Papers by Mark Dras
When the Model Said ‘No Comment’, We Knew Helpfulness Was Dead, Honesty Was Alive, and Safety Was Terrified (2026.eacl-long)
Copied to clipboard
| Challenge: | Existing work uses SFT and MoE to align Large Language Models, but these work face challenges in multi-objective settings. |
| Approach: | They propose a framework that uses prompt-injected fine-tuning to extract axis-specific task features . it deploys a MoCaE module that calibrates expert routing using fractal and natural geometry . |
| Outcome: | The proposed framework achieves significant gains on Alpaca, BeaverTails, TruthfulQA and TruthfulQ with +171.5% win rate and +110.1% truthfulness-informativeness. |
VITAL: A New Dataset for Benchmarking Pluralistic Alignment in Healthcare (2025.acl-long)
Copied to clipboard
| Challenge: | Existing approaches to align Large Language Models with human values model an averaged or monolithic preference, despite progress in pluralistic alignment, no prior work has focused on health . |
| Approach: | They propose a benchmark dataset to assess and benchmark pluralistic alignment methodologies. |
| Outcome: | The proposed model can model pluralistic views within health domains. |
AlignCultura: Towards Culturally Aligned Large Language Models? (2026.acl-long)
Copied to clipboard
| Challenge: | Existing benchmarks represent early steps toward cultural alignment, yet no benchmarks currently enables systematic evaluation of cultural alignment in line with UNESCO’s principles of cultural diversity w.r.t HHH paradigm. |
| Approach: | Align-Cultura aims to evaluate cultural alignment in large language models . it uses a Query Construction pipeline to reclassify prompts and expand underrepresented domains . response generation pairs prompts with culturally grounded responses . |
| Outcome: | Empirically, culturally fine-tuned models improve joint HHH by 4%–6%, reduce cultural failures by 18%, achieve 10%–12% efficiency gains, and limit leakage to 0.3%. |
Fairness Evaluation and Inference Level Mitigation in LLMs (2026.findings-acl)
Copied to clipboard
| Challenge: | Large language models display undesirable behaviors embedded in their internal representations, undermining fairness, inconsistency drift, and the propagation of unwanted patterns during extended dialogues. |
| Approach: | They propose a pruning-based framework that detects context-aware neuron activations and applies adaptive masking to modulate their influence during generation. |
| Outcome: | The proposed framework detects context-aware neuron activations and applies adaptive masking to modulate their influence during generation. |
Here’s a Free Lunch: Sanitizing Backdoored Models with Model Merge (2024.findings-acl)
Copied to clipboard
| Challenge: | democratization of pre-trained language models brings significant security risks, including backdoor attacks. |
| Approach: | They propose to merge a backdoored model with other homogeneous models to remediate backdoor vulnerabilities. |
| Outcome: | The proposed model merging approach outperforms other models on classification tasks without additional resources or specific knowledge. |
SafeConstellations: Mitigating Over-Refusals in LLMs Through Task-Aware Representation Steering (2026.acl-long)
Copied to clipboard
| Challenge: | Current safety alignment methods fail to identify intended benign task before refusing to respond. |
| Approach: | They propose a method that uses inference-time trajectory-shifting to guide model behavior . they show that LLMs persist in refusing inputs containing harmful content . |
| Outcome: | The proposed approach reduces over-refusals with minimal impact on utility. |
Large Scale Author Obfuscation Using Siamese Variational Auto-Encoder: The SiamAO System (2020.starsem-1)
Copied to clipboard
| Challenge: | Existing approaches to author obfuscation are largely heuristic, but they can be used to attack author identification. |
| Approach: | They propose a deep learning architecture for constructing adversarial examples against similarity-based learners and explore its application to author obfuscation. |
| Outcome: | The proposed architectures show that they can be used to attack author obfuscation . the proposed architecture shows that it can be applied to obliquacy of text . |
Too Helpful, Too Harmless, Too Honest or Just Right? (2025.emnlp-main)
Copied to clipboard
| Challenge: | Existing methods optimize for individual alignment dimensions in isolation, leading to trade-offs and inconsistent behavior. |
| Approach: | They propose a modular alignment framework that integrates a Mixture of Calibrated Experts (MoCaE) within the Transformer architecture. |
| Outcome: | The proposed framework outperforms baselines on three alignment benchmarks, achieving 32.5% win rate, 33.9% safety score, and 28.4% truthfulness. |
VnCoreNLP: A Vietnamese Natural Language Processing Toolkit (N18-5)
Copied to clipboard
| Challenge: | Using word segmenters and POS taggers, Vietnamese NLP pipelines are no longer considered SOTA models for Vietnamese. |
| Approach: | They propose a Java NLP annotation pipeline for Vietnamese that provides rich linguistic annotations. |
| Outcome: | The proposed toolkit provides rich linguistic annotations to facilitate research work on Vietnamese NLP. |
Mention Flags (MF): Constraining Transformer-based Text Generators (2021.acl-long)
Copied to clipboard
| Challenge: | Constrained decoding algorithms produce hypotheses satisfying all constraints, but they are computationally expensive and can lower the generated text quality. |
| Approach: | They propose a Mention Flag mechanism which traces whether lexical constraints are satisfied in outputs of an S2S decoder. |
| Outcome: | The proposed models maintain higher constraint satisfaction and text quality than baseline models and other constrained decoding algorithms. |
Framing Political Bias in Multilingual LLMs Across Pakistani Languages (2026.acl-long)
Copied to clipboard
| Challenge: | Large Language Models (LLMs) shape public discourse, yet most evaluations of economic and political bias focus on high-resource Western languages and contexts. |
| Approach: | They propose to use a culturally adapted Political Compass Test to evaluate political bias in 13 state-of-the-art LLMs across five Pakistani languages. |
| Outcome: | The proposed framework captures ideological stance (economic/social axes) and stylistic framing (content, tone, emphasis) in 13 state-of-the-art LLMs across five Pakistani languages. |
A Fast and Accurate Vietnamese Word Segmenter (L18-1)
Copied to clipboard
| Challenge: | Experimental results show that our approach outperforms previous state-of-the-art approaches in terms of accuracy and performance speed. |
| Approach: | They propose a method where rules are stored in an exception structure and new rules are only added to correct segmentation errors. |
| Outcome: | The proposed approach outperforms existing methods on Vietnamese treebank benchmarks. |
Predicting accuracy on large datasets from smaller pilot data (P18-2)
Copied to clipboard
| Challenge: | obtaining training data is often the most difficult part of an NLP or ML project . obtaining data is important to estimate how much training data a system will require to achieve a target accuracy. |
| Approach: | They propose a performance extrapolation task to evaluate extrapolations on larger training sets. |
| Outcome: | The proposed method can predict accuracy on larger training datasets. |
Few-shot fine-tuning SOTA summarization models for medical dialogues (2022.naacl-srw)
Copied to clipboard
| Challenge: | Abstractive summarization of medical dialogues is a challenge for standard training approaches due to the paucity of suitable datasets. |
| Approach: | They propose to use medical dialogues to generate abstractive summaries using transformer-based models with zero-shot and few-shot learning strategies. |
| Outcome: | The proposed models were compared with a medical dialogue dataset with 143 snippets and a general domain and dialogue-specific text to assess their performance. |
Seeing the Forest through the Trees: Data Leakage from Partial Transformer Gradients (2024.emnlp-main)
Copied to clipboard
| Challenge: | Recent studies show that distributed machine learning is vulnerable to gradient inversion attacks . a recent study demonstrated the possibility of reconstructing private textual training data using partial gradients . |
| Approach: | They propose to use partial gradients to reconstruct training data using a Transformer model. |
| Outcome: | The proposed method is vulnerable to gradient inversion attacks, the authors show . they show that applying differential privacy on gradients during training offers limited protection . |