acceptodds
Under review as a conference paper at ICLR 2027

Fair Fact-Checking: Closing the Cross-Lingual Gap in LLM Factual Judgement with RoSh

Abstract

Misinformation on social media remains a critical problem, and more and more people settle it by asking a language model instead of a fact checker, about anything from a claim circulating on a platform to a plain question of fact. Whether models judge such claims reliably is still debated; whether they judge them equally well in every language people ask in has gone almost unasked. We test eight models from five families, 3B to 70B, on 1,500 claims that mix fact-checked news with encyclopedic facts and exist in identical form in eight languages. On every model English is judged better than the other languages, and the gap is widest on the smallest models, where Llama-3B on Arabic is no better than guessing. Existing remedies either retrain the model on more multilingual data or fit an unconstrained map between language representations, and neither asks whether the model already holds the answer and simply fails to say it. It largely does: a linear probe recovers the truth from the very activations the model fails to express. We propose RoSh, a per-language shift and rotation of the residual stream, computed in closed form at three layers, with no training and no weight modified. It improves every model and closes 75% of the gap on average, and it helps most where the model was worst: the two smallest models end close to English, Arabic on Llama-3B goes from chance to nearly the English level, and a fifth fewer of the claims answered correctly in English are lost in translation. What remains is no longer a read-out failure: afterwards the head recovers at least as much of what is encoded outside English as it does in English. An unconstrained map fitted on the same pairs falls below the untouched baseline, so the orthogonality constraint is doing the work, and every model clears a scrambled-correspondence control and ten further controls. On the two benchmarks of the closest inference-time method, latent-space intervention, run with its own data and metric code, RoSh's gains are five to thirteen times larger.

Primary area:
interpretability and explainable AI
Keywords:
multilingual language models, cross-lingual consistency, fairness, activation steering, orthogonal Procrustes, fact verification, training-free intervention

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_csDwOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

Here the authors examine the discrepancy in factual claim verification performance between English and other languages using LLMs. They explore whether this reflects a lack of factual knowledge or difficulties in retrieving information already encoded in the models' internal weights.

The authors compare the performance of the models' original read-out mechanisms with linear probes trained on intermediate representations. They introduce RoSh, an inference-time technique which applies orthogonal rotations to the residual stream at selected transformer layers. These transformations are estimated using Orthogonal Procrustes, without altering the original model weights.

The authors evaluate across eight languages and eight models from five families. The authors report improvements in factual verification performance, particularly for smaller models and languages with larger initial performance gaps. The evaluation includes comparisons with existing methods, component ablations and controls to examine the importance of geometric transformations, layer selection and cross-lingual correspondences.

Critical strengths

1. Interesting research question The paper addresses an interesting question regarding multilingual factual knowledge in LLMs: whether differences in factual verification performance across languages reflect missing knowledge or difficulties in recovering information from internal representations. The distinction between linearly decodable information and the model's actual predictions provides a useful perspective for investigating multilingual model behaviour. The proposed RoSh method is simple and well motivated.

2. Extensive experimental validation and well-designed controls The paper provides a several experimental evaluation across multiple languages, model families, and model sizes. The reported improvements are particularly notable for smaller models. They also include several ablation studies and control experiments. The comparisons involving shift-only and rotation-only interventions, random transformations, scrambled correspondences, and alternative layer choices help establish the importance of different components of the proposed method. These experiments strengthen the empirical evidence supporting RoSh beyond simply reporting improvements over the baseline.

Critical weaknesses

1. Interpretation of the read-out gap My main concern relates to the interpretation of the difference between the performance of linear probes and the models' original read-out mechanisms. The authors argue that factual knowledge is already encoded in the models' internal representations but is not adequately recovered in languages other than English. While the probe results support the presence of linearly decodable information, this does not necessarily establish that the original model can functionally access or use that knowledge. Similarly, the improvements obtained through RoSh provide evidence that cross-lingual representation alignment is beneficial, but they do not uniquely establish the mechanism responsible for the original performance gap.

The authors acknowledge some of these limitations but certain statements about recovering existing knowledge and eliminating read-out failures appear stronger than the evidence directly supports. A more cautious interpretation would help distinguish the demonstrated empirical improvements from the proposed explanation of their underlying mechanism.

2. Dependence on the read-out formulation and limited comparison with simpler alternatives The evaluation relies on a fixed binary verdict formulation, using the output logits associated with the English tokens TRUE and FALSE across all languages. It remains unclear to what extent the observed performance gaps depend on this particular choice of verbalizers. Authors may evaluate translated verdict labels or alternative prompt formulations in order to determine whether part of the multilingual gap can be reduced without modifying internal activations.

Questions for authors
  1. Have the authors investigated whether the multilingual performance gap and the improvements obtained with RoSh remain consistent when using alternative verbalizers, particularly verdict labels translated into each target language? This would help clarify whether the observed read-out gap is partly dependent on the fixed TRUE/FALSE output formulation.
  2. How does RoSh compare with directly adapting the read-out mechanism, for example through a lightweight linear classification head trained on the same fitting data? Given that linear probes already recover substantial factual information from the representations, such a comparison could help clarify the advantages of correcting intermediate activations rather than replacing or adapting the final decision mechanism.
2 comments

Thank you. Both of your questions led to new experiments in the revision, and one of them changed what we believe.

Q1 and W2: does the gap depend on the English TRUE and FALSE words? No.

We changed the prompt in each language to ask for that language's own words (VERDADEIRO/FALSO, PRAWDA/FAŁSZ, WAHR/FALSCH, صحيح/خطأ, 真/假, ВЕРНО/НЕВЕРНО, DOĞRU/YANLIŞ), scored as in Eq. 7. Separately we kept the English instruction and left only the claim in its own language (Section 5.2, Table 4; every model in Appendix Q, Table 22).

The mean gap is 0.047 with native words and 0.049 with the English instruction, against 0.050 with our prompt. Single models move a lot: native words halve Llama-3B's gap, from 0.111 to 0.051, and widen Mistral-7B's, from 0.130 to 0.147. On average they barely move it.

We then refitted RoSh from scratch with the native words. It narrows the gap on all eight models, from 0.047 to 0.021, against 0.015 with English words. So on average neither the gap nor the correction depends on the English answer tokens, although the correction does recover somewhat less without them.

W1: we were claiming more than the evidence supports.

You are right, and the revision now says only what the results show.

Section 3.1 presents the account as our working hypothesis, which the results support but do not prove: the activations carry the truth value in every language, but in a different frame from the English one, meaning a different orientation and position of the activation cloud. The abstract and Section 1 now speak of what a probe recovers against what the verdict expresses, by the probe's measure. They no longer speak of knowledge the model already holds, and "no longer a read-out failure" is gone. Section 5.1 no longer calls read-out inefficiency a general property of the model, and Appendix A says the probe estimates rather than bounds what a linear reader can extract. Section 6 has a new part, "What the probe shows, and what it does not". A probe shows that a linear reader can extract the information, not that the model uses it. The split is an estimate, because the probe is trained on the same templates and the head is not. And the change-of-frame account is offered as the reading most consistent with the results, not as an established mechanism. The Limitations say we do not identify which part of the head fails, or what puts non-English activations in a different frame.

What the evidence does support, we kept. The gain needs the claim-level pairing: scrambling it drops the score to 0.621 and a random rotation gives 0.658, against 0.853 for the fitted map. The map is invertible, so the gain cannot come from deleting information (Proposition E.1). In a re-run the fitted map beats all 20 scrambled pairings in each of 20 model and split runs. Corrected activations land near their own English counterparts without becoming copies (Figure 4).

Q2: how does RoSh compare with a lightweight linear head trained on the same data?

Our probe is exactly that head, fitted at an intermediate layer rather than on the final state: an L2-regularised logistic regression on the same 750 fitting claims, at one layer (Table 11). On held-out claims it scores higher than RoSh, 0.910 outside English against 0.855, and we say so rather than leaving it out. We use it as the measuring instrument that splits the gap, not as a competitor.

Correcting the activations differs from replacing the read-out in two ways. The model's own head still produces the verdict, so the gain is a statement about the model rather than about a new classifier. And the two use the labels differently: a head fits all of its weights to the labels, while RoSh fits its map from the claim pairing alone and uses labels only to rank candidate layers. Section 6 now makes this comparison in the paper.

Reviewer_csDwOpenReview ↗

Assessment after rebuttal

I would like to thank the authors for their response and efforts. The new TRUE/FALSE experiments are useful for analyzing the multilingual performance gap. I also appreciate the clarification regarding the comparison with linear probes.

Furthermore, the revised interpretation of the read-out gap addresses my main conceptual concern. Presenting the proposed explanation as a hypothesis brings the conclusions closer to the experimental evidence. Overall, I believe the rebuttal substantially addressed my concerns and strengthened the paper. Therefore, I maintain my positive recommendation.

Reviewer_aHwUOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The main goal of the paper is to improve the performance of non-English factual judgement of LLMs (or in other terms close the cross-lingual gap to English). The authors propose the RoSH method which does a per-language shift and rotation of the residual stream, but no model weight updates.

Critical strengths
  • Relevant research question of the multilingual accuracy gap
  • Simple method
  • Positive evaluation results (but baselines missing)
Critical weaknesses

A critical weakness is the readability of the paper. I list several things below (not an exhaustive list but meant to illustrate the point that the paper is hard to digest as terms are used but not properly introduced and one cannot easily linearly read through the paper):

  • Some parts of the paper are hard to read, e.g.: “What remains is no longer a read-out failure: afterwards the head recovers at least as much of what is encoded outside English as it does in English.” and “so the orthogonality constraint is doing the work” - That’s not really understandable at this point of the paper (the abstract).
  • Intro: “Across 48 markets” => What is a market here?
  • AUROC performance numbers are used in in the intro without saying what benchmark or at least benchmark type they actually refer to (later it becomes clear that it’s the authors derived benchmark and that it’s binary which btw. means using accuracy would be intuitive in addition to AUROC).
  • I do not understand “It is also asked about everything, so our claims mix fact-checked news with encyclopedic facts, and the gap is as wide on one half as on the other.” - Which claim? Which gaps? Would be really good to make this more readable.
  • “Comparing probe against head within each language splits the gap in two: 0.026 is a shortfall in what the model encodes, and 0.021 is the head reading non-English less efficiently than English” - Sorry, what? Those numbers come out of the blue.
  • Related work: Text like “Who receives a good judgement.” should be clearly marked as an intermediate heading (e.g. bold) - otherwise it looks like an incomplete sentence.
  • Method section: “We observe that the truth is linearly readable in every language we test” should be explained - it’s really unclear at this point of the paper.

Method:

  • Assuming that the rotation introduced in the method acts as an implicit “translated to English”, then why don’t you run a baseline which translates the question to English?
  • There seems to be an assumption that questions are binary, but it is not stated (or only late).

Data construction:

  • “The set mixes claims of the kind fact-checkers adjudicate with encyclopedic facts, and the non-English versions are taken as released from BMLAMA-53 (Qi et al., 2023); Appendix B lists the claim types and gives examples.” - The paper has no appendix and BMLAMA-53 does not contain binary statements, so it’s not really clear what the authors did.

Metrics:

  • The paper has a non-standard definition of accuracy (“Accuracy and consistency are read at each language’s median operating point, fitted on the training half.”). People use an LLM and look at its output. Can you report accuracy as how often the model correctly outputs TRUE or FALSE? That straightforward measurement seems to be missing.
  • You say that a linear probe takes the internal activation at some layer and outputs a true/false score. This probe is then also a fact checker, right? Does “Across all eight models the probe reaches 0.910” means that his fact checker beats your own method which seems to achieve 0.859/0.823 in Table 2 but 0.853/0.820 in Section 5.3.

Ambiguous parts of the paper:

  • Llama-3B, as used in the abstract and elsewhere, is not a proper model name - the authors should spell out the version i.e. Llama 3.2 3B. (Same for Qwen-32B and Llama-70B, but here it’s even worse because there are actually multiple realistic options for what those models could be e.g. Llama 3 70B, Llama 3.1 70B, Llama 3.3 70B.)
Questions for authors
  • Can you run a baseline which translates a question to English and measue its performance?
  • Is the benchmark representative? Does it resemble the social media fakes news scenario mentioned in the introduction?
2 comments

Thank you. Every item you listed was specific enough to act on, and all of them are fixed in the revision.

Readability. Both abstract sentences you quoted are gone. "Across 48 markets" is now "Across the 48 countries of the Reuters Institute survey". Section 1 describes the task, defines AUROC (0.5 chance, 1 perfect) and names the benchmark before any number appears. The sentence beginning "It is also asked about everything" is removed; it described data we do not have. The 0.026 / 0.021 split is now 0.026 / 0.025, after a correction to one model's row, and both are defined where they first appear: the probe scores 0.026 lower outside English, and the drop from probe to verdict is 0.025 larger outside English than inside it. "Who receives a good judgement." is now a paragraph heading. Section 3 says what "linearly readable" means where it says it: a logistic-regression probe on one intermediate layer's hidden states separates true from false claims (Table 1). Section 4 names all eight models in full (Llama-3.2-3B-Instruct, Qwen2.5-32B-Instruct, Llama-3.3-70B-Instruct and the rest), with exact checkpoints in Table 5. The binary task is stated in the second paragraph of Section 1.

The appendix. You were right that it was missing: it sat in the supplementary PDF and is now in the main PDF. You were also right that BMLAMA-53 has no true/false statements. We build them. A true claim states a BMLAMA fact; a false claim keeps the subject and relation and takes a distractor object from BMLAMA's candidate list. Appendix B lists the 18 relation templates with counts and examples.

Q1, translating into English first. We ran it as a user would, with the model as its own translator. In one chat the user writes "Translate this text to English:" and the claim, the model replies, and the user then asks our English question about that reply (Section 5.2, Table 4, Appendix Q).

It widens the gap rather than closing it. Outside English, AUROC falls from 0.820 to 0.780. Four models lose 0.03 to 0.21 (Mistral-24B, 0.866 to 0.660), three move by under 0.01, and Llama-3B rises. The gap grows from 0.050 to 0.090; with RoSh it is 0.015. The replies show why: asked to translate a false claim, the model sometimes corrects it instead. Mistral-24B answers the Portuguese for "Slovakia is located in Oceania" with "It seems there is a mistake in the text." This happens to 13% of false claims against 6% of true ones, and 28% against 5% on Mistral-24B, so the step alters mainly the claims the verdict should reject.

Your reading of the rotation as an implicit translation is fair, with one difference: RoSh never rewrites the claim. Corrected activations land near their own English counterparts without becoming copies (mean cosine on Llama-3B rises from 0.05-0.19 to 0.66-0.78, Figure 4), English is left exactly unchanged (Proposition E.4), and the cost is one d by d multiply at three layers, on one token.

Q2, representativeness. No, and the paper no longer claims it. The claims are encyclopedic facts judged without evidence, the paper is framed as closed-book factual judgement, and real-world claims are named as a limitation (Section 6.1). We use parallel claims because they isolate the language; existing multilingual fact-checking sets are not parallel.

Plain accuracy. Added. Section 5 opens with raw TRUE/FALSE accuracy: 0.728 in English, and 0.672 before and 0.693 after outside English, per model in Table 21. AUROC stays primary because seven of the eight models answer TRUE for at most 36% of English claims although half are true, so raw accuracy mixes that bias with how well the model ranks claims.

The probe. Yes, it is a classifier, and on held-out claims it beats RoSh, 0.910 against 0.855 (Table 11). We say so in the paper and do not propose RoSh as a better classifier. The probe is our measuring instrument: it shows what a linear reader can extract, which is what splits the gap in two. RoSh asks whether the model's own unchanged head can express that once the activations sit in the English frame. They also use labels differently: a probe fits all of its weights to the labels and replaces the verdict, while RoSh fits its map from the claim pairing alone and uses labels only to rank 12 to 27 candidate layers.

0.859 / 0.823 against 0.853 / 0.820. A real inconsistency, now fixed. Same method at two depths: Table 2 corrects three layers per language, the control run corrected two. The revision states the depth wherever it applies, and Table 18 repeats every control at three layers. After the DeepSeek-67B correction the numbers are 0.819 before, 0.855 after at three layers, and 0.853 at two.

Every weakness you listed concerned how the paper reads, where the data description sat, or a missing baseline. All are addressed, and the baseline you asked for became a new experiment with a result we did not expect. We would ask you to look at the soundness score again against the revision.

Thank you for fixing the abstract issue, adding the plain accuracy metric and fixing the inaccuracy. I think this does improve the paper. Nevertheless, I still maintain concerns about readability and potential further inconsistencies or soundness issues which may be hidden due to the paper being hard to follow. It's also a significant pivot of the paper during the rebuttal phase, which means a full set of reviews at another venue seems more appropriate to me. Overall, I maintain my rating.

Reviewer_Po6POpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper finds that LLMs judge factual claims less accurately outside English. Across 8 models (3B–70B) and 1,500 claims in 8 languages, English leads by +0.046 AUROC on average, and Llama-3B on Arabic is at chance. A linear probe shows that the truth is still encoded in the model’s internal activations; the model’s output layer just fails to read it out. The authors propose RoSh, a per-language shift and rotation of the activations at up to three layers. It is fitted in closed form from matched English/target-language claims, with no training. On held-out claims it closes 75% of the gap on average. Controls with scrambled pairings, random rotations and an unconstrained map all score lower, and on two external benchmarks it reports larger gains than prior work. The paper frames this as fair fact-checking, but all claims are encyclopedic BMLAMA templates, so the validation covers cross-lingual factual consistency, not fact-checking.

Critical strengths

S1. Careful controls. Fitting the rotation on mismatched claim pairs or using a random rotation scores below the untouched baseline, while RoSh scores above it. This suggests the gain is tied to the claim pairing rather than to generic perturbation.

S2. An interesting diagnostic. On Llama-3B Arabic, a linear probe on the model’s activations scores well above the model’s own answer (0.871 vs 0.523). This hints that part of the gap lies in the read-out, though it is shown on encyclopedic facts only and builds on existing probing methods.

Critical weaknesses

W1: neither the data nor the task setup supports the paper’s fact-checking framing.

(a) The data contains no fact-checked claims. The abstract says the 1,500 claims “mix fact-checked news with encyclopedic facts.” Appendix A says one half consists of “claims of the sort that circulate publicly and are adjudicated by fact-checking organisations” and that “Section 5 reports it separately for each half.” Appendix B contradicts this: “The 1,500 claims come from 18 BMLAMA relation templates… Three relations, twin cities, shared borders and official language, make up two thirds of the set.” A false claim is a true one with a “distractor object” swapped in. Table 6’s example is “Colorado shares border with Wyoming.” Section 5 contains no per-half analysis, and the authors note that neither external benchmark contains news claims.

(b) The task is not fact-checking. Fact-checking involves collecting evidence, assessing sources, reasoning about context, and giving a graded, explained verdict such as “misleading” or “missing context.” The paper instead scores the TRUE/FALSE logit difference from a single forward pass on a short sentence, with no evidence, retrieval or reasoning. This measures whether a fact is stored in the model’s weights. It also doesn’t match the chatbot scenario used as motivation, since deployed assistants generate explanations and often retrieve sources.

The title, the misinformation motivation and the deployment argument in Section 6 all assume the method makes fact-checking across languages fairer. What the paper actually shows is narrower: a closed-form rotation improves cross-lingual recall of encyclopedic facts, measured from a single token’s logits. Whether this transfers to real claims, or to answers that involve evidence and reasoning, is untested. In addition, the paper’s own appendices contradict each other on what the dataset contains. The central claim is therefore unsupported as written. Could the authors clarify where the fact-checked claims come from, or reframe the paper as a study of cross-lingual factual consistency?

Questions for authors
  1. The abstract states that the claims “mix fact-checked news with encyclopedic facts,” and Appendix A describes the first kind as “claims of the sort that circulate publicly and are adjudicated by fact-checking organisations.” Appendix B, however, states that “the 1,500 claims come from 18 BMLAMA relation templates.” Does the dataset contain any fact-checked claims, and if so, where do they come from?
1 comment

Thank you. You were right on both parts of W1, and the revision does what you asked.

Does the dataset contain fact-checked claims? No. All 1,500 claims are encyclopedic facts from BMLAMA-53. A true claim states a BMLAMA fact; a false claim keeps that fact's subject and relation and replaces the object with a distractor from BMLAMA's candidate list (Section 4; Appendix B, Table 8). The submitted abstract and Appendix A were wrong, so the per-half analysis they promised could not exist. Every such sentence is corrected. We apologise for the confusion.

The task is not fact-checking. We agree, and the paper no longer says it is. Fact-checking gathers evidence, weighs sources and returns a graded verdict. We score a TRUE versus FALSE logit difference from one forward pass on a short sentence. The paper now calls that what it is:

The PDF is titled "Closing the Cross-Lingual Gap in LLM Factual Judgement with RoSh". The fact-checking and misinformation argument is gone from the abstract and introduction. One sentence remains, saying that people ask AI assistants whether a claim is true. Section 6, "Scope and use": the results show that the same model judges the same claim differently by language, and that much of the difference can be removed inside the model. They do not show that evidence-based fact-checking becomes more equal across languages. Section 6.1, Limitations: the claims are short templated encyclopedic statements, so the results concern stored facts, not real claims that need evidence, context or reasoning.

Why this setting is still worth studying. It isolates the language. The claim, the model and the question stay fixed, and only the language changes. Existing multilingual fact-checking sets are not parallel across languages, so they cannot do this (Section 2). And this judgement is what any evidence-based use of a model rests on.

What the reframe does not change. No result depended on the old framing. Every model judges English better than the other seven languages on average, by 0.050 AUROC across the eight models, with Arabic on Llama-3B at chance. A probe splits the gap into two parts of similar size. RoSh closes 68% of it in closed form, with the controls you note in S1.

On S2, the probe. Agreed that probing is standard, and we claim no credit for it: we use it as a measuring instrument, not as a contribution. What is new is comparing the probe against the model's own verdict on the same parallel claims, across eight models and eight languages, which splits the gap into 0.026 that the activations lack and 0.025 that only the verdict fails to express (Section 5.1). Then a closed-form correction removes the second part on average and passes eleven controls. The two new experiments also show that, on average, the gap is not an artefact of the English answer words, and that asking the model in a plain chat to translate the claim first does not remove it (Section 5.2).

We hope the reframed version answers your main objection, and we would be grateful if you would reconsider your assessment in its light.

Comments on the paper

1 comment

We thank the three reviewers. The revised PDF is uploaded, it now contains the appendix, and the changed passages are highlighted. Four things changed.

  1. The paper now says what it measures. The old framing claimed fact-checking. The task is one short claim, no evidence, no retrieval, and a TRUE or FALSE verdict, so we call it closed-book factual judgement. The paper on the PDF is titled "Closing the Cross-Lingual Gap in LLM Factual Judgement with RoSh". The abstract and Sections 1, 2, 6 and 7 are rewritten, and misinformation is now one sentence of motivation.

  2. We were wrong about the data, and we say so. The submitted abstract and Appendix A said the claims mix fact-checked news with encyclopedic facts. They do not. All 1,500 claims are encyclopedic facts built from BMLAMA-53. A true claim states a BMLAMA fact; a false claim keeps the subject and relation and swaps the object for a distractor from BMLAMA's candidate list. Every sentence that said otherwise is corrected, and Section 4 explains the construction. We apologise for the error.

  3. One headline number was wrong. The submitted Table 2 averaged DeepSeek-67B over six languages and left out Arabic. With all seven the mean gap is 0.050, not 0.046, and RoSh closes 68% of it, not 75%. The whole DeepSeek-67B row is replaced and every number in the revision uses the corrected values.

  4. Two new experiments, both on all eight models and the same 750 held-out claims (Section 5.2, Table 4, Appendix Q). Held-out AUROC outside English, and the gap, averaged over the eight models:

Paper prompt, asked in the claim's language: 0.820, gap 0.050 Answer words in each language: 0.823, gap 0.047 Answer words in each language, RoSh refitted for them: 0.849, gap 0.021 Model translates the claim to English first, in one chat: 0.780, gap 0.090 Paper prompt with RoSh: 0.855, gap 0.015

The English answer words are not the cause of the gap. Letting the model translate the claim first makes matters worse, because asked to translate a false claim the model sometimes corrects or comments on it instead of translating it.

We also added plain TRUE or FALSE accuracy (0.728 in English, 0.672 to 0.693 outside English, Table 21), confidence intervals from a re-run with per-claim scores (mean gain +0.036, interval +0.032 to +0.039, excluding zero on every model), and a weaker statement of the mechanism, which is now a working hypothesis that the results support but do not prove.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.