicitly crossed with the label inside each template. For role and containment, the relation between first mention and label can reverse between templates. For temporal order, some templates keep event mention order fixed and change “before/after”, others express the label through the sequence of mentions.

We agree that the context condition is a familiar-construction reference. Word 1–2-grams alone reach .832, .937, and .996 AUC for order, role, and containment there. We will foreground held constructions and the joint construction/vocabulary shift when interpreting generalization. Even those splits need explicit shortcut tests, your proposed binding interaction solves all original held families (AUC 1.000 without embeddings). In a post-review test using equal numbers of rule-consistent and rule-breaking training scenes, GELU beats matched linear on held rule-breaking families (.670 vs .547, known vocabulary), but averages only .332 on the held anaphoric family. We report this qualification alongside the original gain. Focal terms remain fixed throughout and neither test establishes generalization to a new focal entity.

On your Q2, below .5 means that the frozen score tends to rank the negative above the positive; moving a classification threshold cannot repair that ordering. Table 14 has an interpretable example: “When the survey began, the briefing had already finished” mentions the later event first. This perfect-aspect family reverses the predominant training cue and has GELU AUC .384 (.429 with held vocabulary). The cue analysis is consistent with that reversal, but did not explain the below-chance containment Q/A or clock-order results in Table 17. Clock expressions also require comparing two numerical operands. We now distinguish the observed reversals from hypotheses about their cause and retain their original score orientation.

On Q3, yes, for individual-model comparisons. We removed the instruction entirely for Qwen and SFR, re-encoded the same sentences, and refit the same heads on the same folds. Held-construction affine order AUC changed from .891 to .974 for Qwen and from .933 to .874 for SFR and even their ordering reverses. Binding's GELU advantage over matched linear survives removing the instruction in both models, whereas the original Gemma–Qwen GELU binding ordering can reverse when Qwen's prompt changes. We now qualify rankings by encoding recipe, moved the targeted model comparisons to the appendix, and use the main-text space for the question table and the construction results.

Reviewer_Byg3OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies shared semantic structure across seven frozen text encoders. It tests six semantic distinctions using controlled sentences, unseen constructions, and held-out background vocabulary. The analysis compares recoverability, response geometry, and cross-model rule transfer. Outcome and polarity remain easy to recover. Relational distinctions depend more on sentence structure and readout choice. Nonlinear readouts substantially improve binding recovery. However, affine binding rules transfer poorly through the tested linear maps, despite corresponding response geometry.

Critical strengths

-The paper separates recoverability, geometric correspondence, and rule transfer. This distinction provides a useful framework for studying shared semantic structure across text encoders.

-The study evaluates seven frozen encoders on six semantic distinctions. It uses 50 templates and 20 construction-family holdouts. Construction shifts and background vocabulary shifts are tested separately.

-GELU improves binding AUC from 0.667 to 0.938 under construction shift. All seven encoders show gains. The average gain remains positive across three random seeds.

Critical weaknesses

W1: Uselis et al. (ICML 2026, How Can Embedding Models Bind Concepts?) study a closely related problem. They show that object information is recoverable from individual CLIP encoders despite cross-modal binding failures. They also study multiplicative binding functions in controlled models. The present paper does not cite this work. Its text-only setting differs, but the additional insight from its binding results needs clearer explanation.

W2: Table 6 admits a simple positional rule. The label is the XNOR of “doctor before lawyer” and “red before blue.” A nonlinear readout can solve this rule using only two mention-order indicators. The GELU improvement from 0.667 to 0.938 therefore does not, by itself, rule out a positional shortcut. The reported lexical baselines do not directly test this interaction. The paper also does not establish that the held-out families break this rule.

W3: Transfer is tested only with affine readouts through ridge-fitted linear maps. The maps use 5,000 paired texts, mostly from Dolma. The paper does not refit a readout on the mapped vectors. It also does not test transfer of the GELU binding readout. These missing controls make it difficult to separate mapping limitations from source-rule misalignment. The transfer gap therefore remains difficult to interpret.

Questions for authors

Q1: How do the binding findings extend Uselis et al. (ICML 2026)? What new insight follows from the linear/GELU comparison and the cross-model transfer tests?

Q2: Can a nonlinear baseline using only entity and color positions solve the binding benchmark? Does the GELU advantage persist on constructions that break this positional rule?

Q3: Does the transfer gap persist with more alignment data or stronger maps? Can the authors refit a readout on mapped vectors to test what information remains accessible? Please also test GELU-rule transfer against a matched nonlinear target readout.

2 comments

Thank you for identifying two controls that most changed our interpretation! We ran both (breaking the mention-order rule and fitting a new readout after mapping) after receiving the reviews.

On Q1

We missed a directly relevant paper that you pointed out. Uselis et al. (ICML 2026) show that object information recoverable within CLIP encoders can coexist with poor cross-modal retrieval. They also study multiplicative binding functions in controlled models. We now credit them for that dissociation. Our additional evidence concerns independently trained text encoders. We can change the linguistic construction while keeping the underlying assignment fixed, and change the readout while keeping the mapped vector fixed. The first comparison exposes the positional shortcut and the anaphoric failure, the second shows that poor source-rule transfer need not mean the mapped vector lacks readable information. A product that works in our readout is not evidence that the encoder implements the binding function they studied.

On Q2

Yes. Across all eight original binding templates, including every held-out family, the doctor has the red badge exactly when “doctor before partner” and “red before blue” have the same truth value. A text-only classifier given these two bits and their product achieves AUC 1.000. The original n-gram baselines missed this interaction. The two bits are also separately decodable from the embeddings (held-construction AUCs .997 and .993). A classifier using those two scores and their product reaches .926, close to the original GELU readout's .938. The original gain alone cannot distinguish positional decoding from entity–attribute binding.

We tested the original frozen heads on ten new templates. For example, in “The doctor and the lawyer have badges: the latter has the blue badge and the former has the red badge,” doctor and blue appear first, yet the doctor has red. The literal order rule is wrong on every sentence in six new templates and correct on every sentence in four others. With known vocabulary, frozen GELU scores .930 on the new rule-consistent templates and .615 on the rule-breaking ones (matched linear .616 and .537). The latter average hides an important split that is 0.9 for sentences contrasting two badges, .699 for nested clauses, and .245 for anaphoric reference such as “latter/former.”

We also refit with equal numbers of rule-consistent and rule-breaking training scenes, keeping the original 288-sentence budget and holding out one construction family at a time. On held rule-breaking families, GELU reaches .670 versus .547 for matched linear (paired difference +.124, 95% scene interval [.118, .128], positive in five of seven encoders). The anaphoric family remains below chance (.332). These results go beyond the literal two-bit rule for some constructions, they do not establish a general binding mechanism. Rule type and syntax vary together, and the intervals condition on the tested families and encoders. We have replaced the unqualified binding interpretation with this family-level result.

On Q3

We held the map fixed and compared three readouts on the original binding task. The refit uses target training labels on mapped vectors, the native control uses the same labels on unmapped target vectors. Below are held-construction AUCs with known vocabulary, averaged across all 42 directions, for full-response ridge maps:

Matched readoutFrozen source rule on mapped vectorsRefit on those same mapped vectorsTarget's own rule on native vectors
Centered affine.534.647.711
Centered GELU128.723.904.939

Some nonlinear rule transfer succeeds. Nevertheless, its deficit against the matched GELU target is .216 (95% scene interval [.208, .222]), changing only the readout raises AUC from .723 to .904. Much of the task signal thus remains readable after mapping even when the particular source rule is a poor fit. This is a diagnostic using target supervision, not a label-free transfer improvement. Increasing the alignment pool from 5k to 20k texts did not close the gaps. On eight directions we also tested residual and free nonlinear maps. The ordinary-text residual fit selected the initial ridge solution in all eight, and the free map was still improving at the epoch cap. This bounds the result to the tested interfaces; it is not evidence that stronger maps cannot work.

Thank you for the thoughtful response and for running the additional experiments. The new results addressed some of my questions about word order and cross-model transfer. After considering these results, I feel that my current score reflects my overall assessment of the revised paper.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.