acceptodds
Under review as a conference paper at ICLR 2027

Predicting Linear Relational Decodability Without the Relation

Abstract

Linear relational embeddings (LREs) approximate a language model's subject-to-object mapping with an affine function. Their faithfulness varies enormously across relations, from near zero for `person_father` to the ceiling for `country_capital_city`. We investigate the factors that contribute to this variation and demonstrate that it can be significantly predicted from a readout that is not aware of the underlying relation. To do so, we propose a novel form of the Jacobian lens, the *entity lens*. Whereas the Jacobian lens is estimated across all tokens and concepts, the entity lens is trained to predict what the model will output for an entity of type under the neutral prompt ` is known for`; e.g., we train ELs for countries and persons. After estimating an entity lens, we measure how highly it ranks the model's answer for a relation without using any relation-specific information in estimation and prediction. Across 28 relations of the LRE dataset on OLMo-3-7B this relation-agnostic score correlates with LRE faithfulness at (). This suggests that linear decodability is largely a property of what a subject's representation already carries, rather than of the relation being decoded. Moving the unit of analysis from the relation to the individual fact, we conclude that new relation datasets and analysis methods are needed to address several problems in existing datasets, most prominently semantically weak first tokens (e.g., most predicted universities' first token is “University”) and reference competition (e.g., the LLM confuses obscure actor names with competing non-actor entities that have the same last name). Our full pipeline is available [here](https://anonymous.4open.science/r/entity-lens-8CFD/).

Primary area:
interpretability and explainable AI
Keywords:
language models, mechanistic interpretability, linear representations, factual knowledge

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_dsC1OpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper examines why linear relational embeddings (LREs) are faithful for some relations but not others. It introduces the entity lens, a Jacobian-based readout from same-type entities under a neutral prompt, to rank relation-conditioned answers. Across 28 relations in OLMo-3-7B, lens scores strongly correlate with LRE faithfulness.

Critical strengths

It would be meaningful to explore why the linear approximation is accurate for some relations but completely inaccurate for others.

Critical weaknesses

The entity lens is defined as NEL(e)=mean⁡i[Jise]\mathrm{NEL}(e) = \operatorname{mean}_i[J_i s_e], whereas the original Jacobian computation is (mean⁡iJi)s(\operatorname{mean}_i J_i)s. Since mean⁡i[Jise]≡(mean⁡iJi)se\operatorname{mean}_i[J_i s_e] \equiv (\operatorname{mean}_i J_i)s_e, the two expressions are algebraically identical apart from the placement of parentheses and therefore do not represent different approximations. The actual differences should be clarified more carefully.

Questions for authors
  1. The paper conducts experiments on only 28 relations, so the conclusions drawn may not be reliable.
  2. Section 5 states that 8 (subject, object) examples will be provided. Does this not implicitly convey relation information, and thereby run counter to the original intent of using a neutral prompt?
  3. Table 6 shows that ρ is highly sensitive to neutral prompts; under such circumstances, drawing any conclusion about the correlation ρ is highly unreliable.
3 comments

Thank you for reading our paper and raising important questions.

(Q1): "The entity lens is defined as Jˉτ(s−sˉτ)\bar{J}_\tau(s - \bar{s}_\tau), whereas the original Jacobian computation is Jˉs\bar{J}s. Since Jˉτ(s−sˉτ)=Jˉτs−Jˉτsˉτ\bar{J}_\tau(s - \bar{s}_\tau) = \bar{J}_\tau s - \bar{J}_\tau \bar{s}_\tau, the two expressions are algebraically identical apart from the placement of parentheses and therefore do not represent different approximations. The actual differences should be clarified more carefully."

For NEL (mean⁡iJise\operatorname{mean}_i J_i s_e), we do not claim an algebraic difference from Jˉs\bar{J} s in Gurnee et al. (2026): both have the same functional form and differ only in where JiJ_i is evaluated (at same-type reference entities under "{} is known for" rather than arbitrary corpus tokens; lines 131-134 and 172-175).

For CEL (mean⁡iJi(se−si)=mean⁡iJise−mean⁡iJisi\operatorname{mean}_i J_i(s_e - s_i) = \operatorname{mean}_i J_i s_e - \operatorname{mean}_i J_i s_i), it is algebraically different from Jˉs\bar{J} s because the subtracted term mean⁡iJisi\operatorname{mean}_i J_i s_i is a vector, not a scalar constant. When projected onto the vocabulary, it subtracts a different value from each token (demoting attributes shared across reference entities) and changes the token ranking. Empirically on OLMo-3-7B, evaluating JiJ_i on same-type entities in "{} is known for" improves the mean log-rank from −3.48-3.48 to −2.59-2.59 and the correlation from ρ=0.630\rho = 0.630 (Gurnee et al.'s corpus Jˉs\bar{J} s) to ρ=0.735\rho = 0.735 (CEL) and ρ=0.765\rho = 0.765 (NEL). We have revised Section 4 (lines 164-167) to make this distinction clearer: after the definition of NEL, the manuscript now states that NEL shares the form of the Jacobian lens and differs in where the Jacobians are evaluated, and that CEL subtracts a vector that changes the token ranking.


(Q2): "The paper conducts experiments on only 28 relations, so the conclusions drawn may not be reliable."

To test a larger set, we added the BEAR benchmark (Wiland et al., 2024). Factual recall of OLMo-3-7B on BEAR is much lower than on the original LRE dataset (38.0%38.0\% vs. 76.6%76.6\% gold first-token accuracy), which leaves too few facts on which faithfulness to the model's own answer is informative. We therefore ran the evaluation on the larger Qwen3-32B, scoring, as in Hernandez et al. (2024), only facts whose gold first token the model predicts. Across all 49 entity relations, the correlation with LRE faithfulness is ρ=0.597\rho = 0.597 for CEL and 0.5440.544 for NEL. On this larger and harder relation set, the correlation remains in the range of the seven models in Table 1. We have added this analysis to Appendix C and plan to extend the BEAR evaluation to the other models for the camera-ready version.


(Q3): "Section 5 states that 8 (subject, object) examples will be provided. Does this not implicitly convey relation information, and thereby run counter to the original intent of using a neutral prompt?"

As stated in lines 196-199 and line 241, the 8 (subject, object) examples are only used as per definition for the LRE and to record the model's top-1 target token mem_e; the entity lens itself only ever sees "{subject} is known for" with no relation template and no objects.


(Q4): "Table 6 shows that ρ\rho is highly sensitive to neutral prompts; under such circumstances, drawing any conclusion about the correlation ρ\rho is highly unreliable."

In Table 6 (lines 768-781), the low values on Qwen3-8B come from position-0 activation norm spikes (discussed in lines 764-766) when {s} is placed at token index 0 without a prefix, and from bare prompts like "{s}" or "{s} is". Across all prefix-guarded neutral prompts in Table 6, Qwen3-8B stays within ρ∈[0.665,0.733]\rho \in [0.665, 0.733]. Separately, across 6 neutral prompts on OLMo-3-7B it stays within ρ∈[0.729,0.777]\rho \in [0.729, 0.777] (all p≤1.1×10−5p \le 1.1 \times 10^{-5}). While some variance across prompt phrasings is expected, it does not meaningfully change the results or correlation.

Thanks for your response. I have raised the score.

We appreciate that the reviewer took the time to work through our response and reconsider their assessment of our submission.

Thank you!

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

This paper studies why linear relational embeddings (LREs; Hernandez et al., 2024), which approximate a language model's subject-to-object mapping with an affine function, are faithful for some relations but not others (e.g., 87.5% for COUNTRY_CAPITAL_CITY vs. 7.5% for PERSON_FATHER on OLMo-3-7B). The authors argue that this variation can largely be predicted without using the relation. To test this, they propose the entity lens, a variant of the Jacobian lens that averages Jacobians at reference entities of the same type under the neutral prompt "{} is known for", in contrastive (CEL) and noncontrastive (NEL) forms. For each relation, the lens score measures how highly the lens ranks the model's own answer, without using any relation template, objects, or (subject, object) pairs in estimation. Across the 28 LRE-dataset relations with named-entity subjects on OLMo-3-7B, this score correlates with LRE faithfulness at Spearman ρ = 0.7355. The correlation persists after controlling for object cardinality and answer-token frequency and when each subject domain is held out in turn; results for six additional models are also reported. Further analyses examine the residuals in relation to object cardinality, check salience with the model's free generation, show that CEL is sensitive to the reference set, and recombine the LRE's relation-specific anchor with different subject terms. A fact-level analysis identifies problems in existing relation datasets, notably reference competition between entities with similar names and semantically weak first tokens (e.g., "University"). The authors conclude that linear decodability largely reflects what the subject representation already carries.

Critical strengths

1. An interesting question and a new perspective. The paper asks a clear and relevant question: why LRE faithfulness varies so much across relations. It offers a new perspective by shifting the analysis from the relation to the subject representation. The relation-agnostic diagnostic is a neat design: it uses no relation template, objects, or (subject, object) pairs, yet it is informative about LRE faithfulness, and it could be computed before fitting any LRE.

2. Careful controls and excellent reproducibility. The main result is tested against object cardinality and answer-token frequency, under leave-one-domain-out evaluation, and across seven models. The reporting is transparent: the paper acknowledges that LRE hyperparameters are selected in sample, discloses and corrects a numeric-tokenization artifact (Appendix D), and reports bootstrap intervals for the CEL–NEL comparison. The full pipeline is released, and a single command reproduces every number, table, and figure, with pinned seeds and configurations and a reported compute budget.

3. Useful insights into existing relation datasets. The fact-level analysis identifies concrete problems that affect how relational knowledge in language models is evaluated. These include reference competition (e.g., an obscure actor's name resolving to an unrelated entity with the same surname) and semantically weak first tokens (e.g., "University" is the model's answer for 54 of 91 subjects in PERSON_UNIVERSITY, and 6 of 47 relations have a first-token collapse rate above 25%). The paper also gives actionable recommendations for future datasets and evaluation protocols.

Critical weaknesses

1. Long-tail knowledge and pretraining frequency. The paper attributes the variation in LRE faithfulness to the subject representation, but this variation may largely reflect how often each fact appears in the pretraining data. A fact such as France–capital–Paris is extremely frequent, so the model learns it robustly and the answer becomes salient in the subject representation. In contrast, a specific person's father is long-tail knowledge that rarely appears in the training data, so the model learns it poorly and the answer is neither salient nor linearly decodable. Data frequency matters on both sides: how well a relation is learned depends on the amount of training data for it (https://arxiv.org/abs/2601.02931), and how salient an entity is depends on its frequency (https://arxiv.org/abs/2504.12459). However, the paper reports corpus frequency only as separate baselines in Appendix A.1, based on literal string counts, and the answer-frequency control in Section 6 measures the model's prior over the answer token rather than corpus co-occurrence. Without a joint analysis, such as a partial correlation or regression that controls for co-occurrence frequency, it remains unclear whether the entity lens captures anything beyond how well the model has learned a fact, which may be largely determined by how often that fact appears in the data.

2. The decoupling experiment is too small. Section 7.3 presents a decoupling experiment between the linear relation and the "is known for" readout, but I remain skeptical of its conclusions because the experiment is too small (two relation pairs, over 22 countries and 24 athletes).

3. The role of the relation is underestimated. The paper claims that linear decodability is mainly a property of the subject representation rather than of the relation. However, the paper's own swap experiment in Section 7.3 shows that the answer to be read out is determined by the bias term, which depends only on the relation: for the country relations, the bias term alone reads out none of the answers, while replacing the subject term leaves the answers largely unchanged. Moreover, the supposedly relation-agnostic lens score is computed precisely on the answer to that relation.

Questions for authors

see weaknesses

2 comments

Thank you for your thoughtful comments and time. Responses to the individual questions below.

(Q1): "Without a joint analysis, such as a partial correlation or regression that controls for co-occurrence frequency, it remains unclear whether the entity lens captures anything beyond how well the model has learned a fact, which may be largely determined by how often that fact appears in the data."

This is an important consideration. In Appendix A.1 (Table 5), we therefore compared baselines such as subject frequency csc_s (ρ=0.138\rho = 0.138), object frequency coc_o (ρ=0.571\rho = 0.571), joint co-occurrence csoc_{so} (ρ=0.457\rho = 0.457), and conditional probabilities P(o∣s)P(o \mid s) (ρ=0.449\rho = 0.449). To control for these corpus statistics directly, we added two analyses: a joint partial correlation analysis and a nested OLS analysis.

Control Set (OLMo-3-7B, n=28n = 28)CEL Partial ρ\rho (pp)NEL Partial ρ\rho (pp)
None (marginal)0.735 (8.2×10−68.2 \times 10^{-6})0.765 (2.1×10−62.1 \times 10^{-6})
Joint co-occurrence, mean of log⁡(1+cso)\log(1 + c_{so})0.666 (1.5×10−41.5 \times 10^{-4})0.711 (3.3×10−53.3 \times 10^{-5})
Forward conditional P(o∣s)P(o \mid s)0.664 (1.6×10−41.6 \times 10^{-4})--
All 5 corpus features (csc_s, coc_o, log⁡cso\log c_{so}, P(o∣s)P(o \mid s), P(s∣o)P(s \mid o)) + cardinality kk + answer prior0.634 (0.0020)0.722 (2.2×10−42.2 \times 10^{-4})

In nested OLS, adding CEL or NEL to a baseline with corpus co-occurrence, conditionals, kk, and the answer prior (R2=0.687R^2 = 0.687) increases R2R^2 to 0.8310.831 (ΔR2=+0.144\Delta R^2 = +0.144, p=3.8×10−4p = 3.8 \times 10^{-4}) and 0.9110.911 (ΔR2=+0.224\Delta R^2 = +0.224, p=3.8×10−7p = 3.8 \times 10^{-7}), respectively. Attribute importance may confound these associations: attributes more important (as per co-occurrence) to an entity may also exhibit stronger readout. Nevertheless, CEL and NEL explain significant additional variance. We have added this analysis to Appendix A.2 and point to it from Section 6.

(Q2): "2. The decoupling experiment is too small. Section 7.3 presents a decoupling experiment between the linear relation and the 'is known for' readout, but I remain skeptical of its conclusions because the experiment is too small (two relation pairs, over 22 countries and 24 athletes)."

In the original LRE dataset, only two relation pairs shared the exact same subjects while having a working LRE operator (Appendix G, lines 890-893). To address this, we built a 14×14\times larger paired dataset across 6 domains, 12 relations, and 644 held-out test subjects (1,288 evaluations) at layer 17, β=3\beta = 3: pooled accuracy is 10.5%10.5\% for the anchor bβrb_\beta^r alone, 31.6%31.6\% (34.5%34.5\% on gold-correct predictions) when combined with the neutral "is known for" readout, 34.3%34.3\% with a foreign-relation subject term, and 58.3%58.3\% (63.6%63.6\% on gold-correct predictions) for the unmodified LRE (outperforming the anchor alone on 10 of 12 relations). Appendix H describes the dataset, and Section 7.3 now reports the result.

(Q3): "3. The role of the relation is underestimated. The paper claims that linear decodability is mainly a property of the subject representation rather than of the relation. However, the paper's own swap experiment in Section 7.3 shows that the answer to be read out is determined by the bias term, which depends only on the relation... Moreover, the supposedly relation-agnostic lens score is computed precisely on the answer to that relation."

We agree that the relation is essential for selecting which attribute is decoded, and we have revised our wording in the Introduction, Section 5 (Measurement) and Section 7.3 to clarify its role: the relation anchor bβrb_\beta^r selects the target attribute space (achieving only 10.5%10.5\% alone since it is constant across subjects), while the subject term supplies the entity-specific attributes (31.6%31.6\% with the neutral readout, and 58.3%58.3\% with the relation's own Jacobian WrW_r). By "relation-agnostic," we mean that the entity lens readout EL(s)\text{EL}(s) is computed without any relation prompt or (s,o)(s, o) pairs, yielding a single vocabulary ranking per entity ss; evaluating that single ranking on a relation's answer tests whether that attribute is already salient in ss before any relation is given.

I appreciate the authors’ response. I am inclined to maintain my original score.

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

The paper proposes a relation-independent entity lens to measure the linear decodability of relations from subject representation. By showing a statistically significant correlation between LRE faithfulness (a relation-specific measure of linear decodability) and the score based on the proposed relation-agnostic entity lens, the paper argues that linear decodability can be detected from subject representation itself, without using any other relation-specific information. The paper also sheds light on problems in existing relation datasets.

Critical strengths

The paper contains many studies, findings, and insights that make this paper valuable for the interpretability research community. The main strength of the paper is that it addresses a relevant question, and the main takeaway from the paper, i.e., the observation that the linearity of a relation can be predicted from subject representation without any information about the relation, is very impressive. The claims are reasonably substantiated by using 28 relations.

Critical weaknesses

One of the main weaknesses of the paper is that its argument mainly relies on experimental results from one model - OLMo-3-7B. The appendix of the paper contains some study on 6 more models and it is curious to know why the paper is not forefronting that result in the main text. Overall, it was a bit challenging for me to appreciate the work, especially the last part. I believe the clarity of presentation can be improved.

Questions for authors

Faithfulness is not related to the gold answer, and therefore a simple model with very predictive behaviour can have very good faithfulness. In such a case, any reasonably good lens can have the same faithfulness as LRE. Given the possibility of such an extreme case, do you think the power and performance of the language model you used is a confounding factor in your observations? (I can see a low correlation from GPT-J-6B in your appendix table.) How sensitive are your findings to the model selection?

What I understood is that the LRE faithfulness score is calculated by doing a grid search; however, the selected values of layer and beta are not mentioned in the main text. Similarly the layer that used to calculate lens agreement scores is also not mentioned in the main text. Are both the same? How sensitive is the LRE score (and its correlation with lens agreement) to these hyperparameters? Since you already observed that the small dataset size is a limitation, such details are helpful to appreciate your main findings.

Can you extend this entity lens to other LRE relations? How difficult is it to modify your neutral prompt template for that?

The paper observes that CEL may struggle with relations that map many subjects to the same object (to motivate NEL). I can see some discussion about the object cardinality in the paper and some comparative study between CEL and NEL. The paper observes that NEL is a good predictor of LRE faithfulness in the middle of the object cardinality range; however, I couldn’t notice any comparison with CEL. Can we say, in general, that NEL is preferable to CEL if object cardinality increases, or is it only in the middle range? Also, did you observe that object cardinality has any relation to linear decodability? This is in the context where ‘Analysing the Linearity of Linguistic Relations in Language Model Embedding Spaces (https://openreview.net/forum?id=AzrY8LT2Tz)’ observed that relations that are not one-to-one (in this case, relations with higher object cardinality) are less linear.

3 comments

Thank you for the detailed assessment of our work. Below we open the dialogue based on your comments.

(Q1): "One of the main weaknesses of the paper is that its argument mainly relies on experimental results from one model - OLMo-3-7B. The appendix of the paper contains some study on 6 more models and it is curious to know why the paper is not forefronting that result in the main text."

We agree that showing model generalization upfront directly conveys confidence in our findings. Accordingly, we have moved the seven-model comparison into the main text (now Table 1 in Section 6), and Appendix B examines how factual recall and model scale affect the correlation (see Q2).

(Q2): "Faithfulness is not related to the gold answer, and therefore a simple model with very predictive behaviour can have very good faithfulness... Given the possibility of such an extreme case, do you think the power and performance of the language model you used is a confounding factor in your observations? (I can see a low correlation from GPT-J-6B in your appendix table.) How sensitive are your findings to the model selection?"

This is a good point and the reason why Hernandez et al. (2024) filtered for matches to the first token of the gold label. Note that faithfulness can still be inflated under this filter when multiple objects share a generic first token (e.g., "University"). In our main text, we scored all rows unfiltered. When we apply the gold-correct filter to GPT-J-6B, which has weak factual recall (50.1%), the correlation increases from ρ=0.435\rho = 0.435 to 0.6340.634 (p=5.1×10−4p = 5.1 \times 10^{-4}).

To test the effect of model size directly, we ran a sweep across the Qwen3 family (0.6B to 32B):

Qwen3 ScaleLM Gold AccuracyMean FaithfulnessRaw CEL ρ\rho (pp)Gold-Correct CEL ρ\rho (pp)
0.6B57.4%0.6510.210 (0.303)0.318 (0.113)
1.7B66.4%0.6350.658 (2.6×10−42.6 \times 10^{-4})0.679 (1.4×10−41.4 \times 10^{-4})
4B74.0%0.6060.548 (0.0038)0.569 (0.0024)
8B78.7%0.6280.665 (2.1×10−42.1 \times 10^{-4})0.613 (8.7×10−48.7 \times 10^{-4})
14B81.5%0.6490.417 (0.0339)0.474 (0.0145)
32B78.2%0.5680.538 (0.0046)0.562 (0.0028)

From this, we conclude that our results are robust across different model families and sizes. We have added this table to Appendix B (Table 7). In the camera-ready version we will apply the gold-correct filter uniformly to all models, since our question is not whether a fact is retrieved but how it is retrieved when it is.

(Q3): "What I understood is that the LRE faithfulness score is calculated by doing a grid search; however, the selected values of layer and beta are not mentioned in the main text. Similarly the layer that used to calculate lens agreement scores is also not mentioned in the main text. Are both the same? How sensitive is the LRE score (and its correlation with lens agreement) to these hyperparameters?"

This is a (again) good point, and we should have stated these details directly in the main text. We adopted the same setup as Hernandez et al. (2024) as stated in lines 202-206. For the entity lens, we used layer 17 for all relations, meaning that the layer used for LRE approximation can differ from the layer used for the entity lens. To understand how sensitive our results are to these hyperparameters, we analyzed (1) fixing the LRE settings to layer 17 and β=2\beta = 2 for all relations without a grid search (CEL ρ=0.711\rho = 0.711, p<0.01p < 0.01, compared with ρ=0.735\rho = 0.735 with a grid search) and (2) varying the entity lens layer across all six layers, as shown in the table below:

Entity Lens Layer5111417 (ours)2023
CEL ρ\rho0.6230.7580.6780.7350.6600.685
NEL ρ\rho0.6550.7700.7530.7650.6570.521

We interpret this as not having a significant impact on overall predictive power. The manuscript now states the lens layer in Section 5 (Measurement), lists the selected (layer, β\beta) per relation in Table 11 (Appendix I), and reports the fixed-setting result and the table above in Appendix E.

Continued from Part 1.

(Q4): "Can you extend this entity lens to other LRE relations? How difficult is it to modify your neutral prompt template for that?"

The remaining LRE relations include morphological tasks such as big→bigger\text{big} \to \text{bigger} and break→broke\text{break} \to \text{broke}, as well as identifying a word's first letter. These concern transformations or properties of words themselves rather than factual attributes of named entities, so "{} is known for" is not a natural shared prompt for them. We also explicitly state this restriction in the Limitations section (line 514).

Extending the lens to these tasks would require a prompt suited to lexical information. The main difficulty is preserving neutrality: a prompt tailored to elicit a comparative form, past tense, or first letter would reveal the target relation, undermining the relation-agnostic nature of the diagnostic.

(Q5): "Can we say, in general, that NEL is preferable to CEL if object cardinality increases, or is it only in the middle range? Also, did you observe that object cardinality has any relation to linear decodability? This is in the context where 'Analysing the Linearity of Linguistic Relations in Language Model Embedding Spaces' observed that relations that are not one-to-one (in this case, relations with higher object cardinality) are less linear."

Yes: as object cardinality kk increases, NEL is preferable to CEL (whereas at k≈1k \approx 1, CEL is slightly stronger). Comparing both readouts across kk tertiles shows that CEL degrades at high kk due to reference cancellation (Section 7.2), while NEL remains predictive:

Cardinality Tertile (kk = subjects/object)nnCEL ρ\rho (pp)NEL ρ\rho (pp)
Low kk (1.0-1.1, 1-to-1)90.9830.983 (1.9×10−61.9\times 10^{-6})0.9330.933 (2.4×10−42.4\times 10^{-4})
Medium kk (1.1-6.2)90.6610.661 (0.0530.053)0.5270.527 (0.1450.145)
High kk (9.2-189.4, many-to-1)100.3820.382 (0.2760.276)0.7820.782 (0.00760.0076)

Regarding kk and linear decodability: as noted in Section 6 (lines 252-254), kk correlates weakly positively with LRE faithfulness (ρ=0.355,p=0.064\rho = 0.355, p = 0.064).

Thank you for addressing my comments. I will maintain my original score.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.