Singapore

Language models (LMs) judges are widely used to evaluate the quality of LM outputs. Despite many advantages, LM judges display concerning biases that can impair their integrity in evaluations. One such bias is self-preference: LM judges preferring their own answers over those produced by other LMs or humans. The bias is hard to eliminate as frontier LM judges can distinguish their own outputs from those of others, even when the evaluation candidates are not labeled with their sources. In this paper, we investigate strategies to mitigate self-preference by reducing the LM judges&#39; ability to recognize their own outputs. We apply black-box perturbations to evaluation candidates in pairwise comparison to obfuscate the authorship and reduce self-recognition. We find that perturbations as simple as synonym replacement for a few words predictably reduce self-preference. However, we also uncover fundamental challenges to eliminating the bias: when we extrapolate our perturbations to a more complete neutralization of stylistic differences between the evaluation candidates, self-preference recovers. Our findings suggest that self-recognition and self-preference can happen on many semantic levels, and a complete mitigation remains challenging despite promising initial results.

AAAI 2026

Mitigating Self-Preference by Authorship Obfuscation

llm-bias

llm-judge

safety and robustness

## Welcome!
"I am excited to welcome you to this year’s edition of the Conference on Empirical Methods in Natural Language Processing! Importantly, it marks the 30th edition of EMNLP. With over 8,000 submissions, more than 3,000 accepted papers, and thousands of attendees, we have come a long way from that first
workshop, which had 14 accepted papers. As the field looks ahead, Suzhou is the fitting location for celebrating this milestone: rooted in a long literary tradition, yet modern and forward-looking, and home to a large share of the NLP community." 

*Message from the General Chair, Dirk Hovy*

EMNLP 2025

Celebrate 30 Years of EMNLP! 
EMNLP 2025 will be held in Suzhou, China from November 5th to November 9th, 2025.

Language models (LMs) judges are widely used to evaluate the quality of LM outputs. Despite many advantages, LM judges display concerning biases that can impair their integrity in evaluations. One such bias is self-preference: LM judges preferring their own answers over those produced by other LMs or humans. The bias is hard to eliminate as frontier LM judges can distinguish their own outputs from those of others, even when the evaluation candidates are not labeled with their sources. In this paper, we investigate strategies to mitigate self-preference by reducing the LM judges' ability to recognize their own outputs. We apply black-box perturbations to evaluation candidates in pairwise comparison to obfuscate the authorship and reduce self-recognition. We find that perturbations as simple as synonym replacement for a few words predictably reduce self-preference. However, we also uncover fundamental challenges to eliminating the bias: when we extrapolate our perturbations to a more complete neutralization of stylistic differences between the evaluation candidates, self-preference recovers. Our findings suggest that self-recognition and self-preference can happen on many semantic levels, and a complete mitigation remains challenging despite promising initial results.

poster

We are pleased to announce the Fortieth AAAI Conference on Artificial Intelligence (AAAI-26), which will be held in Singapore EXPO from January 20 to January 27, 2026.

The purpose of the AAAI conference series is to promote research in Artificial Intelligence (AI) and foster scientific exchange between researchers, practitioners, scientists, students, and engineers across the entirety of AI and its affiliated disciplines. AAAI-26 will feature technical paper presentations, special tracks, invited speakers, workshops, tutorials, poster sessions, senior member presentations, competitions, and exhibit programs, and a range of other activities to be announced.

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

Please log in

To access this event page, you are required to register.
Please complete your registration to continue.

We recommend reading [**the registration information**](https://aaai.org/conference/aaai/aaai-26/registration/) first.

**Online Registration Form**: https://aaai.getregistered.net/conference-2026 

Registration Required

We are pleased to announce the Fortieth AAAI Conference on Artificial Intelligence (AAAI-26), which will be held in Singapore EXPO from January 20 to January 27, 2026.

This study aimed to investigate the impact of a data-driven
teaching approache on students’ conceptual understanding of
machine learning (ML). To this end, an exemplary
intervention was designed and evaluated using a pre-post
test design and a German-language Concept Inventory on
Machine Learning. A total of 83 German ninth-grade students
participated in the study. The results revealed significant
learning gains related to data handling and the ML
workflow. In contrast, conceptions about the inner workings
of ML models largely persisted. The effectiveness of the
intervention varied depending on context, with greater
gains observed in the text generation domain than in facial
recognition, highlighting challenges in cross-contextual
transfer of understanding. A regression analysis showed no
significant influence of students’ pre-instructional
conceptions on learning outcomes. These findings
demonstrate both the potential and the limitations of
data-driven teaching approaches and emphasize the need for
more explicit engagement with learners' misconceptions to
foster deeper conceptual change.

Impact of a Data-driven Teaching Approach on 9th Graders Conceptual Understanding of Machine Learning

Recent work sheds light on the interests of Machine
Learning (ML) learners, both to support their success and
to broaden participation. Yet, do learning resources like
textbooks ‘fit’ with these interests? We analyzed
introductions of ten openly available ML textbooks using
thematic analysis to identify motivational strategies and
compare them with student interests in ML reported in prior
work. We found that textbooks frequently motivated learners
by setting learning goals, previewing core ML topics (e.g.,
classification, regression), showcasing applications and
current successes, and, less often, by using
learner-centered strategies such as reassurance or
curiosity prompts. We grouped these into three overarching
themes: theoretical, practical, and learner-centered. These
motivations largely align with student interest,
particularly in theory and applications, even in textbooks
published before the recent surge of ML and Artificial
Intelligence (AI). Our findings offer insight into how
textbooks frame the value of ML and provide guidance for
creating future materials that engage and support a wider
range of learners.

Analysis of Motivations in Machine Learning Textbooks

Undergraduate research experiences are often limited to
small-scale apprenticeship models, leaving many students
without accessible entry points into research practice.
This paper presents the design and evaluation of a
semester-long course for undergraduates to gain research
experience in Machine Learning. The course, led by one
faculty instructor, enables nearly a hundred students to
engage in structured research through a scaffolded
replication-and-extension project, where students first
replicate a published research project and then implement
novel additions. The course integrates instructional
modules (e.g., guided paper reading, proposal writing,
public presentation) with project milestones (e.g.,
replication, extension, poster) to support research
learning for students with diverse backgrounds. Every
component of research is visited several times, with each
iteration having progressively increased autonomy coupled
with simultaneously decreased scaffolding. We find that the
scaffolding modules help students develop foundational
conceptual and procedural understanding of doing research,
and the project milestones on replication and extension
help them gain execution skills gradually. Students also
report developing a researcher mindset and feeling like
they understand the research process better.
We discuss the principles used to design a scalable
research-based course: balancing scaffolding to provide
foundational understanding with autonomy for students to
“feel like real researchers”.

Balancing Scaffolding and Autonomy: A Case Study in Designing a Scalable Undergraduate Machine Learning Research Course

As machine learning (ML) becomes integral in more
disciplines, introductory courses in the field are
attracting increasingly diverse audiences. Design of these
introductory ML courses needs to be theoretically sound,
but also intuitive, engaging, and accessible to a range of
students. Effective teaching of ML must go beyond teaching
the theoretical or practical mechanics of algorithms. In
this paper, we synthesize effective teaching strategies
from 6 experienced ML instructors across 5 institutions to
help students define appropriate ML problems, build
intuition, develop reasoning skills, and apply models
responsibly. We organize these strategies into eight
thematic areas: preparing students for success, motivating
learners through real-world relevance, integrating ethics
and societal impact, avoiding common methodological
pitfalls in model evaluation, guiding students on design
decisions, adapting effective classroom practices,
assessing student learning, and preparing for the future.
Each section offers practical examples of classroom-tested
activities (or references to existing resources), and in
many cases, reflections on our experiences with the
strategies. Our aim is for this paper to be a starting
point for instructors aiming to improve learning in
introductory ML courses. We hope this is a resource-rich
guide for teaching ML to diverse learners, grounded in both
pedagogy and practice.

Effective Strategies for Teaching Machine Learning

Large language models (LLMs) have been shown to exhibit social bias however, bias towards non-protected stigmatized identities remain understudied. Furthermore, what social features of stigmas are associated with bias in LLM outputs is unknown. From psychology literature, it has been shown that stigmas contain six shared features: aesthetics, concealability, course, disruptiveness, origin, and peril. In this study, we investigate if human and LLM ratings of the features of stigmas, along with prompt style and type of stigma, have effect on bias towards stigmatized groups in LLM outputs. We measure bias against 93 stigmatized groups across three widely used LLMs (Granite 3.0-8B, Llama-3.1-8B, Mistral-7B) using SocialStigmaQA, a benchmark that includes 37 social scenarios about stigmatized identities; for example deciding whether to recommend them for an internship. We find that stigmas rated by humans to be highly perilous (e.g., being a gang member or having HIV) have the most biased outputs from SocialStigmaQA prompts (60% of outputs from all models) while sociodemographic stigmas (e.g. Asian- American or old age) have the least amount of biased outputs (11%). We test if the amount of biased outputs could be decreased by using guardrail models, models meant to identify harmful input, using each LLM’s respective guardrail model (Granite Guardian 3.0, Llama Guard 3.0, Mistral Moderation API). We find that bias decreases significantly by 10.4% , 1.4%, and 7.8%, respectively. However, we show that features with significant effect on bias remain unchanged post-mitigation and that guardrail models often fail to recognize the intent of bias in prompts. This work has implications for using LLMs in scenarios involving stigmatized groups and we suggest future work towards improving guardrail models for bias mitigation.

Identifying Features Associated with Bias Against 93 Stigmatized Groups in Language Models and Guardrail Model Safety Mitigation

The existence of multiple, equally accurate models for a given predictive task—the Rashomon set—leads to predictive multiplicity, where models achieve similar global accuracy but diverge in their individual predictions. This inconsistency undermines trust in high-stakes applications, where reliable decision-making at the individual level is critical. Existing reconciliation methods attempt to resolve this issue by enforcing global agreement across models, but often fall short in ensuring consistency for specific instances. We first introduce Rocile (Robust Reconciliation), a model-agnostic framework that systematically reduces ensemble disagreement by iteratively adjusting model predictions through a momentum-based batching procedure. Rocile guarantees convergence to a globally consistent ensemble, yet, like other global methods, it overlooks localized inconsistencies that impact individual predictions. To address this limitation, we propose AdaRocile, an extension of Rocile that incorporates local calibration into the reconciliation process. For each test instance, AdaRocile identifies a context-sensitive neighborhood using an adaptive nearest-neighbor strategy, computes empirical correction terms for each model based on residuals in the local neighborhood, and applies Rocile to reconcile the locally adjusted predictions into a globally coherent output. The reconciled predictions are then distilled into a single, transparent decision rule for real-world deployment. Empirical results on high-stakes benchmarks, such as COMPAS and Adult, show that AdaRocile significantly improves the accuracy-reliability trade-off, reducing local calibration error by up to 27.1% over the global-only reconciliation, while driving key multiplicity metrics (variance, ambiguity, discrepancy, and disagreement rate) to near zero, maintaining global performance on key benchmarks. AdaRocile delivers both interpretability and individual-level reliability, offering a scalable, practical pipeline for building trustworthy and aligned AI systems.

Resolving Predictive Multiplicity for the Rashomon Set

Beyond simple text generation, Large Language Models (LLMs) have evolved into agentic systems capable of planning and interacting with external tools to solve complex tasks. This evolution involves fine-tuning LLMs on agent-specific tasks to enhance their proficiency. However, safety concerns are frequently overlooked during this fine-tuning process. In this work, we show that aligned LLMs can become unintentionally misaligned, leading to a higher likelihood of executing harmful tasks and a reduced tendency to refuse them when fine-tuned to execute agentic tasks. To address these safety challenges, we propose Prefix INjection Guard (PING), a simple yet effective method that prepends automatically generated natural language prefixes to agent responses, guiding them to refuse harmful requests while preserving performance on benign tasks. Specifically, we introduce an iterative approach that alternates between (1) generating candidate prefixes and (2) selecting those that optimize both task performance and refusal behavior. Experimental results demonstrate that PING significantly enhances the safety of fine-tuned LLM agents without sacrificing their effectiveness. PING consistently outperforms existing prompting approaches across diverse benchmarks in both web navigation and code generation tasks. Our analysis of internal hidden states via linear probes reveals that prefix tokens are crucial for behavior modification, explaining the performance gains. WARNING: This paper contains contents that are unethical or offensive in nature.

Unintended Misalignment from Agentic Fine-Tuning: Risks and Mitigation

Large language models require consistent behavioral patterns for safe deployment, yet their personality-like traits remain poorly understood. We present PERSIST (PERsonality Stability in Synthetic Text), a comprehensive evaluation framework testing 25+ open-source models (1B-685B parameters) across 500,000+ responses. Using traditional (BFI-44, SD3) and novel LLM-adapted personality instruments, we systematically vary question order, paraphrasing, personas, and reasoning modes.
Our findings challenge fundamental deployment assumptions: (1) Even 400B+ models exhibit substantial response variability (SD > 0.4); (2) Minor prompt reordering alone shifts personality measurements by up to 20%; (3) Interventions expected to stabilize behavior, such as chain-of-thought reasoning, detailed personas instruction, inclusion of conversation history, can paradoxically increase variability; (4) LLM-adapted instruments show equal instability to human-centric versions, confirming architectural rather than translational limitations.
This persistent instability across scales and mitigation strategies suggests current LLMs lack the foundations for genuine behavioral consistency. For safety-critical applications requiring predictable behavior, these findings indicate that personality-based alignment strategies may be fundamentally inadequate.

Persistent Instability in LLM’s Personality Measurements: Effects of Scale, Reasoning, and Conversation History

Large Language Models (LLMs) are prone to generating fluent but incorrect content, known as confabulation, which poses increasing risks in multi-turn or agentic applications where outputs may be reused as context. In this work, we investigate how in-context information influences model behavior and whether LLMs can identify their unreliable responses. We propose a reliability estimation that leverages token-level uncertainty to guide the aggregation of internal model representations. Specifically, we compute aleatoric and epistemic uncertainty from output logits to identify salient tokens and aggregate their hidden states into compact representations for response-level reliability prediction. Through controlled experiments on open QA benchmarks, we find that correct in-context information improves both answer accuracy and model confidence, while misleading context often induces confidently incorrect responses, revealing a misalignment between uncertainty and correctness. Our probing-based method captures these shifts in model behavior and improves the detection of unreliable outputs across multiple open-source LLMs. These results underscore the limitations of direct uncertainty signals and highlight the potential of uncertainty-guided probing for reliability-aware generation.

Can LLMs Detect Their Confabulations? Estimating Reliability in Uncertainty-Aware Language Models

Prompt-based essay writing is an effective and common way to assess students' critical thinking skills.
Recent work has evaluated the impressive capabilities of Large Language Models (LLMs) on this task. However, most studies focus primarily on English. Those examining LLMs' performance in Chinese often rely on coarse-grained text quality metrics, overlooking the structural and rhetorical complexities of Chinese essays, particularly across diverse genres. We therefore propose EssayBench, a multi-genre benchmark specifically designed for Chinese essay writing, along with a fine-grained, genre-specific scoring framework that hierarchically aggregates scores to better align with human preferences.
The dataset comprises 728 real-world prompts across four major genres (Argumentative, Narrative, Descriptive, and Expository), and includes both Open-Ended and Constrained types.
Our evaluation protocol is validated through a comprehensive human agreement study. The results show that our protocol aligns well with human judgments, achieving a highest Spearman's correlation of 0.816 and outperforming coarse-grained evaluation methods by an average of 8.6\%. Finally, we benchmark 15 large LLMs, analyzing their strengths and limitations across genres and instruction types. We believe EssayBench offers a more reliable framework for evaluating Chinese essay generation and provides valuable insights for improving LLMs in this domain.

Downloads

Next from AAAI 2026

Impact of a Data-driven Teaching Approach on 9th Graders Conceptual Understanding of Machine Learning

Similar lecture

Beyond the Surface: Measuring Self-Preference in LLM Judgments

.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 AAAI 2026

Impact of a Data-driven Teaching Approach on 9th Graders Conceptual Understanding of Machine Learning

Similar lecture

Beyond the Surface: Measuring Self-Preference in LLM Judgments

Downloads