Papers by Seraphina Goldfarb-Tarrant
Intrinsic Bias Metrics Do Not Correlate with Application Bias (2021.acl-long)
Copied to clipboard
| Challenge: | a recent survey of bias in natural language processing found that a coreference system makes more errors in an anti-stereotypical coreferent than in a pro-sterereotype one. |
| Approach: | They compare intrinsic and extrinsic bias metrics across hundreds of trained models . they urge researchers to focus on extrindic measures of bias, not easy to measure . |
| Outcome: | a new intrinsic metric and an annotated test set on gender bias in hate speech are tested . authors urge researchers to focus on extrinsic measures of bias, and to make them more feasible . |
Safer or Luckier? LLMs as Safety Evaluators Are Not Robust to Artifacts (2025.acl-long)
Copied to clipboard
| Challenge: | Large Language Models (LLMs) are increasingly employed as automated evaluators to assess the safety of generated content. |
| Approach: | They evaluate 11 LLM judge models across critical safety domains . apologetic language artifacts alone can skew evaluator preferences by up to 98% . |
| Outcome: | The findings show that apologetic language artifacts alone can skew evaluator preferences by up to 98% . |
How Gender Debiasing Affects Internal Model Representations, and Why It Matters (2022.naacl-main)
Copied to clipboard
| Challenge: | Existing studies of gender bias in NLP focus on extrinsic or intrinsic bias, but the relationship between extrindic and intrinsic bias is relatively unknown. |
| Approach: | They propose a framework to measure extrinsic and intrinsic bias together and propose metric to measure debiasing and intrinsic debiases. |
| Outcome: | The proposed framework provides a comprehensive perspective on bias in NLP models, which can be applied to deploy NLP systems in a more informed manner. |
The Multilingual Alignment Prism: Aligning Global and Local Preferences to Reduce Harm (2024.emnlp-main)
Copied to clipboard
null Aakanksha, Arash Ahmadian, Beyza Ermis, Seraphina Goldfarb-Tarrant, Julia Kreutzer, Marzieh Fadaee, Sara Hooker
| Challenge: | Existing approaches to safety alignment focus on homogeneous monolingual settings . preference training and safety measures often overfit to harms common in Western-centric datasets . |
| Approach: | They propose to use human annotated red teaming prompts to identify global and local harms. |
| Outcome: | The proposed approach can address and optimize for a non-homogeneous set of languages and cultural preferences while minimizing both global and local harms. |
MAPS: A Multilingual Benchmark for Agent Performance and Security (2026.findings-eacl)
Copied to clipboard
Omer Hofman, Jonathan Brokman, Oren Rachmil, Shamik Bose, Vikas Pahuja, Toshiya Shimizu, Trisha Starostina, Kelly Marchisio, Seraphina Goldfarb-Tarrant, Roman Vainshtein
| Challenge: | Existing benchmarks do not provide a comprehensive, multi-domain, security-aware evaluation of multilingual agentic AI systems. |
| Approach: | They propose a multilingual benchmark suite to evaluate agentic AI systems across languages and tasks. |
| Outcome: | The proposed framework evaluates agentic AI systems across languages and tasks. |
Content Planning for Neural Story Generation with Aristotelian Rescoring (2020.emnlp-main)
Copied to clipboard
| Challenge: | Current approaches to narrative composition are plagued by difficulty in mastering structure, will veer between topics, and lack long-range cohesion. |
| Approach: | They propose a plot-generation language model and a set of rescoring models that implement an aspect of good story-writing as detailed in Aristotle's Poetics. |
| Outcome: | The proposed system improves the quality of the narrative generated from the proposed model and improves its relevance to a given prompt and quality of stories written with our principled plot structure. |
Cross-lingual Transfer Can Worsen Bias in Sentiment Analysis (2023.emnlp-main)
Copied to clipboard
| Challenge: | Existing evidence of demographic bias in SA systems is limited to a handful of languages, and it is costly to create supervised training data in a new language. |
| Approach: | They use counterfactual evaluation to test whether gender or racial biases are imported when using cross-lingual transfer . r&r is much more prevalent than gender biase . |
| Outcome: | The proposed model is compared with monolingual systems in five languages and shows that it is biased more than monolingual ones. |
Bias Beyond English: Counterfactual Tests for Bias in Sentiment Analysis in Four Languages (2023.findings-acl)
Copied to clipboard
| Challenge: | Sentiment analysis systems are used in hundreds of products and languages . Gender and racial biases are well-studied in English, but understudied elsewhere . |
| Approach: | They build a counterfactual evaluation corpus for gender and racial/migrant bias in four languages. |
| Outcome: | The evaluation corpus reveals which models have less bias and pinpoints changes in model bias behaviour, enabling more targeted mitigation strategies. |
Plan, Write, and Revise: an Interactive System for Open-Domain Story Generation (N19-4)
Copied to clipboard
| Challenge: | a neural narrative generation system interacts with humans to generate stories . a recent resurgence of interest in collaborative storytelling has led to new approaches . |
| Approach: | They propose a neural narrative generation system that interacts with humans to generate stories. |
| Outcome: | The proposed system improves story quality and user engagement under time constraints. |
Lost in Simulation: LLM-Simulated Users are Unreliable Proxies for Human Users in Agentic Evaluations (2026.acl-long)
Copied to clipboard
| Challenge: | Agentic benchmarks rely on LLM-simulated users to evaluate agent performance . however, the robustness, validity, and fairness of this approach remain unexamined . |
| Approach: | They investigate whether LLM-simulated users are reliable proxies for real human users . they find that agent success rates vary up to 9 percentage points across different LLMs . |
| Outcome: | The results show that simulated users underestimate success on challenging tasks while miscalibrate performance on moderately difficult tasks. |
A SMART Mnemonic Sounds like “Glue Tonic”: Mixing LLMs with Student Feedback to Make Mnemonic Learning Stick (2024.emnlp-main)
Copied to clipboard
Nishant Balepur, Matthew Shu, Alexander Hoyle, Alison Robey, Shi Feng, Seraphina Goldfarb-Tarrant, Jordan Boyd-Graber
| Challenge: | a new study shows that mnemonics are not effective at matching student learning to a standardized learning model. |
| Approach: | They build a keyword mnemonic generator that finds mnemonics students favor in a flashcard app . they use expressed and observed preferences to find out what students think is helpful . |
| Outcome: | The proposed mnemonics outperform existing models in keyword mnemonics . the human writer outperformed both models in terms of keyword simplicity and explanation quality . |
A Good Plan is Hard to Find: Aligning Models with Preferences is Misaligned with What Helps Users (2025.emnlp-main)
Copied to clipboard
Nishant Balepur, Matthew Shu, Yoo Yeon Sung, Seraphina Goldfarb-Tarrant, Shi Feng, Fumeng Yang, Rachel Rudinger, Jordan Lee Boyd-Graber
| Challenge: | We test alignment methods to ensure LLMs are helpful, but they train or evaluate on what users prefer . |
| Approach: | They test alignment methods to ensure LLMs generate plans that help users . they get 4388 plan executions and 5584 comparisons to measure user preferences . |
| Outcome: | The proposed approach can be applied to the problem of user preferences and helpfulness. |
This prompt is measuring <mask>: evaluating bias evaluation in language models (2023.findings-acl)
Copied to clipboard
| Challenge: | a growing body of work uses prompts and templates to assess bias in language models . authors examine the scope of possible bias types and identify those under-researched . |
| Approach: | They draw on a measurement modelling framework to create a bias taxonomy . they show that bias tests are often unstated or ambiguous, carry implicit assumptions . |
| Outcome: | The proposed taxonomy shows that bias tests are often unstated or ambiguous . the analysis illuminates the scope of possible bias types the field can measure . |