HNR-DAC: Hard-Negative Reranking and Distribution-Aligned Classification for Scientific Claim Verification

# HNR-DAC: Hard-Negative Reranking and Distribution-Aligned Classification for Scientific Claim Verification
Source: [https://arxiv.org/html/2608.07204](https://arxiv.org/html/2608.07204)
11institutetext:Southern University of Science and Technology, Shenzhen, China
11email:12533253@mail\.sustech\.edu\.cn22institutetext:Institute of Artificial Intelligence, Shenzhen University of Advanced Technology, Shenzhen, China
22email:250701451@stu\.suat\-sz\.edu\.cn, \{zhangluoxi,nishiwen\}@suat\-sz\.edu\.cn33institutetext:Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences, Shenzhen, China
33email:min\.yang@siat\.ac\.cn###### Abstract

Scientific claim verification over a cited paper requires predicting the claim–paper relation and identifying the paragraphs that justify that prediction\. This setting poses two linked challenges: within\-paper distractors often resemble genuine evidence, while a classifier trained on gold evidence must operate on retrieved evidence at inference\. We present HNR\-DAC, a two\-stage framework that trains each stage on the cases it will actually encounter\. Hard\-Negative Reranking \(HNR\) quantifies evidence confusability using a base reranker’s scores on non\-gold paragraphs and contrasts gold evidence against the most confusable candidates\. Distribution\-Aligned Classification \(DAC\) trains on the Top\-1 paragraph produced by the same frozen HNR used to construct inference inputs, while HNR’s Top\-3 paragraph identifiers provide the evidence output\. On the NLPCC 2026 Task 10 Track 2, the final configuration obtains 97\.21% Hit@3, 95\.79% Macro\-F1, 94\.47% Joint@3, and an average score of 95\.13%\. The corresponding submission ranks third on the official Track 2 leaderboard while achieving the highest overall Macro\-F1 of 93\.05%, alongside 70\.16% Joint@3 and an average score of 81\.61%\.

## 1Introduction

Large language models \(LLMs\) can generate literature\-review drafts and citation\-supported scientific text\[[1](https://arxiv.org/html/2608.07204#bib.bib23),[2](https://arxiv.org/html/2608.07204#bib.bib12),[16](https://arxiv.org/html/2608.07204#bib.bib24)\], but their claims may remain unsupported by or inaccurately represent the cited sources\[[15](https://arxiv.org/html/2608.07204#bib.bib4),[17](https://arxiv.org/html/2608.07204#bib.bib6)\]\. The reliability of AI\-assisted scientific reporting therefore depends not only on fluent generation, but also on whether a claim remains faithful to the evidence it cites\. NLPCC 2026 Shared Task 10 is designed around this problem\[[5](https://arxiv.org/html/2608.07204#bib.bib1)\]\. This paper focuses on Track 2, where each instance provides an atomic AI\-generated claim and the structured full text of its cited paper\. A system must determine whether the paper supports, overstates, is only topically related to, or is irrelevant to the claim, while also returning ranked evidence paragraph identifiers\.

The task follows the retrieve\-and\-verify paradigm established by FEVER and SciFact\[[11](https://arxiv.org/html/2608.07204#bib.bib7),[12](https://arxiv.org/html/2608.07204#bib.bib8)\], but considers a more constrained retrieval setting: the cited paper is already fixed, evidence must be identified at paragraph level, and the relation label depends on whether the selected paragraphs support the exact scope of the claim\.

This fixed\-paper setting creates two stage\-specific difficulties\. For evidence reranking, topical overlap is pervasive within the cited paper: many non\-gold paragraphs share the claim’s entities, terminology, and experimental setting\. The most informative negatives are therefore not arbitrary non\-gold paragraphs, but those that the current reranker itself treats as plausible evidence\. We refer to this claim\-specific and model\-conditioned tendency as evidence confusability\. For relation classification, training only on gold evidence exposes the classifier to cleaner contexts than the automatically retrieved paragraphs available at inference\. The reranker therefore requires evidence\-confusable within\-paper negatives, whereas the classifier requires contexts produced by the retrieval pipeline it will encounter at inference\.

We propose HNR\-DAC, a two\-stage framework that trains each stage on the cases it will actually encounter\. Hard\-Negative Reranking \(HNR\) operationalizes evidence confusability using the scores assigned by a base reranker to non\-gold paragraphs\. It then trains a cross\-encoder to distinguish gold evidence paragraphs from the most confusable candidates using a group\-wise contrastive objective\. Distribution\-Aligned Classification \(DAC\) freezes the trained HNR and fine\-tunes a four\-way classifier on its top\-ranked paragraph, aligning the retrieval process used to construct classifier inputs across training and inference\. This construction applies to all four classes, including non\-empty Top\-1 inputs forIrrelevantclaims\. At inference, the DAC relation label is paired with HNR’s Top\-3 paragraph identifiers for evidence output\. On the development set, the final configuration reaches 97\.21% Hit@3, 95\.79% Macro\-F1, 94\.47% Joint@3, and a 95\.13% Average Score, and the corresponding submission ranks third on the official Track 2 leaderboard\[[4](https://arxiv.org/html/2608.07204#bib.bib2)\]\.

The main contributions of this work are as follows:

- •We introduce HNR, which quantifies evidence confusability using base\-reranker scores and trains against the most confusable non\-gold paragraphs within each cited paper\.
- •We present DAC, which trains the relation classifier on the Top\-1 paragraph produced by the same frozen HNR used at inference, while retaining the Top\-3 paragraph identifiers for evidence output\.
- •We demonstrate that confusability\-guided negatives consistently outperform random negatives and that HNR Top\-1 provides the strongest classifier context; the full system ranks third on the official leaderboard with the highest overall Macro\-F1\.

## 2Related Work

### 2\.1Citation Faithfulness and Scientific Claim Verification

Citation faithfulness asks whether a generated statement can be traced to an identified source and whether that source supports the statement\[[2](https://arxiv.org/html/2608.07204#bib.bib12),[10](https://arxiv.org/html/2608.07204#bib.bib11)\], while fine\-grained citation evaluation shows that support is not purely binary\[[20](https://arxiv.org/html/2608.07204#bib.bib13)\]\. Thorne et al\. introduced FEVER for evidence\-based fact verification over textual sources, while Wadden et al\. introduced SciFact for scientific claim verification and SciFact\-Open for the open\-domain setting\[[11](https://arxiv.org/html/2608.07204#bib.bib7),[12](https://arxiv.org/html/2608.07204#bib.bib8),[13](https://arxiv.org/html/2608.07204#bib.bib17)\]\. Unlike open\-domain verification, NLPCC 2026 Task 10 Track 2 fixes the cited paper and requires both a four\-way relation label and ranked evidence paragraph identifiers\[[5](https://arxiv.org/html/2608.07204#bib.bib1)\]\. The main challenge therefore shifts from document retrieval to distinguishing topically related paragraphs from those that support the comparison or scope expressed by the claim\.

### 2\.2Evidence Reranking and Retrieval\-Conditioned Classification

Evidence reranking and relation classification are commonly combined in scientific claim verification\. Karpukhin et al\. proposed DPR, while Nogueira et al\. used BERT and T5 for neural reranking\[[3](https://arxiv.org/html/2608.07204#bib.bib9),[6](https://arxiv.org/html/2608.07204#bib.bib18),[7](https://arxiv.org/html/2608.07204#bib.bib10)\]\. Hard\-negative training methods, including ANCE and RocketQA, use high\-scoring incorrect candidates to improve retrieval\[[18](https://arxiv.org/html/2608.07204#bib.bib19),[9](https://arxiv.org/html/2608.07204#bib.bib20),[19](https://arxiv.org/html/2608.07204#bib.bib21)\]\. For scientific claim verification, VerT5erini and MultiVerS connect evidence selection and label prediction through a pipeline and joint prediction, respectively\[[8](https://arxiv.org/html/2608.07204#bib.bib15),[14](https://arxiv.org/html/2608.07204#bib.bib16)\]\. In the fixed\-paper setting, HNR mines high\-scoring non\-gold paragraphs as hard negatives to improve within\-paper evidence discrimination\. DAC constructs classifier training inputs from the outputs of the same frozen HNR used at inference, reducing the discrepancy between gold evidence at training and retrieved evidence at inference\.

## 3Methodology

![Refer to caption](https://arxiv.org/html/2608.07204v1/x1.png)Figure 1:Overview of HNR\-DAC\. HNR mines evidence\-confusable non\-gold paragraphs for reranker training; its top\-ranked paragraph conditions DAC, while the top\-three paragraph identifiers bypass DAC for the final evidence output\.### 3\.1Framework Overview

Track 2 considers claim verification over a fixed cited paper\. Given an atomic claimqqand a paper segmented into paragraphsP=\{pi\}i=1nP=\\\{p\_\{i\}\\\}\_\{i=1\}^\{n\}, the system must predict a relation label

y∈𝒴=\{Supported,Irrelevant,Topical Match,Overstate\},y\\in\\mathcal\{Y\}=\\\{\\textsc\{Supported\},\\textsc\{Irrelevant\},\\textsc\{Topical Match\},\\textsc\{Overstate\}\\\},together with an ordered list of up to three evidence paragraph identifiers\. In the training data, each non\-Irrelevant claim is paired with one or more human\-annotated gold evidence paragraphs, whereas Irrelevant claims have no gold evidence annotations\. These annotations are used only to construct training supervision and are unavailable at inference\.

Although relation prediction and evidence identification are evaluated jointly, the two stages face different learning conditions\. Evidence reranking must distinguish annotated evidence from evidence\-confusable distractors within the same paper, whereas relation classification must judge the fine\-grained claim–paper relation from automatically retrieved context\. Moreover, the context most useful for classification need not contain the same number of paragraphs required for evidence output\. These differences motivate stage\-specific training connected through an explicit ranked\-evidence interface\.

As illustrated in Figure[1](https://arxiv.org/html/2608.07204#S3.F1), HNR\-DAC consists of two specialized stages\. HNR assigns a relevance score to each claim–paragraph pair and produces an ordered paragraph ranking\. DAC predicts the relation label from the claim and HNR’s highest\-ranked paragraph,

y^=fDAC​\(q,E^1​\(q\)\),\\widehat\{y\}=f\_\{\\mathrm\{DAC\}\}\\bigl\(q,\\widehat\{E\}\_\{1\}\(q\)\\bigr\),\(1\)whereE^1​\(q\)\\widehat\{E\}\_\{1\}\(q\)denotes the top\-ranked paragraph produced by HNR\. In parallel, the top\-three ranked paragraphsE^3​\(q\)\\widehat\{E\}\_\{3\}\(q\)bypass DAC and are returned as the evidence output; following the task protocol, the evidence list is set to empty when the predicted label isIrrelevant\.

### 3\.2Hard\-Negative Reranking via Evidence Confusability

Paragraphs within the same scientific paper often share entities, terminology, methods, and experimental settings\. Consequently, a paragraph may appear highly relevant to a claim while failing to support its specific finding, comparison, or scope\. We refer to the tendency of such a non\-gold paragraph to be ranked as genuine evidence by the current reranker as*evidence confusability*\.

For a training claimqq, let𝒢​\(q\)⊆P\\mathcal\{G\}\(q\)\\subseteq Pdenote its set of gold evidence paragraphs\. A base reranker assigns a relevance scores0​\(q,p\)s\_\{0\}\(q,p\)to each non\-empty paragraph\. For each non\-gold paragraphp∈P∖𝒢​\(q\)p\\in P\\setminus\\mathcal\{G\}\(q\), we use this score as a claim\-specific and model\-dependent proxy for evidence confusability, definingc0​\(q,p\)=s0​\(q,p\)c\_\{0\}\(q,p\)=s\_\{0\}\(q,p\)\. A largerc0​\(q,p\)c\_\{0\}\(q,p\)indicates that the base reranker ranksppas a stronger competitor to gold evidence\. HNR uses only the within\-paper ordering of this proxy and selects

𝒩K​\(q\)=TopKp∈P∖𝒢​\(q\)c0​\(q,p\)\.\\mathcal\{N\}\_\{K\}\(q\)=\\operatorname\*\{TopK\}\_\{p\\in P\\setminus\\mathcal\{G\}\(q\)\}c\_\{0\}\(q,p\)\.\(2\)
For each gold evidence paragraphp\+∈𝒢​\(q\)p^\{\+\}\\in\\mathcal\{G\}\(q\), we form a training group containingp\+p^\{\+\}and theKKmined hard negatives\{pj−\}j=1K\\\{p\_\{j\}^\{\-\}\\\}\_\{j=1\}^\{K\}from𝒩K​\(q\)\\mathcal\{N\}\_\{K\}\(q\)\. The reranker is optimized with a group\-wise contrastive objective:

ℒHNR=−log⁡exp⁡sθ​\(q,p\+\)exp⁡sθ​\(q,p\+\)\+∑j=1Kexp⁡sθ​\(q,pj−\),\\mathcal\{L\}\_\{\\mathrm\{HNR\}\}=\-\\log\\frac\{\\exp s\_\{\\theta\}\(q,p^\{\+\}\)\}\{\\exp s\_\{\\theta\}\(q,p^\{\+\}\)\+\\sum\_\{j=1\}^\{K\}\\exp s\_\{\\theta\}\(q,p\_\{j\}^\{\-\}\)\},\(3\)wheresθ​\(q,p\)s\_\{\\theta\}\(q,p\)denotes the relevance score assigned by the fine\-tuned reranker\. This objective encourages gold evidence to outrank the non\-gold paragraphs ranked highest by the base reranker\. When multiple gold evidence paragraphs are available, each forms a separate training group\.Irrelevantinstances are excluded from HNR training because they contain no gold evidence\.

At inference, HNR scores all non\-empty paragraphs and ranks them in descending order, yieldingE^1​\(q\)\\widehat\{E\}\_\{1\}\(q\)for DAC andE^3​\(q\)\\widehat\{E\}\_\{3\}\(q\)for evidence output\.

### 3\.3Distribution\-Aligned Classification

DAC trains the relation classifier on evidence produced by the frozen HNR\. We describe its retrieval\-conditioned training construction and label\-aware resampling strategy below\.

#### 3\.3\.1Retrieval\-Conditioned Classifier Training

Relation classification is performed on automatically retrieved evidence at inference\. A classifier trained only on gold paragraphs, however, observes oracle contexts that are cleaner than those available in the complete pipeline\. We therefore construct its training inputs from the outputs of the frozen HNR, aligning the retrieval process used to form classifier contexts across training and inference\.

Specifically, after HNR training, we freeze the reranker and apply it to the cited paper of every training claim\. LetE^1​\(qi\)\\widehat\{E\}\_\{1\}\(q\_\{i\}\)denote the highest\-ranked paragraph for claimqiq\_\{i\}\. DAC is trained on the retrieval\-conditioned input

xi=\[qi;E^1​\(qi\)\],x\_\{i\}=\[q\_\{i\};\\widehat\{E\}\_\{1\}\(q\_\{i\}\)\],with the corresponding relation labelyiy\_\{i\}\. This construction is applied to all four classes\. In particular, Irrelevant instances also receive a non\-empty HNR top paragraph, preventing the classifier from using the absence of evidence as a trivial label cue, an input pattern that never occurs at inference, where every claim yields a non\-empty top\-ranked paragraph\.

DAC is optimized through supervised fine\-tuning with standard token\-level negative log\-likelihood over the target relation label\. Let𝒟′\\mathcal\{D\}^\{\\prime\}denote the resampled training set\. The model is instructed to generate exactly one of the four relation labels, while evidence identifiers remain determined by HNR rather than generated by DAC\.

At inference, DAC receives the same input structure,\[q;E^1​\(q\)\]\[q;\\widehat\{E\}\_\{1\}\(q\)\], and predicts the claim–paper relation\. Here, distribution alignment refers to matching the retrieval process used to construct classifier inputs across training and inference, rather than assuming that their empirical distributions are identical\.

#### 3\.3\.2Label\-Aware Resampling

The four\-way classification data are imbalanced: as shown in Table[2](https://arxiv.org/html/2608.07204#S4.T2),Supportedaccounts for nearly half of the original training set, whereas each remaining class contributes approximately 17%\. We therefore apply sampling multipliers of11,22,33, and33toSupported,Irrelevant,Topical Match, andOverstate, respectively\. This resampling changes only the frequency of existing training instances; the development and test distributions remain unchanged\.

## 4Experimental Results

### 4\.1Experimental Setup

#### 4\.1\.1Dataset\.

We use the official training and development splits of NLPCC 2026 Task 10 Track 2\. The training set contains 1,945 instances, including 926 Supported, 344 Irrelevant, 338 Topical Match, and 337 Overstate examples\. The development set contains 217 instances, with 103 Supported examples and 38 examples for each of the other three classes\. Retrieval metrics are evaluated on the 179 development instances with non\-empty gold evidence, while classification and joint metrics are computed on all 217 instances\.

#### 4\.1\.2Metrics\.

The official evaluation uses Macro\-F1 and Joint@3\. Macro\-F1 is the unweighted average of the F1 scores over the four relation classes\. Joint@3 counts an instance as correct only when the relation label is correct and at least one gold evidence paragraph appears in the predicted top three\. For Irrelevant instances, the predicted evidence list must be empty\. The overall Score is the average of Macro\-F1 and Joint@3\.

For evidence\-reranking analysis, we additionally report Hit@kkand mean reciprocal rank \(MRR\)\. Hit@kkmeasures whether at least one gold paragraph occurs within the topkkpositions, while MRR is computed from the rank of the first retrieved gold paragraph\.

#### 4\.1\.3Implementation Details\.

HNR uses Qwen3\-Reranker\-8B and is trained with 15 mined hard negatives per positive paragraph\. The trained reranker is then frozen to construct DAC inputs and to produce the final top\-three evidence ranking\. Unless otherwise specified, DAC uses Qwen3\.5\-27B with LoRA fine\-tuning on HNR Top\-1 paragraphs\. Class resampling expands the classifier training set to 3,639 instances, and training uses seed 42, 2,275 optimization steps, and an effective batch size of 8\. For the classifier scaling analysis, Qwen3\.5\-0\.8B, 2B, 9B, and 27B are trained under the same data, optimization, and evaluation settings\.

Table 1:Development\-set comparison with baseline systems and component variants of HNR\-DAC\. Vanilla denotes the standard two\-stage pipeline that uses the original model without hard\-negative adaptation and trains the relation classifier on gold evidence\. Vanilla \+ HNR replaces the base reranker with HNR while retaining gold\-evidence classifier training\. Vanilla \+ DAC retains the base reranker but constructs classifier training inputs from its Top\-1 outputs\.MethodHNRDACClass\-wise F1OverallSup\.Over\.Top\.Irr\.Macro\-F1Joint@3ScoreBaselinesDeepSeek\-V4\-Pro––67\.6558\.956\.4565\.7549\.7042\.8646\.28GPT\-4o\-mini––66\.6726\.093\.575\.1325\.3735\.0230\.19GPT\-5\.4––61\.6253\.330\.0051\.7241\.6732\.7237\.19OursVanilla––97\.1294\.5974\.0053\.8579\.8982\.0380\.96Vanilla \+ HNR✓\\checkmark–95\.8194\.5984\.7173\.3387\.1188\.0287\.57Vanilla \+ DAC–✓\\checkmark99\.0496\.0093\.3397\.3796\.4493\.0994\.76HNR\-DAC✓\\checkmark✓\\checkmark99\.0496\.0092\.1196\.0095\.7994\.4795\.13

### 4\.2Main Results and Component Comparison

Table[1](https://arxiv.org/html/2608.07204#S4.T1)compares HNR\-DAC with general\-purpose LLM baselines and controlled component variants\. The prompted LLM baselines perform poorly, particularly onTopical Match, indicating that general semantic reasoning alone is insufficient for fine\-grained evidence verification\. The Vanilla pipeline achieves a Score of 80\.96\. Adding HNR improves the Score to 87\.57, with notable gains onTopical MatchandIrrelevant, showing that hard\-negative training improves discrimination among confusable within\-paper paragraphs\.

DAC provides the largest improvement, increasing Macro\-F1 from 79\.89 to 96\.44 and Score to 94\.76\. This supports training the classifier on retrieved rather than gold evidence\. The full HNR\-DAC system achieves the highest Joint@3 of 94\.47 and the highest overall Score of 95\.13\. Although its Macro\-F1 is slightly lower than that of Vanilla \+ DAC, the improvement in Joint@3 shows that HNR contributes stronger evidence identification\. Overall, DAC primarily improves relation classification, while HNR complements it by improving joint label–evidence correctness\.

### 4\.3HNR Analysis

![Refer to caption](https://arxiv.org/html/2608.07204v1/x2.png)Figure 2:HNR analysis on the development set\. We compare Random\-Neg and HNR under different numbers of negatives per positive paragraph\.#### 4\.3\.1Number of Hard Negatives\.

We first examine the number of mined hard negatives included with each positive paragraph\. Figure[2](https://arxiv.org/html/2608.07204#S4.F2)reports the results forK∈\{5,10,15\}K\\in\\\{5,10,15\\\}\. IncreasingKKdoes not produce a monotonic improvement across all retrieval metrics\. HNR withK=5K=5achieves the highest Hit@3, whereasK=15K=15gives the strongest Hit@1 and MRR\. Since DAC consumes only the highest\-ranked paragraph, we selectK=15K=15for the final system, prioritizing evidence quality at the first rank rather than top\-three coverage alone\.

#### 4\.3\.2Hard versus Random Negatives\.

We next compare HNR and randomly sampled negatives under the same values ofKK\. Random\-Neg uniformly samples non\-gold paragraphs from the cited paper, whereas HNR selects high\-scoring non\-gold paragraphs produced by the base reranker\. HNR consistently improves Hit@1 and MRR over Random\-Neg for all evaluated values ofKK, showing that high\-scoring within\-paper distractors provide more informative supervision than uniformly sampled paragraphs\.

The advantage is most pronounced atK=5K=5, where HNR improves Hit@1 by 26\.26 points and MRR by 21\.11 points\. AtK=15K=15, it improves Hit@1, Hit@3, and MRR by 30\.17, 13\.41, and 23\.33 points, respectively\. These results support hard\-negative mining as the primary source of HNR’s early\-rank improvement\.

### 4\.4DAC Analysis

![Refer to caption](https://arxiv.org/html/2608.07204v1/x3.png)Figure 3:DAC analysis on the development set\. \(a\) Effect of the number of HNR paragraphs supplied to DAC\. \(b\) Effect of DAC backbone scale\.We analyze three design choices of DAC: evidence width, classifier scale, and label\-aware resampling\. All configurations use the sameK=15K=15HNR and the same top\-three evidence output for Joint@3, so the reported differences arise from relation classification\.

#### 4\.4\.1Evidence Width\.

We vary the number of rank\-ordered HNR paragraphs supplied to DAC, using the same evidence width during training and inference\. As shown in Figure[3](https://arxiv.org/html/2608.07204#S4.F3)\(a\), Top\-1 achieves the highest Macro\-F1, Joint@3, and Score\. Expanding the input to Top\-2 or Top\-3 reduces Score by 0\.47 and 0\.35 points, respectively, suggesting that lower\-ranked paragraphs provide no additional benefit under the evaluated setting\. We therefore use HNR Top\-1 as the DAC input while retaining the top\-three paragraph identifiers for evidence output\.

#### 4\.4\.2Effect of Model Scale\.

We further examine the effect of classifier capacity by varying the DAC backbone from 0\.8B to 27B parameters\. As shown in Figure[3](https://arxiv.org/html/2608.07204#S4.F3)\(b\), performance generally improves with classifier scale\. The 0\.8B and 2B variants obtain comparable results, while increasing the backbone to 9B yields clearer gains in both Macro\-F1 and Joint@3\. The 27B model performs best, exceeding the 9B variant by 2\.09 Macro\-F1 points and 1\.97 Score points\. These results indicate that DAC benefits from greater classifier capacity, although the smaller variants remain competitive under the same retrieved\-evidence interface\.

#### 4\.4\.3Label\-Aware Resampling\.

Table 2:Training label distributions and corresponding DAC performance before and after label\-aware resampling\. Distribution cells report instance counts with percentages in parentheses\. Score is the average of Macro\-F1 and Joint@3\.Sup\., Over\., Top\., and Irr\. denote Supported, Overstate, Topical Match, and Irrelevant, respectively\.

The original DAC training set is dominated bySupported, which accounts for 47\.60% of the instances, whereas each of the remaining classes contributes approximately 17%\. Since Macro\-F1 assigns equal importance to all relation classes, we apply label\-aware resampling to increase the training exposure of the minority labels\.

As shown in Table[2](https://arxiv.org/html/2608.07204#S4.T2), however, resampling does not improve aggregate development\-set performance\. The original distribution obtains a Macro\-F1 of 95\.82 and a Score of 95\.15, compared with 95\.79 and 95\.13 after resampling, while Joint@3 remains unchanged\. The near\-identical results suggest that the 27B DAC backbone is already able to fit the observed class imbalance on the development set\. Accordingly, we do not regard resampling as a primary source of the development\-set improvement\.

### 4\.5Official Results

Table[3](https://arxiv.org/html/2608.07204#S4.T3)reports the official Track 2 results on the hidden test sets\. Our submission ranks third overall, achieving an Average Score of 81\.61, a Macro\-F1 of 93\.05, and a Joint@3 of 70\.16\. Notably, HNR\-DAC obtains the highest Macro\-F1 in both evaluation phases and overall, demonstrating strong generalization in fine\-grained relation classification\.

Compared with the development results, Macro\-F1 decreases only moderately, whereas Joint@3 exhibits a substantially larger drop\. This gap suggests that relation classification transfers more robustly than evidence ranking to the hidden test distribution, and that identifying annotated evidence across heterogeneous paper structures remains the primary bottleneck\. Label\-aware resampling was retained in the submitted system to increase exposure to minority relations; however, because no official no\-resampling result is available, its independent contribution to test\-set generalization cannot be isolated\.

Table 3:Official Track 2 top\-three results \(%\)\. Each block reports Score \(Average Score\), Macro\-F1, and Joint@3; column bests are bold, second\-best underlined\.

## 5Conclusion

We presented HNR\-DAC, a two\-stage framework for scientific claim verification over cited papers\. HNR improves evidence ranking through within\-paper hard negatives, while DAC trains the classifier on HNR’s top\-ranked paragraph to better match inference conditions\. Experiments show that mined negatives outperform random negatives, Top\-1 evidence provides the strongest classifier input, and larger classifier backbones further improve performance\. The final system achieves a 95\.13% development Score and ranks third in the official evaluation, with the highest overall Macro\-F1\. Nevertheless, the official Joint@3 results indicate that evidence\-ranking generalization remains more challenging than relation classification\.

## References

- \[1\]X\. Chen, J\. Wu, S\. Yang, R\. Zhan, Z\. Wu, Z\. Luo, D\. Wang, M\. Yang, L\. S\. Chao, and D\. F\. Wong\(2025\)RepreGuard: detecting LLM\-generated text by revealing hidden representation patterns\.Transactions of the Association for Computational Linguistics13,pp\. 1812–1831\.External Links:[Link](https://aclanthology.org/2025.tacl-1.81/),[Document](https://dx.doi.org/10.1162/tacl.a.61)Cited by:[§1](https://arxiv.org/html/2608.07204#S1.p1.1)\.
- \[2\]T\. Gao, H\. Yen, J\. Yu, and D\. Chen\(2023\)Enabling large language models to generate text with citations\.InProceedings of the 2023 Conference on Empirical Methods in Natural Language Processing,pp\. 6465–6488\.External Links:[Document](https://dx.doi.org/10.18653/v1/2023.emnlp-main.398),[Link](https://aclanthology.org/2023.emnlp-main.398/)Cited by:[§1](https://arxiv.org/html/2608.07204#S1.p1.1),[§2\.1](https://arxiv.org/html/2608.07204#S2.SS1.p1.1)\.
- \[3\]V\. Karpukhin, B\. Oguz, S\. Min, P\. Lewis, L\. Wu, S\. Edunov, D\. Chen, and W\. Yih\(2020\)Dense passage retrieval for open\-domain question answering\.InProceedings of the 2020 Conference on Empirical Methods in Natural Language Processing,pp\. 6769–6781\.External Links:[Document](https://dx.doi.org/10.18653/v1/2020.emnlp-main.550)Cited by:[§2\.2](https://arxiv.org/html/2608.07204#S2.SS2.p1.1)\.
- \[4\]NLP2CT Lab\(2026\)NLPCC 2026 shared task 10 track 2 leaderboard\.Note:https://nlp2ct\.github\.io/NLPCC\-2026\-Task10\-Science/leaderboard\.htmlAccessed: 2026\-07\-14Cited by:[§1](https://arxiv.org/html/2608.07204#S1.p4.1)\.
- \[5\]NLP2CT Lab\(2026\)NLPCC 2026 shared task 10: reliability of ai\-assisted scientific reporting\.Note:https://nlp2ct\.github\.io/NLPCC\-2026\-Task10\-Science/Accessed: 2026\-06\-27Cited by:[§1](https://arxiv.org/html/2608.07204#S1.p1.1),[§2\.1](https://arxiv.org/html/2608.07204#S2.SS1.p1.1)\.
- \[6\]R\. Nogueira and K\. Cho\(2019\)Passage re\-ranking with BERT\.External Links:1901\.04085,[Document](https://dx.doi.org/10.48550/arXiv.1901.04085),[Link](https://arxiv.org/abs/1901.04085)Cited by:[§2\.2](https://arxiv.org/html/2608.07204#S2.SS2.p1.1)\.
- \[7\]R\. Nogueira, Z\. Jiang, R\. Pradeep, and J\. Lin\(2020\-11\)Document ranking with a pretrained sequence\-to\-sequence model\.InFindings of the Association for Computational Linguistics: EMNLP 2020,T\. Cohn, Y\. He, and Y\. Liu \(Eds\.\),Online,pp\. 708–718\.External Links:[Link](https://aclanthology.org/2020.findings-emnlp.63/),[Document](https://dx.doi.org/10.18653/v1/2020.findings-emnlp.63)Cited by:[§2\.2](https://arxiv.org/html/2608.07204#S2.SS2.p1.1)\.
- \[8\]R\. Pradeep, X\. Ma, R\. Nogueira, and J\. Lin\(2021\)Scientific claim verification with VerT5erini\.InProceedings of the 12th International Workshop on Health Text Mining and Information Analysis,pp\. 94–103\.External Links:[Link](https://aclanthology.org/2021.louhi-1.11/)Cited by:[§2\.2](https://arxiv.org/html/2608.07204#S2.SS2.p1.1)\.
- \[9\]Y\. Qu, Y\. Ding, J\. Liu, K\. Liu, R\. Ren, W\. X\. Zhao, D\. Dong, H\. Wu, and H\. Wang\(2021\)RocketQA: an optimized training approach to dense passage retrieval for open\-domain question answering\.InProceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies,pp\. 5835–5847\.External Links:[Document](https://dx.doi.org/10.18653/v1/2021.naacl-main.466),[Link](https://aclanthology.org/2021.naacl-main.466/)Cited by:[§2\.2](https://arxiv.org/html/2608.07204#S2.SS2.p1.1)\.
- \[10\]H\. Rashkin, V\. Nikolaev, M\. Lamm, L\. Aroyo, M\. Collins, D\. Das, S\. Petrov, G\. S\. Tomar, I\. Turc, and D\. Reitter\(2023\)Measuring attribution in natural language generation models\.Computational Linguistics49\(4\),pp\. 777–840\.External Links:[Document](https://dx.doi.org/10.1162/coli%5Fa%5F00486),[Link](https://aclanthology.org/2023.cl-4.2)Cited by:[§2\.1](https://arxiv.org/html/2608.07204#S2.SS1.p1.1)\.
- \[11\]J\. Thorne, A\. Vlachos, C\. Christodoulopoulos, and A\. Mittal\(2018\)FEVER: a large\-scale dataset for fact extraction and verification\.InProceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies,pp\. 809–819\.External Links:[Document](https://dx.doi.org/10.18653/v1/N18-1074)Cited by:[§1](https://arxiv.org/html/2608.07204#S1.p2.1),[§2\.1](https://arxiv.org/html/2608.07204#S2.SS1.p1.1)\.
- \[12\]D\. Wadden, S\. Lin, K\. Lo, L\. L\. Wang, M\. van Zuylen, A\. Cohan, and H\. Hajishirzi\(2020\)Fact or fiction: verifying scientific claims\.InProceedings of the 2020 Conference on Empirical Methods in Natural Language Processing,pp\. 7534–7550\.External Links:[Document](https://dx.doi.org/10.18653/v1/2020.emnlp-main.609)Cited by:[§1](https://arxiv.org/html/2608.07204#S1.p2.1),[§2\.1](https://arxiv.org/html/2608.07204#S2.SS1.p1.1)\.
- \[13\]D\. Wadden, K\. Lo, B\. Kuehl, A\. Cohan, I\. Beltagy, L\. L\. Wang, and H\. Hajishirzi\(2022\)SciFact\-open: towards open\-domain scientific claim verification\.InFindings of the Association for Computational Linguistics: EMNLP 2022,pp\. 4719–4734\.External Links:[Document](https://dx.doi.org/10.18653/v1/2022.findings-emnlp.347),[Link](https://aclanthology.org/2022.findings-emnlp.347/)Cited by:[§2\.1](https://arxiv.org/html/2608.07204#S2.SS1.p1.1)\.
- \[14\]D\. Wadden, K\. Lo, L\. L\. Wang, A\. Cohan, I\. Beltagy, and H\. Hajishirzi\(2022\)MultiVerS: improving scientific claim verification with weak supervision and full\-document context\.InFindings of the Association for Computational Linguistics: NAACL 2022,pp\. 61–76\.External Links:[Document](https://dx.doi.org/10.18653/v1/2022.findings-naacl.6),[Link](https://aclanthology.org/2022.findings-naacl.6/)Cited by:[§2\.2](https://arxiv.org/html/2608.07204#S2.SS2.p1.1)\.
- \[15\]W\. H\. Walters and E\. I\. Wilder\(2023\)Fabrication and errors in the bibliographic citations generated by ChatGPT\.Scientific Reports13,pp\. 14045\.External Links:[Document](https://dx.doi.org/10.1038/s41598-023-41032-5)Cited by:[§1](https://arxiv.org/html/2608.07204#S1.p1.1)\.
- \[16\]J\. Wu, S\. Yang, R\. Zhan, Y\. Yuan, L\. S\. Chao, and D\. F\. Wong\(2025\-03\)A survey on LLM\-generated text detection: necessity, methods, and future directions\.Computational Linguistics51\(1\),pp\. 275–338\.External Links:[Link](https://aclanthology.org/2025.cl-1.8/),[Document](https://dx.doi.org/10.1162/coli%5Fa%5F00549)Cited by:[§1](https://arxiv.org/html/2608.07204#S1.p1.1)\.
- \[17\]K\. Wu, E\. Wu, K\. Wei, A\. Zhang, A\. Casasola, T\. Nguyen, S\. Riantawan, P\. Shi, D\. Ho, and J\. Zou\(2025\)An automated framework for assessing how well LLMs cite relevant medical references\.Nature Communications16,pp\. 3615\.External Links:[Document](https://dx.doi.org/10.1038/s41467-025-58551-6)Cited by:[§1](https://arxiv.org/html/2608.07204#S1.p1.1)\.
- \[18\]L\. Xiong, C\. Xiong, Y\. Li, K\. Tang, J\. Liu, P\. N\. Bennett, J\. Ahmed, and A\. Overwijk\(2021\)Approximate nearest neighbor negative contrastive learning for dense text retrieval\.InInternational Conference on Learning Representations,External Links:[Link](https://openreview.net/forum?id=zeFrfgyZln)Cited by:[§2\.2](https://arxiv.org/html/2608.07204#S2.SS2.p1.1)\.
- \[19\]J\. Zhan, J\. Mao, Y\. Liu, J\. Guo, M\. Zhang, and S\. Ma\(2021\)Optimizing dense retrieval model training with hard negatives\.InProceedings of the 44th International ACM SIGIR Conference on Research and Development in Information Retrieval,pp\. 1503–1512\.External Links:[Document](https://dx.doi.org/10.1145/3404835.3462880),[Link](https://dl.acm.org/doi/10.1145/3404835.3462880)Cited by:[§2\.2](https://arxiv.org/html/2608.07204#S2.SS2.p1.1)\.
- \[20\]W\. Zhang, M\. Aliannejadi, Y\. Yuan, J\. Pei, J\. Huang, and E\. Kanoulas\(2024\)Towards fine\-grained citation evaluation in generated text: a comparative analysis of faithfulness metrics\.InProceedings of the 17th International Natural Language Generation Conference,pp\. 427–439\.External Links:[Document](https://dx.doi.org/10.18653/v1/2024.inlg-main.35),[Link](https://aclanthology.org/2024.inlg-main.35/)Cited by:[§2\.1](https://arxiv.org/html/2608.07204#S2.SS1.p1.1)\.