Singapore

When analyzing a mental health conversation between a
counselor and his/her client, one examine the semantics
underlying the utterances of conversation to understand if
the counselor has practiced the appropriate psychotherapy
techniques at different points of the conversation. Despite
the many breakthroughs in solving NLP tasks,
state-of-the-art large language models (LLMs) still perform
poorly on this utterance label prediction task. While a
simple supervised learning architecture combining an
utterance encoder with a linear softmax layer can yield
better accuracy, the trained classifiers still suffer from
poor quality ground truth labels assigned by human
annotators. Motivated by this observation, we propose a
quality-aware framework that derives quality weights of
ground truth utterance labels, trains a target classifier
in two stages, and evaluates the target classifier with
quality weights. Our experiments on three mental health
conversation datasets show that target classifiers trained
using our framework yield significantly improved accuracy
over classifiers trained not using quality weights, even
outperforming the strong LLMs using direct prompting.

AAAI 2026

Not All Labels Are Equal: On Predicting Utterance Labels in
Mental Health Conversation Data

utterance label quality

utterance label assignment

mental health conversations

workshop 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.<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.

Recent studies in pathology foundation models have shown
that scaling training data, diversifying cancer types, and
increasing model size consistently improve their
performance. However, giga-scale foundation models, which
are trained on hundreds of thousands of slides covering
tens of cancer types and contain billions of parameters,
pose significant challenges for practical use due to their
tremendous computational costs in both development and
deployment. In this work, we present a novel strategy,
named the G2L framework, to increase the performance of
large-scale foundation models, which consist of only 15% of
the parameters of giga-scale models, to a comparable
performance level of giga-scale models in cancer-specific
tasks. Our approach applies knowledge distillation,
transferring the capabilities of a giga-scale model to a
large-scale model, using just 1K pathology slides of a
target cancer (e.g., breast, prostate, etc.). The resulting
distilled model not only outperformed state-of-the-art
models of the same size (i.e., large-scale) across several
benchmarks but also, interestingly, surpassed the
giga-scale teacher and huge-scale models in some
benchmarks. In addition, the distilled model exhibited a
higher robustness index, indicating improved resilience to
image variations originating from multiple institutions.
These findings suggest that the proposed distillation
approach for a large-scale model is a data- and
parameter-efficient way to achieve giga-scale-level
performance for cancer-specific applications without
prohibitive computational burden.

G2L: From Giga-Scale to Cancer-Specific Large-Scale
Pathology Foundation Models via Knowledge Distillation

Lung adenocarcinoma is the most common histological subtype
of lung cancer. The pathological diagnosis of invasion is
very important to predict patients’ prognoses, but
interobserver variability among pathologists remains a
major issue. In this study, we propose a two-step deep
learning framework for prognostic and clinicopathological
prognostic factor prediction, in which dot-style images are
generated from high-resolution EVG-stained pathological
images and then used as model inputs. Prognosis and seven
prognostic factors are each treated as independent binary
classification tasks. The generated dot-style images
exhibit severe class imbalance. To address class imbalance
in classification tasks, we extend the existing (Cross
Entropy–based) Logit-Adjusted Loss (LA-CE) and propose a
new loss function called Logit-Adjusted Dice Loss
(LA-Dice). LA-Dice integrates the batch-level direct
optimization of the F1-score property of Dice loss with a
logit-adjustment mechanism based on the class prior
distribution of the training data, thereby enhancing the
learning of the minority class. As a result, LA-Dice
simultaneously overcomes (i) the majority-class bias
inherent in CE-based losses caused by pointwise
minimization and (ii) the inability of Cross Entropy-based
losses to directly optimize the F1-score, which is critical
for medical image classification. In our experiments, the
proposed framework achieved an average accuracy of 89.2% in
dot-style image generation. Furthermore, by using the
generated dot-style images, binary classifications for
prognosis and the seven prognostic factors were evaluated.
When averaged across all eight tasks, the proposed LA-Dice
achieved the highest mean F1-score (×100) of 58.0, showing
improvements of +6.0 over Cross Entropy, +3.3 over Dice,
and +3.3 over LA-CE. Similarly, the proposed LA-Dice
achieved the highest average mean accuracy of 59.0%,
outperforming CE by +3.7 points, Dice by +2.5 points, and
LA-CE by +2.4 points.

Prognosis and Clinicopathological Prognostic FactorPrediction in Lung Adenocarcinoma Using the Logit-AdjustedDice Loss: A New Loss Function for Addressing ClassImbalance

The failure to address environmental factors, which account
for a significant portion of autoimmune disease risk,
represents a critical gap in personalized medicine. This is
an intractable challenge stemming from the vast,
heterogeneous nature of clinical and toxicological data. We
present the Exposome Interpreter, a multi-modal
architecture that bridges this translational gap by
systematically mapping patient-specific immunological
dysregulation to environmental exposures. The framework
utilizes Vision-Language Models (VLMs), including Gemini
2.5 Flash and PaliGemma, to automatically structure data
from complex clinical lab reports. Concurrently, a
domain-adapted Gemma model within a Retrieval-Augmented
Generation (RAG) system synthesizes complex biomedical
literature into an actionable knowledge base. The final
stage provides targeted, evidence-based hypotheses, ranking
potential environmental triggers and linking abnormal
biomarkers directly to specific consumer products in a
personalized output.

Bridging the Translational Gap: A Multi-Modal Framework for
Exposome-Driven Autoimmune Management

Estimating heterogeneous treatment effects (HTE) from
electronic health records (EHR) is critical for precision
medicine, yet challenging with small clinical cohorts,
confounding bias, and high-dimensional feature space.
Pretrained EHR foundation models encode rich patient
medical history into feature embeddings that capture
patients’ characteristics; however, whether these
embeddings can be effectively utilized for HTE estimation
remains unknown. We leverage embeddings generated by a
large pretrained EHR foundation model, CLMBR, and apply the
R-learner framework for HTE estimation. We develop an
outcome-specific dimensionality reduction strategy that
adapts to small sample sizes through principal component
analysis (PCA) and strong regularization.We compare the
foundation model embedding with five baseline methods.
Evaluated on comparing GLP-1 receptor agonists versus
insulin/sulfonylureas for type 2 diabetes patients across
four clinical outcomes on Stanford University EHRs dataset,
our approach achieves rank #1 on all four outcomes and
averages 7.9% improvement over the best baseline per
outcome. Our results demonstrate that pretrained EHR
foundation model embedding can significantly enhance HTE
estimation with limited data when properly adapted.
Furthermore, we develop
interpretable tools that connect clinical event codes to
HTE predictions by mapping dimension-level importance to
tokenlevel importance, providing actionable insights for
personalized treatment decisions.

Foundation Models for Heterogeneous Treatment Effect
Estimation: Adapting Pretrained EHR Embeddings in Small
Clinical Cohorts

Multi-organ segmentation is often challenged by partially
annotated datasets and domain shifts across different
imaging sources. To address these limitations, we propose a
two-stage learning framework that efficiently leverages
partial supervision.
In the first stage, the model learns from available
annotations to produce accurate segmentations of annotated
organs, establishing robust feature representations. In the
second stage, we introduce learnable organ prototypes and a
Sinkhorn-triplet loss to enforce organ-wise feature
consistency across datasets. This encourages latent
embeddings of the same organ to remain close, while
increasing separation between different organs, even when
annotations are missing.
Our approach achieves performance comparable to
state-of-the-art methods on the BTCV dataset, while
remaining computationally efficient.
By explicitly aligning feature distributions rather than
relying solely on pseudo-labels, the framework effectively
mitigates domain shift, making it particularly suitable for
medical image segmentation tasks with limited annotation
resources.

Optimal Transport Metric Learning for Feature Alignment in
Partially Supervised Segmentation

While large language models (LLMs) with Chain-of-Thought (CoT) reasoning excel in mathematics and coding, their potential for systematic reasoning in chemistry, a domain demanding rigorous structural analysis for real-world tasks like drug design and reaction engineering, remains untapped. Current benchmarks focus on simple knowledge retrieval, neglecting step-by-step reasoning required for complex tasks such as molecular optimization and reaction prediction. To address this, we introduce ChemCoTBench, a reasoning framework that bridges molecular structure understanding with arithmetic-inspired operations, including addition, deletion, and substitution, to formalize chemical problem-solving into transparent, step-by-step workflows. By treating molecular transformations as modular "chemical operations", the framework enables slow-thinking reasoning, mirroring the logic of mathematical proofs while grounding solutions in real-world chemical constraints. We evaluate models on two high-impact tasks: Molecular Property Optimization and Chemical Reaction Prediction. These tasks mirror real-world challenges while providing structured evaluability. By providing annotated datasets, a reasoning taxonomy, and baseline evaluations, ChemCoTBench bridges the gap between abstract reasoning methods and practical chemical discovery, establishing a foundation for advancing LLMs as tools for AI-driven scientific innovation.

Beyond Chemical QA: Evaluating LLM’s Chemical Reasoning with Modular Chemical Operations

We present a training-free graph-based approach for solving interactive reasoning tasks in the ARC-AGI-3 benchmark. ARC-AGI-3 comprises game-like tasks where agents must infer task mechanics through limited interactions, and adapt to increasing complexity as levels progress. Success requires forming hypotheses, testing them, and tracking discovered mechanics. The benchmark has revealed that state-of-the-art LLMs are currently incapable of solving these tasks. Our method combines frame processing with systematic state-space exploration using graph-based representations. It segments visual frames into meaningful components, prioritizes actions based on visual salience, and maintains a graph of explored states. By tracking visited states and tested actions, the method prioritizes actions that provide the shortest path to untested state-action pairs. The approach scored 3rd in the number of solved levels on the private leaderboard of ARC-AGI-3 Preview Challenge, significantly outperforming LLM-based solutions. This demonstrates the effectiveness of structured exploration and suggests that graph representations of novel tasks can guide exploration and enable progress in unknown environments where LLMs fail to understand task dynamics.

Graph-Based Exploration for ARC-AGI-3 Interactive Reasoning Tasks

This is the third paper of the CayleyPy project applying artificial intelligence methods to problems in group theory. We announce the first public release of CayleyPy, an open-source Python library for computations with Cayley and Schreier (coset) graphs. Compared with state-of-the-art systems based on classical methods, such as GAP and Sage, CayleyPy handles significantly larger graphs and performs many orders of magnitude times faster. Using CayleyPy we obtained about 200 new mathematical conjectures on Cayley and Schreier graphs which can be turned into an efficient benchmarks for both RL and LLM models. For many Cayley graphs of symmetric groups $S_n$ we observe quasi-polynomial diameter formulas: a small set of quadratic or linear polynomials indexed by $n \mod s$ and conjecture that it is general phenomenon. We conjecture improved Babai-type bounds of the diameters by $\frac12 n^2 + 4n$ for undirected case, by $\frac34 n^2 + O(n)$ for directed cases, and by $\frac14 n^2 + O(n)$ for certain Schreier graphs, comparing to prior conjectural bounds of $O(n^2)$. For nilpotent groups we conjecture an improvement of J.S. Ellenberg's results on the diameter of the upper-triangular matrices over $Z/p$, presenting a phenomenon of linear dependence of the diameter with respect to $p$. Moreover, the growth for nilpotent groups is conjectured to follow Gaussian distributions, that is, to exhibit a central limit phenomenon similar to results of P. Diaconis for $S_n$.

ResearchArena-CayleyBench: RL/LLM Benchmark challenges which can advance mathematical research

While AI can accelerate text-intensive research, unsupervised autonomous agents risk compounding errors and cannot be held accountable for the integrity of their findings. Accordingly, we propose a human-supervised, multi-agent model for Human-AI collaboration in science that reconciles the power of Large Language Models (LLMs) with the core principles of transparency and accountability. In our vision, a human researcher supervises a team of specialized AI agents that execute discrete tasks within a rigorously defined scientific methodology. Each step in the process is governed by human expert validation, in order to prevent error propagation and maintain conceptual integrity. We instantiate this model in an end-to-end framework for systematic taxonomy generation by operationalizing the Nickerson et al. (2013) methodology. We provide a proof-of-concept (PoC) implementation and a qualitative case study that validates our framework by re-deriving a recently published taxonomy. Our analysis documents the framework's fidelity to the original method and illustrates how human-gated oversight is crucial for ensuring the scientific rigor of the final taxonomy. To support reproducibility, we release our prompts, agent roles, and workflow design as a generalizable blueprint for structuring human-AI teams. This work argues that method-bounded human-in-command (HIC) supervision offers a principled and practical path to developing reliable and accountable AI research assistants.

Human-in-Command Governance for Multi-Agent Scientific Workflows

Comparing neural network representations is essential for understanding and validating models in scientific applications. Existing methods, however, often provide a limited view. We propose the Triangle of Similarity, a framework that combines three complementary perspectives: static representational similarity (CKA/Procrustes), functional similarity (Linear Mode Connectivity or Predictive Similarity), and sparsity similarity (robustness under pruning). Analyzing a range of CNNs, Vision Transformers, and Vision-Language Models, our initial findings suggest that: (1) architectural family is a primary determinant of representational similarity, forming distinct clusters; (2) CKA self-similarity and task accuracy are strongly correlated during pruning, though accuracy often degrades more sharply; and (3) for some model pairs, pruning appears to regularize representations, exposing a papershared computational core. This framework offers a more holistic approach for assessing whether models have converged on similar internal mechanisms, providing a useful tool for model selection and analysis in scientific research.

Content not yet available

Downloads

Next from AAAI 2026

G2L: From Giga-Scale to Cancer-Specific Large-Scale Pathology Foundation Models via Knowledge Distillation

Stay up to date with the latest Underline news!

PRESENTATIONS

CONFERENCES

COMPANY

RESOURCES