United States

This paper describes our submission to SemEval-2026 Task 8 on multi-turn retrieval-augmented generation (RAG). We propose a hybrid multi-stage pipeline that combines high-recall lexical retrieval, dual-embedding dense re-ranking with reciprocal rank fusion, LLM-based relevance judging, and strictly constrained evidence-grounded generation. Our design emphasizes robustness and faithfulness across the full retrieval-to-generation pipeline. Our results suggest that relevance-aware filtering and constrained generation are important for improving faithfulness and overall RAG performance.

ACL 2026

GenAIus at SemEval-2026 Task 8: Beyond Retrieval with Relevance-Aware RAG for Faithful Multi-Turn Generation

multi-turn question answering

retrieval-augmented generation (rag)

faithful generation

It is my great pleasure and honor to welcome you to the 64th Annual Meeting of the Association for Computational Linguistics (ACL 2026), held in San Diego, California, USA, July 2–7, 2026. ACL 2026 continues our fields tradition of excellence in scholarship, innovation, and inclusivity, and I am deeply grateful to the many volunteers who have worked tirelessly to bring this event to life.

It is exciting to see the many developments in the field reflected at our conference, such as the advancement of large language models (born out of our community) into many application directions. This is not without challenges, partly due to unsolved problems, and partly due to misuse of the technology. All this is reflected well at ACL 2026. The conference also reflects that the field is not only about advancing technology, but also the science of fundamentally understanding language and how it is use by many different speaker communities. <br>
[Full Message from General Chair Philipp Koehn](https://drive.google.com/file/d/1QiJFOmMjvDgsDo3i1XzlQnImxe3RU_Cy/view?usp=sharing) <br><br><br>
<html><button style="display: inline-flex; align-items: center; justify-content: center; white-space: nowrap; border-radius: 9999px; font-weight: bold; background: #7c3aed; color: white; font-family: 'Space Grotesk', sans-serif; height: 40px; font-size: 16px; padding: 0 20px; border: none; cursor: pointer" onclick="window.open('https://drive.google.com/file/d/1QiJFOmMjvDgsDo3i1XzlQnImxe3RU_Cy/view?usp=sharing')">Conference Handbook</button></html>
<br><br><br><br>

To access this event page, you need to log in with the **email address you registered with**. <br>Access credentials will be sent to your email from Underline -  subject line "Welcome to ACL 2026". Please be sure to check your spam email folder if you do not see an email confirmation right away.

To access this event page, you are required to register.
Please complete your registration to continue.

We recommend reading [**the registration information**](https://2026.aclweb.org/registration/) first.

**Online Registration Form**: https://acl.swoogo.com/acl2026

Registration Required

The 64th Annual Meeting of the Association for Computational Linguistics (ACL 2026) will take place in San Diego, California, from July 2 - 7, 2026.

workshop paper

We first use Qwen2.5-32B-Instruct model to generate hard negatives from three
narrative dimensions. We then train a Qwen3-Embedding-8B model with a multi-negative
contrastive objective and use self-distllation.

Team HITS at SemEval-2026 Task 4:Enhancing narrative text embedding model training with hard negatives generation and self-distillation

We describe our system for predicting sense plausibility in short narratives. Our approach centers on task decomposition: instead of predicting a score directly, we break the problem into simpler subtasks and combine their outputs. We further improve performance by ensembling complementary signals, including word sense disambiguation and fine-tuned embedding models. We also find empirical support for the one-homonym-per-translation principle of Hauer and Kondrak (2020a). Our best ensemble system achieves competitive performance in the official evaluation. Our code and data are available on GitHub.

UAlberta at SemEval-2026 Task 5: Disambiguating Stories via Task Decomposition

Our submission presented in this paper is for SemEval-2026 Task 9: Multilingual Text Classification Challenge - Polarization Detection and it covers all three subtasks: (1) binary polarization detection, (2) polarization type classification and (3) polarization manifestation identification. We adopt a systematic approach of research on short designed prompts by considering twelve designed prompts that are different in terminology clarity, detail of the definition, guidance of reasoning and in-context examples use. The experiments are conducted using aya-101 and Gemma3-27B, with the latter chosen for the submission at the end of the development through performance considerations. Our system has an average macro level \textbf{F1-score of 0.762 on Subtask 1, 0.587 on Subtask 2 and 0.444 on Subtask 3} with the average accuracy of 0.819, 0.678 and 0.498, respectively, on the official test set averaged among 22 languages, respectively. With cross-task and cross-lingual analysis, we demonstrate that prompt-based approaches can be used effectively to detect coarse-grained polarization but encounter more and more difficulties as far as fine-grained and multi-label sociolinguistic classification is concerned.

Lingo_Research_Group at SemEval-2026 Task 9: Evaluating Prompt Variants for Polarization Detection

We present a hybrid system for SemEval 2026
Task 5: Rating Plausibility of Word Senses in
Ambiguous Stories. Our approach reframes
LLMs as feature generators rather than direct
predictors. We combine two subsystems: one
that appends LLM-generated hints to the in-
put context and trains an encoder-based regres-
sion model, and another that feeds structured
hints from multiple LLM configurations into a
lightweight regression ensemble. We generate
multilingual enrichments to probe LLMs for
complementary signals and take advantage of
the fact that translation into certain languages
implicitly disambiguates word senses, making
the encoder more robust. The 50/50 ensem-
ble achieves 859/930 (92.37%) accuracy with
Spearman ρ= 0.8384 on the test set, exceed-
ing the estimated human ceiling of 89.2%. The
same LLM enrichments, processed through fun-
damentally different paradigms (tabular regres-
sion vs. full-text encoding), produce comple-
mentary errors that cancel under ensembling.
Notably, simple 50/50 averaging captures this
gain without any learned combiner, suggest-
ing that

Tifin India at SemEval-2026 Task 5: Semantic Bridge: Augmented Encoding for Word Sense Plausibility

Automatically determining whether source code is human written or produced by a specific family of large language models is becoming essential for reliable assessment, provenance tracking, and dataset curation. We present a lightweight yet competitive system for SemEval 2026 Task 13 Subtask B, which requires attributing each snippet to one of eleven classes: human or one of ten LLM families. Our method repurposes code oriented instruction tuned backbones from the Qwen2.5 Coder series as sequence classifiers and adapts them using QLoRA, combining frozen low precision weights with low rank trainable adapters to reduce memory and compute overhead. The core design choice addresses long snippets without losing evidence. Instead of truncating to a fixed context, we apply an overlapping sliding window strategy that expands long examples into multiple fixed length windows during training, all sharing the same label. For validation and test, windows are generated on the fly and their evidence is aggregated by averaging logits to yield a single prediction per snippet, enabling token complete use of the input while keeping inference stable. Our final submission ranked 8th on the official Subtask B test set leaderboard.

CodeDet-NITS at SemEval-2026 Task 13: AI Code Authorship Detection Beyond Truncation

This paper presents the NIT-Agartala-NLPTeam's submission to SemEval-2026 Task 9
on polarization detection in textual data. The
task comprises two subtasks: (i) binary classification to distinguish polarized from nonpolarized content, and (ii) multi-label classification to identify the specific type(s) of polarization. We propose a weighted soft-voting
ensemble framework that integrates multiple
fine-tuned large language models (LLMs). The
probabilistic outputs of the individual models are combined using weighted averaging
to effectively leverage their complementary
strengths and enhance overall performance.
Our system achieved a test macro F1-score of
78.6 (26th out of 44 teams) in Subtask 1 and
46.0 (18th out of 29 teams) in Subtask 2.

NIT-Agartala-NLP-Team at SemEval-2026 Task 9: A Weighted Soft-Voting Ensemble Framework of Fine-Tuned LLMs for Binary and Multi-Label Polarization Detection

We present our coarse-to-fine Valence-Arousal (VA) ensemble system for subtask 1 of task 3 (DimABSA) which covers aspect-level VA prediction. We use a pair of trained Qwen 3 8B LoRA-tuned LLMs to predict coarse bins between 1 and 8, providing ordinal VA guidance signals along with distributional features. We then train an instruction-style, multilingual E5 encoder model with a multitask head using these LLM-derived guidance features to produce continuous VA predictions. At inference time, the same guidance signals are generated for the test set by the trained LLMs and fed into the trained encoder. This approach leverages the LLM as a high-level prior while relying on the encoder for precise calibration across languages and domains. Our system achieves an RMSE_VA of 1.20 across six languages and five domains. We compare the joint VA model to separated valence and arousal models trained on coarsened ground truth data, showing that it outperforms them, particularly on arousal correlations.

RPI Team at SemEval-2026 Task 3: An LLM-Encoder Ensemble for Coarse-to-Fine Valence-Arousal Sentiment Prediction

We present a Span-Consistency Network (SCN) for conspiracy marker extraction in English social media text. The task requires identifying character-level spans for five marker types (Actor, Action, Effect, Evidence, and Victim) under overlap-based Macro F1 evaluation. Standard token-level classifiers often produce fragmented spans, ignore inter-marker dependencies, and struggle with severe class imbalance.

Our approach addresses these challenges through three components. First, a Span Consistency Layer (SCL) propagates span-level confidence signals to encourage coherent boundary formation. Second, Cross-Marker Attention (CMA) models co-occurrence patterns between marker types via a learned correlation matrix. Third, we introduce Span Count Regularization (SCR), a total-variation-based constraint that aligns soft token probabilities with the expected number of discrete spans, mitigating prediction collapse under threshold decoding.

Built on DeBERTa-v3-large and trained with a recall-biased Tversky loss, our system is ensembled across five stratified folds. It achieved a Macro F1 of 0.24 on the official test set, placing second among participating teams. Ablation studies show that SCR plays a critical role in maintaining span structure, particularly for low-frequency and long-span markers.

CredenceAI at SemEval-2026 Task 10: A Span-Consistency Network with Cross-Marker Attention for Conspiracy Marker Extraction

This paper presents our findings for SemEval-2026 Task 9. We submit to all three subtasks using an LLM-as-an-annotator strategy, simulating the data annotation process with large language models. We created 30 LLM annotators using persona injection (also known as sociodemographic prompting) and experimented with various annotation aggregation methods, including Dawid-Skene and MACE. To further increase the variability in annotator responses, we used the hatefulness detection task as proxy for identifying polarisation. Our findings indicate that this reframing of the problem is effective for the binary classification of polarisation, but is less effective for finer-grained polarisation detection. For subtasks 2 and 3, majority voting yielded the best overall performance. While our unsupervised approach does not rank as highly as supervised ones, this work provides insight into the utility of persona-based prompting and the issue of LLM annotators exhibiting high intra-model agreement.

ShefFriday at SemEval-2026 Task 9: LLM-Based Annotation Methods for Detecting Multilingual, Multicultural and Multievent Online Polarisation

Conspiracy believers use significantly more psycholinguistic markers per post than nonbelievers (Cohen's d = 0.56, p < 10⁻⁸⁰), a pattern we term narrative density, suggesting that belief manifests as structurally denser conspiratorial frames distributed across the full text rather than concentrated in specific lexical cues.

We present Truth Gradient's system for SemEval-2026 Task 10 Subtask 2 (Samory et al., 2026): a DeBERTaV3-large model with mean pooling and a 5-seed probability-averaging ensemble achieving macro F1 = 0.829 on the 77-sample development set and 0.75 on the official test set. The 5-fold CV estimate (0.734 ± 0.007) proves the more reliable predictor of test performance, and we recommend it as standard practice for low-resource shared tasks.

Two convergent tests support the narrative density account: masking annotated marker spans drops F1 by 5.3 pp, and direct marker-count fusion recovers +0.9 pp, though we note these are not conclusive given the small dev set. Cross-validated ablation identifies encoder fine-tuning as the dominant design factor (−7.2 pts), and layer-wise probing confirms belief information peaks at mid-stack layers (layer 16/24).

Downloads

Next from ACL 2026

Team HITS at SemEval-2026 Task 4:Enhancing narrative text embedding model training with hard negatives generation and self-distillation

Similar lecture

IIMAS-RAG at SemEval-2026 Task 8: Hybrid Sparse-Dense Retrieval and Answerability-Conditioned Generation for Multi-Turn RAG

.css-70qvj9{display:-webkit-box;display:-webkit-flex;display:-ms-flexbox;display:flex;-webkit-align-items:center;-webkit-box-align:center;-ms-flex-align:center;align-items:center;}Downloads

Next from ACL 2026

Team HITS at SemEval-2026 Task 4:Enhancing narrative text embedding model training with hard negatives generation and self-distillation

Similar lecture

IIMAS-RAG at SemEval-2026 Task 8: Hybrid Sparse-Dense Retrieval and Answerability-Conditioned Generation for Multi-Turn RAG

Downloads