Singapore

Small language models (SLMs) are increasingly deployed on edge devices, making their safety alignment crucial yet challenging. 
Current shallow alignment methods that rely on direct refusal of malicious queries fail to provide robust protection, particularly against adversarial jailbreaks. 
While deliberative safety reasoning alignment offers deeper alignment for defending against sophisticated attacks, effectively implanting such reasoning capability in SLMs with limited capabilities remains an open challenge. 
Moreover, safety reasoning incurs significant computational overhead as models apply reasoning to nearly all queries, making it impractical for resource-constrained edge deployment scenarios that demand rapid responses. 
We propose EASE, a novel framework that enables practical and Efficient Safety Alignment for Small languagE models. 
Our approach first identifies the optimal safety reasoning teacher that can effectively distill safety reasoning capabilities to SLMs.
We then align models to selectively activate safety reasoning for dangerous adversarial jailbreak queries while providing direct responses to straightforward malicious queries and general helpful tasks. 
This selective mechanism enables small models to maintain robust safety guarantees against sophisticated attacks while preserving computational efficiency for benign interactions. Experimental results demonstrate that EASE reduces jailbreak attack success rates by up to 17% compared to shallow alignment methods while reducing inference overhead by up to 90% compared to deliberative safety reasoning alignment, making it practical for SLMs real-world edge deployments.

AAAI 2026

EASE: Practical and Efficient Safety Alignment for Small Language Models

safety and robustness

(large) language models

reasoning

## 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.

technical paper

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.

With the growing use of Large Language Model (LLM)-based
Question-Answering (QA) systems in education, it is
critical to evaluate their performance across individual
pipeline components. In this work, we introduce EduMod-LLM,
a modular function-calling LLM pipeline, and present a
comprehensive evaluation along three key axes: function
calling strategies, retrieval methods, and generative
language models. Our framework enables fine-grained
analysis by isolating and assessing each component. We
benchmark function-calling performance across LLMs, compare
our novel structure-aware retrieval method to vector-based
and LLM-scoring baselines, and evaluate various LLMs for
response synthesis. This modular approach reveals specific
failure modes and performance patterns, supporting the
development of interpretable and effective educational QA
systems. Our findings demonstrate the value of modular
function calling in improving system transparency and
pedagogical alignment.

EduMod-LLM: A Modular Approach for Designing Flexible and Transparent Educational Assistants

Automated Essay Scoring (AES) and Automatic Essay Feedback
(AEF) systems aim to reduce the workload of human raters in
educational assessment. However, most existing systems
prioritize numeric scoring accuracy over feedback quality
and are primarily evaluated on school-level writing. This
paper presents Multi-Agent Argumentation and Grammar
Integrated Critiquer (MAGIC), a framework using five
specialized agents to evaluate prompt adherence,
persuasiveness, organization, vocabulary, and grammar for
both holistic scoring and detailed feedback generation. To
support evaluation at the college level, we collated a
dataset of Graduate Record Examination (GRE) practice
essays with expert-evaluated scores and feedback. MAGIC
achieves substantial to near-perfect scoring agreement with
humans on the GRE data, outperforming baseline LLM models
while providing enhanced interpretability through its
multi-agent approach. For feedback quality evaluation, we
employ human annotators using a structured rubric and
report inter-annotator agreement.

MAGIC: Multi-Agent Argumentation and Grammar Integrated Critiquer

Effective collaborative learning requires both dynamic
interaction and systematic pedagogical planning, yet
existing AI tutoring systems focus primarily on one-on-one
interactions. We introduce SAGE (Scaffolded Agent-Guided
Education), a novel compositional two-phase framework that
combines automated pedagogical planning with proactive
multi-agent collaboration. SAGE first generates optimized
lesson plans through specialized planning agents, then
executes them via autonomous conversational agents in
structured dialogue with students.
This approach ensures that dynamic, multi-agent
interactions are grounded in a pedagogically sound
foundation. We evaluate the conversational phase of SAGE,
demonstrating improved performance against a next-speaker
prediction baseline (72.13\% win rate) and effective group
coordination in a study with real students. Specifically,
our study with students reveals high role adherence from AI
agents, a balanced progression between task-oriented and
socio-emotional interactions, and a clear scaffolding
effect where instructional support fades as learner
autonomy increases.
Our findings highlight the significant potential of
synergizing automated instructional design with autonomous
conversational execution for collaborative learning.

SAGE: A Compositional Multi-Agent LLM Framework with Pedagogical Reasoning for Structured Collaborative Problem Solving

Automated short answer grading with feedback (ASAG-F) systems currently face challenges in transparency, pedagogical alignment, and cost-effectiveness that limit their real-world deployment. We introduce GraphRAG, a knowledge graph-based retrieval-augmented generation framework that addresses these limitations by grounding all large language model (LLM)-generated feedback and scores in instructor-curated atomic facts, ensuring traceability and verifiability. Using the Short Answer Feedback (SAF) dataset with 31 topics, we evaluate GraphRAG on unseen-question and unseen-answer splits. Our systematic evaluation demonstrates that GraphRAG achieves grading accuracy comparable to vector-based RAG and generally superior to a fine-tuned LLM baseline model while providing more transparent source attribution. Additional findings include: (1) Instructing the LLM to discretize continuous scores to match pedagogical rubrics, such as the 0.25 increments common in SAF, improves grading accuracy; (2) LLM-generated feedback exhibits length-dependent quality variations when unconstrained; prompt-based length control substantially enhances feedback quality and its stability, achieving optimal balance of instructional richness and conciseness; (3) Performance scaling analysis reveals that basic models like GPT-4o-mini offer cost-effective performance, while premium models like Claude-Opus-4 show diminishing returns. These results demonstrate that GraphRAG offers a robust, explainable, pedagogy-aligned, and cost-effective solution for large-scale educational applications, enabling transparent automated grading with effective pedagogical feedback and practical deployment costs.

Graph RAG for Automated Short Answer Grading with Feedback: Bridging Pedagogical Needs and Technical Capabilities

Classroom exercises are imperative for reinforcing
learning; however, in conventional instruction, students
frequently lack timely and personalized feedback. To
address this, we present TacpAgent(Teaching Agent for
Classroom Exercises), a generative LLM-based agent that
delivers reflective, individualized feedback through a
prompt-template framework. TacpAgent takes as input the
teacher-prepared exercises and reference answers, together
with students’ submitted responses, confidence levels, and
brief explanations, then leverages the DeepSeek LLM with
designed prompt templates to generate reflective feedback
and recommend relevant textbook sections for targeted
review. A three-month quasi-experimental study was
conducted, involving two high school classes (N=87). The
study compared TacpAgent-supported exercises with
traditional paper-based exercises. The findings indicated
that the experimental group exhibited significantly higher
practice quiz scores (F=18.516, p<0.001, η²=0.181) and
demonstrated notable improvements in emotional (p<0.001)
and behavioral engagement (p=0.003). In contrast, the
control group demonstrated no significant changes.These
findings highlight TacpAgent’s potential to enable
scalable, personalized formative assessment in classroom
settings and provide practical guidance for integrating
generative AI into everyday teaching.

TacpAgent: Enhancing Student Engagement in Classroom Exercises Through LLM-Generated Feedback

AI-supported tools have entered K-12 classrooms in recent
years to reshape student learning and skill-building. We
are particularly interested in AI’s application in literacy
subjects, such as English, where students are expected to
hone their critical thinking and public speaking skills
through AI interactions. This report details the pilot
implementation of Debate Guru, an AI-enhanced debate
education platform, across two secondary schools with
varying instructional contexts. Over the course of a summer
school course, educators integrated Debate Guru in one of
two ways: 1) by using Debate Guru’s complete curriculum; or
2) by combining platform resources with their own
instruction, such as a literary text. The pilot was
implemented with approximately 50 8th-11th grade students.
Findings suggest significant increases in students’
confidence, argumentative reasoning, and engagement.
Teachers reported high usability and strong pedagogical
value, while students responded positively to interactive
features and AI feedback integration.

“Debate Guru”: Honing Public Speaking Skills Among Secondary School Students with AI Tutoring Systems

Effective classroom teaching requires instructors to be
responsive to their students, such as by pivoting their
lectures in real-time to address common misconceptions that
their students may have developed. Classroom response
systems such as multiple-choice "clicker" systems are one
method by which instructors can gauge their students’
understanding during classroom lectures, but open-ended
questions that prompt students to engage in
self-explanation are better suited to promoting critical
thinking. Additionally, analyzing students’ natural
language responses typically requires time-consuming manual
analysis, which makes it challenging to implement in a
classroom setting. To address this challenge, we present an
LLM-driven method for automatically assessing students'
responses and generating an aggregated summary of LLM-based
evaluations for their self-explanations during
undergraduate classroom lectures. Our approach extracts
relevant knowledge components for a given question, tags
students’ responses according to whether they correctly
address each knowledge component, and generates class-level
summaries that highlight common misconceptions and gaps in
knowledge to support instructors in pivoting their lectures
in real time. We evaluate the system’s effectiveness at
these tagging and summarization tasks on data from an
undergraduate computer science course, using quantitative
and qualitative metrics such as relevance, sufficiency,
hallucination rate, and alignment with instructional goals
and desired feedback format gathered through instructor
interviews. Results suggest that the explanation-based
classroom response system can accurately analyze students’
natural language explanations.

An Explanation-Based Classroom Response System for Real-Time Analysis of Undergraduate Students’ Natural Language Explanations

Automated scoring plays a crucial role in education by reducing the reliance on human raters and offering scalable and immediate evaluation of student work. While large language models (LLMs) have shown strong potential in this task, their use as end-to-end raters faces challenges such as low accuracy, prompt sensitivity, limited interpretability, and rubric misalignment, which hinder practical implementation. To address the limitations, we propose AutoSCORE, a multi-agent LLM framework enhancing automated scoring via rubric-aligned Structured COmponent REcognition. With two agents, AutoSCORE first extracts rubric-relevant components from student responses and encodes them into a structured representation (i.e., Scoring Rubric Component Extraction Agent), which is then used to assign final scores (i.e., Scoring Agent). This design ensures that model reasoning follows a human-like grading process, enhancing interpretability and robustness. We evaluate AutoSCORE on four benchmark datasets from the ASAP benchmark, using both proprietary and open-source LLMs (GPT-4o, LLaMA-3.1-8B, LLaMA-3.1-70B). Across diverse tasks and rubrics, AutoSCORE predominantly improves scoring accuracy, human-machine agreement (QWK, correlations), and reduces error metrics (MAE, RMSE) compared to single-agent baselines, with particularly strong benefits on complex, multidimensional rubrics, and especially large relative gains on smaller LLMs. These results demonstrate that structured component recognition combined with multi-agent design offers a scalable, reliable, and interpretable solution for automated scoring.

AutoSCORE: Enhancing Automated Scoring with Multi-Agent Large Language Models via Structured Component Recognition

This paper introduces a novel system for in-home cognitive health assessment using ambient sensors and a machine learning technology that can robustly detect mild cognitive impairment (MCI) despite its noisy and sparsely limited available data. The learned model can transparently explain the aspects of individuals' daily lives led to the prediction, while reliably predicting MCI, providing more insights to healthcare workers for further clinical interventions. We developed the robust transparent machine learning model, based on fusion adaptive resonance theory (Fusion ART) neural network to learn individuals' daily patterns of activity from continuous sensor data in terms of a suite of digital biomarkers reflecting four key domains: physical activity, daily routines, cognitive engagement, and sleep patterns. 
Based on a longitudinal study of over one hundred participants, deployed with non-intrusive sensors in their homes to undergo parallel clinical evaluation across a period of five years, our model successfully identified individuals with MCI, achieving high predictive accuracy regardless the noisy and sparse availability of data. As a transparent neural network, the learned model can also serve as classification rules to distinguish MCI from normal cognition (NC) cases based on the digital biomarkers. These results demonstrate that passively collected, sensor-derived digital biomarkers can be leveraged to indicate cognitive status and potentially providing clinically meaningful insights on the impairment conditions. We also discuss the practical challenges and lessons learned from this real-world deployment to inform future large-scale implementations of such AI-driven health monitoring systems.

Downloads

Next from AAAI 2026

EduMod-LLM: A Modular Approach for Designing Flexible and Transparent Educational Assistants

Similar lecture

Reasoning-to-Defend: Safety-Aware Reasoning Can Defend Large Language Models from Jailbreaking