Singapore

Large language models are increasingly influencing human moral decisions, yet current approaches focus primarily on evaluating rather than actively steering their moral decisions. 
We formulate this as an out-of-distribution moral alignment problem, where LLM agents must learn to apply consistent moral reasoning frameworks to scenarios beyond their training distribution. 
We introduce Moral-Reason-QA, a novel dataset extending 680 human-annotated, high-ambiguity moral scenarios with framework-specific reasoning traces across utilitarian, deontological, and virtue ethics, enabling systematic evaluation of moral generalization in realistic decision contexts.
Our learning approach employs Group Relative Policy Optimization with composite rewards that simultaneously optimize decision alignment and framework-specific reasoning processes to facilitate learning of the underlying moral frameworks. 
Experimental results demonstrate successful generalization to unseen moral scenarios, with softmax-normalized alignment scores improving by +0.757 for utilitarian and +0.450 for deontological frameworks when tested on out-of-distribution evaluation sets. 
The experiments also reveal training challenges and promising directions that inform future research.
These findings establish that LLM agents can be systematically trained to internalize and apply specific moral frameworks to novel situations, providing a critical foundation for AI safety as language models become more integrated into human decision-making processes.
Code and data will be open-sourced.

AAAI 2026

MoralReason: Generalizable Moral Decision Alignment for LLM Agents Using Reasoning-Level Reinforcement Learning

nlp: ethics — bias

peai: morality & value-based ai

nlp: (large) language models

transparency & privacy

fairness

## 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.

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 talk surveys my research journey toward building
reliable machine learning systems that behave safely and
predictably in the open world. While modern machine
learning models—including foundation models (FMs)—have
demonstrated unprecedented capabilities, they often suffer
from reliability failures under distribution shift, leading
to overconfident mispredictions, hallucinated generations,
or susceptibility to adversarial prompts. My research
rethinks reliability not as an afterthought, but as a
first-class algorithmic principle, to be optimized
alongside accuracy with minimal human supervision.


The talk is organized around three key threads. To respect
the allotted 20-30 minutes, the first and second parts
will be briefly discussed.


1. Unknown-Aware Learning via Outlier Synthesis.
I introduce a class of learning algorithms that synthesize
“virtual outliers” in representation or pixel space to
explicitly teach models what they don’t know. This includes
the VOS, NPOS, and Dream-OOD frameworks, which shape the
energy landscape around in-distribution data to avoid
overconfidence on OOD.



2. Learning in the Wild with Unlabeled Data.
I present theoretical insights and practical algorithms for
leveraging unlabeled in-the-wild data to improve
reliability. This includes SAL framework, which uses a
gradient-based spectral method to separate potential
outliers, and SCONE, which handles semantic and covariate
shifts via constrained optimization. These results turn
unlabeled data contamination into a learning signal.



3. Reliable Foundation Models.
I explore reliability failures in LLMs and multimodal
systems. I introduce HaloScope for hallucination detection
via subspace separation on LLM representations, and TSV
that performs LLM latent steering for improved
hallucination detection. I will also briefly cover the LLM
security and alignment, which includes VLMGuard for
detecting malicious prompts in vision-language models and a
data-centric paradigm for AI alignment through source-aware
feedback cleaning.

Throughout the talk, I highlight how representation
learning, data generation, and theoretical guarantees
intersect to produce scalable, label-efficient reliability
methods. I will also reflect on my broader vision:
designing proactive and collaborative AI systems that
anticipate uncertainty and support rich human-AI
interaction—especially for underrepresented communities and
emerging scientific domains.



This talk will be accessible to a broad AAAI audience,
combining foundational algorithmic insights with real-world
applications and forward-looking perspectives on the future
of responsible AI.

Teach AI What It Doesn’t Know

Large-scale tuberculosis (TB) screening is limited by the high cost and operational complexity of traditional diagnostics, creating a need for artificial-intelligence solutions. We propose DeepGB-TB, a non-invasive system that instantly assigns TB risk scores using only cough audio and basic demographic data. The model couples a lightweight one-dimensional convolutional neural network for audio processing with a gradient-boosted decision tree for tabular features. Its principal innovation is a Cross-Modal Bidirectional Cross-Attention module (CM-BCA) that iteratively exchanges salient cues between modalities, emulating the way clinicians integrate symptoms and risk factors. To meet the clinical priority of minimizing missed cases, we design a Tuberculosis Risk-Balanced Loss (TRBL) that places stronger penalties on false-negative predictions, thereby reducing high-risk misclassifications. DeepGB-TB is evaluated on a diverse dataset of 1,105 patients collected across seven countries, achieving an AUROC of 0.903 and an F1-score of 0.851, representing a new state of the art. Its computational efficiency enables real-time, offline inference directly on common mobile devices, making it ideal for low-resource settings. Importantly, the system produces clinically validated explanations that promote trust and adoption by frontline health workers. By coupling AI innovation with public-health requirements for speed, affordability, and reliability, DeepGB-TB offers a tool for advancing global TB control.

DeepGB-TB: A Risk-Balanced Cross-Attention Gradient-Boosted Convolutional Network for Rapid, Interpretable Tuberculosis Screening

Counterfactual Regret Minimization (CFR) algorithms are widely used to compute a Nash equilibrium (NE) in two-player zero-sum imperfect-information extensive-form games (IIGs). Among them, Predictive CFR$^+$ (PCFR$^+$) is particularly powerful, achieving an exceptionally fast empirical convergence rate via the prediction in many games. However, the empirical convergence rate of PCFR$^+$ would significantly degrade if the prediction is inaccurate, leading to unstable performance on certain IIGs. To enhance the robustness of PCFR$^+$, we propose Asymmetric PCFR$^+$ (APCFR$^+$), which employs an adaptive asymmetry of step sizes between the updates of implicit and explicit accumulated counterfactual regrets to mitigate the impact of the prediction inaccuracy on convergence. We present a theoretical analysis demonstrating why APCFR$^+$ can enhance the robustness. To the best of our knowledge, we are the first to propose the asymmetry of step sizes, a simple yet novel technique that effectively improves the robustness of PCFR$^+$. Then, to reduce the difficulty of implementing APCFR$^+$ caused by the adaptive asymmetry, we propose a simplified version of APCFR$^+$ called Simple APCFR$^+$ (SAPCFR$^+$), which uses a fixed asymmetry of step sizes to enable only a single-line modification compared to original PCFR$^+$. Experimental results on five standard IIG benchmarks and two heads-up no-limit Texas Hold’em (HUNL) Subagems show that (i) both APCFR$^+$ and SAPCFR$^+$ outperform PCFR$^+$ in most of the tested games, (ii) SAPCFR$^+$ achieves a comparable empirical convergence rate with APCFR$^+$, and (iii) our approach can be generalized to improve other CFR algorithms, e.g., Discount CFR (DCFR).

Faster Game Solving via Asymmetry of Step Sizes

Overestimation in evaluating large language models (LLMs) has become an increasing concern. Due to the contamination of public benchmarks or imbalanced model training, LLMs may achieve unreal evaluation results on public benchmarks, either intentionally or unintentionally, which leads to unfair comparisons among LLMs and undermines their realistic capability assessments. Existing benchmarks attempt to address these issues by keeping test cases permanently secret, mitigating contamination through human evaluation, or repeatedly collecting and constructing new samples. However, these approaches fail to ensure reproducibility, transparency, and high efficiency simultaneously. Moreover, the extent of overestimation in current LLMs remains unquantified. To address these issues, we propose ArxivRoll, a dynamic evaluation framework inspired by one-time pad encryption in cryptography. ArxivRoll comprises two key components: i) SCP (Sequencing, Cloze, and Prediction), an automated generator for private test cases, and ii) Rugged Scores (RS), metrics that measure the proportion of public benchmark contamination and training bias. Leveraging SCP, ArxivRoll constructs a new benchmark every six months using recent articles from ArXiv and employs them for one-time evaluations of LLM performance. Extensive experiments demonstrate the high quality of our benchmark, and we provide a systematic evaluation of current LLMs. The source code is available at https://github.com/liangzid/ArxivRoll/.

How Much Do Large Language Model Cheat on Evaluation? Benchmarking Overestimation Under the One-Time-Pad-Based Framework

The rapid advancement of text-to-image generative models has catalyzed widespread applications. However, persistent model biases continue to pose significant challenges to their ethical and effective deployment, often resulting in adverse outcomes across many use cases. Previous research has primarily addressed bias in narrowly defined scenarios, typically involving single-subject generation with limited contextual variability. Such simplified tasks fall short of serving as meaningful model evaluations in more complex real-world settings. For example, the prompt ``an assistant wearing a pink hat'' may reflect female-inclined biases associated with a pink hat. The neglected joint effects of the semantic binding in the prompts cause significant failures in current debiasing approaches. This work investigates **how bias manifests under semantic binding**, where contextual associations between objects and attributes influence generative outcomes. We demonstrate that the underlying bias distribution can be amplified based on these associations. To address this, we introduce a bias adherence score that quantifies how specific object-attribute bindings activate bias. Using this score, we develop a training-free context-bias control framework that decouples the underlying bias from the semantic bindings, improving over 10% biases in compositional generation tasks. Our analysis of bias scores across various attribute-object bindings and token decorrelation highlights a fundamental challenge: reducing bias without disrupting essential semantic relationships. These findings expose critical limitations in current debiasing approaches when applied to semantically bound contexts, underscoring the need to reassess prevailing bias mitigation strategies.

How Bias Binds: Measuring Hidden Associations for Bias Control in Text-to-Image Compositions

Refusal on harmful prompts is a key safety behaviour in instruction‑tuned large language models (LLMs), yet the internal causes of this behaviour remain poorly understood. We study two public instruction tuned models—Gemma‑2-2B‑IT and LLaMA‑3.1-8B‑IT using sparse autoencoders (SAEs) trained on residual‑stream activations. Given a harmful prompt, we search the SAE latent space for feature sets whose ablation flips the model from refusal to compliance, demonstrating causal influence and creating a jailbreak. Our search proceeds in three stages: 1. Refusal Direction - Finding a refusal mediating direction and collecting SAE features close to that direction, followed by 2. Greedy Filtering - to prune this set to obtain a minimal set and finally 3. Interaction Discovery - a factorization‑machine (FM) model that captures non‑linear interactions among the remaining active features and the minimal set. This pipeline yields a broad set of jailbreak-critical features, offering insight into the mechanistic basis of refusal. Moreover, we also find evidence of redundant features which remain dormant unless earlier features are suppressed. Our findings highlight the potential for fine-grained auditing and targeted intervention in safety behaviours by manipulating the interpretable latent space.

Beyond I’m Sorry, I Can’t: Dissecting Large-Language-Model Refusal

Reinforcement learning (RL) has recently become a powerful yet resource-intensive approach for post-training large language models (LLMs). Incorporating curriculum learning (CL) into RL has been shown to significantly improve training efficiency, particularly in reasoning tasks. However, existing CL methods face substantial challenges in multi-objective RL (MORL) settings, including: (1) difficulty in evaluating model capabilities online, (2) challenges in assessing sample importance under diverse objectives, and (3) inherent trade-offs between online training and offline inference in dynamically designing the curriculum. To address these issues, we propose a **M**ulti-**R**eward space guided **A**daptive **C**urriculum **L**earning framework (**MRACL**), which is the first to incorporate curriculum learning into multi-objective RL. MRACL first constructs a multi-dimensional reward space via offline inference to establish initial reward profiles for each training sample. During training, based on reward space, it estimates the evolving model capabilities by computing the centroid of the space and calculates the sample priority score through its capability distance, optimization direction and historical evolution, which enables adaptive selection of the most informative training samples at each step, independent of the specific RL algorithm. After each RL training iteration, the reward space is dynamically updated to reflect the model's evolving capabilities and the shifting distribution of sample priorities. Experiments on multi-objective alignment tasks demonstrate that MRACL achieves 1.62× faster convergence compared to state-of-the-art curriculum methods and 2.55× faster than non-curriculum methods. Furthermore, it consistently outperforms all baselines in both win rate and rule-based evaluation metrics. We further provide an in-depth analysis of the key factors contributing to MRACL's effectiveness, and summarize its advantages, applicable scenarios, and generalization across diverse experimental settings.

MRACL: Multi-Reward Space Guided Adaptive Curriculum Reinforcement Learning for LLMs

Large Language Models (LLMs) commonly rely on explicit refusal prefixes for safety, making them vulnerable to prefix injection attacks. We introduce HumorReject, a novel data-driven approach that reimagines LLM safety by decoupling it from refusal prefixes through humor as an indirect refusal strategy. Rather than explicitly rejecting harmful instructions, HumorReject responds with contextually appropriate humor that naturally defuses potentially dangerous requests. Our approach effectively addresses common "over-defense" issues while demonstrating superior robustness against various attack vectors. Our findings suggest that improvements in training data design can be as important as the alignment algorithm itself in achieving effective LLM safety. The code and dataset are available at https://github.com/wooozihui/HumorReject.

HumorReject: Decoupling LLM Safety from Refusal Prefix via a Little Humor

In recent years, recommendation systems have evolved from providing a single list of recommendations to offering a comprehensive suite of topic-focused services. To better accomplish this task, conversational recommendation systems (CRS) have progressed from basic retrieval-augmented LLM generation to agentic systems with advanced reasoning and self-correction capabilities. However, agentic systems come with notable response latency—a longstanding challenge for conversational recommendation systems. To balance the trade-off between handling complex queries and minimizing latency, we propose AdaptJobRec, the first conversational job recommendation system that leverages autonomous agent to integrate personalized recommendation algorithm tools. The system employs a user query complexity identification mechanism to minimize response latency. For straightforward queries, the agent directly selects the appropriate tool for rapid responses. For complex queries, the agent uses the memory processing module to filter chat history for relevant content, then passes the results to the intelligent task decomposition planner, and finally executes the tasks using personalized recommendation tools. Evaluation on Walmart’s real-world career recommendation scenarios demonstrates that AdaptJobRec reduces average response latency by up to 53.3\% compared to competitive baselines, while significantly improving recommendation accuracy.

AdaptJobRec: Enhancing Conversational Career Recommendation Through an LLM-Powered Agentic System

There has been a lot of exciting recent progress on new and powerful machine learning algorithms and architectures: how to learn. But for autonomous agents acting in the dynamic, uncertain world, it is at least as important to be able to identify which concepts and subproblems to focus on: what to learn.
This talk presents methods for identifying what to learn within the framework of reinforcement learning, focusing especially on applications in multiagent systems and robotics.

Content not yet available

Next from AAAI 2026

Teach AI What It Doesn’t Know

Similar lecture

Tales of Morality: Comparing Human- and LLM-Generated Moral Stories from Visual Cues

Stay up to date with the latest Underline news!

PRESENTATIONS

CONFERENCES

COMPANY

RESOURCES