Singapore

Direct Preference Optimization (DPO) typically relies on a
fixed inverse-temperature β that controls divergence from a
reference model. Fixed β is brittle: too small causes
underregularization
(verbosity, safety drift); too large causes
overregularization
(underfitting). I propose an adaptive per-token
KL controller using EMA smoothing, deadband filtering, and
clipping to dynamically adjust β throughout training.
Initial
results on a 7B model show 72% win rate vs. base SFT and
60% vs. fixed-β DPO. The goal is a practical recipe for
stable,
compute-efficient DPO with reduced manual tuning.

AAAI 2026 Main Conference

Adaptive KL Control for Direct Preference Optimization in Instruction-Following LLMs

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.<br><br>

To access this event page, you need to log in with the **email address you registered with**. <br>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.

Generative AI shows strong capabilities in language,
reasoning, and code but remains prone to
hallucinations—outputs that are fluent yet incorrect. In
cybersecurity, such errors pose serious risks, from
misleading analysts to potential adversarial exploitation.
This project investigates hallucinations in three
directions: (1) creating benchmarks and interpretability
tools to characterize them in security contexts; (2)
developing mitigation strategies such as
retrieval-augmented generation, symbolic-neural hybrids,
and uncertainty-aware decoding; and (3) integrating these
methods into real-world workflows like vulnerability
assessment, malware analysis, and penetration testing,
while exploring how attackers might exploit hallucinations.
Evaluation will combine accuracy metrics, human-in-the-loop
studies, and red-team simulations. By bridging theory and
applied system design, the work aims to advance
understanding of hallucinations and improve the reliability
of AI in cybersecurity, with broader implications for other
high-stakes areas such as healthcare and law.

Hallucinations at the Firewall

Large language models (LLMs) have rapidly advanced, but
their growing compute and memory demands make them
unsustainable and limit accessibility, especially in
under-resourced regions such as Southeast Asia (SEA). While
recent hybrid architectures combining attention and
state-space models (SSMs) have shown promise, most adopt
sequentially interleaving attention and Mamba layers,
leaving parallel-head mixing of attention and Mamba heads
largely unexplored. Hence, I propose investigating
Hymba-style parallel-head hybrid architecture as a
foundation for efficient, multilingual LLMs in SEA. My
short-term goal is to perform continual pre-training (CPT)
of the released Hymba-1.5B weights on SEA corpora to
evaluate their adaptability across diverse languages. In
the longer term, I plan to study scaling strategies beyond
1.5B parameters, assessing whether parallel-head hybrids
maintain efficiency and performance at larger scales.
Evaluation will combine standard perplexity and benchmark
tasks with SEA-specific benchmarks, alongside profiling for
inference throughput and deployability on
resource-constrained devices. The expected outcomes are
twofold: (1) demonstrating that parallel-head hybrids can
be effectively adapted to SEA multilingual contexts, and
(2) providing evidence that this underexplored architecture
scales efficiently. Success would broaden the design space
of efficient LLMs while advancing equitable access to AI by
enabling practical, low-cost, and locally relevant models
for SEA communities.

Adapting Hybrid Parallel-Head Large Language Models for Southeast Asia

Modern generators often violate basic physics—e.g.,
inconsistent shadows, geometry, and measurement models
limiting trust for video synthesis and computational
imaging. We propose finite-time Schrödinger-Bridge (SB)
world models that cast generation as entropy-regularized
optimal transport from a simple prior to a physics- and
data-consistent distribution. Unlike post-hoc corrections,
our approach injects structure along the transport path:
multi-view geometry (reprojection/epipolar constraints,
homographies, depth-aware warps) for video, and
differentiable optical operators (PSF-based defocus,
lightweight Fourier propagation for coherent/partially
coherent settings) for imaging. With known poses, we
penalize reprojection and warp-aligned photometric/feature
errors; with unknown poses, a compact head estimates
motion/flow with cycle-consistency. Compact UNet/ViT
backbones and short SB horizons target efficiency.
Evaluation spans 3D consistency metrics, physics fidelity
via forward-simulation error, and generative
quality/efficiency (FID/KID, FVD) against strong diffusion
baselines, plug-and-play data-consistency, and
unconstrained SB. By constraining the path rather than only
the endpoint, the method aims to shorten sampling while
improving cross-view coherence and physical plausibility
across sensors (cameras, microscopes, medical scanners).

Physics Consistent World Models via Schrödinger-Bridge Optimal Transport for Computational Imaging and 3D-Consistent Video Generations

I present a compact, testable architecture that endows learning agents with continuous proto-emotional dynamics and interpretable modulators (Persona, Ego, Shadow, Self). The design grounds these modulators in a computational interpretation of Jung’s Map of the Soul, mapping each archetype to a differentiable control that modulates policy selection via a bounded, low-dimensional affect vector. I describe concrete modular implementations, a staged experimental program (toy domains → multi-agent/social tasks → nonstationary transfer), baselines, ablations, and reproducible evaluation metrics.

Persona, Ego, Shadow, and Self: A Map of the Soul Framework for Proto-Emotional Homeostasis in AI

Sexual trauma leaves wounds that science cannot see, yet
survivors live with them every day. Traditional tools rely
on words or self-reports, often forcing survivors to “bleed
in silence” when their pain is doubted or dismissed.
Trauma, however, is not one-dimensional. It disrupts
multiple brain networks and produces states fear,
vigilance, detachment that cannot be captured by words
alone. This creates the need for approaches that reveal
trauma’s complexity in ways that are both objective and
interpretable.
We propose a framework that combines fMRI, EEG, and
interpretable self-supervised AI (DINO) to uncover hidden
patterns of trauma in the brain. Instead of producing
abstract scans or opaque predictions, the system will
generate exploratory measures of trauma response that
support therapists’ understanding while guiding future
research. These measures will be presented through a simple
dashboard that summarizes three indices (TPI, DI, RBS)
alongside heatmaps and plain language notes. By turning
complex data into clear, anonymized session snapshots, the
dashboard provides researchers with a practical output that
can be compared across participants and refined in future
work

Understanding the Management of Rape Trauma with AI and Neuroimaging

The widespread adoption of artificial intelligence(AI) in
cybersecurity has led to the emerging of intelligent
threats, such as Advanced Persistent Threats (APTs),
challenging the conventional deception defense mechanisms.
My work aims to fill this critical gap by developing a game
theoretic defense agent capable of confronting these
intelligent threats. In this proposal, we formalize the
attacker-defender interactions as a Bayesian game model
between AI agents so as to derive equilibrium defense
strategies. Simulation based experiments and real-world
implementations would be conducted to evaluate the proposed
framework. This study is potential to revolutionize cyber
defense methodologies by shifting from bilateral
decision-making to game theoretic strategy evolution.

When AI Meets AI: A Game-Theoretic Defense Framework Against AI Empowered Cyber Threats

RNA 3D structure prediction remains a fundamental challenge
due to limited experimental data, conformational
heterogeneity, and complex folding landscapes. Inspired by
breakthroughs in protein modeling, a data-efficient deep
learning framework that integrates RNA language embeddings,
geometric constraints, and physics-informed priors is
proposed. This approach leverages self-supervised
pretraining, contrastive learning, and SE(3)-equivariant
architectures to capture higher-order structural
relationships from scarce data. Predicted structures are
evaluated using benchmark datasets, thermodynamic
plausibility, and docking-based functional assessments,
ensuring both structural accuracy and biophysical
relevance. By advancing RNA structure prediction and
design, this work aims to accelerate the development of
RNA-based therapeutics, catalytic ribozymes, and precision
medicine applications, while providing an open-source
framework for the broader scientific community.

Towards Data-Efficient Deep Learning for RNA 3D Structure Prediction and Design

This research proposes an extension to the Program Lattice
Transformer (PLT) , a neuro-symbolic framework for program
induction that embeds programs into a structured latent
space. The current PLT model, which uses a flat lattice, is
computationally inefficient when modeling invariant
programs—operations that return to an initial state after a
set number of applications (e.g., a 360° rotation). To
address this, we propose embedding the program space onto a
cylindrical manifold instead of a plane. This approach is
grounded in the principle that only isometric
transformations preserve the lattice's compositional
structure, limiting valid manifolds to developable surfaces
like cylinders . A cylindrical geometry naturally
represents invariant programs as closed loops, enhancing
efficiency. The proposed method will be evaluated on
synthetic tasks like Rubik's Cube and the Abstraction and
Reasoning Corpus (ARC) to demonstrate improved performance
and efficiency. This work serves as a step toward models
that can autonomously configure their own geometric latent
spaces, connecting to future research in geometric deep
learning and meta-learning.

Cylindrical Lattice Embedding for Program Induction

With the increase in the number of open sourced models
available to the average consumer, Low-Rank Adaptation
(LoRA) have become essential tools for adapting large
language models with limited computational resources (Hu et
al., 2021). LoRA works by introducing trainable low-rank
matrices A and B to update pre-trained weights efficiently,
significantly reducing memory and compute requirements
compared to full fine-tuning. The original motivation was
that the weight update during fine-tuning could be
estimated by ∆W ≈ AB, and thus could drastically reduce the
number of trainable parameters while still capturing
task-specific adaptations, significantly reducing the
number of parameters that needs to be stored in memory.
While recent work has shown that LoRA is not strictly
equivalent to full fine-tuning (Shuttleworth et al., 2024),
it remains an efficient and practical method for adapting
models to specific tasks, playing a crucial role in the
democratization of AI. Despite being efficient, LoRA has
been shown to be suboptimal for finetuning models with
large embedding dimensions, due to differences in the
magnitudes of the values of A and B (Hayou et al., 2024;
Yen et al., 2024; Zhang & Pilanci, 2024). While there are
existing LoRA variants that claim to handle these problems
(Hayou et al., 2024; Yen et al., 2024; Bensaïd et al.,
2025; Zhang & Pilanci, 2024), we would like to rigorously
evaluate the performance of these LoRA variants on a
variety of tasks and models, and investigate an alternative
and novel approach, which introduces a penalty.

Scale Regularization for Stable Low-Rank Adaptation

With the development of Large Language Models (LLMs), there
is growing interest in how we can apply the knowledge of
LLMs to tasks beyond text generation and question
answering. Speech processing, a field of interest for
decades, has recently seen successful applications of LLM
architectures following the release of the Transformer
architecture such as in the Whisper models. While
significant development has occurred in improving LLM
capabilities through Supervised Fine-Tuning (SFT),
reasoning, and alignment with Reinforcement Learning (RL),
their application to the speech domain remain somewhat
underexplored. Recent work has demonstrated that fine
tuning LoRA (Low-Rank Adaptation) adapters for LLMs can
perform Automatic Speech Recognition (ASR) tasks natively,
leveraging existing LLM capabilities and bypassing the
pre-training stage. However, no approaches have yet to
successfully apply LLM knowledge in a similar fashion to
other speech processing tasks like speaker diarisation.
Current approaches utilise LLMs as a post-processing step
on the outputs of a speaker diarisation model, but no model
based on LLMs has yet to be able to natively perform
speaker diarisation. Therefore, this research proposal
explores how we can create LoRA adapters for LLMs to
perform speaker diarisation tasks natively, and explores
how we can also overcome the speech domain’s reliance on
annotated data.

Downloads

Next from AAAI 2026 Main Conference

Hallucinations at the Firewall

Stay up to date with the latest Underline news!

PRESENTATIONS

CONFERENCES

COMPANY

RESOURCES