White-box method ReXTrust achieves highest AUC (peak 93.0) on Gut-VLM across five VLMs, outperforming alternatives by statistically significant margins while black-box and some gray-box methods collapse on certain models.
hub Tool reference
DeBERTaV3: Improving DeBERTa using ELECTRA-Style Pre-Training with Gradient-Disentangled Embedding Sharing
Tool reference. 80% of classified Pith citations use this work as a method, library, or software dependency, not as a substantive claim.
59
Pith papers citing it
397
external citations · Pith
Method reference
80% of classified citations
abstract
This paper presents a new pre-trained language model, DeBERTaV3, which improves the original DeBERTa model by replacing mask language modeling (MLM) with replaced token detection (RTD), a more sample-efficient pre-training task. Our analysis shows that vanilla embedding sharing in ELECTRA hurts training efficiency and model performance. This is because the training losses of the discriminator and the generator pull token embeddings in different directions, creating the "tug-of-war" dynamics. We thus propose a new gradient-disentangled embedding sharing method that avoids the tug-of-war dynamics, improving both training efficiency and the quality of the pre-trained model. We have pre-trained DeBERTaV3 using the same settings as DeBERTa to demonstrate its exceptional performance on a wide range of downstream natural language understanding (NLU) tasks. Taking the GLUE benchmark with eight tasks as an example, the DeBERTaV3 Large model achieves a 91.37% average score, which is 1.37% over DeBERTa and 1.91% over ELECTRA, setting a new state-of-the-art (SOTA) among the models with a similar structure. Furthermore, we have pre-trained a multi-lingual model mDeBERTa and observed a larger improvement over strong baselines compared to English models. For example, the mDeBERTa Base achieves a 79.8% zero-shot cross-lingual accuracy on XNLI and a 3.6% improvement over XLM-R Base, creating a new SOTA on this benchmark. We have made our pre-trained models and inference code publicly available at https://github.com/microsoft/DeBERTa.
hub tools
citation-role summary
method 4
other 1
citation-polarity summary
representative citing papers
SciTraj is the first claim-grounded typed citation graph with 32,559 papers and 573,126 edges across six relation types, plus a temporally split link-prediction benchmark.
SlotGCG uses Vulnerable Slot Score (VSS) to identify and target the most vulnerable prompt positions for adversarial token insertion, delivering 14% higher ASR than standard GCG and 42% higher against defenses.
Face-Feature Tuning is a label-free logit remapping method that reduces FPR/TPR gaps across groups in deepfake detection while preserving overall accuracy.
NEI competence does not transfer reliably across evidence constructions in fact verification; mixed training narrows but does not close the gap, and aggregate scores can mask specific weaknesses.
Belief Engine is a configurable belief-update mechanism for multi-agent LLM systems that uses structured argument extraction and log-odds stance updates to make evidence-grounded deliberation inspectable and controllable.
BOOKMARKS introduces searchable bookmarks as reusable answers to storyline questions, enabling active initialization and passive synchronization for more consistent role-playing agent memory than recurrent summarization.
A framework jointly models annotator-specific NLI labels and explanations using conditioned representations and two explainer architectures, improving predictive performance over baselines.
RAGognizer adds a detection head to LLMs for joint training on generation and token-level hallucination detection, yielding SOTA detection and fewer hallucinations in RAG while preserving output quality.
Dual Triangle Attention achieves effective bidirectional attention with built-in positional inductive bias via dual triangular masks, outperforming standard bidirectional attention on position-sensitive tasks and showing strong masked language modeling results with or without positional embeddings.
BBQ is a new benchmark dataset showing that QA models often default to social stereotypes, achieving up to 3.4 points higher accuracy when the correct answer aligns with bias.
Tenant comments on Dutch housing association Facebook posts cluster into six discourse types, and organizational context—not post style—predicts which type appears.
Dialectal robustness and generation are dissociated in LLMs: benchmarks are driven by pretraining and SFT while alignment reshapes generation invisibly to benchmarks, and the method maximizing dialectal reward is least preferred by human evaluators.
MARGO mitigates thinking-induced hallucination in large reasoning models by using mixed-mode GRPO rollout groups that compare thinking trajectories against same-model non-thinking references.
Proposes a source-data-free transfer learning framework for sparse single-index models that transfers generalized Stein's lemma summaries and uses a guided MLP for nonlinear adaptation.
ConvMemory v3 introduces a dual-evidence gate for target-conditioned memory validity verification, reporting 90.12% accuracy on synthetic benchmarks, 98.8% transfer to real data, and H@1 improvement from 45.1% to 95.7% in deployment while preserving safety contracts.
Query-driven table integration that uses Steiner-tree search to choose which joins LLMs must verify, reporting 30%+ accuracy gains at 5x lower LLM cost.
Introduces claim-conditioned re-scoring (SIFT) and warranted supports proportion (WSP) metric, reporting accuracy recovery up to 27.6 points and WSP calibration at AUC 0.92 on FEVER, SciFact and other benchmarks.
RidgeFT enables replay-free lifelong MGT attribution via frozen encoder, class-wise sufficient statistics, covariance calibration, and closed-form ridge regression updates, outperforming baselines on macro-F1 and retention-adaptation balance.
RISC reformulates self-consistency answer selection as a ranking task solved by a lightweight LambdaRank model with five hand-designed features, yielding better accuracy-efficiency trade-offs than majority voting on QA benchmarks.
The paper creates InsightVQA, a 725K QA-pair benchmark with perception, grounded-understanding, and cognition levels for emotion-cognitive visual question answering, plus a 30K-sample evaluation set and InsightNet baseline.
ICPT converts a few reference images of a personalized concept into an adaptive-length visual prompt plus a label embedding, letting a frozen LVLM add and reason about multiple concepts on the fly.
SAVER proposes a conformal groundability gate plus submodular image selector that activates vision only when needed for multimodal named entity recognition and relation extraction, improving F1 while lowering compute.
A 149M-parameter distributional energy-based verifier with low-rank adapter ensemble reduces constraint violations in structured LLM reasoning and outperforms or matches much larger models on five benchmarks.
citing papers explorer
-
A Benchmark for Hallucination Detection in VLMs for Gastrointestinal Endoscopy
White-box method ReXTrust achieves highest AUC (peak 93.0) on Gut-VLM across five VLMs, outperforming alternatives by statistically significant margins while black-box and some gray-box methods collapse on certain models.
-
How Does Research Evolve? Tracing Cross-Domain Trajectories in NLP, ML, and CV with Claim-Grounded Typed Citations
SciTraj is the first claim-grounded typed citation graph with 32,559 papers and 573,126 edges across six relation types, plus a temporally split link-prediction benchmark.
-
SlotGCG: Exploiting the Positional Vulnerability in LLMs for Jailbreak Attacks
SlotGCG uses Vulnerable Slot Score (VSS) to identify and target the most vulnerable prompt positions for adversarial token insertion, delivering 14% higher ASR than standard GCG and 42% higher against defenses.
-
Toward Calibrated, Fair, and accurate Deepfake Detection
Face-Feature Tuning is a label-free logit remapping method that reduces FPR/TPR gaps across groups in deepfake detection while preserving overall accuracy.
-
Evidence Absence Is Not Evidence Insufficiency: Diagnosing NEI Construction Artifacts in Fact Verification
NEI competence does not transfer reliably across evidence constructions in fact verification; mixed training narrows but does not close the gap, and aggregate scores can mask specific weaknesses.
-
Belief Engine: Configurable and Inspectable Stance Dynamics in Multi-Agent LLM Deliberation
Belief Engine is a configurable belief-update mechanism for multi-agent LLM systems that uses structured argument extraction and log-odds stance updates to make evidence-grounded deliberation inspectable and controllable.
-
BOOKMARKS: Efficient Active Storyline Memory for Role-playing
BOOKMARKS introduces searchable bookmarks as reusable answers to storyline questions, enabling active initialization and passive synchronization for more consistent role-playing agent memory than recurrent summarization.
-
Fine-Grained Perspectives: Modeling Explanations with Annotator-Specific Rationales
A framework jointly models annotator-specific NLI labels and explanations using conditioned representations and two explainer architectures, improving predictive performance over baselines.
-
RAGognizer: Hallucination-Aware Fine-Tuning via Detection Head Integration
RAGognizer adds a detection head to LLMs for joint training on generation and token-level hallucination detection, yielding SOTA detection and fewer hallucinations in RAG while preserving output quality.
-
Dual Triangle Attention: Effective Bidirectional Attention Without Positional Embeddings
Dual Triangle Attention achieves effective bidirectional attention with built-in positional inductive bias via dual triangular masks, outperforming standard bidirectional attention on position-sensitive tasks and showing strong masked language modeling results with or without positional embeddings.
-
BBQ: A Hand-Built Bias Benchmark for Question Answering
BBQ is a new benchmark dataset showing that QA models often default to social stereotypes, achieving up to 3.4 points higher accuracy when the correct answer aligns with bias.
-
Does online sustainability communication shape public discourse? Insights from six years of tenant-housing provider interactions
Tenant comments on Dutch housing association Facebook posts cluster into six discourse types, and organizational context—not post style—predicts which type appears.
-
DiaLLM: An Investigation into the Robustness-Generation Gap in English Dialect Adaptation
Dialectal robustness and generation are dissociated in LLMs: benchmarks are driven by pretraining and SFT while alignment reshapes generation invisibly to benchmarks, and the method maximizing dialectal reward is least preferred by human evaluators.
-
Mitigating Factual Hallucination in Large Reasoning Models via Mixed-Mode Advantage Regularization
MARGO mitigates thinking-induced hallucination in large reasoning models by using mixed-mode GRPO rollout groups that compare thinking trajectories against same-model non-thinking references.
-
Multi-Source Transfer Learning of Sparse Single-Index Models
Proposes a source-data-free transfer learning framework for sparse single-index models that transfers generalized Stein's lemma summaries and uses a guided MLP for nonlinear adaptation.
-
ConvMemory v3: A Validity Context Layer for Conversational Memory via Target-Conditioned Relation Verification
ConvMemory v3 introduces a dual-evidence gate for target-conditioned memory validity verification, reporting 90.12% accuracy on synthetic benchmarks, 98.8% transfer to real data, and H@1 improvement from 45.1% to 95.7% in deployment while preserving safety contracts.
-
EcoTable: Cost-effective Table Integration in Data Lakes for Natural Language Queries
Query-driven table integration that uses Steiner-tree search to choose which joins LLMs must verify, reporting 30%+ accuracy gains at 5x lower LLM cost.
-
The Warrant Gap: Claim-Conditioned Re-scoring for Fact-Checking
Introduces claim-conditioned re-scoring (SIFT) and warranted supports proportion (WSP) metric, reporting accuracy recovery up to 27.6 points and WSP calibration at AUC 0.92 on FEVER, SciFact and other benchmarks.
-
When New Generators Arrive: Lifelong Machine-Generated Text Attribution via Ridge Feature Transfer
RidgeFT enables replay-free lifelong MGT attribution via frozen encoder, class-wise sufficient statistics, covariance calibration, and closed-form ridge regression updates, outperforming baselines on macro-F1 and retention-adaptation balance.
-
Boosting Self-Consistency with Ranking
RISC reformulates self-consistency answer selection as a ranking task solved by a lightweight LambdaRank model with five hand-designed features, yielding better accuracy-efficiency trade-offs than majority voting on QA benchmarks.
-
InsightVQA: High-Dimensional Emotion-Cognitive Visual Question Answering Benchmark
The paper creates InsightVQA, a 725K QA-pair benchmark with perception, grounded-understanding, and cognition levels for emotion-cognitive visual question answering, plus a 30K-sample evaluation set and InsightNet baseline.
-
Personalize Your Large Vision-language Models With In-context Prompt Tuning
ICPT converts a few reference images of a personalized concept into an adaptive-length visual prompt plus a label embedding, letting a frozen LVLM add and reason about multiple concepts on the fly.
-
SAVER: Selective As-Needed Vision Evidence for Multimodal Information Extraction
SAVER proposes a conformal groundability gate plus submodular image selector that activates vision only when needed for multimodal named entity recognition and relation extraction, improving F1 while lowering compute.
-
Distributional Energy-Based Models for Uncertainty-Aware Structured LLM Reasoning
A 149M-parameter distributional energy-based verifier with low-rank adapter ensemble reduces constraint violations in structured LLM reasoning and outperforms or matches much larger models on five benchmarks.
-
SynIB: Informational Bottleneck for Maximizing Synergy in Multimodal Learning
SynIB is an information-theoretic objective that adds a penalty for unimodal confidence to standard task loss, improving accuracy on synergy-dependent examples by up to 7.8% across synthetic XOR tasks and five real-world multimodal benchmarks.
-
A Cascaded Generative Approach for e-Commerce Recommendations
A cascaded generative merchandising framework with placement theme generation, constrained keyword generation, and teacher-student fine-tuning achieves a 2.7% lift in cart adds per page view over a strong baseline in online e-commerce experiments.
-
ProCompNav: Proactive Instance Navigation with Comparative Judgment for Ambiguous User Queries
ProCompNav builds a candidate pool from ambiguous queries then uses pool-splitting binary questions for disambiguation, improving success rate and shortening responses on CoIN-Bench and TextNav.
-
SHIELD: A Diverse Clinical Note Dataset and Distilled Small Language Models for Enterprise-Scale De-identification
SHIELD is a new diverse clinical note dataset paired with distilled small language models that achieve 0.89 span-level precision and 0.88 recall for on-premise PHI de-identification.
-
Effective Performance Measurement: Challenges and Opportunities in KPI Extraction from Earnings Calls
Encoder models trained on SEC filings struggle with earnings calls due to domain shift, while LLMs enable open-ended KPI extraction with 79.7% human-verified precision on newly introduced benchmarks.
-
TwinGate: Stateful Defense against Decompositional Jailbreaks in Untraceable Traffic via Asymmetric Contrastive Learning
TwinGate deploys a stateful dual-encoder system with asymmetric contrastive learning to detect decompositional jailbreaks in untraceable LLM traffic at high recall and low false-positive rate with negligible latency.
-
ADE: Adaptive Dictionary Embeddings -- Scaling Multi-Anchor Representations to Large Language Models
ADE scales multi-anchor word representations to transformers via Vocabulary Projection, Grouped Positional Encoding, and context-aware reweighting, achieving 98.7% fewer trainable parameters than DeBERTa-v3-base while matching or exceeding it on two text-classification benchmarks and compressing the
-
Mind the Unseen Mass: Unmasking LLM Hallucinations via Soft-Hybrid Alphabet Estimation
SHADE adaptively combines coverage and spectral signals to estimate semantic alphabet size from few LLM samples, yielding better performance than baselines in low-sample regimes for alphabet estimation and QA error detection.
-
Sensory-Aware Sequential Recommendation via Review-Distilled Representations
ASER distills sensory attributes extracted from reviews into item embeddings that improve sequential recommendation metrics by an average of 7.9% HR@10 and 11.2% NDCG@10 across 20 Amazon domain-backbone tests.
-
GigaCheck: Detecting LLM-generated Content via Object-Centric Span Localization
GigaCheck detects LLM-generated text at both document and span levels by combining fine-tuned language-model embeddings with a DETR-like architecture that treats generated intervals as detectable objects.
-
AdaLoRA: Adaptive Budget Allocation for Parameter-Efficient Fine-Tuning
AdaLoRA uses SVD-based pruning to allocate the parameter budget for low-rank fine-tuning updates according to per-matrix importance scores, yielding better performance than uniform allocation especially under tight budgets.
-
moBERTo: A Modern Encoder for Portuguese via Continued Pretraining of ModernBERT
Continued pretraining of ModernBERT on curated Portuguese data produces moBERTo, which reports top results on Portuguese retrieval reranking and PLUE-PT benchmarks.
-
Can News Predict the Market? Limits of Zero-Shot Financial NLP and the Role of Explainable AI
Zero-shot NLP models do not outperform baselines for short-term stock prediction from news, but multi-layered explainability reliably separates trustworthy from unreliable outputs.
-
StanceNakba Shared Task: Actor and Topic-Aware Stance Detection in Public Discourse
Introduces StanceNakba 2026 shared task and 2,606-post dataset for Pro-Palestine/Pro-Israel/Neutral stance in English posts and Favor/Against/Neither stance in Arabic posts on normalization and refugees, with top Macro F1 scores of 0.9620 and 0.8724 from transformer models.
-
IR3DE: A Linear Router for Large Language Models
IR3DE is a ridge regression router for domain-expert LLMs that matches or exceeds baselines in language modeling and reasoning tasks while allowing dynamic expert addition or removal without retraining.
-
CasualSynth: Generating Structurally Sound Synthetic Data
CausalSynth combines structural causal models with LLMs and iterative verification to produce synthetic data that respects given causal structures while remaining linguistically natural.
-
Feature-Augmented Transformers for Robust AI-Text Detection Across Domains and Generators
Feature-augmented DeBERTa-v3-base with attention-based fusion reaches 85.9% balanced accuracy on the multi-domain M4 benchmark under fixed-threshold evaluation, outperforming zero-shot baselines by up to 7.22 points.
-
MKJ at SemEval-2026 Task 9: A Comparative Study of Generalist, Specialist, and Ensemble Strategies for Multilingual Polarization
Per-language architecture selection among generalists, specialists, and ensembles achieves 0.796 macro F1 across 22 languages in SemEval-2026 Task 9.
-
ZSG-IAD: A Multimodal Framework for Zero-Shot Grounded Industrial Anomaly Detection
ZSG-IAD is a zero-shot multimodal system that uses language-guided two-hop grounding and rule-based reinforcement learning to produce anomaly masks and explainable reports from industrial sensor data.
-
AGSC: Adaptive Granularity and Semantic Clustering for Uncertainty Quantification in Long-text Generation
AGSC combines NLI neutral probabilities for adaptive granularity with GMM semantic clustering to improve uncertainty quantification in long-text LLM generation, claiming SOTA factuality correlation and 60% faster inference.
-
Cross-modal Consistency Guidance for Robust Emotion Control in Auto-Regressive TTS Models
An emotion TTS system adjusts Classifier-Free Guidance strength according to text-style semantic mismatch; it shows small emotion-accuracy gains, but headline baselines and subjective results are absent from the main text.
-
Talk Less, Fly Lighter: Autonomous Semantic Compression for UAV Swarm Communication via LLMs
LLM-based autonomous semantic compression in four 2D UAV swarm simulations shows potential for efficient collaborative communication under bandwidth constraints.
-
LTRR: Learning To Rank Retrievers for LLMs
LTRR learns to rank a pool of retrievers by their expected contribution to RAG answer correctness and shows that query-dependent selection beats the best single retriever on QA benchmarks.
-
Fact-Checking with Contextual Narratives: Leveraging Retrieval-Augmented LLMs for Social Media Analysis
CRAVE is a new framework that clusters retrieved text and image evidence into narratives and uses an LLM judge to produce explained fact-checking verdicts.
-
Defending against Backdoor Attacks via Module Switching
Module-switching defense disrupts backdoors more effectively than weight averaging with fewer models and remains robust even when some models share the same backdoors.
-
Semantic Embeddings of Chemical Elements for Enhanced Materials Inference and Discovery
ElementBERT generates literature-derived semantic embeddings for chemical elements that outperform empirical descriptors in alloy property prediction and optimization tasks with up to 23% accuracy gains.